A Review of the Applications of Machine Learning for Prediction and Analysis of Mechanical Properties and Microstructures in Additive Manufacturing

Parts can be manufactured using various manufacturing techniques. Subtractive manufacturing, forming, and molding processes are some of the most used manufacturing processes. Parts are manufactured by removing material in subtractive manufacturing, while parts are manufactured by applying heat and force in forming processes. Along with these processes, there is another manufacturing process called additive manufacturing (AM), which manufactures parts by adding material layer by layer [1]. Additive manufacturing is gaining popularity because of its ability to manufacture parts with complex geometries, customized parts, parts of hard-to-form materials, and multimaterial parts. Additive manufacturing processes are classified into seven categories [2], which are powder bed fusion (PBF) [3,4], directed energy deposition (DED) [5,6], binder jetting (BJ) [7], material extrusion (ME) AM [8], material jetting (MJ) AM (MJ-AM) [9], vat polymerization [10], and sheet lamination [11]. The four types of AM technologies, namely, powder bed fusion, directed energy deposition, binder jetting, and material extrusion AM, are the most trending AM technologies. The additive manufacturing processes like PBF, DED, and ME melt the material in either powder or wire form by supplying heat and depositing the melted material in a layer-by-layer manner or moving the tailored heat source on the material powder to melt it in a layer-by-layer manner in order to build the three-dimensional (3D) objects of metal, polymer, or ceramic. The vat polymerization process manufactures 3D parts by curing and solidifying a photo-sensitive material layer by layer using a light source or radiation. In the material jetting additive manufacturing process, liquid resin droplets are selectively deposited and cured by exposing them to ultraviolet light to build 3D parts layer by layer. The binder jetting process manufactures parts in a layer-by-layer manner by selectively bonding powder particles together using a bonding agent or binder. This article primarily discusses PBF, DED, ME, BJ, and material jetting additive manufacturing. The general classification of AM process types and the types discussed in this article have been shown in Fig. 1.

Because of the complex functioning of AM processes, it is hard to control them. AM technologies are the combination of various technologies, namely, materials technologies, computer technologies, electronics technologies, thermal technologies, and mechanical technologies. Because of this complex nature, various factors, including process parameters, working environment, stability of AM machine, and material properties, affect the various attributes of the final product, such as quality, properties, and strength [12,13].

Machine learning (ML) is the set of various algorithms that finds the correlation between input and output entities and the hidden patterns in the provided data without using any existing empirical or science-based mathematical equations. The ML models find the correlations between input–output variables and the hidden patterns by analyzing only the data provided to them [14]. The performance of the machine learning models relies upon the quality and quantity of the data provided to the model for training and testing.

ML models are classified into three groups: supervised learning, unsupervised learning, and reinforcement learning (RL), based on the nature of the provided data and functioning method, as shown in Fig. 2. Supervised learning models need labeled data for training [15], while unsupervised learning models can get trained with unlabeled data [16]. In reinforcement learning, the model learns continuously from the actions and maximizes the reward of action [17]. Further supervised and unsupervised ML models can be classified into three groups based on their tasks: classification, regression, and clustering. The classification of the ML models is presented in Table 1.

ML is a part of artificial intelligence (AI), and deep learning (DL) is a special part of machine learning that mimics the human brain’s functioning. It finds hidden patterns and extracts features from unstructured data to perform tasks like classification and regression [18]. The relation between AI, ML, and DL is shown in Fig. 3. Some examples of DL models are convolutional neural networks (CNNs) [19], long short-term memory (LSTM) networks [20], recurrent neural networks [21], generative adversarial networks (GANs) [22], and multilayer perceptrons (MLPs) [23].

According to the existing research, machine learning techniques can handle complex problems in various fields such as computer science [24], robotics [25], aviation [26], biomedical science [27], materials science [28], and manufacturing [29–31]. The ability of machine learning techniques to perform tasks such as regression, classification, clustering, recognition of unseen patterns, and image feature extraction attracts researchers. Because of the tremendous progress in computing and data storage systems, the application of ML has become easy now. Machine learning is applied in various manufacturing processes to control the process and improve the process performance and product quality [32,33]. As discussed earlier, the functioning of the AM processes is complicated; hence, the application of machine learning in AM can simplify the complex nature of AM processes. Researchers have used ML techniques in various sections of AM processes, such as process control, design for AM, process monitoring, process parameter optimization, quality control, defect detection, and prediction of properties [34–36], as shown in Fig. 4.

