wl-hydrophilic-polymer/task2/task2-chunks/Application of machine learning in polymer additive manufacturing_ A review.json

[
    {
        "id": 1,
        "chunk": "# Application of machine learning in polymer additive manufacturing: A review  \n\nTahamina Nasrin1 | Farhad Pourkamali-Anaraki² | Amy M. Peterson $\\mathbf{1}_{\\mathbb{O}}$  \n\n'Department of Plastics Engineering, University of Massachusetts Lowell, Lowell, Massachusetts, USA ²Department of Mathematical and Statistical Sciences, University of Colorado Denver, Denver, Colorado, USA",
        "category": " Introduction"
    },
    {
        "id": 2,
        "chunk": "# Correspondence  \n\nAmy M.Peterson, Department of Plastics Engineering, University of Massachusetts Lowell, Lowell, MA, USA. Email: amy_peterson@uml.edu",
        "category": " References"
    },
    {
        "id": 3,
        "chunk": "# Abstract  \n\nAdditive manufacturing (AM) is a revolutionary technology that enables production of intricate structures while minimizing material waste. However, its full potential has yet to be realized due to technical challenges such as the dependence of part quality on numerous process parameters, the vast number of design options, and the occurrence of defects. These complications may be magnified by the use of polymers and polymer composites due to their complex molecular structures, batch-to-batch variations, and changes in final part properties caused by small alterations in process settings and environmental conditions. Machine learning (ML), a branch of artificial inteligence, offers approaches to tackle these challenges and significantly reduce the experimental and computational time and expense. This review provides a comprehen sive analysis of existing research on integrating ML techniques into polymer AM. It highlights the challenges involved in adopting ML in polymer AM, proposes potential solutions, and identifies areas for future research.",
        "category": " Abstract"
    },
    {
        "id": 4,
        "chunk": "# KEYWORDS  \n\nin-situ monitoring, machine learning, polymer additive manufacturing, process optimization, property prediction",
        "category": " Abstract"
    },
    {
        "id": 5,
        "chunk": "# 1 INTRODUCTION  \n\nAdditive manufacturing(AM) involves fabricating three dimensional objects, often in a layer-by-layer manner, from computer-aided design (CAD) models. Over the past few decades, AM has transformed from a technology primarily used for prototyping into a robust tool capable of producing functional end-use parts.l One of the key advantages of AM over traditional subtractive manufacturing methods is its ability to manufacture complex geometries that would be challenging or impossible to achieve using conventional processes.² Furthermore,  \n\nAM typically results in less material waste than subtractive methods due to its minimal post-processing steps and near-net-shape output.² However, the widespread adoption of AM is still hindered by a number of obstacles, such as anisotropy of final part properties,4 limitations in dimensional accuracy and resolution,5 slow manufacturing speeds relative to mass production techniques,% and a limited selection of printable materials compared to conventional manufacturing.?  \n\nThe history of machine learning (ML) dates back to 1957. Inspired by the human nervous system, psycholo gist Frank Rosenblatt and his team created an alphabet letter-recognition machine. This device was dubbed the “perceptron” and is regarded as the basis for modern artificial neural networks. Although the concept of ML has been around for over half a century, it has only recently gained  significant popularity due to advancements in computing power, data storage capabilities, open-source software, and ML libraries such as TensorFlow,9 Scikit-Learn,10 PyTorch,1l and Keras.1² The fundamental operating principle of ML models entails learning from existing data to unveil patterns and relationships, which can then be used to make predictions or decisions about new and unseen cases.13 This process can handle various types of data, including numeric, categorical, text, image, audio, and video data.14 With advances in data acquisition techniques and data storage technologies, ML has attracted significant interest in a wide range of fields including AM.  \n\nAM techniques exhibit high levels of complexity due, in part, to an extensive number of process parameters.15-19 The microstructural and macrostructural properties of the printed parts are significantly influenced by these parameters. For instance, cure depth, irradiation time, and irradiation power are all crucial factors for dictating the final properties of vat photopolymerized parts. There are additional pre-printing and post-processing variables, too. Continuing the vat photopolymerization example, the printing process may be affected by resin moisture content. Additionally, the extent and conditions of post curing substantially affect the mechanical properties. Therefore, in order to ensure that the final part is of high quality, an in-depth understanding of the material-process-structure-property relationships is essential.  \n\nThe most straightforward approach for understanding material-process-structure-property relationships entails conducting physical experiments wherein a single parameter is altered while the remaining parameters are held constant, thereby enabling the observation of variations in the quality of the printed parts. However, this is not practical due to the sheer number of process parameters involved. Therefore, modeling approaches are frequently adopted. Mathematical modeling is one approach for understanding the AM process. However, it is challenging to develop these models because multiple processing parameters dictate resulting properties, and the relative importance of parameters changes across process steps and with different materials.20 Physics-based models can be used instead, but they require comprehensive domain expertise and substantial computational resources.2- ML techniques provide an alternative approach for constructing predictive models capable of simultaneously handling multiple process and material parameters. Additionally, ML techniques exhibit flexibility in their ability to learn from various data types. Therefore, ML techniques have been increasingly employed in AM, including property prediction, defect detection, quality control, material development, and design for AM.20-23  \n\nAM enables fabrication of parts using many classes of materials. Numerous reviews have concentrated on the application of ML to metal-based AM techniques, while few polymer AM aspects are covered.21-24 Polymer-based AM techniques present distinct challenges. For instance, the molecular structures of polymers exhibit greater complexity compared to metals due to the presence of varying chain lengths within the same material. Part consistency is, therefore, a challenge. In addition, thermal and environmental factors are more likely to alter the properties of polymers compared to metals. To achieve adequate predictive accuracy when implementing ML techniques in polymer AM, it is necessary to account for these nuances. This article provides an overview of the applications of ML techniques specifically for polymer AM. The article is structured as follows: Section 2 provides a brief overview of polymer-based AM techniques. In Section 3, the ML techniques used in polymer AM research are discussed. Section 4 explains the importance of ML techniques to polymer AM. Section 5 provides a comprehensive summary of the existing research that has applied ML to various polymer AM techniques. In Section 6, the challenges associated with the application of ML in polymer AM are discussed, along with types of opportunities in this area.",
        "category": " Introduction"
    },
    {
        "id": 6,
        "chunk": "# 2 | POLYMER ADDITIVE MANUFACTURING TECHNIQUES  \n\nAM techniques have gained considerable attention for their ability to fabricate parts with complex geometries and substantial reductions in material waste. In AM processes, three-dimensional (3D） objects are typically created in a layer-by-layer fashion, directed by a computer aided design （CAD） file. ASTM 52910 classifies according to seven types of additive manufacturing processes: material extrusion (ME), vat photopolymerization (VP), material jetting (MJ), binder jetting (BJ), powder bed fusion (PBF), direct energy deposition (DED), and sheet lamination (SL). The most common polymer-based AM techniques are ME, VP, and MJ. Polymer-based PBF requires powdered feedstocks, limiting material choice because few polymers are available in powder form with the required sintering window.25 BJ is another AM method that uses powdered materials. The powder can be metal, ceramic, or polymer and the powders are bound using a polymer binder.26 Figure 1 includes a summary of the categories and most common subcategories of polymer AM techniques.  \n\n![](images/ec623e901269d444e9f105d2d05cc20ae1f031412814abf3ab6579829cb7f8a4.jpg)  \nFIGURE1 Classification of polymer additive manufacturing techniques.  \n\nFor further information on the development and use of polymer materials in AM, readers are referred to the comprehensive review by Tan et al.27 The following sections include short overviews of polymer-based AM techniques.",
        "category": " Introduction"
    },
    {
        "id": 7,
        "chunk": "# 2.1 Material extrusion (ME)  \n\nME involves dispensing materials selectively through a nozzle or orifice onto a build surface layer-by-layer to create a 3D structure. A wide range of polymer and polymer composites are used as feedstocks for ME. Fused fila ment fabrication (FFF), a prevalent form of ME, uses thermoplastic filaments. Initially, FFF was limited to a few polymers such as acrylonitrile butadiene styrene (ABS) and polylactic acid (PLA).28 However, the technique has evolved to allow for use of many engineered polymers and polymer composites and is capable of producing printed parts with enhanced mechanical,29 thermal,30 and/or electrical performance.31 Big area additive manufacturing (BAAM) is a large format ME method that uses thermoplastic pellets instead of filaments as its feedstock and is designed to fabricate large-scale structures.32 The working principle of BAAM and FFF are similar. However, BAAM retains heat longer than FFF, allowing for better interlayer bonding through diffusion and weld formation, but presents challenges such as material sagging or slumping due to their significant scale differences.33 Direct ink  writing(DIW） uses viscoelastic inks and can print a wide range of polymeric materials including thermoplastics, thermosets, elastomers, hydrogels, and polymer composites. Essentially, any material that has suitable rheological properties can be printed using DIW.34 A relatively new ME technology is ambient reactive extrusion (ARE), in which reactive thermosetting polymers that cure after deposition at room temperature are printed.35 ME applications have greatly advanced, evolving from solely rapid prototyping to producing durable and functional end-use-parts. These parts are found in fields including gaerospace,36 automotive,37 and biomedical sectors.38-40  \n\nThe properties of ME parts are influenced by part design and process parameters.41 For example, typical process parameters in FFF include extruder temperature, layer height, material extrusion rate, raster orientation, raster width, build orientation, and infill. With numerous parameters involved, optimization becomes time-consuming and resource-intensive. Additionally, processstructure-property correlations are often complex and nonlinear, further complicating the optimization process.42",
        "category": " Materials and methods"
    },
    {
        "id": 8,
        "chunk": "# 2.2 | Vat photopolymerization (VP)  \n\nVP consists of selective curing of liquid photo-sensitive material using light/radiation. Light is projected onto the photopolymer, following a pattern defined by a CAD file. This process selectively cures the photopolymer, resulting in the formation of the final part.43 There are many variants of VP, which are categorized based on factors such as light source, speed and resolution, layering method, and build platform position. The two main VP methods are stereolithography (SLA） and digital light processing (DLP). Both techniques use UV light to cure the photo polymer in a layer-by-layer manner to fabricate the final part. The main distinction lies in the method of curing: DLP cures an entire layer simultaneously using digital micro-mirrors devices, while SLA cures pixel-by-pixel within a layer.16 Therefore, DLP offers faster print speed compared to SLA. Other VP techniques include twophoton polymerization (2PP), Volumetric AM, and continuous liquid interface production (CLIP). 2PP is capable of very high resolutions ( ${\\sim}100\\ \\mathrm{nm})$ and is best suited to very small structures.44 Volumetric AM and continuous liquid interface production (CLIP） are both layerless VP approaches.45,46 Comparative discussions of different VP techniques are presented in review articles by Zhang et al.47 and Rashid et al.16  \n\nVP is limited to photo-sensitive polymers. It finds application in the electrical and biomedical fields.16,48 VP is capable of printing a wide range of photopolymer and photopolymer suspensions with high resolution.47 Solid particle reinforcement in vat photopolymerization has gained significant interest due to its potential as an alternative to processes such as PBF or BJ.49-52 This approach involves incorporating metal or ceramic particles into the photopolymer, enabling improved structural and mechanical properties and fabrication of ceramic or metal green bodies.53-55 Typical material systems consist of acrylate or epoxy-based monomers, photoinitiators, diluents, light absorbers, and radical inhibitors.56-58 Low viscosity $(0.25\\mathrm{-}10\\ \\mathrm{Pa\\cdots})$ resins are recommended to avoid damaging printed features during the recoating process.59  \n\nThe structural and surface properties of VP parts are significantly influenced by factors including layer height build orientation, exposure time, and light source intensity.60-62 Resin properties,such as viscosity, reactivity, and photoinitiator concentration, also have great impact on the final part properties.63 Introducing solid particles into the formulation further complicates the rheological behavior.64 This can lead to increased viscosity,65 reduced cure depth due to scattering,66 inhomogeneous particle distribution due to sedimentation or creaming,67 which may cause print failures. Achieving defect-free parts requires adjusting both process and material parameters, which often relies on experimentally costly trial-and-error and/or  intuition-based methods.",
        "category": " Materials and methods"
    },
    {
        "id": 9,
        "chunk": "# 2.3 Material jetting (MJ)  \n\nMJ is a popular AM method known for its ability to print highly complex structures with multiple materials, offering high dimensional accuracy and low surface roughness. The process may use UV-curable inks in a manner similar to an inkjet printer. Wax can also be printed with the MJ process and is commonly used as a support material.68.69 Droplets of inks are ejected through multiple nozzles onto a substrate and, if necessary, subsequently cured using UV-light. Based on the droplet dispensing method, MJ processes can be classified as either continuous inkjet (CI） printing or droplet-on-demand (DoD) printing.7° It is important to note that commonly used commercially available photopolymer inks have poor mechanical and thermal properties.7l Therefore, MJ cannot be used to print structures for applications involving heavy loads.72 However, the technology is valuable in the realm of functionally graded materials, making it well-suited for sophisticated applications such as bioprinting73 and printed electronics.74,75 For further exploration of recent developments and advanced applications of MJ, readers are referred to review articles by Elkaseer et al.72 and Guilcan et al.71  \n\nMJ is an intricate process that requires careful optimization of process parameters to achieve high quality prints. The existing body of literature on process parameter optimization in the field of interest is limited. Bass et al. investigated the impact of part orientation on the MJ-produced parts, highlighting its significance in the printing process.76 Pugalendhi et al. explored the effect of MJ process parameters on the mechanical properties of the printed objects.77 Achieving high resolution printed structure requires precise control of ink droplet size as well as the droplet spread on the surface. Moreover, the optimization of curing strategies is also crucial, yet it has not been extensively explored. This aspect becomes more complicated for multi-material MJ process, as different materials may require different curing strategies.",
        "category": " Results and discussion"
    },
    {
        "id": 10,
        "chunk": "# 2.4 I Powder bed fusion (PBF)  \n\nPBF selectively fuses powdered material from a powder bed in a layer-by-layer manner, typically using thermal energy provided by a laser beam. Selective laser sintering (SLS) is a more common term for processing polymer powders. The key advantage of SLS is the design freedom attribued to not requiring any support structure.78,79 Custom designed automotive and aerospace parts as well as functional implants have been fabricated via SLS.80-82  \n\nAs compared to other polymer-based AM techniques, SLS has a limited range of available materials. The primary material used for SLS is polyamide 12 (PA12) due to its appropriate sintering window (between melting and crystallization） and free flowing behavior facilitated by highly spherical particles with precise particle size distributions (PSDs).25 Other commercially available polymers used in SLS include other polyamides, ABS, polyaryletherketones (PAEK), thermoplastic elastomers (TPE), polypropylene (PP), and polystyrene (PS). Numerous process parameters control the final part quality in SLS. Han et al. categorized SLS experimental parameters as either laser or build parameters while discussing the effects of these parameters on part property and quality.79 Similar to other AM processes, SLS optimization is complicated due to the high dimensionality of its parameter space.",
        "category": " Materials and methods"
    },
    {
        "id": 11,
        "chunk": "# 2.5 Binder jetting (BJ)  \n\nBJ is a powder-based AM technique in which a liquid binder is selectively dispensed onto a powder bed in the X-Y plane to create a two-dimensional (2D） pattern. These patterns are then repeated in the Z direction to construct a complete 3D structure of bonded powder par ticles, known as a“green\" structure. The green part can be used directly, sintered, or infiltrated with other materials.83 BJ is frequently used to fabricate sand molds and cores with complex geometries.84,85 The porous nature of the printed structures is particularly favorable because the pores may act as channels for gas transport and simple release of the molded part.85 Early works also showed the efficacy of BJ for fabricating drug delivery devices.86  \n\nThe key to BJ's success is its versatility with a variety of powdered materials, including metal, ceramic, and polymer. Metal and ceramic powders are commonly used.87,88 Research on polymer powders in BJ is limited because powder forms are not commonly used in the production of polymer structures.26 The choice of binder material is crucial because it directly influences print success and the properties of the final part. The binder needs to have sufficient wettability to the powder. Many polymeric materials have been used as binders, including polyvinyl alcohol (PvA),89,90 polyacrylicacid (PAA),% cellulose derivatives,91 and waxes.92  \n\nNumerous process and material parameters influence the dimensional precision and final properties of BJ parts. First and foremost, it is essential to maintain the powder bed density across all layers in order to produce parts with uniform properties. Uniformity in powder beds is crucial for consistent binder jetting quality, but variations can occur due to roller movement, leading to differences in green part density and shrinkage during debinding. 93 Fine powders, with micron and sub-micron dimensions, frequently aggregate.94 In addition, layer thickness, drying power level and drying time of the binder, and powder spreading speed affect printed part quality.95,96",
        "category": " Results and discussion"
    },
    {
        "id": 12,
        "chunk": "# 3 | MACHINELEARNING TECHNIQUECATEGORIESAND TASKSINPOLYMERADDITIVE MANUFACTURING  \n\nML is a sub-field of artificial intelligence (AI) that employs a data-based modeling approach to uncover patterns within a dataset.97 By extracting these patterns, ML algorithms can make predictions for previously unseen cases. AM techniques are complex, and the end part quality is influenced by numerous material and pro cessing parameters as discussed in Section 2. This complexity creates an opportunity for ML techniques to reduce the time and resources required to evaluate process-structure-property-performance relationships as compared to purely trial-and-error-based experiments, numerical, and analytical models. While ML has found extensive applications in metal AM, its use in polymerbased AM techniques is still emerging and will be discussed in detail in Section 5.98-104",
        "category": " Introduction"
    },
    {
        "id": 13,
        "chunk": "# 3.1 Machine learning tasks  \n\nML models can perform tasks such as regression, classification, clustering, and dimensionality reduction.Each of these types of tasks is described in greater detail below, including applications in AM.  \n\nRegression tasks focus on predicting numerical or quantitative outcomes. Therefore, the main objective of regression-based ML modeling is to minimize the difference (error) between the predicted and actual outputs, thereby finding accurate input-output relationships.105 Regression-based modeling approaches have been used to predict properties of AM structures, including compressvestrngthnileis rougness,12 and hardness.113 In theseexamples, process parameters were used as model inputs.  \n\nClassification tasks focus on generating decision boundaries between predefined classes based on patterns learned from the training data.l14 Classification-based modeling techniques have been extensively used in real-time/in situ monitoring of AM processes.115 In situ monitoring with classification techniques includes defect detection,16 process anomaly detection,1,i18 and product quality prediction.119 These tasks often involve training ML models with images.  \n\nClustering tasks involve grouping similar data points together based on their intrinsic properties.120 Clusteringbased ML techniques have been applied to in-situ monitoring of AM process for defect detection,12l failure mode detection,122 and process monitoring.123  \n\nLastly, dimensionality reduction entails reducing the number of input features, or data dimensionality, while preserving as much variance or information as possible.124 Polymer AM-related problems may have numerous input variables, but not all of them have the same level of impact on the parameter being investigated. Hence, dimensionality reduction techniques can be useful for enhancing the efficiency and interpretability of models by identifying the most important or relevant input features. These techniques have been used to investigate process parameter-property correlations42 and the selection of printed formulations.125",
        "category": " Results and discussion"
    },
    {
        "id": 14,
        "chunk": "# 3.2| Categorization of machine learning techniques  \n\nThere are various ways to categorize ML techniques. In this review, we categorize ML techniques according to data supervision and model complexity for aiding further discussion of their applications in polymer AM.",
        "category": " Results and discussion"
    },
    {
        "id": 15,
        "chunk": "# 3.2.1 Data supervision  \n\nUsing data supervision as the metric, ML models can be categorized as being supervised, unsupervised, or semi-supervised, as shown in Figure 2. Supervised ML techniques train models using labeled data.126 The term \"labeled\" means that the data points include known targets or outcomes. Supervised models are primarily used for regression and classification tasks. In contrast to supervised models, unsupervised models learn from unlabeled data without the guidance of predefined labels or outcomes.127 Unsupervised techniques discover  the hidden pattern in the dataset. They are popular for tasks such as data clustering and reducing dimensionality in the dataset.  \n\nSemi-supervised ML techniques use principles of both supervised and unsupervised approaches.128 Semisupervised models start with a dataset where a small number of data points are labeled, and a large portion is often unlabeled. The labeled data are first used to train the ML model. Then, the model learns iteratively from the unlabeled data. Based on interactions with labeled and unlabeled data, semi-supervised techniques can be categorized as active learning (AL), passive learning, or self-training. AL queries an external source, typically a human expert, for labeling the most informative data points.129 Passive learning, on the other hand, uses both labeled and unlabeled data without actively searching for new labels, while self-training uses its own predictions to label and iteratively retrain using unlabeled data.130 Semi-supervised approaches are typically used for regression and classification tasks. Semi-supervised approaches are used in AM when dealing with large design spaces, when labeling processes are impractical, or when experiments are costly.131",
        "category": " Results and discussion"
    },
    {
        "id": 16,
        "chunk": "# 3.2.2 Model complexity  \n\nDepending on the complexity of the architecture, ML techniques can be classified as being either shallow or deep, as shown in Figure 3. In general, shallow ML models have simple architectures, whereas deep learning models have complex architectures that are capable of identifying complex data patterns. The ability to discern intricate and hierarchical patterns from data is a key distinction between shallow and deep models. A shallow model's input features are manually extracted, while deep models can automatically learn feature representations from raw input data.132 One additional distinction between shallow and deep learning models is the types of data they can process. Shallow models are typically limited to structured data such as tabular data. Deep learning models can handle both structured and unstructured data, including but not limited to tabular data, image data, time series or sequential data, text data, audio data, and video data.133  \n\n![](images/696bbe6e5e3fbf9540571927a8607e789b4999f14487b709a295a938fd1a51d6.jpg)  \nFIGURE 2 Classification of machine learning models based on data supervision approach.  \n\n![](images/37f995507d11086f23edc25dad94df36b8395f9f73a0dd1f18fa4f54c0f126a4.jpg)  \nFIGURE3 Classification of machine learning models based on model complexity.",
        "category": " Results and discussion"
    },
    {
        "id": 17,
        "chunk": "# Shallow models  \n\nShallow models can be parametric or non-parametric. Parametric models are characterized by their strong assumptions regarding the relationship between input features and target variables or outcomes. These assumptions typically involve a fixed number of coefficients or weights. Consequently, the flexibility of parametric models in capturing complex patterns is limited. Nevertheless, these models exhibit simplicity and efficiency, especially when data availability is limited. Several parametric models commonly used in polymer AM include linear regression (LiR), multivariate linear regression (MLR), multiple regression analysis (MRA), ridge regres sion (RR), logistic regression (LoR), polynomial regression (PR), Naive Bayes (NB), and Gaussian mixture model (GMM).  \n\nLiR, MLR, MRA, RR, LoR, PR, and NB are supervised techniques, whereas GMM is unsupervised. However, they can all be adapted to semi-supervised techniques. Typically, adaptation involves iterative methods that use labeled data to infer labels for unlabeled data. The newly labeled data are then incorporated into the model's training process.  \n\nLiR, MLR, MRA, RR, and PR are commonly employed to conduct regression analysis.134 LiR, MLR, RR, and MRA model the relationship between input features and output variables in a linear manner. MLR and MRA are straightforward extensions of LiR, while RR and PR introduce additional complexity. The primary distinction among LiR,MLR,and MRA lies in the number of variables used. LiR uses a single independent variable (input) to make predictions about a corresponding dependent variable (output). MRA employs two or more independent variables in order to make predictions about a single dependent variable. MLR, in contrast, uses multiple independent variables to make predictions about multiple dependent variables. PR is an extension of LiR.However, PR incorpo rates higher-degree polynomial terms of the input features, allowing the model to capture more complex, non-linear relationships in the data that would be impossible to capture with linear models. RR, another extension of LiR, uses a regularization term to address multicollinearity (highly correlated input variables) and overfitting issues. RR can capture nonlinear inputoutput relationships by employing polynomial features and nonlinear transformations.  \n\nLoR and NB algorithms are classification techniques. LoR assumes a linear relationship between the input features and the logarithm of the output class's probabilities. It is frequently applied to continuous data. NB calculates class probabilities based on the distribution of input features within each class, under the assumption that all input features are independent. It is especially useful for categorical data.  \n\nGMM is a probabilistic technique employed for clustering tasks. Typically, it is used to model continuous data. GMM assumes that the data are derived from a mixture of several Gaussian distributions. Each distribution has its own mean,variance (or covariances in the case of multivariate distributions), and weight. These parameters are estimated using the expectationmaximization algorithm.  \n\nSupport vector methods are known for their versatility because they are capable of effectively handling both regression and classification tasks. These methods are especially suited for continuous or categorical data with high dimensionality. The designation “Support Vector Machine\" (SVM) denotes its use as a classification method, while “Support Vector Regression” (SVR) signifies its use in regression analysis. Support vector methods seek an optimal hyperplane that best separates (classification task) or fits (regression task) the data.  \n\nIn contrast to parametric models, non-parametric models do not assume a predefined mapping or distribution for a given problem, which makes them more flexible, adapting to the characteristics of the data. Common non-parametric techniques used in polymer AM research include decision trees (DT),135 K-nearest neighbors (KNN),136 K-means clustering (KMC),137 Gaussian process (GP),138 principal component analysis (PCA),139 and t-distributed stochastic neighbor embedding (t-SNE).125 KMC, PCA, and t-SNE are unsupervised ML techniques, while DT and KNN are supervised ML techniques. GP can be used in both supervised and unsupervised settings. Although all models mentioned can be used for semisupervised learning with proper adaptation, only KNN is inherently suited for this task.  \n\nDTs are algorithms that support both continuous and categorical data for classification and regression tasks. They are capable of capturing complex and nonlinear relationships, which is why they are widely used in for polymerAMIDstais edly based on input features until a termination criterion is met.Each split corresponds to a decision node and terminal nodes (leaves). Leaves represent predicted outputs, which may be class labels or continuous values. DTs are intuitive and resemble human decision-making. They are versatile and powerful tools that can be used effectively as base learners in ensemble learning algorithms (ELA), where the predictions of multiple models can be combined to improve the overall performance and robustness of final prediction. DT-based ensemble methods used in polymer AM and discussed in Section 5 include random forest (RF), extremely randomized trees (EXTr), AdaBoost (ADA), gradient boosting (GB), extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). Notably, while RF and EXTr are required to use DTs as the base learner, other ensemble techniques are not.  \n\nGPs are probabilistic modeling techniques that are used to model continuous data for regression (GPR） and classification (GPC) tasks. GPs assume that the underlying data-generating process is a Gaussian process, which provides a distribution over all possible functions.142 Therefore, when this distribution is used to make predictions for new data points, it can also provide a measure of the uncertainty associated with those predictions.  \n\nKMC and KNN share similarities in terms of handling data, as both use distance metrics. In addition, both methods are applied to continuous and categorical data. Despite similarities, their applications are distinct. KMC is used for clustering tasks, while KNN is used for classification and regression tasks. KMC uses unlabeled data to find patterns, whereas KNN uses labeled data to make predic tions. KNN bases its predictions on the majority class or average value of the K nearest training data points. In contrast, KMC divides the data into K clusters where the center of a cluster is the average of its data points.  \n\nPCA and t-SNE are two methods for reducing the dimensionality of a dataset while retaining as much information as possible.143 However, their approaches to the task vary. PCA uses linear transformations to project data onto a subspace with fewer dimensions. It seeks to preserve the global structure and variation of the dataset. t-SNE, on the other hand, models nonlinear relationships between data points by preserving local structure and relationships in a lower-dimensional space.",
        "category": " Results and discussion"
    },
    {
        "id": 18,