Properties like tensile strength, hardness, elasticity, and toughness play essential roles in the performance of the manufactured product. The properties of the as-built AM part depend on various AM process factors such as process parameters, solidification rate, microstructure, defects, printing orientation, and material properties. Mapping the relation between these factors and final part properties is challenging. Many published articles have proposed numerical simulation models based on mathematical equations to predict the properties and microstructure of AM parts. [37–45]. Different simulation models such as phase field, cellular automata, and kinetic Monte Carlo have been adopted by researchers for microstructure modeling in AM. For mechanical properties modeling in AM, models like general finite element method, crystal plasticity–finite element model, and cellular automata–finite difference model have also been employed. The accuracy of the simulation models is acceptable, but they are time consuming and need high-performance computational facilities. Also, these mathematical models are based on assumptions and neglect some real-time factors to reduce the computational complexity, for instance, the finite difference-Monte Carlo model used for simulation of microstructure in AM does not consider more detailed physical phenomena like vaporization and Marangoni convection [42,46]. This can affect the accuracy of the final results of these models. Simulation models require high computational time to simulate complex phenomena, e.g., the phase-field model used to simulate grain evolutions in AM can take more than 300 h to complete [47]. Because of their high computational time, simulation models are unsuitable for real-time operations, which demand quick output during the live process. The main advantage of ML models is that they give quick output once they are trained, and because of this, ML models can be used in real-time operations. ML can find the complex relations between the input and output entities. Hence, ML techniques can be used to map the relation between the affecting factors of the AM process and the final AM part’s properties.

Critical and systematic review articles play an essential role in analyzing and studying the application of ML techniques in AM. Some existing review articles discuss the applications of ML in AM, such as topology optimization [48], AM material design [49], AM process control [50,51], and AM defect detection [52,53]. Abdelhamid et al. [54] provided a review on the application of ML in the process–structure–property modeling for material extrusion AM. Toprak and Dogruer [55] presented a review on the application of ML for prediction of properties in the metal PBF process. Thus, the existing literature reviews are focused on specific AM processes. In this review article, we have attempted to present the applications of ML methods in several AM processes (see Fig. 1), such as powder bed fusion, directed energy deposition, material extrusion, binder jetting, and material jetting for prediction, analysis, and amelioration of mechanical properties and microstructures.

The article has been organized as follows. In Sec. 2, we discuss the studies that have used ML techniques to predict mechanical properties and also optimize the process parameters based on the predictions. We have grouped the reviews based on the AM process types. In Sec. 3, we have reported articles that have applied ML methods for predicting and analyzing microstructures of the AM-manufactured parts. We have classified these articles based on the AM type. Section 4 discusses the challenges in generating data for ML in AM. In Sec. 5, we present our conclusions.

AM technologies like PBF, DED, and ME involve various stages. Stages such as melting the powder material by the heat source or fusing the powder particles using a heat source and solidification of the heated layers are involved in PBF processes. DED and ME processes include stages such as melting the material using a heat source, deposition of the melted material, and solidification of the deposited layers. Because of the complexity of the AM processes and the dependency of the properties of the AM product on various factors such as process parameters, AM machine stability, AM machine characteristics, and powder characteristics, the prediction of the AM product properties becomes complicated. As AM, mostly metal AM, is a costly process, it is inefficient to manufacture each part with different process parameters and conditions and measure its properties. Prediction of properties of the final AM product can help to save costs and improve the properties. According to the literature, ML techniques show satisfying performance in predicting the properties of AM parts in terms of prediction accuracy and usefulness in amelioration of AM parts’ mechanical properties [55,56].

Factors like AM process parameters, initial material properties, microstructural features, insights into the solidification phenomenon, and thermal histories of the AM process are the most common inputs, and properties like tensile strength, hardness, toughness, and defects are common outputs for ML models (see Fig. 5).