        "chunk": "# Deep learning models  \n\nBased on architecture, input data type, and application, deep models can be categorized as multilayer perceptron (MLP), recurrent neural network (RNN), or convolutional neural network (CNN). These models are generally described as artificial neural networks (ANN). The architectures of each of these model types are shown in Figure 4.  \n\nANN models are capable of capturing highly complex and non-linear relationships between inputs and outputs.144,145 The models are inspired by the structure and functionality of the human brain. ANN models generally consist of three types of layers: an input layer, one or more hidden layers, and an output layer.146 Each layer is comprised of several nodes, also known as neurons. The information passes from one layer to another through the connections that link the neurons. Each connection in a neural network has a numerical parameter, known as its weight, representing the emphasis given to a partic ular input feature. These weights are iteratively updated by an optimization algorithm during the training process to improve prediction accuracy. The neurons also contain an additional parameter called bias, which allows the model to fit the data flexibly.147  \n\n![](images/f703122f12899bc2e7cdb769fb67e596e6de5e39870424379ff8ee4dd56d731d.jpg)  \nFIGURE 4Generalarchitectures ofdifferentdeeplearming models.(A)Multilayer perceptronfor predicting polymerAMprint properties;(B)RecurenturaletwrkforqualitycontolinpolymerAMprints;(C)Convolutionalnuraletworkfordefectdetectionin a polymer AM process.  \n\nMLPs are fully interconnected feedforward networks As shown in Figure 4A, every neuron in one layer is connected to every neuron in the following layer. MLP is well-suited for structured data such as tabular data. MLP is also capable of working with unstructured data that does not contain spatial or temporal structure, such as flattened images.  \n\nIn contrast to MLPs, which are completely feedforward networks, RNNs have connections that loop back on themselves, as shown in Figure 4B, allowing them to maintain a “memory\" of previous inputs in their internal state. RNNs are, therefore, ideally suited for sequential data with temporal structures. However, long sequences make it challenging for a network to transmit information from one end of the architecture to the other, resulting in unstable training. Long short-term memory (LSTM), a subtype of RNN, has a unique structure for remembering long sequences.148 The choice between standard RNN and LSTM depends on the sequence type of the dataset. Standard RNN is adequate when the output is dependent on the most recent elements in the sequence. On the other hand, LSTM may be considered for long-term dependencies, that is, when the output depends on elements that appeared in the sequence along timeago.149  \n\nCNNs are designed to process data with spatial structures, such as images. The arrangement of pixels in rows and columns, as well as the relative position of different colors and intensities, convey crucial information in an image. As depicted in Figure 4C, CNNs are composed of convolutional layers that apply convolutional filters (also known as kernels) to the input data, capturing local image features such as edges and textures to produce feature maps. After the convolutional layers, CNNs often incorporate pooling layers, which reduce the size and complexity of the feature maps to save computational time and energy.  \n\nANNs exhibit a high degree of customization. These tools have the potential to be modified to suit a diverse range of tasks. MLP, RNN, CNN have been widely employed for various tasks such as regression, classification, and dimensionality reduction.150-157 However, their application in clustering tasks is not as prevalent. Autoen coders (AE) are frequently employed in conjunction with MLP, RNN, and CNN to carry out tasks related to dimensionality reduction.158 The utilization of self-organizing maps (SOMs), a variant of ANN, is prevalent in the domain of clustering tasks.159 sOM is not classified within the overarching framework of deep learning methods due to the absence of multilayered architectures.",
        "category": " Results and discussion"
    },
    {
        "id": 19,
        "chunk": "# 3.3 | Performance metrics for machine learning models  \n\nVarious performance metrics are used to evaluate predic tive performance, modify hyperparameters, and make decisions regarding the selection of ML models. The selection of evaluation metrics is dependent upon the task being carried out. Some of these performance metrics will be introduced in this section to facilitate discussion in subsequent sections.  \n\nRoot mean squared error (RMsE),mean absolute error (MAE), coefficient of determination $(\\mathrm{R}^{2})$ ,and relative error (RE) are used to evaluate the performance of ML models for regression tasks. RMSE and MAE both quantify the disparity between predicted and observed values.However, RMSE is more sensitive than MAE to outliers.160 Low RMSE and MAE values signify performance excellence. ${\\mathrm{R}}^{2}$ values range from O to 1 and represent how well the model fits the dataset. A value closer to 1 represents a superior fit. RE is beneficial when the measured quantities vary greatly in magnitude. Similar to RMSE and MAE values, a low RE value denotes excellent predictive performance.  \n\nF1 score is commonly used as a classification performance metric. F1 score is used when class distributions are unbalanced. Similar to $R^{2}$ values, Fl scores range from O to 1, with values closer to 1 indicating a higher degree of accuracy when identifying a particular class (true positive or true negative).",
        "category": " Results and discussion"
    },
    {
        "id": 20,
        "chunk": "# 4丨THENEEDFORMACHINE LEARNINGINPOLYMER ADDITIVE MANUFACTURING  \n\nDespite advances in polymer AM, several roadblocks continue to hinder its broader adoption. One major limitation is the relatively narrow range of available materials, particularly when compared to subtractive polymer manufacturing processes, which limits possible applica tions. Additionally, the success of polymer AM processes relies heavily on numerous process and material parameters, making optimization a complex and timeconsuming task. Compared to metals, polymers are less recyclable. Therefore, a key objective during the design phase is to minimize the use of support structures to reduce material waste; this often depends on trialand-error and/or simulations, both of which are time consuming.161 ML techniques are gradually being incorporated to help navigate these challenges. In this section, we discuss the potential of ML techniques to accelerate advances in polymer AM.  \n\nAs AM techniques have progressed from rapid proto typing methods to large scale manufacturing processes capable of creating functional end-use parts, there has been a significant increase in research focused on using a wider range of polymers and polymer composites. Over the past decade, these efforts have been aimed at broadening the application in sectors such as automotive, aerospace and biomedical industries.162 A common aspect of material discovery or screening materials for the AM techniques is that it requires a series of experiments and/or physics-based simulations in order to characterize and validate the choice. Hence, only a few among the vast array of available polymer materials have been employed in AM.  \n\nDue to continued progress in experimental and simulation-based approaches, numerous material databases, such as ChemSpider and MatWeb, now contain an enormous amount of material data for physical, mechanical, and chemical properties.163,164 However, the data are mostly unsorted and accompanied by a large amount of variance, making analysis challenging using only subject-matterexpertise. ML techniques can be used to extract meaningful structure-property correlation from these data, thereby providing recommendations for new materials for AM techniques that are likely to result in improved part quality. This approach can be useful for both the development of novel materials and screening of existing ones.  \n\nAM techniques, which involve many process parameters, are difficult to optimize for achieving desired part properties. Traditional full factorial design of experiments (DoE) is infeasible for exploring the design space due to high experimental cost. Advanced DoEs, such as Taguchi,i65 response surface methodology (RSM),166 and Latin hypercube sampling (LHS)l67 are useful in reducing the total number of experiments. However, these methods only consider the input design space, not the process parameter-property correlations that are crucial for optimization tasks. ML techniques, on the other hand, take into account input-output correlations and therefore have the potential to accelerate the optimization process. Theoretical models derived from fundamental principles are too complex to solve analytically when an array of material and process parameters is involved. Additionally, physics-based models may capture the behavior of the AM system accurately due to considering the under lying physics. However, they are computationally intensive and often time-consuming. ML can be combined with physics-based models to reduce the computational cost and streamline the optimization process.  \n\nWhile ML techniques can extract meaningful information from high-volume data, they can also be useful when it comes to exploring a high-dimensional parameter space, particularly in solving an optimization problem cost-effectively by guiding the sampling only from the promising regions, rather than sampling the entire space, such as AL. Lookman et al. provided a comprehensive guide to using AL in material science.168 The overall goal of AL is to find an accurate predictive model without needing to train with a large volume of data, which is typically required for supervised ML techniques, as the model can intelligently learn from most informative instances.  \n\nThe use of AM techniques to generate complex structures is widespread. For printing to be successful, complex designs require support structures. During the design optimization phase, one objective is to minimize the support structure in order to reduce post-processing steps. Topology optimization (TO) can be undertaken to generate an optimal design with reduced material while still ensuring substantial performance.169 However, TO often suffers from being computationally expensive and may recommend designs that are impractical to implement.17o,171 ML techniques have great potential to be synergistically applied with TO to attain good performance with reduced computational cost.For example, ANN techniques such as CNN-integrated TO were reported to be more efficient than traditional TO without compromising accuracy.172",
        "category": " Results and discussion"
    },
    {
        "id": 21,
        "chunk": "# 5 | APPLICATIONOF MACHINE LEARNING IN POLYMER ADDITIVE MANUFACTURING  \n\nThis section summarizes the application of ML techniques to various polymer AM-related tasks. The organization is as follows: 1. Section 5.l summarizes the literature that employed ML techniques in ME; 2. Section 5.2 summarizes the ML-related research in VP; 3. Section 5.3 summarizes the literature based on PBF; and 4. Section 5.4 summarizes the application of ML techniques in inkjet-based AM, including BJ and MJ. Common ML-related tasks for polymer AM include property prediction, process optimization, and in-situ monitoring. Figure 5 provides a comprehensive guide for employing ML techniques for the aforementioned tasks. The selection of ML techniques is primarily determined by the collected data type and dataset size. Small tabular and sensor-based datasets (manual pre-processing) are frequently analyzed with shallow models for property prediction and process optimization tasks. In these instances, the input features are used directly.However, when datasets are large and appropriate features must be extracted from the raw dataset, deep learning models, specifically MLPs, can be employed. CNNs are specially designed for processing image-based data and are primarily used for in-situ monitoring, although process parameter optimization tasks have been performed in certain scenarios.173-176 For in-situ monitoring and process parameter optimization, RNNs are commonly used to process spatiotemporal data, such as data collected from recorded videos. Nevertheless, image segmentation tasks for processing video data are handled uniquely by convolutional layers, which is why RNNs are frequently combined with CNNs. Notably, sequence-based data can be processed by RNNs for any task.",
        "category": " Results and discussion"
    },
    {
        "id": 22,
        "chunk": "# 5.1 | Machine learning for material extrusion  \n\nAmong all ME-based AM techniques, FFF has made the most extensive use of ML. This is likely due to the wide use of FFF. Additionally, FFF feedstock materials are commercially available and can be used as received. ML applications in DIW are relatively new.177-179 In this review, literature reporting the use of ML for ME are classified according to their broad motivation and are summarized in Table 1. The primary focus has been on predicting mechanical properties and surface quality of the printed parts, reducing experimental effort for process parameter optimization, and in-situ monitoring for online defect detection. ML has also been proven useful in detecting cyber security attacks and reducing the computational time and cost of physics-based simulations.  \n\n![](images/f76297d4395b08d6c2f84dc96167ae8aa2e2696b26ea3d7a40cd74fa90871bfa.jpg)  \nFIGURE5Aguide forapplying machinelearning techniques tocommon types of tasks assciated withpolymer additive manufacturing based on the type and amount of data available.",
        "category": " Results and discussion"
    },
    {
        "id": 23,
        "chunk": "# 5.1.1 Predicting properties  \n\nThe mechanical and surface properties of ME-based printed parts largely depend on process and design parameters such as extrusion temperature, print bed temperature, print speed, infill density, infill pattern, raster orientation, and layer height. In order to optimize the performance and appearance of the printed parts, it is impractical to depend solely on trial-and-error based methods to tune the process and design parameters. Researchers have taken various ML-based approaches to reduce the experimental effort required to understand the correlations between process and/or design parameters and part properties. In the domain of predicting part properties, training datasets typically consist of empirical data from physical experiments. Data acquired from sensors such as thermocouples and accelerometers can also be used to train. Experiments are often conducted systematically based on advanced DoEs such as Taguchi, RSM, and LHS so that the defined design space is explored efficiently and the data points are informative toward the training process of the ML models.  \n\nShallow and deep learning models have been used for predicting mechanical properties of the ME-based printed parts. ANN models often yield higher predictive accuracy due to their capacity to capture complex non-linear process parameter-property relationships, which are common in AM. However, due to the high level of design freedom, ANN models are typically more complex and computationally expensive than shallow models. Sharma et al. compared the accuracy of shallow models such as RF, KNN, ADA, and DT with LSTM for predicting tensile and flexural strength of polydopamine (PDA)-coated poly(lactic acid)(PLA） bone plates fabricated with FFF.135 The authors investigated the effect of process parameters including infill density of the base PLA structure, immersion time of the PLA structure in a PDA solution, incubator shaking speed for the coating solution, and coating solution concentration. The ML models used an experimentally generated training dataset consisting of 100 data points, where the process parameters were considered as input features and corresponding tensile and flexural strengths were considered as outputs. The work reported a significant gain in prediction accuracy for the LSTM model compared to the traditional ML models. For example, predictive performance for tensile strength as $R^{2}$ for RF, KNN, ADA, and DT were 0.7425, 0.7217, 0.7191, and 0.695l, respectively,whereas $R^{2}$ for LSTM was 0.9242. However, training ANN models with limited data may result in overfitting due to lack of generalizability for unseen data points.196  \n\nSurface roughness is an important property since it greatly affects mechanical and optical properties of AM parts. Thus, multiple studies have investigated the surface roughness of ME parts. Surface roughness of printed parts depends on parameters such as layer thickness, raster orientation, print speed, print bed temperature, infill pattern, and infill density. Several experimental studies197 and analytical models19s,199 have been used to estimate surface roughness of AM parts. However, ML-based modeling methods present an alternative approach to roughness determination since they are computationally less demanding than numerical techniques such as finite element analysis (FEA） and are capable of reducing experimental effort.  \n\nTraining of ML models for predicting surface roughness of ME parts has been achieved with both sensor signals and experimental data acquired by varying multiple process parameters. Li et al. proposed a data-based hybrid modeling technique for FFF, where the training process takes place offline with temperature and vibration data from in-situ sensors while the prediction of the surface roughness takes place online.182 The authors used RF to select important features from the processed sensor data to reduce computational time and to avoid overfitting. The RF selected features were then fed into an ELA consisting of six ML models to develop a predictive model for surface roughness, which was later employed in online prediction. Based on RMSE and RE values, the authors reported that the overall predictive performance of the ensemble algorithm is higher than its base models. Nevertheless, the simpler constituent models, RR and SVR, had very close RMSE and RE values to the ELA, indicating the potential to bypass the computational complexity. It is worth noting that ELAs can be important for providing improved accuracy and better generalizability compared to the constituent models. However, they are often computationally intensive and somewhat less interpretable than the simpler models. Hence, the tradeoff between accuracy and interpretability must be considered carefully when ML techniques are being employed for property prediction in AM structures.  \n\n(teseentn   \nreneaeeeenereeeneereeeenereeeereeeieeee TT   \n\n\n<html><body><table><tr><td> Broad motivation</td><td colspan=\"3\">AM technique ML technique</td><td> Remarks</td><td>References</td></tr><tr><td rowspan=\"10\">Property prediction</td><td>FFF</td><td>RF, KNN, ADA, DT, and LSTM</td><td>Infill density, submersion time, shaker speed, and coating solution concentration</td><td> Predicting tensile and flexural strengths</td><td> Sharma et al.135</td></tr><tr><td>FFF</td><td>LiR, GPR, RR, and KNN</td><td>Extruder temperature and layer height</td><td>Predicting five critical tensile properties (Young's modulus, yield stress, yield strain, tensile stress,</td><td> Nasrin et al.129</td></tr><tr><td>FFF</td><td>LiR, DT, RF, and</td><td> Infill density, layer thickness, print orientation,</td><td>and tensile strain) Predicting hardness</td><td> Veeman et al.113</td></tr><tr><td>FFF</td><td>ADA RF</td><td>and raster orientation Infill pattern, infill density, and the number of sprayed layers</td><td>Predicting ultimate flexural strength, fracture flexural strength, strain at peak and strain at</td><td> Ranjan et al.180</td></tr><tr><td>FFF</td><td> MLP</td><td>Bed temperature, printing speed, layer thickness, and orientation angle</td><td>break Predicting surface roughness</td><td>Malleswari et al.181</td></tr><tr><td>FFF</td><td>ELA</td><td>Build plate temperature and vibrations, extruder temperature and vibration, and temperature of</td><td>Predicting surface roughness</td><td>Li et al.182</td></tr><tr><td>FFF</td><td>KMC, LiR, and MLP</td><td>the deposited material Layer thickness, fan speed, and infill density</td><td>Predicting surface roughness</td><td>Si et al.183</td></tr><tr><td>FFF</td><td> MLP</td><td>Extruder temperature, infill percentage, and layer thickness</td><td>Predicting toughness, part thickness, and production cost</td><td> Meiabadi et al.184</td></tr><tr><td>FFF</td><td>MLP</td><td>Printing speed, extrusion temperature, infill density, extruded filament thickness, extrusion orientation</td><td>Predicting tensile strength</td><td> Silva et al.185</td></tr><tr><td>Process parameter FFF optimization</td><td>CNN and RF</td><td>Material extrusion rates and extrusion temperatures</td><td>Correlate the process parameters with the quality of printed parts (surface roughness, hardness,</td><td> Butt et al.173</td></tr><tr><td rowspan=\"5\"></td><td>FFF</td><td>CNNs</td><td> Infill type, density, material, wallthickness, layer</td><td>and tensile strength) Predicting optimal input parameters for user defined mechanical properties</td><td> Ratnavel et al.174</td></tr><tr><td>FFF</td><td> LiR and PR</td><td>Layer height, printing speed and printing bed temperature for the coating layer</td><td>Finding most influential coating layer print parameter for ultimate shingle-lap shear</td><td> Belei et al.186</td></tr><tr><td>FFF</td><td> GPR</td><td>Tensile testing and surface imaging data</td><td>strength Optimizing print parameters to achieve superior surface quality</td><td> Liu et al.187</td></tr><tr><td>FFF</td><td>RF, SVM, LoR,</td><td>Formulation data from literatures</td><td> Predicting processng parameters and printability</td><td>Castro et al.136</td></tr><tr><td></td><td> KNN, and MLP</td><td></td><td>for additively manufactured drugs based on literature-mined formulations</td><td></td></tr></table></body></html>  \n\n(panunuon) TTRE   \n\n\n<html><body><table><tr><td>Broad motivation</td><td> technique</td><td>ML technique</td><td>Model inputs</td><td>Remarks</td><td> References</td></tr><tr><td></td><td> FFF</td><td>GP</td><td>Temperature history from heat transfer analysis</td><td>Optimizing print parameters for reducing geometrical inaccuracy</td><td>Nath et al.138</td></tr><tr><td></td><td> FFF</td><td> SOM</td><td> Laser scanning data from the surface of test parts</td><td>Understanding correlations between process parameters and geometric accuracy</td><td>Khanzadeh et al.188</td></tr><tr><td></td><td> DIW</td><td>PCA and SVM</td><td>Pivotal lines (areas providing mechanical support) and transition points (areas not bearing stress after printing)</td><td>Optimizing print parameters for newly developed ink </td><td> Zhu et al.139</td></tr><tr><td></td><td> DIW</td><td>KMC, SVM, GPR</td><td> Sheath gas flow rate andcarrergas flowrate</td><td>Improving the quality of aerosol jet printing by optimizing the deposited droplet morphology</td><td> Zhang et al.137</td></tr><tr><td>In-situ monitoring for anomaly/defect detection</td><td>FFF</td><td>CNN</td><td>Image data from printhead</td><td>On-site monitoring system for detecting defects (under-extrusion and over-extrusion) and make corrections in real-time</td><td>Goh et al.189</td></tr><tr><td></td><td>FFF</td><td>RNN and ADA</td><td> Sensor signal from side channel</td><td>Sensor-based anomaly detection during unintended process/product alterations caused by cyber-security attacks</td><td>Shi et al.190</td></tr><tr><td></td><td>FFF</td><td> SVM, NBC, and DT</td><td>Combined feature vector of acoustic emission and point cloud data</td><td>Real-time monitoring based on acoustic emission and laser technology for monitoring warpage defect</td><td>Xu et al.118</td></tr><tr><td></td><td>FFF</td><td>RNN</td><td> Simulation results from digital twin</td><td>Automated clogging detection by predicting extrusion rate, extrudate temperature, and</td><td> Rossi et al.191</td></tr><tr><td></td><td>FFF</td><td> SVM and CNN</td><td>Digital imaging</td><td>compression force acting on the filament Detecting anomalies at the topographic level</td><td> Rossi et al.192</td></tr><tr><td></td><td>FFF</td><td>SVM, KNN, and CNN</td><td>Image data from 3D laser scanning</td><td>Laser-based process monitoring system for assuring print quality</td><td>Lyu et al.193</td></tr><tr><td></td><td>FFF</td><td>MLP</td><td>Thermal data from a physics-based model</td><td>Surrogate modeling replicating thermal profile simulation</td><td> Roy et al.194</td></tr><tr><td></td><td>FFF</td><td>PCA, SVM, and CNN</td><td> Image data from video recording during printing</td><td> Automated classification of print quality during</td><td>Narayanan et al.195</td></tr></table></body></html>  \n\nSurface roughness has been more commonly predicted based on process parameters such as layer thickness, infill density, raster orientation, and print speed.181183Asthedatasetisoftenexperimentally acquired, researchers leverage advanced statistical DoE techniques to reduce the number of total experiments. For example, Malleswari et al. used Taguchi method, which was useful to reduce experimental time and cost for generating the dataset by using orthogonal arrays. S.181 In this work, the authors explored the effect of four processing factors on surface roughness: print bed temperature, printing speed, layer thickness, and raster orientation. Each processing factor included three levels, which would require conducting a total of 81 experiments with full factorial DoE. The Taguchi method helped to reduce the required experiments by three-fold. The acquired data was informative toward the training process evidenced by $R^{2}$ values of 0.993 and 0.994 for RSM and MLP, respectively.",
        "category": " Results and discussion"
    },
    {
        "id": 24,
        "chunk": "# 5.1.2 | Process parameter optimization  \n\nML techniques have been proved to be useful to optimize process parameters to obtain superior qualities in the printed parts. For example, dimensional inaccuracies are a common problem in AM parts. ML techniques have been used to tune process parameters to reduce dimensional inaccuracies.i38,188 Khanzadeh et al. reported that geometric deviations are correlated with extruder temperature and infill percentage. Thus, these parameters can be optimized to attain higher geometric accuracy.188 This work used a large dataset consisting of laser-scanned coordinates of the parts printed at varying extruder temperature and infill percentage. In order to measure the geometric deviations, the laser scanned coordinates were compared to the original CAD design of the part. The geometric deviations were then clustered using SOM based on their similarity in shape deviations, both in direction and magnitude. The clusters were then ranked based on the severity of their geometric deviations. Upon analyzing the most critical clusters, the authors were able to adjust the extruder temperature and layer height to reduce the geometric inaccuracies.  \n\nML techniques have also been used for optimizing process parameters for obtaining superior surface quality173,17anddesidmecanicalpprtieshe overall goal is to reduce reliance on trial-and-error-based approaches to find the best parameter setting for the desired part quality. Liu et al. proposed a GP-based non parametric Bayesian framework for optimizing process parameters (extruder temperature, print speed, and layer thickness) for improving surface quality of graphene filled nanocomposite FFF parts.187 In this work, Bayesian optimization (BO） guided the search for the process parameters yielding the lowest surface roughness. GPR served as the surrogate model and was initially trained with just four data points, establishing a prior distribution. The model was then iteratively refined with new data points corresponding to the acquisition function maximum value, updating the prior to form a posterior distribution. The optimization process was terminated after only the fifth iteration because the predicted surface roughness for the recommended process parameter settings converged with the experimental value.  \n\nML-based techniques have been used to optimize printing parameters for reducing defects in printed parts.139 The modeling approach taken in this study used data from in-situ imaging. A camera was attached to the printing nozzle to record the printing process in realtime. Nine videos were recorded with different layer thicknesses and nozzle speeds. The videos were converted into individual frames, which were then fed to PCA to reduce dimensionality and extract relevant features. After dimensionality reduction, SVM was used to classify the transformed data into two distinct categories: pivotal lines and transition points. The knowledge gained from PCA and SVM provided the foundation for the opti mization step, where the aim was to detect the processing parameters (layer thickness and nozzle speed) that produce the fewest defects. The model performance was assessed through a three-fold cross validation process, and it was found that slower nozzle speed and smaller layer thickness are beneficial for reducing print defects.",
        "category": " Results and discussion"
    },
    {
        "id": 25,
        "chunk": "# 5.1.3 In-situ monitoring  \n\nIn-situ monitoring is largely applied in AM for defect and anomaly detection. Sensorsi18,190 and high-resolution cameras189195arecommonlyused todetectdefectsin real-time. Data acquired from sensor signals and processed images are often combined with ML techniques for predicting defects in the early stages of printing process. An example of using sensor signal-based ML for insitu monitoring was demonstrated by Xu et al.l18 They combined acoustic emission with laser scanning technology to develop a real-time monitoring system for detecting warpage in FFF parts. The acoustic emission sensors were placed along the print bed to capture platform vibrations during the printing process, while the warpage in each layer of the printed structure was quantified using point cloud data from laser scanning. The combined data were then used to train three ML algorithms: SVM, NB, and DT. It is worth noting that sensor signals consist of numerous features. In order to ensure good predictive accuracy, selecting the appropriate features is crucial. Xu et al. extracted voltage distribution information from the raw acoustic emission signal because it was highly correlated with structure warpage. Overall, DT outperformed the other two ML models for classifying the printed parts based on the amount of warpage.  \n\nFor in-situ monitoring involving image-based data, CNN models using different architectures are preferred.189,19195For instance,Gohet al.usedCNN with various“You Only Look Once\"(YOLO） architectures to analyze images captured during the printing process in order to detect print anomalies (over under-extrusion/ over-extrusion).189 Using YOLO-based architectures with CNN has the advantage of processing images in a single pass, which enables simultaneous detection of multiple objects.200,201  \n\nPhysics-based models can offer a comprehensive understanding of the printing process and can be useful for in-process monitoring. However, their in-process application is frequently hindered by high computational cost. ML models have been used as surrogates because they are more computationally efficient than physicsbased models. In this circumstance, ML-based surrogate models are built using data collected from physics-based models. They replace the computationally intensive models when they achieve sufficient accuracy. For example, Rossi et al.19l and Roy et al.194 used physics-based models to collect the initial data for training deep learning models, which were subsequently used for in-situ monitoring of the FFF process. Rossi et al. used an RNN model to simulate a complex extrusion process simulation. This model was then used to detect clogging events by analyzing the deviations between predicted and realtime values of extrusion rate, extrudate temperature, and filament compression force, as measured by various sensors during printing. Roy et al. used an MLP-based surrogate model to replicate the thermal profile of various geometries, allowing for inline monitoring of the FFF process.",
        "category": " Results and discussion"
    },
    {
        "id": 26,
        "chunk": "# 5.2 | Machine learning for vat photopolymerization  \n\nVP techniques are widely favored due to their rapid print ing speeds and ability to achieve high resolutions. The use of ML techniques in VP has emerged more recently in comparison to ME. ML-based research in VP tends to fall into three distinct groups: material and process optimization, in-situ monitoring, and metamaterial design The optimization of material and process parameters through ML techniques focuses on minimizing the need for extensive experimentation. In-situ optimization aided by ML methods focuses on defect detection. Therefore, the primary workflows are examined within the context of VP techniques. ANN models have been the prevailing approach in research pertaining to VP. Table 2 presents a comprehensive overview of the literature relating to MLassisted VP. This table includes details such as the ML techniques employed in each study, the model inputs that reflect the source of training data, and the specific objectives of each work.",
        "category": " Results and discussion"
    },
    {
        "id": 27,