Process parameters have a prominent impact on the mechanical properties of the final product of AM [44,57,58]. The selection of the wrong combination of process parameters like laser power, scanning speed, and layer thickness can result in the formation of defects in AM parts, such as lack of fusion and keyhole defects, and, thereby, cause degradation in their mechanical properties. The mapping between process parameters and mechanical properties can help to select the right combination of process parameters. Researchers have applied machine learning techniques to map the relationship between AM process parameters and mechanical properties of AM parts, and it was observed that the developed ML models showed good performance in predicting mechanical properties with respect to the input process parameters [56,59–68].

Microstructure parameters, crystal defect, phase morphology, and atomic and phase arrangements strongly influence the mechanical properties of materials [69]. Microstructural features such as size, shape, chemistry and orientation of grains, grain boundaries, transition and arrangement of phases, vacancies, dislocations, and microstructure defects are the indicators of the variation in the mechanical properties of materials [69–73]. Hence, the mechanical properties can be predicted by taking microstructural features as input for ML models [74,75]. Defects in AM such as lack of fusion, keyhole, balling, and cracks influence the mechanical properties of the final part [76–78]. Defect and microstructural features combined can be used to predict mechanical properties more precisely [79].

Thermal conditions formed during the AM process have an impact on the microstructure and mechanical properties of fabricated parts. The changes in temperature with respect to time at a particular location are termed as thermal history. The thermal history leads to the formation of dissimilar microstructures during the AM process, thereby causing variations in the mechanical properties of fabricated parts. Hence, thermal history data can be used to predict the mechanical properties [80,81].

In AM processes, the orientation and positioning of the part also influence its mechanical properties [82,83]. Some researchers have included other inputs like temperature data, material compositions, layer-wise AM process data, and infill characteristics along with process parameters and microstructural features in their training dataset to train the ML model for mechanical properties prediction [84–87].

In the following subsections, we will discuss the current state of application of ML for the prediction of mechanical properties for different AM types.

Lesko et al. [59] used regression-based ML models, namely, ordinary least squares (OLS), partial least squares (PLS), Lasso and Ridge, to set up the relation between laser powder bed fusion (LPBF) process parameters and Vicker’s hardness of nickel-based superalloy 718 parts manufactured by the LPBF process.

Eshkabilov et al. [61] trained various ML models (K-nearest neighbor (KNN), decision tree (DT), and support vector machines (SVM)) to predict the mechanical properties of LPBF parts using the data collected from various published literature, which contains the mechanical properties corresponding to their LPBF process parameters. According to them, laser power, laser energy density, and scanning speed have more impact on the mechanical properties than the other process parameters.

Ali et al. [88] trained an artificial neural network (ANN) model to predict the evolution of local strains, plastic anisotropy, and failure during tensile deformation of selective laser melting (SLM)-fabricated AlSi10Mg. The input data in this study consist of the parameters associated with the tensile test and porosity, such as print direction, tensile deformation time, porosity volume fraction, and size of pores.

Kusano et al. [79] used microstructural as well as defect-characterizing features to train a multiple linear regression (LR) model for the prediction of tensile properties of heat-treated SLMed Ti–6Al–4V alloys. The microstructural features were extracted from scanning electron microscopy (SEM) images using a random forest model, and defect features were obtained from X-ray computed tomography (CT). They proposed that alpha-grain size influenced yield and ultimate tensile strength and presented a mathematical relation between them. Their prediction model showed less than 2% error for the prediction of yield strength (YS) and ultimate tensile strength (UTS) but gave a noticeable scatter for the prediction of elasticity.

Wang et al. [56] applied MLP, KNN, DT, and SVM techniques for optimization of AM process parameters to improve mechanical properties. They demonstrated their methodology on Ti–6Al–4V alloy processed by electron beam melting (EBM) process and obtained a 5% improvement in UTS after using optimized parameters obtained from the developed ML model.

To find a processing window that gives high density in LPBF-fabricated AlSi10Mg, Liu et al. [89] developed a Gaussian process regression (GPR) model, which established a relationship between the processing parameters, namely, laser power and scan speed, and the relative density.

Minkowitz et al. [68] implemented machine learning models based on decision tree regression (DTR) and extra trees regression (ETR), to predict mechanical properties, namely, density, ultimate tensile strength, and hardness of AlSi10Mg parts fabricated by the LPBF process. They used process parameters (laser power, laser spot size, hatching distance, layer height, and scan speed) as input to train the ML models. In this study, ETR models are used to predict density, ultimate tensile strength, and hardness, and the DTR model is used to predict density. For density prediction, the best $R2$ values of ETR and DTR models are 98.54% and 97.26%, respectively, and the best $R2$ values of ETR models for the prediction of ultimate tensile strength and hardness are 88.55% and 83.81%, respectively.

Wang et al. [84] predicted the tensile strength of Ti–6Al4V AM parts using process parameters, annealing temperatures, and microstructural features as the input to the MLP model. In their study, an MLP model employed for the prediction of tensile strength shows a $R2$ value of 0.804 when microstructural features are not included in the input data and shows a $R2$ value of 0.907 when microstructural features are included in input data. They observed that the MLP model can capture the influence of laser parameters, build orientations, and annealing temperatures after including microstructural features in the input training data, and it leads to improving the performance of the MLP model for tensile strength prediction.

Pashmforoush and Seyedzavvar [90] predicted the mechanical properties of SLM specimens with different metallic materials using a transfer learning-based artificial neural network. The authors have reported a significant improvement in the prediction accuracy of the ML model after using the transfer learning technique, even though using a small training dataset.

Some published articles have included orientation and positioning of the part in the training data of ML models, such as DT, gradient-boosting regression (GBR), and Ada Boost, for predicting mechanical properties [91].

Johnson et al. [92] have modeled porosity defects in specimens using the finite element method and generated a training dataset for a CNN-based deep learning model to predict the failure of the specimen measured in metrics of peak load (84.7% accuracy), critical displacement(84.7% accuracy), equivalent plastic strain along with critical load (89.2% accuracy), and equivalent plastic strain along with critical displacement (91.9% accuracy).

He et al. [93] have used a regression-based ML model to identify a processing window for manufacturing TiCN/AlSi10Mg composite with configurable yield strength at fixed density.

Yu et al. [94] created a machine learning-based framework for the prediction of the mechanical properties like hardness and relative mass density of AlSi10Mg nanocomposites by using microstructural texture features data. They used machine learning models such as AdaBoost, gradient tree boosting, KNN, DT, and ETR.

The performance of an ML model varies depending on factors such as the model type, the inputs, and the material used to fabricate a part. A variety of inputs are used to train the ML models for property predictions. The performance of Ridge, Lasso, OLS, and PLS models are below par for the hardness predictions of AM-fabricated Inconel 718. ETR model shows satisfactory performance for the hardness prediction of AM-fabricated AlSi10Mg. For UTS predictions, ETR and MLP models show good performance. The inclusion of microstructural features in the dataset of process parameters boosted the performance of the ML model. ETR and DTR models showed excellent performance in the prediction of relative density. Except for the LR model, the ML models, namely, DT, GBR, and AdaBoost, showed good performance for the prediction of elastic constant. We have summarized the performance of ML applications in PBF as per the $R2$ scores in Fig. 6(a).

Fang et al. [80] presented the one-dimensional CNN model, which extracts features from thermal history data obtained from the validated computational thermal model and predicts the mechanical properties of the final Inconel 718 part built by the DED process. They proposed that the CNN model predicts the UTS accurately with an $R2$ score of 0.96 and 0.67 for the training and testing, respectively.

Xie et al. [81] predicted location-dependent mechanical properties with the help of a CNN model trained over the data consisting of thermal histories.

Chigilipalli and Veeramani [95] have included applied load, build orientation, and positioning of the part in the training data to predict the tensile deformation behavior of wire arc additive manufacturing (WAAM) fabricated Incoloy 825 using various ML models, namely LR, DT, SVM, KNN, and random forest (RF).