        "chunk": "# 5.2.1 Process parameter optimization  \n\nFor VP, ML models are used to optimize different printing and material parameters such as light dosage, exposure time, printing speed, and material compositions for printing structures with desired properties. Existing research has used various types of image-based data, captured during printing process, totrain ML models.175agesofo of large quantities of data. ANN models tend to be more generalizable when trained with a large amount of data.2ll In addition, after determining a suitable architecture, ANN techniques can automatically extract important features from a large volume of data. As a result, many studies have used ANN techniques for optimizing the VP process. For example, Guan et al. used ANN to enhance the quality of celloaded bioprinting.176 They printed structures using digital masks and imaged those structures. A genetic algorithm was used to calibrate a mathematical 3D printer simulator using digital masks and images of the structures. This procedure aimed to replicate the effects of cell loading on light scattering during printing. 400o training data points (digital maskimage pairs) were produced using the calibrated simulator, which was not possible using only the printing process. The trained CNN model was then used to create a mask that accounted for light scattering consisting of a grayscale image that represented the light exposure dose for any given target structure. ML-optimized masks increased print fidelity for the highly scattering cellloaded material.  \n\nWhen the training dataset is small, shallow models are typically chosen to avoid overfitting and improve interpretability.212 For instance, Tagami et al. investigated the effect of light exposure time and material composition on drug release from VP-based poly(ethylene glycol) diacrylate (PEGDA) tablets with MRA and SVM trained on a dataset containing only 108 data points.203 In this study, six input features were analyzed, five of which were associated with material composition and one with the processing condition, specifically light exposure time. MRA with a sequential forced entry method elucidated which parameters significantly affected drug release. The results of the MRA analysis indicated that excluding the “light exposure time\" variable from the dataset improved the accuracy of drug release predictions. Later, the MRA analysis served as the foundation for the development of an SVM-based drug release kinetics model.  \n\nrreeeeaereeneeeerenerneereeeereareeaererenaeen AR   \n\n\n<html><body><table><tr><td> Broad motivation</td><td>AM technique</td><td> ML technique</td><td> Model inputs</td><td>Remarks</td><td> Reference</td></tr><tr><td rowspan=\"6\">Process-parameter optimization</td><td>SLA and DLP</td><td>CNN</td><td>High-resolution images of microneedle patches</td><td>Optimizing lithium phenyl (2,4,6-trimethylbenzoyl) phosphinate (LAP) concentration, water concentration, and exposure time for controlling needle morphology and</td><td>Bagde et a)</td></tr><tr><td>DLP</td><td>LSTM</td><td>Grayscale values of each element in finite element model</td><td>Optimizing grayscale distributions for obtaining varying deformations in printed structures</td><td>Zhao et al.</td></tr><tr><td>DLP</td><td>CNN</td><td>Digital mask image data and simulator generated structures</td><td>Optimizing grayscale value to compensate for scattering effect of cell-loaded bioink</td><td>Guan et al</td></tr><tr><td>DLP</td><td>MRA and SVM</td><td>Composition of ink and printing parameter (light exposure time during printing)</td><td>Investigating the effects of ink composition and printing conditions on drug release of printed tablets</td><td>Tagami et</td></tr><tr><td>CLIP</td><td>Shallow models (DT, NB, KNN, SVM), ELAs (RF, GB, ADA), and deep learning (MLP-based Siamese</td><td>Material properties, geometric and physical parameters, printing dynamics, surface texture and quality, hardware and setup</td><td>Optimizing printing speed</td><td>He et al.204</td></tr><tr><td>In-situ monitoring SLA and DLP</td><td>neural network) GPR</td><td>Thermistor data</td><td>Real-time failure detection and area prediction</td><td> Shan et al.</td></tr><tr><td></td><td>2PP</td><td>3D-CNN and CNN-LSTM</td><td>Image data from printing while light dosage varies</td><td>Identifying the optimal light dosage and part defect detection</td><td>Lee et al.20</td></tr><tr><td rowspan=\"5\"> Metamaterial design</td><td>SLA</td><td>MLP</td><td>Control points from Bezier curves</td><td>Designing novel beam elements with varying cross-sections</td><td>Lee et al.20</td></tr><tr><td>SLA</td><td>Variation AE (Combined with CNN and MLP)</td><td>Binary images of representative volume elements</td><td>Designing optimal representative volume elements with specified macroscopic elastic moduli</td><td>Xu et al.208</td></tr><tr><td>DLP</td><td>MLP</td><td>FEA simulation results for mechanical properties while varying the design aspects</td><td>Designing novel metamaterial with varying stiffness in three spatial directions</td><td>Fleisch et </td></tr><tr><td>SLA</td><td>MLP</td><td>Length and orientation angle</td><td>Optimizing the design of SLA-printed</td><td>Tak et al.21</td></tr><tr><td></td><td></td><td></td><td>W-band slotted waveguide array antenna</td><td></td></tr></table></body></html>",
        "category": " Results and discussion"
    },
    {
        "id": 28,
        "chunk": "# 5.2.2 In-situ monitoring  \n\nIn-situ monitoring is frequently used in VP for monitoring unexpected process variations, which lead to thermal distortions, print failure, and other defects in printed parts. ML techniques have the capacity to predict print outcomes at the initial stages of the printing process. In bioprinting, where feedstock is expensive, this is particularly useful for reducing material waste. Camera systems are commonly used in in-situ monitoring of AM. Using video data captured during the printing process in twophoton lithography (TPL) printing, Lee et al. developed a spatiotemporal ML-based process for detecting part defects.206 The work entailed the creation of a comprehensive dataset consisting of raw videos from four sets of experiments: three with commercially available photoresists and one with a custom photoresist. Experiments included varying structures (cuboids and truncated cones), discretization (log-pile rectangular Cartesian grid and grid of concentric circles), scan paths, and experimental parameters such as discretization period in the X-Y plane, write speed, and laser power. Due to the spatiotemporal and high dimensional nature of the dataset, which is typical of video-extracted data, two different ANN models were trained: 1. 3D-CNN and 2. CNNLSTM. While 3D-CNN only captured the spatial relationship in three dimensions from the dataset, CNN-LSTM captured both spatial and temporal relationship due to its hybrid nature (CNN for feature extraction from images and LSTM for sequence prediction tasks). Consequently, CNN-LSTM had higher accuracy in defect detection $(\\sim95\\%)$ compared to 3D-CNN $(\\sim91\\%)$ .  \n\nThe printing mechanisms and optical systems used in VP techniques make it challenging to monitor printingrelated changes using traditional image-based analysis.205 Thermal cameras have been reported in some studies.213,214 However, the use of thermal cameras is constrained by their high cost and limited accuracy. To address these concerns, Shan et al. developed a compact intelligent vat system that monitors the printing process through the analysis of temperature data acquired from thermistors that are evenly distributed and affixed to the edges of the resin vat.205 The system was designed to monitor temperature fluctuations throughout the resin vat during printing and these fluctuations were used as indicators of the degree of polymerization occurring. The authors trained a GPR model with temperature data from thermistor to predict the printed area of each layer since the rise in temperature is related to photopolymerization. Subsequently, the trained model was used to predict potential feature defects that could arise throughout the printing procedure. The study found that the predictive model had low accuracy, which was attributed to overfitting resulting from the limited size of the training dataset. The choice of temperature fluctuations as an input feature, which may not be a good indicator of the degree of polymerization in VP, may also contribute to low predic tive accuracy.",
        "category": " Results and discussion"
    },
    {
        "id": 29,
        "chunk": "# 5.2.3 Metamaterial design  \n\nMechanical metamaterials are comprised of periodic architectures capable of producing exceptional macroscale level properties such as auxetic behavior (negative Poisson's 。ratio),215 programmable mechanical response,216controlledbulkingbehavior,217shape morphing,218andacousticband gaps.219VPtechniques are useful for fabricating polymer-based metamaterials due to their capacity to offer high speed, good resolution and excellent surface quality.47 However, designing the lattice units for metamaterials based on past knowledge and intuition is insufficient for next generation designs. Furthermore, using FEA for exploring the large design space of possible lattice structures is computationally expensive. Hence, research has focused on applying ML techniques for creating new designs for lattice structures for metamaterials and optimization of the existing ones. In particular, ANNs have become important for metamaterial design because they can generate equivalent solutions to FEA, which is a crucial technique for determining the feasibility of structural designs, while reducing the computational time and cost by many orders of magnitude.  \n\nDesign optimization of lattice structures starts from the root, which is the beam element of the unit cells. A uniform cross-sectional area is typically assumed for the root, which limits the achievable strength-to-weight ratio. Moreover, the junction where the beam elements meet often suffers from lower strength compared to other regions. As a result, the lattice structures with a uniform beam element cross-section often fail to yield the desired performance without increasing the density of the lattice structure. To address this, Lee et al. introduced a workflow that integrates ANN with genetic optimization (GO)to enable more diverse and advanced selection of beam element shapes.207 An interesting aspect of this work is applying Bezier curves for generating the designs of beam elements. Bézier curves include control points that determine the curve's shape and orientation.220 A population of initial unit cell structures was generated by randomly modifying the control points of the beam elements and the initial designs underwent assessment for elastic modulus and strength using FEA. The GO step then chose the cell designs with superior performance and mixed their control points to generate new cell designs. After a significant amount of data (cell designs and their FEA-evaluated mechanical performances) was generated through the GO process, a dataset containing control points as input and mechanical performance as output was curated,which was used to train an MLP model. In order to verify the efficacy of the entire workflow and the resulting designs, a selection of the highest performing lattice structures (identified by MLP and subsequently confirmed by FEA) were manufactured using SLA and tested experimentally. The results of the experimental validation demonstrated that the lattice design optimized using ML exhibited superior performance compared to two benchmark models, namely the gradeddensity beam and the cylindrical beam.  \n\nComposite materials can achieve properties that would be impossible to achieve with a single base material. Designs of metamaterials based on composite materials present unique challenges due to the large design space, which includes a wide variety of possible configurations for the unit cells and fine-tuning structural properties through spatial material distribution. Thus, process automation necessitates the intervention of ML. Xue et al. proposed an ANN-based framework that incorporates multi-material-based unit cell design, also known as representative volume element (RVE).208 The materials consisted of a hard polyurethane-based material and a soft silicone-based material. The focus of the work was to control the elastic moduli of the cumulated lattice structure. A database was developed with combinations of artificially generated binary images. The authors used variation AE to generate potential design structures for RVE using the binary images from the database. The encoder architecture included a combination of convolutional and dense layers (fully connected) and the decoder architecture included a combination of deconvolutional and dense layers. The architecture leveraged CNN's spatial feature extraction and ML's data compression and reconstruction capabilities. The best RVE configuration that met the desired properties was determined with BO and the optimized designs were printed using SLA and subjected to tensile testing to ascertain the elastic moduli. The majority of experimental results were found to be consistent with computational predictions.",
        "category": " Results and discussion"
    },
    {
        "id": 30,
        "chunk": "# 5.3 | Machine learning for powder bed fusion  \n\nPBF is a popular AM technique since it is capable of printing without a support structure. Polymer-based PBF is restricted by its limited material selection. In addition, the current quality control and in-situ optimization techniques heavily rely on human expertise, which hinders process automation. The complexity of existing physicsbased models for characterizing relationships between process parameters and part properties prevents their use in process optimization. To address these issues, current research employs ML techniques, as summarized in Table 3. Literature is classified in accordance with four broad motivations: 1. Material selection; 2. Quality control; 3. In-situ monitoring; and 4. Discrete outcome pre dictions. This section provides a brief discussion of ongoing research outlining the trends in ML for polymerbased PBF.",
        "category": " Results and discussion"
    },
    {
        "id": 31,
        "chunk": "# 5.3.1 Material selection  \n\nIntrinsic properties such as particle size and shape, thermal properties, optical properties, rheological properties, and extrinsic properties such as powder flowability, bulk density, and tapped density are important when selecting polymer powders for SLS.231 Therefore, when introducing a new material, a series of characterization techniques are typically employed to determine its applicability for SLS printing, which is time consuming and requires extensive subject matter expertise. Moreover, printing withnew materialsincludes resource intensive trail-and-error to determine appropriate combinations of processing conditions to achieve successful prints. ML techniques have been used to reduce experimental effort by identifying correlations between the powder's intrinsic properties and its printability.125 Utilizing the power of ML techniques, a simple material-agnostic screening technique was also developed based on extrinsic properties to predict the suitability of new materials for SLS.221  \n\nIn order to identify printable formulations of a powdered pharmaceutical, Abdalla et al. developed a pipeline based on classification techniques that uses the data related to intrinsic properties of the formulations, such as differential  scanning calorimetry (DSC), Fourier-transformed infrared spectroscopy (FTIR), and x-ray diffraction (XRD) results.125 Raw data from each characterization technique were fed to an unsupervised ML model (PCA） for reducing dimensionality before applying the supervised models (RF, LR, SVM, GB, XGBoost, DT, MLP, KNN, and EXTr), thereby decreasing the reliance on human expertise for data analysis and decision making. The predictions from supervised models showed that FTIR spectra resulted in higher accuracy $(84.2\\%)$ than DsC thermograms $(80.1\\%)$ and XRD diffraction data $(81.3\\%)$ due to its higher capacity to handle the diversity and complexity of the formulations, especially when they contained amorphous polymers, which posed challenges for the other two methods. Moreover, combining all the characterization information led to even higher predictive accuracy $(88.9\\%)$ than the spectral data alone. In both cases, the best predictions were provided by RF, an ensemble model developed through combining multiple DTs.  \n\noepetedaeeneteeieareareeoreeneened SRAR   \n\n\n<html><body><table><tr><td>Broad motivation</td><td>AM technique</td><td> ML technique</td><td> Model inputs</td><td> Remarks</td><td> References</td></tr><tr><td>Material selection</td><td>SLS</td><td>LR, KNN, SVM, DT, MLP, EXTr, RF, GB, and XGBoost</td><td>Formulation composition and characterization data</td><td>Predicting printability of formulations using multi-modal data</td><td> Abdalla et al.125</td></tr><tr><td></td><td>SLS</td><td>SVM</td><td>Powder flowability and as-spread surface roughness</td><td>Developing a pre-screening method for SLS materials</td><td> Sassaman et al.221</td></tr><tr><td>Quality control</td><td> SLS </td><td>GPR</td><td>Surface diffusivity and interparticle distance</td><td>Analyzing the relationship between input parameters and the size of the neck region between particles</td><td> Batabyal et al.222</td></tr><tr><td></td><td> SLS</td><td>CNN</td><td>Data obtained from X-ray Computed Tomography (XCT)</td><td>Evaluation of DL vs. traditional methods on low-quality XCT scans</td><td> Bellens et al.223</td></tr><tr><td>Discrete outcome prediction</td><td>SLS</td><td>MLP</td><td>Binary image data from sliced CAD files</td><td>Predicting energy consumption using a knowledge distillation approach</td><td>Li et al.224</td></tr><tr><td></td><td>SLS</td><td>Genetic programming, SVR and MLP</td><td>Layer thickness,laser power, and feed rate</td><td> Predicting open porosity</td><td>Garg et al.225</td></tr><tr><td></td><td>SLS</td><td>Ensemble-based multi- gene genetic programming</td><td>Layer thickness,laser power, and laser scan speed</td><td>Predicting open porosity</td><td>Garg et al.226</td></tr><tr><td></td><td>SLS</td><td>MLP</td><td>Laser power, scan speed, scan spacing and layer thickness</td><td> Predicting density</td><td>Shen et al.227</td></tr><tr><td></td><td>SLS</td><td>MLP, DT, GB, and SVR</td><td> Part orientation</td><td> Predicting part dimension</td><td> Baturynska et al.141</td></tr><tr><td></td><td>SLS</td><td>MLP, SVM, and NB</td><td>Coordinates of the parts within the print volume and print parameters</td><td>Predicting surface roughness in the production planning phase</td><td> Kog et al.228</td></tr><tr><td> In-situ monitoring</td><td> SLS</td><td>CNN</td><td>Images of powder bed samples captured during printing process</td><td>Automatically classifying powder bed defects</td><td>Westphal et al.229</td></tr><tr><td></td><td> SLS</td><td>CNN</td><td> Thermal infrared recordings</td><td> In-situ quality control that detects</td><td> Klamert et al.230</td></tr></table></body></html>  \n\nSassman et al. used a classification algorithm to develop a material-agnostic screening method for SLS materials in an effort to accelerate the process of determining whether a particular powder (nylon) or powder mix (nylon mixed with various amounts of alumina and carbon fibers) would be suitable for a particular applications.21 This method used extrinsic properties, such as powder flowability information extracted from revolution powder analysis (RPA) and as-spread surface roughness, pertaining to the condition of the powder after it has been spread but prior to the SLS process. A SVM classifier was trained separately using RPA and asspread surface roughness data. The prediction results demonstrated that powder systems were correctly classified using RPA information $(93.1\\%)$ , but not surface roughness information $(62.5\\%)$ ， establishing the RPA process as a promising technique for pre-screening materials for the SLS process.",
        "category": " Results and discussion"
    },
    {
        "id": 32,