Era et al. [60] used process parameters of the DED process, namely, layer height, scanning speed, laser power, and energy density, to train the ML models for the prediction of tensile properties of the stainless steel 316L parts manufactured by the DED process. They used extreme gradient-boosting (XGBoost) and RF models for the prediction of tensile properties and mentioned that the XGBoost model comparatively performed better than the RF model with $R2$ values of 76% and 72% for UTS and YS, respectively.

Some researchers have predicted properties using ML and further optimized process parameters to obtain the best combination of process parameters to obtain improved mechanical properties. Barik et al. [64] used support vector regression (SVR) model for optimizing the process parameters for manufacturing low-carbon steel using the WAAM process.

The ML models trained on the thermal history-based input data show comparatively lower performance than those trained on the process parameter-based input data for the prediction of UTS and YS. This shows that the process parameters have more impact on tensile properties than thermal histories. The existing CNN model trained on thermal history data and used for the prediction of elastic constant shows poor performance. Here, there is a scope for research to improve the elastic constant prediction performance. The SVR model shows excellent performance in the prediction of UTS using process parameters as input. CNN and XGBoost models show better performance for the prediction of tensile properties than the RF model. We have summarized the performance of ML models for DED with respect to the reported $R2$ scores in Fig. 6(b).

Along with metal additive manufacturing processes, literature has witnessed the applications of machine learning techniques for the prediction of mechanical properties in polymer AM as well.

Alafaghani et al. [66] used ANN models to examine the influence of process parameters such as layer thickness, nozzle temperature, infill percentage, and infill pattern on the mechanical properties of acrylonitrile butadiene styrene (ABS) parts fabricated by fused filament fabrication (FFF). They found that the ANN models are more accurate for the prediction of mechanical properties than the design of experiment (DOE) regression models. In their study, the $R2$ values of the ANN model for the modulus of elasticity and ductility prediction are 0.968 and 0.873, respectively, while the $R2$ values shown by the DOE model for the modulus of elasticity and ductility prediction are 0.896 and 0.658, respectively.

Machine learning models like Lasso regression, SVR, XGBoost, RF, and KNN showed good performance in predicting the tensile strength of polylactic acid (PLA) parts manufactured by FFF or fused deposition modeling (FDM) [65,67].

Hossain et al. [96] fabricated polymer composite parts in which PLA and high-density polyethylene (HDPE) are the primary materials and optimized process parameters using ML to obtain parts with more reliable and improved mechanical properties.

Mishra and Jatti [63] applied the RL-based Q-learning model to optimize FDM process parameters and achieve improved mechanical properties of FDM parts. In the aforementioned study, the predicted values of tensile strength differ by only 0.66% from experimental values. Deep neural networks have also shown good accuracy in the selection of the best combinations of parameters of the polymer AM process to ameliorate the mechanical properties of the built part [97].

Regalla et al. [98] have applied transfer learning of AlexNet to study defects in parts manufactured using FDM to enhance mechanical properties eventually.

Along with tensile strength, ML models are capable of predicting the hardness of polymer parts fabricated by AM [62]. Machine learning is also capable of predicting anisotropic properties and fiber orientation of extrusion deposition additive manufacturing (EDAM) fabricated fiber-reinforced polymers like short fiber-reinforced polymer (SFRP) [99].

In ME-based AM, machine learning models show satisfactory performance in the prediction of polymer properties, namely, hardness, UTS, elastic constants, and ductility, and properties of composite material like SFRP. The SVR model shows excellent performance for finding properties of SFRP with an $R2$ value of 0.99. The ANN displays strong predictive capabilities for several properties, such as elastic modulus, UTS, and ductility. ANN slightly outperforms the LSTM model in predicting UTS for polymers. This could be due to discrepancies in the dataset or different material types. RF and Adaboost models show good performance for the prediction of the hardness for AM-fabricated ABS. Here, there is a scope for research into property predictions for AM-fabricated PLA using ML to predict properties other than UTS. ML can predict composite properties with satisfactory performance, and this area of research can be investigated for various composites using different ML methods. We have summarized the performance of ML models in ME based on their $R2$ scores in Fig. 6(c).

Yu̇ksel et al. [100] employed a deep learning-based GAN model to improve the mechanical properties of lattice structures fabricated by MJ-AM.

Next, we present some studies in which the authors have not mentioned any specific AM type.

Gu et al. [75] employed the CNN model to predict the mechanical properties of the hierarchical system by taking microstructures containing unit cells as input and validating the generated designs by 3D printing and testing.

Herriott and Spear [74] predicted the mechanical properties of stainless steel 316L manufactured by metal AM (exact type not specified), taking microstructure features and images as input to ML models. This study used three ML models for predictions, namely, ridge regression [101], XGBoost [102], and a custom 3D convolutional neural network (3D CNN) based on VGGNet. Morphological and crystallographic microstructure features were used as input for the ridge regression and XGBoost models, while 3D microstructure images carrying informative data, such as grain ID and crystal orientation, were used as input for the 3D CNN model. This study reported that the performance of the 3D CNN model, which used crystal orientations as input for predicting mechanical properties, was better than the other two models. The 3D CNN model is used in the aforementioned study because the training data are three dimensional, that is, 3D image data of microstructure. The 3D CNN model can analyze 3D images or videos, unlike the conventional 2D CNN model, which can handle only two-dimensional data (see Fig. 7). The convolutional and pooling layers of 3D CNN are also three dimensional [103].

Yan et al. [87] incorporated material composition and process parameters in the ML training dataset to predict mechanical properties using the Gaussian process metamodels.

Hu et al. [104] investigated the effect of micro-defects, such as surface roughness and interior voids, on the mechanical properties of the AM lattice structures prone to stretch and bending phenomena with the help of GBR-based ML model. They found that ML is useful for the analysis of the data as the ML model used in the study clearly highlighted that the surface roughness and deformation mode have a prominent impact on the stiffness and strength of the AM lattice structures. Also, the GBR model shows very high prediction accuracy with $R2$ values of $∼1$ and 0.98 for stiffness and strength, respectively.

Overall, we can notice that the number of studies dealing with the prediction of the microstructure of AM-manufactured parts is small. There is no clear trend or logical progression among the studies, although there is an overarching goal to improve the mechanical performance of the parts. The studies differ in the target microstructural properties—grain neck size, material phase distribution, crystallographic orientation, grain boundaries, porosity, and 3D chord length distribution.

The applications of ML for the prediction and amelioration of mechanical properties in AM are summarized in Tables 2 and 3.

Analysis and study of the microstructure of a material is crucial to understanding its behavior. AM processes involve phase transformations and solidification, which can cause variations in the microstructures of AM-produced samples. During AM, distinct hierarchical microstructures can be created for metallic materials, giving them a variety of superior qualities and properties [105,106]. Hence, it is necessary to have a thorough understanding of the microstructures formed during an AM process in order to obtain materials with excellent properties and qualities. Close scrutiny of microstructural features like grain size, grain structure, pores, and grain orientation is complex and time consuming. Also, it is difficult to predict the microstructure before the fabrication of the part, although some researchers have attempted numerical simulations to this end [37,41–44,107,108]. The applications of ML techniques in AM microstructure prediction and analysis are shown in Fig. 8. Microstructure images can be obtained using SEM and electron backscatter diffraction (EBSD). The microstructural characteristics can be analyzed and extracted from images generated from SEM and EBSD using machine learning algorithms.

In the following subsections, we present the studies targeting the prediction of microstructure for various AM types.

The optimization of microstructural features can help improve the properties of AM-fabricated parts. In the selective laser sintering (SLS) process, the microstructure evolves through distinctive stages, such as the formation of neck and grain boundaries, growth of neck length. Optimizing the neck size is crucial for improving the quality and strength of sintered parts. Batabyal et al. [109] applied the Gaussian process regression model to analyze microstructure evolution obtained from a two-phase field model. They also determine the ideal values of surface diffusivity and interparticle distance to maximize the neck size for equal and unequal particle sizes while also quantifying uncertainty.

Cao et al. [110] used an image-driven conditional generative adversarial network (cGAN) to form the processing parameters—microstructure relationship and predict key microstructural features (e.g., martensite morphology and size) in LPBF-fabricated Ti–6Al–4V. Micrograph image data from samples built with different process parameters were used to train the ML model. The ML model predicted the microstructural features with an error of less than 20%.