        "chunk": "# 5.3.2 |Quality control  \n\nThe complex dynamics of the SLS process, influenced by numerous parameters, often result in final part quality variability.232 A significant challenge in SLS, as with other AM methods, is achieving consistent part quality. Computational modeling such as phase-field microstructure model222 as well as characterization tools such as x-ray computed tomography $\\left(\\mathrm{X-CT}\\right)^{223}$ have been combined with ML techniques for data processing and analysis to gain a better understanding of the critical parameters influencing SLS part quality.  \n\nSLS final part quality depends on factors such as particle size and shape,233 contact area between particles,234 and diffusion kinetics (grain boundary vs volume diffusion).235 Batabyal et al. investigated microstructure variation during sintering based on two parameters: 1. Surface diffusivity between two polymer particles (equal sized particles and unequal sized particles) and 2. Interparticle distance.222 The size of the neck that forms between two particles during sintering was selected as the response quantity of interest (QOI). Training data was obtained from a two-particle phase-field microstructure model simulation. GPR was used as a surrogate model to approximate the underlying relationships between the input features (surface diffusivity and interparticle distance) and QOI. Sensitivity analysis showed that neck size is more sensitive to changes in interparticle distance than surface diffusivity irrespective of the particle size. BO was used to optimize the input features using two acquisition functions: 1. Expected improvement and 2. Probability of improvement. The optimization results from both acquisition functions were in good agreement, validating the optimization approach. Thus, the MLbased framework served as a rapid predictive tool for capturing the complex behavior of the sintering process, paving the way for enhanced quality control.  \n\nIn order to assess the quality of SLS printed parts, numerous non-destructive characterization techniques have been shown to be useful. Such characterization techniques include XCT,236 micro computed tomography (micro-CT),237 eddy current testing,238 and acoustic emission.239 Multiple techniques are necessary to confidently determine the quality of printed parts,which requires time and expertise. ML techniques have been able to accelerate quality assessment.223 For example, Bellens et al. employed CNN using U-net and MultiResUnet architectures, which are particularly useful for image segmentation tasks.240,241 To compare the effciency of the CNN-based image segmentation with the traditional Otsu's global algorithm, several observations were made in terms of quality inspection such as porosity and defect detection. Traditional technique requires 1500-3000 pro jections for accurate segmentation, while CNN showed improved results with as few as 99-1572 projections, substantially reducing the acquisition time.",
        "category": " Results and discussion"
    },
    {
        "id": 33,
        "chunk": "# 5.3.3 In-situ monitoring  \n\nCommon in-situ monitoring techniques applicable to polymer-based PBF include surface defect detection through image-based optical monitoring,242 temperature distribution monitoring using infrared thermography,243 and laser power monitoring to ensure uniform melting.244 Recent studies have centered on incorporating ML techniques into image-based in-situ monitoring for defect detection.29.230 CNNs are commonly used given their ability to handle image-based data.245,246 Westphal et al. reported a detection accuracy of up to $95.8\\%$ for defects using a VGGl6 CNN architecture trained with data extracted from powder bed images.229 Klamert et al. also reported a very high accuracy of $98.54\\%$ for detecting curling defects using the same architecture trained with thermal imaging data.230",
        "category": " Results and discussion"
    },
    {
        "id": 34,
        "chunk": "# 5.3.4 Discrete outcome prediction  \n\nSupervised learning has been used to predict discrete outcomes of SLS structures, such as energy consumption,224 porosity as a percentage of void content in printed structures26eity27atdims, face roughness.228 These are important quantities/ qualities to investigate in order to ensure precision and reliability, quality control, resource optimization, and the safety and functionality of a system.  \n\nExisting research has relied on MLP for discrete outcome predictions. The problems were framed as either regression or classification task. For example, Li et al. trained an MLP model with a teacher-student architec ture using layer-by-layer images of the printed structures to predict the energy consumption during the printing process.224 The correlations of the input and output features were transferred from a complex model (the \"Teacher\" model) to a simpler model (the “Student\" model), thereby reducing the model training time. More over, MLP was used to predict the part density based on scan speed, scan spacing, laser power, and layer thickness.227 MLP was also used for clasification, where it was used to predict surface roughness (set to a categorical output). In this example, MLP performed better than traditional classification methods such as SVM and NB when its hyperparameters were appropriately tuned.228 On the other hand, evolutionary algorithm-based approaches such as genetic programming and ensemblebased multi-gene genetic programming, outperformed MLP for predicting the open porosity of SLS parts.225,226",
        "category": " Results and discussion"
    },
    {
        "id": 35,
        "chunk": "# 5.4 | Machine learning for binder jetting and material jetting  \n\nBJ and MJ are two categories of AM that employ inkjet based material deposition techniques. Therefore, the exploration of ML applications within both BJ and MJ is discussed together. While there have been several reports of ML being used in BJ, particularly for porosity analysis, its application in MJ has been more limited. Conversely, ML techniques have been widely used for process parameter optimization, benefiting both BJ and MJ techniques Table 4 summarizes applications of ML in BJ and MJ, categorizing them into two main groups: one focusing on porosity analysis unique to the BJ process, and the other on process parameter optimization,relevant to both BJ and MJ processes.",
        "category": " Results and discussion"
    },
    {
        "id": 36,
        "chunk": "# 5.4.1 Porosity analysis  \n\nNumerous studies have been conducted to address the issues associated with defect detection and quality control in BJ to ensure the consistency and reliability of end-use parts.249 Due to its direct effect on the final properties, porosity analysis is of particular interest for BJ.254 However, characterization techniques such as SEM and transmission electron microscopy (TEM) provide analysis of local pore morphologies that may not be representative of the entire structure. In addition, traditional image analysis software requires specific conditions and manual inputs, making the porosity analysis process timeconsuming and prone to error. XCT is a non-destructive characterization technique that offers global morphology analysis, but is constrained by its long acquisition time.255 ANNs have been used to reduce the acquisition time for morphological analysis for PBF.223 However, the pore morphology of BJ components is significantly different from that of PBF components (uniform vs. clustered pores), requiring a different analytical approach.254,256 Zhu et al. created a pore evolution pipeline unique to BJ by implementing a CNN-based fast tomography algorithm that efficiently analyzed 3D morphological images from XCT.247 Moreover, they were able to create a database containing the morphological characteristics of ${{10}^{5}}$ pores, which was useful for tracing changes in morphology of different types of pores throughout the various stages of BJ manufacturing. Clustering techniques such as PCA and GMM were employed to extract key morphological descriptors of these pores and classify them into four distinct morphological groups.  \n\nAnother approach was reported by Satterlee et al. for realizing the global morphologies of BJ parts from local cross-sectional analysis using image augmentation.248 Image augmentation is a data expanding technique, which is often used when training data is limited, and data acquisition is time-consuming.257 The study obtained 3966 images (27,294 pores) from 67 SEM images (4545 cross-sectional pores) through image augmentation using generative adversarial ANN. The authors reported that CNN performed poorly for porosity detection, which can be attributed to regional proposal algorithms related to reducing image areas for examining specific sections more closely. Therefore,a Faster R-CNN method and YOLOv5 were also investigated, which demonstrated good performance in porosity detection on the original dataset, with Fl scores of $84\\%$ and $77\\%$ ，respectively.However, YOLOv5 performed better than Faster R-CNN on the augmented dataset $88\\%$ VS. $75\\%$ ). In addition, YOLOv5 was observed to produce an F1 score of $85\\%$ for a test dataset comprised of SEM images from the literature.  \n\neeeereaereeenieeeree AARE   \n\n\n<html><body><table><tr><td>Broad motivation</td><td>AM technique</td><td> ML technique</td><td> Model inputs</td><td>Remarks</td><td> References</td></tr><tr><td rowspan=\"3\"> Porosity analysis</td><td>BJ</td><td>GMM</td><td> XCT images</td><td>Inspection pipeline for defect characterization</td><td>Zhu et al.247</td></tr><tr><td>BJ</td><td>CNN</td><td>Cross section images of post-processed part</td><td>Automated porosity detection with small dataset</td><td>Satterle et al.248</td></tr><tr><td>BJ</td><td> LighGBM</td><td>Scanning electron microscopy (SEM) and XCT images</td><td>Characterizing microstructure of green body through binder and porosity distribution</td><td>Ojea et al.249</td></tr><tr><td rowspan=\"5\">Process parameter optimization</td><td>BJ </td><td> MLP and KNN</td><td>Roller speed, binder level, binder drying time, and layer thickness</td><td>Optimizing input parameters to maximize density, minimize dimensional errors and surface roughness, and obtain defect-free green bodies</td><td> Onler et al.250</td></tr><tr><td>BJ </td><td>MLR and GPR</td><td>Recoat speed, oscillator speed, layer thickness, drying time, roller transverse speed, drying power, and binder saturation level</td><td>Optimizing input parameters for desired green body density</td><td> Jimenez et al.251</td></tr><tr><td>BJ </td><td>Aggregated ANN (MLP-based)</td><td>Layer thickness, delay time between spreading, and print orientation</td><td>Optimizing input parameters for desired compressive strength and porosity</td><td>Asadi-Eydivand et al.252</td></tr><tr><td> MJ</td><td>Drop radius and velocity Pertictiog GB and RF;. prediction: KNN and MLP</td><td>Machine-related configurations and material properties</td><td>Optimizing inkjet printing parameters fo able itigt iding ivel </td><td> Brishty et al.140</td></tr><tr><td> MJ</td><td>ANNs (MLP, CNN, and RNN)</td><td>Droplet evolution captured in both simulated and experimental videos</td><td>Predicting droplet behavior under difetnt mamial and ontrois</td><td> Segura et al.253</td></tr></table></body></html>",
        "category": " Results and discussion"
    },
    {
        "id": 37,
        "chunk": "# 5.4.2 / Process parameter optimization  \n\nSimilar to other AM techniques, numerous process parameters affect the quality of BJ parts.258 Shallow and ANN techniques have been used to understand the effects of process parameters such as roller speed, oscillator speed, recoating speed, binder saturation level, layer thickness, and print orientation on the green part and sintered part properties such as density, surface rough ness, and compressive strength.250-252 Optimal print settings for user-defined properties have been determined using Pareto front252 and genetic algorithm methods.250 For efficient exploration of the parameter space, print conditions for the initial dataset are typically determined by orthogonal array-based DoEs.250,251 However, full factorial DoE has been used with a small number of param eters with limited levels.252  \n\nMJ requires ensuring consistent droplet quality in order to print high-resolution parts.140 The effect of material and process parameters on droplet characteristics has been investigated using ML techniques. Brishty et al. modeled droplet velocity, radius, and jetting regime (single drop/multiple drop/no ejection) using traditional ML models such as ensembles of DTs (boosted DTs and RF), KNN, and DNN.140 Additionally, Segura et al. evaluated droplet morphology using different ANNs.253 Machine configurations such as dwell and echo voltage, dwell and echo time, and material properties such as density, surface tension, and viscosity were considered as model inputs for both works. Both frameworks aimed for greater adaptability beyond the materials they tested, ensuring the feasibility of analyzing new materials.",
        "category": " Results and discussion"
    },
    {
        "id": 38,
        "chunk": "# 6 | CHALLENGES AND OPPORTUNITIES  \n\nIn reviewing the state of the art for applying ML to polymer AM, some common challenges emerge that limit model accuracy and/or broader adoption. These challenges can be related to materials, polymer AM processes, distributed manufacturing, or a combination of factors, which are discussed in more detail below.  \n\nPolymer materials are composed of large, often branching, molecules and are often formulated with many smaller molecule constituents to improve processing and performance. Subtle variations in composition and processing conditions can result in significant changes to a material's properties. In addition, polymer materials are highly sensitive to environmental factors like humidity and storage temperature. Therefore, the data collection procedure must be robust enough to capture intricate details of polymer behavior so that  \n\nML models can make accurate predictions for unseen cases.  \n\nAM processes are significantly slower than conventional polymer manufacturing techniques such as injection molding and extrusion, making it difficult to collect substantial amounts of experimental data. As a result, it is challenging to consider ML approaches that require large datasets for tasks such as process-parameter optimization and property prediction. To tackle the challenge, research aimed at improving model efficiency by developing techniques that decrease the reliance on large training datasets is being conducted.129,259 For example, AL aims to minimize experimental effort by iteratively exploring the design space to identify the most valuable data points for the model's training process168 Data augmentation is another technique that can be used to artificially enlarge a dataset, thereby reducing the effort for on-hand experimentation.Data augmentation techniques are particularly prevalent for image processing tasks such as defect detection and in-situ monitoring.29.260,261 Some techniques for enhancing image data include rotating images by a certain angle,262 adjusting the brightness or contrast of the image,263 translating the image in any direction,264 and injecting noise.265 Synthetic minority over-sampling techniques,266 and feature jittering267 are examples of augmentation techniques for tabular data. Notably, excessive reliance on data augmentation can result in overfitting to the augmented data. Therefore, it is essential to ensure that the augmented data represents accurate information in the context of the problem.  \n\nAs discussed in Section 2, polymer AM techniques have numerous parameters controlling the final part quality. While ML techniques can be useful in mapping the complex correlations between multiple process parameters and part properties, using raw data can lead to models capturing irrelevant noise rather than meaningful relationships, potentially resulting in overfitting Feature engineering, a data preprocessing technique, can be advantageous to tackle this challenge.268 The process includes selecting, transforming, or creating new features from the original dataset to enhance the performance of ML models. Feature engineering often involves utilizing domain knowledge to select more informative and relative signals from the raw data. For example, the raw data from XCT scans typically contains grayscale values corresponding to the densities of the voxels (3D pixels). Instead of relying solely on the raw intensity values, the data can be refined through segmentation, boundary extraction, and statistical analysis to derive quantifiable morphological descriptor of internal defects, enhancing the ability of the ML models to detect and classify defects based on their size and shape. Another example of feature engineering includes using dimensionality reduction techniques such as PCA and t-SNE.269 These techniques reduce the dimensionality of the dataset while retaining significant variance (PCA) or preserving local similarities (t-SNE), thereby making the data more manageable and interpretable for ML models.  \n\nCorrectly labeling data is an important task in the data collection phase for supervised learning. Analyzing the raw data and labeling them correctly is a timeconsuming process. The matter is further complicated by the dynamic nature of AM. For example, as AM processes evolve to accept more materials for manufacturing, the labeling criteria for the same task may shift. Transfer learning may be an approach to circumvent manual data labeling. In this process, a pre-trained model from a related system is used. The model is first fine-tuned with a small set of manually labeled data, which can be then used to label large dataset automatically. AL may also be used to reduce the effort for data labeling. While these methods can help automate data labeling, domainspecific insights are still required to ensure quality. Hence, a hybrid approach combining automatic labeling with expert review could be optimal.  \n\nAnother challenge involves cloud computing, which refers to a variety of services including data processing, analysis, and storage. Cloud computing is advantageous for AM, particularly in industrial settings, for design collaboration, scalable processing power for complex simulation and modeling, and data analysis. However, it presents significant difficulties for tasks involving realtime monitoring and control, such as in-situ monitoring, because sending data to the cloud system for processing and analysis takes time. Adopting ML models with simple architectures that can be implemented using local devices is a viable solution to the problem. Additionally, cloud computing is vulnerable to cybersecurity risks. Any variation in the input data can significantly mislead the model, leading to defective products and compromised structural integrity. Using robust end-to-end encryption and periodically  validating  input data by crossreferencing against benchmarks are recommended.  \n\nDespite these challenges, applications of ML for polymer AM continue to grow. This is due, in large part, to the incredible opportunities enabled by this combination of ML and AM, many of which are discussed in previous sections. Here, we highlight some particularly promising opportunities.  \n\nFeedstocks for polymer AM techniques are limited compared to traditional manufacturing techniques. ML can be a powerful tool in introducing new materials. ML has already been shown to effectively predict mechanical, thermal, and chemical properties, optimiz ing the process conditions, failure analysis, and defect detection in reduced time.20,2247 This indicates that ML techniques can also be useful in analyzing new formulations by predicting whether the materials would be suitable to a particular technique and application or not.  \n\nAM offers incredible design flexibility. ML offers the potential to accelerate design optimization for AM structures. ML models can be trained to generate design suggestions based on performance criteria, which may consequently reveal new design opportunities. MLenabled metamaterial designs are a relatively new area of study.207,27o,271 Proper adaptation of ML techniques will open the door to novel lattice structure designs with features such as wireless energy transfer,272 nonlinear optics,273 and acoustic properties.274 In addition, ML models that have been properly trained can analyze and predict the material distribution within a structure in order to increase its strength and reduce its weight. Lastly, ML allows for the modification of existing designs based on prerequisites without requiring a complete redesign.  \n\nCreating separate ML models for each material and printer on an industrial scale is time-consuming and resource intensive. Transfer learning emerges as a solution, permitting the transfer of knowledge from one task to another. This strategy typically entails utilizing limited training data to adapt a model originally developed for one material or printing system to work with a different but comparable system. Beyond just model adaptation, transfer learning also facilitates the transfer of pertinent data and insights, ensuring smooth integra tion across different systems. By actively sharing existing models and their associated data on platforms such as the Materials Data Facility and GitHub, researchers and developerscan expediteML advancementsin polymer AM.",
        "category": " Results and discussion"
    },
    {
        "id": 39,
        "chunk": "# V CONCLUSIONS  \n\nThe article provides an overview of the use of ML in polymer AM. ML has been applied to a variety of tasks, and primarily for property prediction, process parameter optimization, and quality control via in-situ monitoring. ML has been used more frequently in FFF these printers are widely available and the feedstock requires minimal preparation prior to use. As ML is a data-driven modeling technique, the primary concern is the data source and data collection procedure. Existing studies have focused on training ML models with data collected with minimal experimental effort, while ensuring that the collected data is sufficiently informative to the ML models so that their predictions are reliable. ANNs were the most popular ML model due to their ability to learn from diverse data sources, including tabular, image, video, and sensor data. However, care should be taken when applying ANNs to polymer AM, as they require a large amount of data to achieve sufficient generalizability. Lastly, challenges, potential solutions, and future research opportunities were outlined in an effort to provide readers with research directions.  \n\nML offers unique research prospects for polymer AM in terms of adapting new materials, exploring new designs, accelerating defect detection, and ensuring quality control. The scope of research extends beyond the utilization of established ML methodologies, encompassing the creation of novel algorithms with the objective of expediting advancements in polymer AM. As a result, the appropriate integration of ML techniques will also facilitate the exploration of novel applications.",
        "category": " Conclusions"
    },
    {
        "id": 40,
        "chunk": "# ACKNOWLEDGMENTS  \n\nAmy Peterson acknowledges the Department of Plastics Engineering and the Dandeneau Endowed Professorship.  \n\nORCID   \nAmy M. Peterson $\\textcircled{1}$ https://orcid.0rg/0000-0002-4612-   \n0062",
        "category": " Acknowledgments"
    },
    {
        "id": 41,
        "chunk": "# REFERENCES  \n\n[1] T. Vaneker, A. Bernard, G. Moroni, I. Gibson, Y. Zhang, CIRP Ann. 2020,69,578.   \n[2] T. Pereira, J. V. Kennedy, J. Potgieter, Proc. Manuf 2019, 30,11.   \n[3] M. Javaid, A. Haleem, R. P. Singh, R. Suman, S. Rab, Adv. Ind. Eng.Polym.Res.2021, 4,312.   \n[4] N. S. Hmeidat, R. C. Pack, S. J. Talley, R. B. Moore, B. G. Compton, Addit.Manuf. 2020, 34, 101385.   \n[5] A. Al-Ahmari, M. Ashfaq, S. H. Mian,W. Ameen,Int. J.Mater.Prod.Technol.2019, 58,129.   \n[6] S. Jang,S.Park, C.Bae,Biomed.Eng. Lett.2020, 10,493.   \n[7] T. D. Ngo, A. Kashani, G. Imbalzano, K. T. Q. Nguyen, D. Hui, Compos. B Eng. 2018,143,172.   \n[8]A. L.Fradkov, IFAC-PapersOnLine 2020, 53, 1385.   \n[9] B. Pang, E. Nijkamp, Y. N. Wu, Journal of Educational and Behavioral Statistics 2019, 45,227.   \n[10] E. Bisong, I. B. M. Learning, in Deep Learning Models on Google Cloud Platform: A Comprehensive Guide for Beginners (Ed: E.Bisong),Apress, Berkeley, CA 2019, p.215.   \n[11] S. Imambi,K. B. Prakash, G. R. Kanagachidambaresan,in Programming with TensorFlow: Solution for Edge Computing Applications (Eds: K. B. Prakash, G. R. Kanagachidambaresan), Springer International Publishing, Cham 2o21, p.87.   \n[12] N. Ketkar, in Deep Learning with Python: A Hands-on Introduction (Ed: N.Ketkar), Apress, Berkeley, CA 2017, p. 97.   \n[13] K.-L. Du, M.N. S. Swamy, in Neural Networks and Statistical Learning (Eds: K.-L.Du, M. N. S. Swamy), Springer, London, London 2014, p. 15.   \n[14] J. Mueller, X. Shi,A. Smola,Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, Association for Computing Machinery, New York, NY, USA 2020, p. 3509.   \n[15] M. Ahmadifar, K. Benfriha, M. Shirinbayan, A. Tcharkhtchi, Appl. Compos. Mater. 2021, 28,1335.   \n[16] A. Al Rashid, W.Ahmed, M.Y. Khalid, M. Koc,Addit Manuf 2021, 47, 102279.   \n[17] H. Chen, Y.F. Zhao, Rapid Prototyp J. 2016,22, 527.   \n[18]  T. Kozior, 3D Print Addit. Manuf. 2020, 7, 126.   \n[19] Y. L. Yap, C. Wang, S.L. Sing, V. Dikshit, W. Y. Yeong, J. Wei, Precis. Eng. 2017, 50, 275.   \n[20] J. Qin, F.Hu, Y. Liu, P. Witherell, C. C. L. Wang, D. W. Rosen, T. W. Simpson, Y. Lu, Q. Tang, Addit. Manuf. 2022, 52,102691.   \n[21] G. D. Goh, S. L. Sing, W. Y. Yeong, Artif Intell Rev 2021, 54, 63.   \n[22] X. Qi, G. Chen, Y. Li, X. Cheng, C. Li, Engineering 2019, 5,721.   \n[23] C. Wang, X. P.Tan, S. B.Tor, C.S.Lim, Addit. Manuf. 2020, 36,101538.   \n[24] M. Parsazadeh, S. Sharma, N. Dahotre, Prog. Mater. Sci. 2023, 135,101102.   \n[25] M. Schmid, A. Amado, K. Wegener, J. Mater. Res. 1824, 2014,29.   \n[26] M. Ziaee, N. B. Crane, Addit Manuf 2019, 28, 781.   \n[27] L. J. Tan, W. Zhu, K. Zhou, Adv. Funct. Mater. 2020,30, 2003062.   \n[28]  A. M. Peterson, Addit. Manuf. 2019, 27, 363.   \n[29] J. Lee, H. Lee, K.-H. Cheon, C. Park, T.-S. Jang, H.-E. Kim, H.-D. Jung, Addit. Manuf. 2019, 30, 100883.   \n[30] H. Wu, M. Sulkis, J. Driver, A. Saade-Castillo, A. Thompson, J. H. Koo,Addit. Manuf. 2018, 24,298.   \n[31] H. Pei, S. Shi, Y.Chen, Y.Xiong, Q. Lv,ACS Appl. Mater. Interfaces 2022, 14, 15346.   \n[32] Z. Wang, R. Liu, T.Sparks, F. Liou, 3D Print Addit. Manuf. 2016, 3, 39.   \n[33]T. D'Amico, A. M. Peterson, Addit. Manuf.2020, 34,101239.   \n[34] M. A. S. R. Saadi, A. Maguire, N. T. Pottackal, M. S. H. Thakur, M. M. Ikram, A. J. Hart, P. M. Ajayan, M. M. Rahman, Adv. Mater. 2022, 34, 2108855.   \n[35] O. Rios, W. Carter, B. Post, P.Lloyd, D.Fenn, C. Kutchko, R. Rock, K. Olson, B. Compton, Mater. Today Commun. 2018, 15,333.   \n[36] J. Pierre, F. Iervolino, R. D. Farahani, N. Piccirelli, M. Levesque, D. Therriault, Addit. Manuf. 2023, 61, 103344.   \n[37] N. Vidakis, M. Petousis, N. Mountakis, J. D. Kechagias, Int. J. Adv. Manufact. Technol. 2023, 124, 2931.   \n[38] P. Honigmann, N. Sharma, R. Schumacher, J. Rueegg, M. Haefeli,F.Thieringer,Biomed.Res.Int.2021,2021,30028.   \n[39] C. A. G. Beatrice, K. M. B. Shimomura, E.H. Backes, S. V. Harb, L. C. Costa, F. R. Passador, L. A. Pessan, Polym. Compos. 2021, 42, 1198.   \n[40] J. Zhang, Y. Su, X. Rao, H. Pang, H. Zhu, L.Liu, L. Chen, D. Li, J. He, J. Peng, Y. Jiang, Int. J. Bioprint. 2023, 9, 173.   \n[41] G. D. Goh, Y. L. Yap, H. K. J. Tan, S. L. Sing, G. L. Goh, W.Y. Yeong, Crit. Rev. Solid State Mater. Sci. 2020, 45, 113.   \n[42] D. J. Braconnier, R. E. Jensen, A. M. Peterson, Addit. Manuf. 2020, 31, 100924.   \n[43]M. Pagac, J. Hajnys, Q.-P.Ma,L.Jancar, J.Jansa, P.Stefek, J. Mesicek, Polymers (Basel) 2021, 13, 598.   \n[44] Z.Faraji Rad, P. D. Prewet,G. J. Davies, Microsyst.Nanoeng. 2021, 7, 71.   \n[45] J. R. Tumbleston, D. Shirvanyants, N. Ermoshkin, R. Janusziewicz, A. R. Johnson, D. Kell, K. Chen, R. Pinschmidt, J. P. Rolland, A. Ermoshkin, E. T. Samulski, J. M. DeSimone, Science (1979) 2015, 347, 1349.   \n[46] D. Loterie, P.Delrot, C.Moser, Nat. Commun. 2020, 11,852.   \n[47]F. Zhang, L. Zhu, Z. Li, S.Wang, J. Shi, W.Tang, N. Li, J. Yang, Addit.Manuf.2021, 48,102423.   \n[48] L. Andjela, V. M. Abdurahmanovich, S. N. Vladimirovna, G. I. Mikhailovna, D. D. Yurievich, M. Y. Alekseevna, Dental Mater.2022,38,e284.   \n[49] S. Zakeri, M. Vippola, E. Levanen, Addit. Manuf. 2020, 35, 101177.   \n[50] J. W. Halloran, Annu. Rev. Mater. Res. 2016, 46,19.   \n[51] W. Wang, J. Sun, B. Guo, X. Chen, K. P. Ananth, J. Bai, J.Eur. Ceram. Soc. 2020, 40,682.   \n[52] S. Mubarak, D. Dhamodharan, N. Divakaran, M. B. Kale, T. Senthil, L. Wu, J. Wang, Nanomaterials 2020, 10,79.   \n[53] H. Zhang, Y. Yang, K. Hu, B. Liu, M. Liu, Z. Huang, Addit. Manuf. 2020, 34,101199.   \n[54] H. Gong, C. Wang, H. Wu, X. Luo, J. Liang, N. Li, S. Chen, Y. Long, Mater. Today Commun. 2023, 34, 105149.   \n[55] L. Zhang, H. Liu, H. Yao, Y. Zeng, J. Chen, J. Manuf. Process 2022, 83, 756.   \n[56] I. Gibson, D. W. Rosen, B. Stucker, Additive Manufacturing Technologies, Springer, US, Boston, MA 2010.   \n[57] A. Bagheri, J. Jin, ACS Appl. Polym. Mater. 2019, 1, 593.   \n[58] C. Yu, J. Schimelman, P. Wang, K. L. Miller, X. Ma, S. You, J. Guan, B. Sun, W. Zhu, S. Chen, Chem. Rev. 2020, 120, 10695.   \n[59] R. J. Mondschein, A. Kanitkar, C. B. Williams, S. S. Verbridge,T.E.Long, Biomaterials 2017,140,170.   \n[60] I. Valizadeh, T. Tayyarian, O. Weeger, Addit. Manuf. 2023, 72,103641.   \n[61] D. L. Naik, R. Kiran, Addit. Manuf. 2018, 23, 181.   \n[62] Y. Rudenko, A. Lozovaya, L. Asanova, N. Fedyakova, P. Chapala, Progr. Addit. Manufact. 2023.   \n[63] A. Andreu, P.-C. Su, J.-H. Kim, C. S. Ng, S. Kim, I. Kim, J. Lee, J. Noh, A. S. Subramanian, Y.-J. Yoon, Addit. Manuf. 2021, 44, 102024.   \n[64] I. L. de Camargo, M. M. Morais, C. A. Fortulan, M. C. Branciforti, Ceram. Int. 2021, 47, 11906.   \n[65] B. J. Konijn, O. B. J. Sanderink, N. P. Kruyt, Powder Technol. 2014, 266, 61.   \n[66] C.Qian, K. Hu, J. Li, P.Li, Z. Lu, J. Eur.Ceram. Soc. 201, 41,7141.   \n[67] I. J. Kim, J. G. Park, Y. H. Han, S. Y. Kim, J. F. Shackelford, J. Korean Ceram. Soc. 2019, 56, 211.   \n[68]  H. Yang, J. C. Lim, Y. Liu, X. Qi, Y. L. Yap, V. Dikshit, W. Y. Yeong,J. Wei, Virtual Phys. Prototyp.2017,2,95.   \n[69] L. Giorleo, B. Stampone, G. Trotta, Addit. Manuf. 2022, 56, 102947.   \n[70] A. Sridhar, T. Blaudeck, R. R. Baumann, 2011.   \n[71] O. Guilcan, K. Guinaydin, A. Tamer, Polymers (Basel） 2021, 13,2829.   \n[72] A. Elkaseer, K. J. Chen, J. C. Janhsen, O. Refle, V. Hagenmeyer, S. G. Scholz, Addit. Manuf. 2022, 60, 103270.   \n[73] J.M. Lee. W.Y. Yeong, Mater. Today Proc. 2022, 70.535.   \n[74] F. Zhang, E. Saleh, J. Vaithilingam, Y. Li, C. J. Tuck, R. J. M. Hague, R. D. Wildman, Y. He, Addit. Manuf. 2019, 25,477.   \n[75] M. Jamal, T. K. Dey, T. Nasrin, A. Khosla, K. M. Razeeb, J. Electrochem. Soc. 2022, 169, 057517.   \n[76] L. Bass, N. A. Meisel, C. B. Williams, Rapid Prototyp J. 2016, 22,826.   \n[77] A. Pugalendhi, R. Ranganathan, S. Ganesan, Mater. Today Proc. 2021, 46, 9139.   \n[78] M. Casini, in Construction 4.0 (Ed: M. Casini), Woodhead Publishing, Cambridge, UK 2022, p. 405.   \n[79] W. Han, L. Kong, M. Xu, Int. J. Extreme Manufact.2022, 4, 42002.   \n[80] 3D Printing Now Good Enough for Final & Spare Car Parts.   \n[81]  Airbus Helicopter: Cabin Ventilation Distributor.Prodways.   \n[82] New Jersey Man Receives 3D Printed PEEK Skull Implant3DPrint.com.The Voiceof3D Printing/Additive Manufacturing.   \n[83] N. D. Parab, J.E. Barnes, C. Zhao, R. W. Cunningham, K. Fezzaa, A. D. Rollett, T. Sun, Sci. Rep. 2019, 9, 2499.   \n[84] E. S. Almaghariz, B. P. Conner, L. Lenner, R. Gullapalli, G. P. Manogharan, B. Lamoncha, M. Fang, Int. J. Metalcast. 2016, 10,240.   \n[85] M. Upadhyay, T. Sivarupan, M. El Mansori, J. Manuf. Process 2017, 29, 211.   \n[86] B. M. Wu, S. W. Borland, R. A. Giordano, L. G. Cima, E. M. Sachs, M.J. Cima, J. Controlled Release 1996, 40, 77.   \n[87] A. Lores, N. Azurmendi, I. Agote, E. Zuza, Powder Metall. 2019, 62, 267.   \n[88] X. Lv, F. Ye, L. Cheng, S. Fan, Y. Liu, Ceram. Int. 2019, 45, 12609.   \n[89] S. I. Yanez-Sanchez, M.D.Lennox, D.Therriault, B. D.Favis, J. R. Tavares, Ind. Eng. Chem. Res. 2021, 60,15162.   \n[90] S. Im, R. Batmaz, A. Natarajan, E. Martin, in Light Metals, Vol. 2023 (Ed: S.Broek),Springer Nature, Switzerland, Cham 2023, p. 471.   \n[91]  P.Shakor, S. H. Chu, A. Puzatova, E. Dini, Progr. Addit Manufact. 2022, 7, 643.   \n[92] I. Gibson, D. Rosen, B. Stucker, M. Khorasani, in Additive Manufacturing Technologies (Eds: I. Gibson, D. Rosen, B. Stucker, M. Khorasani), Springer International Publishing, Cham 2021, p. 237.   \n[93] T. Do, C. S. Shin, D. Stetsko, G. VanConant, A. Vartanian, S. Pei, P. Kwon, Proc. Manuf. 2015, 1, 263.   \n[94] J. Ingaglio, J. Fox, C. J. Naito, P. Bocchini, Constr. Build. Mater. 2019, 206, 494.   \n[95] T. Do, T. J. Bauder, H. Suen, K. Rego, J. Yeom, P. Kwon, Volume 1: Additive Manufacturing; Bio and Sustainable Manufacturing, American Society of Mechanical Engineers, New York 2018.   \n[96] H. Miyanaji, S. Zhang, A. Lassell, A. A. Zandinejad, L. Yang, Proc. Manuf. 2016, 5,870.   \n[97] D. Bzdok, M. Krzywinski, N. Altman, Nat. Methods 2017, 14, 1119.   \n[98] T. Herzog, M. Brandt, A. Trinchi, A. Sola, A. Molotnikov, J. Intell. Manuf. 2023.   \n[99] K. Zhu,J.Y.H.Fuh, X. Lin,IEEE/ASME Trans. Mechatron. 2022, 27, 2495.   \n[100] R. Zhou, H. Liu, H. Wang, Int. J. Adv. Manufact. Technol. 2022,121,5693.   \n[101] Y. Fu, A. R. J. Downey, L. Yuan, T. Zhang, A. Pratt, Y. Balogun, J. Manuf. Process 2022, 75, 693.   \n[102] A. Bouabbou, S. Vaudreuil, Virtual Phys. Prototyp. 2022, 17,543.   \n[103] Y. Zhang, W. Yan, J IntellManuf 2022, 34,2557.   \n[104] P. Wang, Y. Yang, N.S. Moghaddam, J. Manuf. Process 2022, 73,961.   \n[105] S. Ray, 2019 International Conference on Machine Learning, Big Data, Cloud and Parallel Computing (COMITCon), pp. 35-39. 2019.   \n[106] R. Agarwal, J. Singh, V. Gupta, Polym. Compos. 2022, 43, 5663.   \n[107] I. Ait-Mansour, N. Kretzschmar, S. Chekurov, M. Salmi, J. Rech, Progr. Addit. Manufact. 2020, 5, 51.   \n[108] R. V. Pazhamannil, P. Govindan, P. Sooraj, Mater. Today Proc. 2021, 46, 9187.   \n[109] J. Lee, H. Kim, H. Kim, T. Lee, J.-H. Kim, A.Andreu, S. Kim, Y.-J. Yoon, Addit. Manuf. 2022, 55, 102799.   \n[110] P.K. Jain, P.M.Pandey, P.V.M. Rao, Virtual Phys. Prototyp. 2008, 3,177.   \n[111] D. Kaweesa, L. Bobbio, A. M. Beese, N. A. Meisel, Rapid Prototyp. J. 2023, 29, 488.   \n[112] A. Selvam, S. Mayilswamy, R. Whenish, K. Naresh, V. Shanmugam, O. Das, Sci. Rep.2022, 12,16887.   \n[113] D. Veeman, S. Sudharsan, G. J. Surendhar, R. Shanmugam, L. Guo, Mater. Today Commun. 2023, 35, 106147.   \n[114] A. A.Soofi, A.Awan,J. Basic Appl. Sci.2017,13,459.   \n[115] Y. Fu, A. Downey, L. Yuan, A. Pratt, Y. Balogun, Addit. Manuf. 2021, 38,101749.   \n[116] H. Hu, K. He, T. Zhong, Y. Hong, Rapid Prototyp J.2020, 26,330.   \n[117] K. Bastani, P. K. Rao, Z. Kong, IIE Trans. 2016, 48,579.   \n[118] K. Xu, J. Lyu, S. Manoochehri, J. Manuf. Process 2022, 84,357.   \n[119] K. Okarma, J. Fastowicz, Pattern Anal. Appl. 2020, 23, 1035.   \n[120] A. E.Ezugwu, A. M. Ikotun, O. O. Oyelade, L. Abualigah, J. O. Agushaka, C. I. Eke, A. A. Akinyelu, Eng. Appl. Artif. Intell.2022,110,104743.   \n[121] L. Bertoli, F.Caltanissetta, B.M. Colosimo, 2021 IEEE 17th International Conference on Automation Science and Engineering (CASE), 2057-2062. 2021.   \n[122] F.Li, Z.Yu, Z.Yang, Rapid Prototyp.J.2020,26,177.   \n[123] H. Wu, Z.Yu, Y. Wang, Volume 3: Joint MSEC-NAMRC Symposia, American Society of Mechanical Engineers. 2016.   \n[124] C. O. S. Sorzano, J. Vargas, A. D. Pascual-Montano, ArXiv abs/1403.2877. 2014.   \n[125] Y. Abdalla, M. Elbadawi, M. Ji, M. Alkahtani, A. Awad, M. Orlu,S. Gaisford, A. W. Basit, Int. J. Pharm. 2023, 633, 122628.   \n[126] A. Singh, N. Thakur, A. Sharma, 20l6 3rd International Conference on Computing for Sustainable Global Development (INDIAC0m),1310-1315.2016.   \n[127] M. Alloghani, D. Al-Jumeily, J. Mustafina, A. Hussain, A. J. Aljaaf, in Supervised and Unsupervised Learning for Data Science (Eds: M. W. Berry, A. Mohamed, B.W. Yap), Springer International Publishing, Cham 2020, p.3.   \n[128]J.E.van Engelen, H.H.Hoos, Mach. Learn.2020,109,373.   \n19 NasII ruura, ruuIkaman-AnarakI, A. IVI. Peterson, Sci. Rep. 2023, 13, 11460.   \n[130] W. Han, E.Coutinho, H. Ruan, H. Li, B. Schuller, X. Yu, X. Zhu, PLoS One 2016,11, e0162075.   \n[131] N. V. Nguyen, A. J. W. Hum, T. Do, T. Tran, Virtual Phys. Prototyp. 2023, 18, e2129396.   \n[132] H. Liang, X. Sun, Y. Sun, Y. Gao, EURASIP J. Wirel. Commun. Netw. 2017, 2017, 211.   \n[133] X. Wang, Y. Zhao, F. Pourpanah, Int. J. Machine Learn. Cybernet. 2020, 11, 747.   \n[134] D. Maulud, A. M. Abdulazeez, J. Appl. Sci. Technol. Trends 2020, 1,140.   \n[135] S. Sharma, V. Gupta, D. Mudgal, V. Srivastava, Eng. Appl. Artif. Intell. 2023, 124, 106587.   \n[136]B.MunizCastro,.ElbadawiJOngTPollardg S. Gaisford, G. Perez, A. W. Basit, P. Cabalar, A. Goyanes, J. Controlled Release 2021, 337, 530.   \n[137] H. Zhang, E. Hong, X. Chen, Z. Liu, ACS Appl. Mater. Interfaces 2023, 15, 14532.   \n[138] P. Nath, J. D. Olson, S. Mahadevan, Y.-T. T. Lee, Addit. Manuf. 2020, 35, 101331.   \n[139] Y. Zhu, T. Kwok, J. C. Haug, S. Guo, X. Chen, W. Xu, D. Ravichandran, Y. D. Tchoukalova, J. L. Cornella, J. Yi, O. Shefi, B. L. Vernon, D.G.Lot, J.N.Lancaster, K.Song, Adv. Mater. Technol. 2023,8,2201421.   \n[140] F. P. Brishty, R. Urner, G. Grau, Flexible Printed Electron. 2022,7,15009.   \n[141] I. Baturynska, K. Martinsen, J. Intell. Manuf. 2021, 32, 179.   \n[142] M. Seeger, Int. J. Neural Syst. 2004, 14, 69.   \n[143] J. Pareek, J. Jacob, in Advances in Information Communication Technology and Computing (Eds: V. Goar, M. Kuri, R. Kumar, T. Senjyu), Springer Singapore, Singapore 2021, p. 327.   \n[144] J. S. Almeida, Curr. Opin. Biotechnol. 20o2, 13, 72.   \n[145] F. Pourkamali-Anaraki, M. A. Hariri-Ardebili, IEEE Access 2021, 9, 15334.   \n[146] M. Uzair, N. Jamil, 2020 IEEE 23rd International Multitopic Conference (INMIC), 1-6.2020.   \n[147] A. Krenker, J. Bester, A. Kos, in Artificial Neural Networks (Ed: K. Suzuki), IntechOpen, Rijeka, Croatia 2011.   \n[148] S. Hochreiter, J. Schmidhuber, Neural Comput.1997, 9, 1735.   \n[149] A. Shewalkar, D. Nyavanandi, S. A. Ludwig, J. Artif. Intell. Soft Comput. Res. 2019, 9, 235.   \n[150] L. Zheng, S. Duffner, K. Idriss, C. Garcia, A. Baskurt, Multimed. Tools Appl. 2016, 75, 5055.   \n[151] F.Murtagh, Neurocomputing 1991, 2, 183.   \n[152] X. Xie, G.Liu, Q. Cai, P.Wei, H. Qu, Neural Networks 2019, 119,151.   \n[153] S. Bodapati, H. Bandarupally, R. N. Shaw, A. Ghosh, in Advances in Applications of Data-Driven Computing (Eds: J. C. Bansal, L. C. C.Fung, M. Simic, A. Ghosh), Springer Singapore, Singapore 2021, p. 49.   \n[154]Y. Zou, J. Lv,Electronics (Basel)2020,9,2205.   \n[155] S. Hershey, S. Chaudhuri, D. P.W. Ellis, J. F. Gemmeke, A. Jansen, R. C.Moore,M. Plakal, D. Platt, R. A. Saurous, B. Seybold, M. Slaney, R. J. Weiss, K. Wilson, 2017 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 131-135. 2017.   \n[156] S. Vellangiri, S. Alagumuthukrishnan, S. I. T. Joseph, Proc. Comput. Sci. 2019, 165, 104.   \n[157] S. Miao, Z. J. Wang, R. Liao, IEEE Trans. Med. Imaging 2016, 35,1352.   \n[158] Y. Wang, H.Yao, S. Zhao, Neurocomputing 20l6, 184, 232.   \n[159] J. Vesanto, E. Alhoniemi, IEEE Trans. Neural Netw. 2000, 11,586.   \n[160] T.Chai, R. R. Draxler, Geosci. Model Dev.2014, 7,1247.   \n[161]T.Li, J.Yeo,Adv. Intell.Syst.2021, 3,2100069.   \n[162] K. S. Prakash, T. Nancharaih, V. V. S. Rao, Mater. Today Proc. 2018,5, 3873.   \n[163] P. S. Ramalhete, A. M. R. Senos, C. Aguiar, Mater. Des. (1980-2015) 2010, 31, 2275.   \n[164] H.E.Pence, A. Williams,J. Chem. Educ.2010,87,1123.   \n[165]  J. J. Pignatiello, IIE Trans. 1988, 20, 247.   \n[166] A. I. Khuri, S. Mukhopadhyay, WIREs Computat. Stat. 2010, 2,128.   \n[167] D. E. Huntington, C. S. Lyrintzis, Probabilistic Eng. Mech. 1998, 13,245.   \n[168] T. Lookman, P. V. Balachandran, D. Xue, R. Yuan, Npj Comput. Mater. 2019, 5, 21.   \n[169]  A. M. Mirzendehdel, K. Suresh, Comput. Aided Design 2016, 81, 1.   \n[170] M. Zhou, Y. Liu, C. Wei, Struct.Multidiscipl. Optim. 2020, 61, 2423.   \n[171] H. T. Kollmann, D. W. Abueidda, S. Koric, E. Guleryuz, N. A. Sobh, Mater. Des. 2020, 196, 109098.   \n[172] F. V. Senhora, H.Chi, Y. Zhang, L. Mirabella, T. L. E. Tang, G. H. Paulino, Comput. Methods Appl. Mech. Eng. 2022, 398, 115116.   \n[173] J. Butt, V. Mohaghegh, Metals (Basel) 2023,13,24.   \n[174] R. Ratnavel, S. Viswanath, J. Subramanian, V. K. Selvaraj, V. Prahasam, S. Siddharth, Micromachines (Basel） 2022, 13, 2231.   \n[175] A. Bagde, S.Dev, L. Madhavi, K. Sriram, S.D. Spencer, A. Kalvala, A. Nathani, O. Salau, K. Mosley-Kellum, H. Dalvaigari, S. Rajaraman, A. Kundu, M. Singh, Int. J. Pharm. 2023, 636, 122647.   \n[176] J. Guan, S. You, Y. Xiang, J. Schimelman, J. Alido, X. Ma, M. Tang, S. Chen, Biofabrication 2022, 14, 15011.   \n[7] M. Piovarci,M.Foshey,J. Xu, T.Erps, V.Babae, P.Didyk, S. Rusinkiewicz, W. Matusik, B. Bickel, ACM Trans. Graphics 2022, 41,1.   \n[178] H. Chen, Y. Liu, S. Balabani, R. Hirayama, J. Huang, Research 2023, 6,197.   \n[179] D. J. Roach, A. Rohskopf, C. M. Hamel, W. D. Reinholtz, R. Bernstein, H. J. Qi, A. W. Cook, Addit. Manuf. 2021, 41, 101950.   \n[180] N. Ranjan, R. Kumar, R. Kumar, R. Kaur, S. Singh, J. Mater. Eng. Perform. 2023, 32, 4555.   \n[181] V. Naga Malleswari, G. Kameswara Manaswy, P. G. Pragvamsa, Mater Today Proc 2023.   \n[182] Z. Li, Z. Zhang, J. Shi, D. Wu, Robot Comput.Integr. Manuf. 2019, 57, 488.   \n[183] H. Si, Z. Zhang, O. Huseynov, I. Fidan, S. R. Hasan, M. Mahmoud, Inventions 2023, 8,24.   \n[184] M. S. Meiabadi, M. Moradi, M. Karamimoghadam, S. Ardabili, M. Bodaghi, M. Shokri, A. H. Mosavi, Polymers (Basel)2021,13,3219.   \n[185] M. A. da Silva, B. Amaro Junior, R. R. B. Medeiros, P. R. Pinheiro, Algorithms 2022, 15, 263.   \n[186] C. Belei, R. Pommer, S.T. Amancio-Filho, Mater.Des. 2022, 219,110776.   \n[187] J. Liu, J. Ye, F. Momin, X. Zhang, A. Li, Addit. Manuf. 2022, 54,102765.   \n[188] M. Khanzadeh, P. Rao, R. Jafari-Marandi, B. K. Smith, M. A. Tschopp, L.Bian,J. Manuf.Sci.Eng.2017,140,0301l.   \n[189] G. D. Goh, N.M. Bin Hamzah, W.Y. Yeong, 3D Print Addit. Manuf. 2022, 10, 428.   \n[190] Z. Shi, A. Al Mamun, C. Kan, W. Tian, C. Liu, J. Intell. Manuf. 2023, 1815, 34.   \n[191] A. Rossi, M. Moretti, N. Senin, J. Manuf. Process 2022, 84, 64.   \n[192] A. Rossi, M. Moretti, N. Senin, J. Manuf. Process 2021, 70,438.   \n[193] J. Lyu, J. A. T. Boroujeni, S. Manoochehri, 4lst Computers and Information in Engineering Conference (CIE), Vol. 2, American Society of Mechanical Engineers, New York 2021.   \n[194] M. Roy, O. Wodo,Addit. Manuf. 2020, 32,101017.   \n[195] B. N. Narayanan, K. Beigh, G. Loughnane, N. Powar, Proc. SPIE 2019,11139,1113913.   \n[196] N. Fiorentini, D. Pellgrini, M. Losa, Transp. Res. Rec. 202, 2677, 1455.   \n[197] L. M. Galantucci, F. Lavecchia, G. Percoco, CIRP Annals 2009, 58, 189.   \n[198] A. Boschetto, L. Bottini, J. Mater. Process. Technol. 2015, 219,181.   \n[199]  A. Boscheto, V. Giordano, F. Veniali, Rapid Prototyp. J.2013, 19, 240.   \n[200] M.A. Al-masni, M.A. Al-antari, J.-M. Park, G. Gi, T.-Y. Kim, P. Rivera, E. Valarezo, M.-T. Choi, S.-M. Han, T.-S. Kim, Comput. Methods Programs Biomed. 2018, 157,85.   \n[201] N. S. Artamonov, P. Y. Yakimov, J. Phys. Conf. Ser. 2018, 1096,12086.   \n[202] B. Zhao, M. Zhang, L. Dong, D. Wang, Compos. Commun. 2022, 36,101395.   \n[203] T. Tagami, C. Morimura, T. Ozeki, Int. J. Pharm.2021, 604, 120721.   \n[204] H. He, Y. Yang, Y. Pan, J. Manuf. Syst. 2019, 50, 236.   \n[205] Y. Shan, A. Krishnakumar, Z. Qin, H. Mao, Volume 1: Additive Manufacturing; Biomanufacturing; Life Cycle Engineering Manufacturing Equipment and Automation, Nano/Micro/- Meso Manufacturing, American Society of Mechanical Engineers, New York 2022.   \n[206] X.Y. Lee, S. K. Saha, S. Sarkar, B. Giera, Addit. Manuf. 2020, 36,101444.   \n[207] S. Lee, Z. Zhang, G. X. Gu, Mater. Horiz. 2022, 9,952.   \n[208] T. Xue, T. J. Wallin, Y. Menguc, S. Adriaenssens, M. Chiaramonte,Extreme Mech.Lett.2020, 41,100992.   \n[209] M. Fleisch, A. Thalhamer, G. Meier, I. Raguz, P.F. Fuchs, G. Pinter, S. Schlogl, M. Berer, Mater. Today Adv. 2021,11, 100155.   \n[210] J. Tak, A. Kantemur, Y. Sharma, H. Xin, IEEE Antennas Wirel. Propag. Lett. 2008, 2018,17.   \n[211] V. Sandfort, K. Yan, P.J. Pickhardt, R. M. Summers, Sci. Rep. 2019, 9,16884.   \n[212] K. Pasupa, W. Sunhem, 2016 8th International Conference on Information Technology and Electrical Engineering (ICITEE), 1-6.2016.   \n[213] P. Garra, A.-H. Bonardi, A. Baralle, A. Al Mousawi, F. Bonardi, C. Dietlin, F. Morlet-Savary, J.-P. Fouassier, J. Lalevee, J. Polym. Sci., Part A: Polym. Chem. 2018, 56, 889.   \n[214] K. Xu, Y. Chen, J. Manuf. Sci. Eng. 20l6, l39, 021002.   \n[215] U. Veerabagu, H. Palza, F. Quero, ACS Biomater. Sci. Eng. 2022,8,2798.   \n[216] P. Sinha, T. Mukhopadhyay, Mater. Sci. Eng. R: Rep. 2023, 155,100745.   \n[217] P. Jiao, T. Chen, Y. Xie, Compos. Struct. 2021, 256, 113053.   \n[218] Z. Zhang,A. O. Krushynska, APL Mater.2022, 10, 080701.   \n[219] Y. Chen, L.Wang, Appl. Phys. Lett.2014, 105,191907.   \n[220] H. N. Fitter, A. B. Pandey, D. D. Patel, J. M. Mistry, Proc. Eng. 2014, 97, 1155.   \n[221] D. Sassaman, T. Phillips, C. Milroy, M. Ide, J. Beaman, JOM 2022, 74, 1102.   \n[222] A. Batabyal, S. Sagar, J. Zhang, T. Dube, X. Yang, J. Zhang, ASCE-ASMEJ.Risk Uncert EngrgSyst.PartBMech.Engrg. 2021,8, 011102.   \n[223] S. Bellens, G. M. Probst, M. Janssens, P. Vandewalle, W. Dewulf, Polym. Test.2022, 110,107540.   \n[224] Y. Li, F. Hu, M. Ryan, R. Wang, Y.Liu, IFAC-PapersOnLine 2022, 55, 390.   \n[225]  A. Garg, J. S. L. Lam, Measurement (Lond) 2015, 75, 210.   \n[226] A. Garg, J. S.L. Lam, M. M. Savalani, Int. J. Adv. Manufact. Technol. 2015,80, 555.   \n[227] X. Shen, J. Yao, Y. Wang, J. Yang, in Advances in Neural Networks-ISNN, Vol. 2004 (Eds: F.-L.Yin, J. Wang, C. Guo), Springer Berlin Heidelberg, Berlin, Heidelberg 2004, p.832.   \n[228] E. Koc, S. Zeybek, B. O. Kisasoz, C. i. Caliskan, M. E. Bulduk, Int.J.Adv.Manufact.Technol.2022,123,3033.   \n[229] E. Westphal, H.Seitz,Addit.Manuf.2021, 41,101965.   \n[230] V. Klamert, M. Schmid-Kietreiber, M. Bublin, Proc. CIRP 2022,111, 317.   \n[231] M. Schmid, A. Amado, K. Wegener, AIP Conf. Proc. 2015, 1664,160009.   \n[232] N. Kumar, H. Kumar, J. S. Khurmi, Procedia Technol. 2016, 23,352.   \n[233] K. Miyake, Y. Hirata, T. Shimonosono, S. Sameshima, Materials 2018,11, 1137.   \n[234] T. Ostrowski, A. Ziegler, R. K. Bordia, J. Rodel, J. Am. Ceram. Soc.1998,81, 1852.   \n[235] R. D. Bagley, I. B. Cutler, D. L. Johnson, J. Am. Ceram. Soc. 1970, 53, 136.   \n[236] B. Sagbas, M. N. Durakbasa, in Proceedings of the International Symposium for Production Research 2019 (Eds: N. M. Durakbasa, M. G. Gencyilmaz), Springer International Publishing, Cham 2020, p. 481.   \n[237] W. Shen, X. Zhang, X. Jiang, L.-H. Yeh, Z. Zhang, Q. Li, B. Li, H. Qin, Proc. Manuf. 2021, 53, 568.   \n[238] M. A. Spurek, V. H. Luong, A. B. Spierings, M. Lany, G. Santi, B. Revaz, K. Wegener, Metals (Basel) 2021, 11, 1376.   \n[239] Z. Li, X. Zou, X. Zhang, F. Gu, A. D. Ball,in Proceedings of TEPEN 2022 (Eds: H. Zhang, Y. Ji, T. Liu, X. Sun, A. D. Ball), Springer Nature Switzerland, Cham 2023, p.255.   \n[240] O. Ronneberger, P. Fischer, T. Brox, CoRR. 2015.   \n[241] N.Ibtehaz, M. S. Rahman, Neural Netw. 2020, 121, 74.   \n[242] M. R. Gardner, A. Lewis, J. Park, A. B. McElroy, A. D. Estrada, S. Fish, J. J. Beaman, T. E. Milner, Opt. Eng. 2018, 57, 1.   \n[243] K. Wudy, S. Greiner, M. Zhao, D. Drummer, Proc. CIRP 2018, 74, 238.   \n[244] T. Phillips, S. Fish, J. Beaman, Addit. Manuf. 2018, 24, 316.   \n[245] F.Pasa, V. Golkov, F. Pfeiffer, D.Cremers, D. Pfeiffer, Sci. Rep. 2019, 9, 6268.   \n[246] M. F. Hashmi, S. Katiyar, A. G. Keskar, N. D. Bokde, Z.W. Geem, Diagnostics 2020, 10, 417.   \n[247] Y. Zhu, Z. Wu, W. D.Hartley, J. M. Sietins, C. B.Williams, H. Z. Yu, Addit. Manuf. 2020, 34, 101183.   \n[248] N. Satterlee, E. Torresani, E. Olevsky, J. S. Kang, J. Intell. Manuf. 2023.   \n[249] S. Bafaluy Ojea, J. Torrents-Barrena, M. T. Perez-Prado, R. Munoz Moreno, F.Sket, J. Mater. Res. Technol. 2023, 23, 3974.   \n[250] R. Onler, A. S. Koca, B. Kirim, E. Soylemez, Int. J. Adv. Manufact. Technol. 2022, 119, 1091.   \n[251] E. Mendoza Jimenez, D.Ding, L.Su, A. R. Joshi, A.Singh,B. Reeja-Jayan, J. Beuth, Addit. Manuf. 2019, 30, 100864.   \n[252] M. Asadi-Eydivand, M. Solati-Hashjin, A. Fathi, M. Padashi, N.A. Abu Osman, Appl. Soft Comput. 2016, 39, 36.   \n[253] L. J. Segura, Z. Li, C. Zhou, H. Sun, Addit. Manuf. 2023, 66, 103461.   \n[254] A. Yegyan Kumar, J. Wang, Y. Bai, S. T. Huxtable, C. B. Williams, Mater.Des.2019,182,108001.   \n[255] B. Zielinski, T.Sadat, R. Guibert, D. Jouaffre, E. Markiewicz, L. Dubar, J. Nondestr. Eval. 2023, 42, 72.   \n[256] A. Mostafaei, E. L. Stevens, E. T.Hughes, S. D. Biery, C. Hilla, M. Chmielus, Mater. Des. 2016, 108, 126.   \n[257] J. Shijie, W. Ping, J. Peiyi, H. Siping, 2017 Chinese Automation Congress (CAC), 4165-4170. 2017.   \n[258] S. Shrestha, G. Manogharan, JOM 2017, 69, 491.   \n[259] F. Pourkamali-Anaraki, T. Nasrin, R. E. Jensen, A. M. Peterson, C. J. Hansen, Eng. Appl. Artif. Intell. 2023, 126, 106983.   \n[260] H. Baumgartl, J. Tomas, R. Buettner, M. Merkel, Progr. Addit. Manufact. 2020, 5, 277.   \n[261] Z. Jin, Z. Zhang, G. X. Gu, Manuf. Lett.2019, 22, 11.   \n[262] K. Maharana, S. Mondal, B. Nemade, Global Trans. Proc. 2022, 3, 91.   \n[263] D. Kim, J.Lee, Mech. Syst. Signal Process 2022, 179,109363.   \n[264] Y. Wang, G. Huang, S. Song, X. Pan, Y. Xia, C. Wu, IEEE Trans. Pattern Anal. Mach. Intell.2022, 44, 3733.   \n[265] P. Chlap, H. Min, N. Vandenberg, J. Dowling, L. Holloway, A. Haworth, J. Med. Imaging Radiat. Oncol. 2021, 65,545.   \n[266] C. Bunkhumpornpat, K. Sinapiromsaran, C. Lursinsap, Appl. Intell. 2012, 36, 664.   \n[267] S. Demir, K. Mincev, K. Kok, N. G. Paterakis, Appl Energy 2021, 304, 117695.   \n[268] D. Dai, T. Xu, X. Wei, G. Ding, Y. Xu, J. Zhang, H. Zhang, Comput.Mater. Sci. 2020, 175, 109618.   \n[269] F.Anowar, S. Sadaoui, B. Selim, Comput. Sci. Rev. 2021, 40, 100378.   \n[270] J.K. Wilt, C. Yang, G. X. Gu, Adv. Eng. Mater.2020, 22, 1901266.   \n[271] M. A. Bessa, P. Glowacki, M. Houlder, Adv. Mater. 2019, 31, 1904845.   \n[272] W. Adepoju, I. Bhattacharya, M. Sanyaolu, M. E. Bima, T. Banik,E.N.Esfahani, O.Abiodun,IEEE Access 2022, 10,42699.   \n[273] N. M. Litchinitser, Adv. Phys. X 2018, 3, 1367628.   \n[274] S. A. Cummer, J. Christensen, A. Alu, Nat. Rev. Mater. 2016, 1,16001.",
        "category": " References"
    },
    {
        "id": 42,
        "chunk": "# AUTHORBIOGRAPHIES  \n\n![](images/6cde9c3f60e921ddff70b9d0e9b0cf1fe78f7dbe046b56d1f3d98bd2171286aa.jpg)  \n\nTahamina Nasrin is a PhD candidate in the Department of Plastics Engineeringat University of Massachusetts Lowell, where she works under the supervision of Amy Peterson. Her research is focused on the application of machine learning techniques to multilayered polymer composites and investigating innovative approaches for additive manufacturing and multilayered packaging materials. She obtained her Bachelor of Science from the Department of Applied Chemistry and Chemical Engineering at the University of Dhaka, Bangladesh.  \n\n![](images/bacac322eb4af1e490e062e0af8929372ef2b6d229db358e82e4830e5d16fdc5.jpg)  \n\nFarhad Pourkamali-Anaraki is an Assistant Professor in the Department of Mathematical and Statistical Sciences at the University of Colorado Denver. Previously, he was an Assistant Professor of Computer Science at the University of Massachusetts Lowell (2018-2022)  \n\nand received his Ph.D. in Electrical Engineering from CU Boulder in 20l7. His main research interest revolves around transitioning machine learning models from controlled lab environments to realworld settings involving unpredictable and changing conditions, such as accelerating the design and discovery of new materials using cost-effective and uncertainty-aware machine learning models.  \n\nAmy M. Peterson is an Associate Professor and Dandeneau Endowed Professor of Plastics Engineering at University of Massachusetts Lowell with expertise in interfacial phenomena and additive manufacturing. Her research group studies processing-structure-property relationships in polymers and polymer composites, with a focus on interfacial phenomena in multilayered systems. She received her PhD in 2011 from Drexel University. She was an Alexander von Humboldt Postdoctoral Fellow while at the Max Planck Institute of Colloids and Interfaces 201l-2013 and was an Assistant Professor of Chemical Engineering at Worcester Polytechnic Institute 2013-2018.  \n\n![](images/89f0144aa31e664b54b8546852ac760e2fbc7f93721ca70dc5e73690dac805df.jpg)  \n\nHow to cite this article: T. Nasrin, F.Pourkamali-Anaraki, A. M.Peterson, J. Polym. Sci. 2024, 62(12),2639.https://doi.org/10.1002/pol. 20230649",
        "category": " References"
    }
]