Wittwer and Seita [111] presented a CNN-based approach to predict crystal orientations by mapping the relationship between directional reflectance microscopy (DRM) signals and Euler angles. A set of Euler angles represents the crystallographic orientation; hence, prediction of Euler angles further helps to predict crystallographic orientations. In this work, an experimental dataset was obtained from samples of Inconel 718 manufactured by a DED process.

Liu et al. [112] compiled a dataset of single-layer deposition experiments conducted by Wire-feed laser additive manufacturing (WLAM) process under regulated processing parameters to train a Gaussian process regression model for the prediction of microstructural features, including alpha phase thickness and beta grain length.

Grain boundary segmentation is necessary to measure the grains in the microstructure images accurately. Segmentation of the grain boundaries can be done manually or using computational methods. Manual segmentation is time consuming, while computational segmentation methods are fast but less accurate. Hence, Warren et al. [113] employed a CNN model for speedy segmentation of the grain boundaries in microstructure images with good accuracy. In this research, the training dataset was generated using manual segmentation and existing computational segmentation methods. Artificial grain images were created by combining Voronoi tessellation, random synthetic noise, and simulated flaws to increase the size of the training dataset. An experimental dataset was also obtained by manufacturing cubical samples of 316L stainless steel using a binder jetting AM machine.

Ojea et al. [114] developed an pixel-wise classification model using a light gradient-boosting machine (LightGBM) to quantify SEM micrographs of green parts and extract microstructural characteristics of binder in the binder jetting AM process.

To calculate porosity and analyze microstructural features like contact surfaces and changes in fiber morphologies, Özen et al. [115] developed a two-dimensional (2D) digital image correlation code using a K-means cluster-based ML model. In their study, the 2D porosity images were converted into 2D arrays and binarized to serve as data points in the K-means cluster algorithm. The binarization converted the pixels to black and white, which allows calculation of porosity by counting the black pixels. The $k$-means cluster algorithm is used to cluster data with similar characteristics, which allows them to count the black pixels and thereby calculate porosity in their study.

In Sec. 3.5, we report studies in which the authors have either not specified the exact AM process type used or their results apply to multiple AM types.

Han et al. [116] introduced a deep learning approach to quantitatively analyze the microstructural changes of metals produced through additive manufacturing in various processing conditions by extracting microstructural descriptors from EBSD patterns. They have employed a regeneration neural network for the prediction of new microstructures. This study regenerated the microstructure images for the input EBSD measurement data and also successfully predicted the new microstructure images for the EBSD measurements, which are not included in the input data.

Along with microstructural features, the prediction of the microstructural images and evolution histories is also possible by using machine learning [116,117]. Qin et al. [117] introduced a sequence-to-sequence LSTM deep neural network for the prediction of the spatiotemporal microstructure evolution history by taking morphology-controlling parameters and initial grain states, which includes temperature gradient, pulling velocity, kinetic anisotropy, number of grains, grain width, and grain orientations.

Ackermann and Haase [118] trained various ML models, namely, polynomial regression (POLY), RF, SVR, and MLP on the physics-based, low-dimensional microstructure data generated from kinetic Monte Carlo microstructure simulations. In this study, the morphology descriptor named 3D chord length distribution was predicted using ML models by taking process parameters (rotation angle per layer, melt pool characteristics, heat-affected-zone width, and velocity) as input. The highest accuracy of the ML model reported was 90.6%.

Some researchers have used machine learning to assist numerical modeling procedures associated with microstructure analysis and predictions. Snider-Simon and Frantziskonis [119] proposed a new workflow that uses digital image processing, machine learning, and spatial statistics to generate two-dimensional spatial models from digital micrographs of material microstructures. These models can then be used for statistical reconstruction modeling in numerical procedures like finite element analysis. In this study, they developed the spatial model of hydrogen porosity for AM-fabricated AlSi10Mg and later used it in the finite element-based reliability assessment of the material.

Montes de Oca Zapiain et al. [120] developed a machine learning model that can accurately predict the nonlinear microstructure evolution of a two-phase mixture during spinodal decomposition to accelerate the high-fidelity phase-field simulations by using the outputs from the ML model in the simulations.

The applications of ML for the prediction and analysis of microstructures in AM are summarized in Table 4.

The characteristics of the dataset used to train the ML model influence its performance [121]. The accuracy of the ML models depends on the size and relevance of the dataset used [122,123]. Adequate and relevant data are required for effective training of the ML models. Hence, it is necessary to discuss the data generation techniques. In the articles that are discussed in this article, the data are generated through experiments and numerical simulations to train the ML models. Also, in some articles, the experimental or simulated data available in the literature are collected to train the ML models.

Different AM machines, testing machines, specialized pieces of equipment, and sensors are used to generate training data by performing experiments, as shown in Fig. 9. Some of these equipment are very expensive, and it may be difficult to mount sensors on the AM machines. Parts are fabricated with different process parameters and tested on testing machines, like universal testing machines and hardness testing machines, to measure mechanical properties. Microstructural features are extracted from SEM, EBSD, DRM, and digital microscope images. IR cameras and temperature sensors are used to generate thermal history data. Acoustic emissions (AE) sensors are also used to monitor the AM process for defects [124–126]. Micro-CT is also an effective method to identify defects in AM [79,127]. Numerical simulations can be performed to generate a training dataset for ML. A finite element method-based thermal simulation model is used to generate the thermal history data [80].

The generation of large experimental datasets on additive manufacturing, mainly metal additive manufacturing, is time consuming and costly. The experimentally generated data or data generated from simulations can be expanded by using data augmentation techniques. Data augmentation is a technique to artificially generate new data from existing data [128]. Especially when the training data are in the form of images, then the dataset can be expanded by performing data augmentation operations like image flipping, image rotation, color adjustments in images, and the addition of noise in images. Hence, data augmentation techniques need to be explored to make attempts to enhance the performance of ML models by expanding the size of small datasets [116].

The recently introduced applications of ML-based methods in additive manufacturing for prediction and amelioration of mechanical properties, as well as prediction and analysis of microstructure, were comprehensively reviewed in this review article. Applications of ML methods for property prediction are present in almost all main types of AM processes, namely, PBF, DED, ME-based AM, and BJ. It can be observed that a majority of the current applications of ML in AM mechanical properties prediction and amelioration are concentrated on a few materials, namely, Inconel 718, Inconel 825, Ti–6Al4V, SS316L, ABS, PLA, ER70S6 steel, AlSi10Mg, SFRP, HDPE, and TiCN. More work is needed on materials like magnesium alloys (like WE43), aluminum alloys (like AlSi12, AlSi7Mg, and A356), steel alloys, and nickel alloys like Inconel 625, which are also extensively fabricated by additive manufacturing. The training dataset of ML models for the prediction of AM properties consists of inputs like process parameters, microstructural features, defect features, design features, thermal histories, or a combination of two or more of these inputs. Of these, process parameters like laser power, scan speed, hatch spacing, layer thickness, and laser spot size are more popular. Build orientation, design feature, topology, and alloy composition are also found as inputs in some studies. Mechanical properties like YS, UTS, hardness, impact strength, stiffness, and compressive strength are the most common outputs used for training ML models. Microstructural features and thermal history are used as inputs to ML models only for the prediction of tensile properties. It is expected that defects in AM parts effect the final mechanical properties. As of now, only features related to porosity defects have been studied using ML for their effect on the mechanical properties, but with little success. The influence of other defect-related features on the tensile properties needs to be investigated further.

It is difficult to compare the ML models present in the articles on the basis of their accuracy because of dissimilar target properties, AM processes, and fabricated materials. Prediction of tensile properties using ML remains the main focus of researchers. ML models show satisfactory performance in metal AM for predicting UTS and YS, but the performance of ML is comparatively lower for predicting the modulus of elasticity than predicting UTS and YS in metal AM processes. The performance of ML models for prediction of modulus of elasticity in DED is relatively poor.

The majority of the reviewed articles have focused on establishing the relationships between process parameters and properties in AM using ML methods, whereas only a few articles use ML methods to establish the relationships between microstructure and properties. As of now, only one study has attempted to find a linkage between all three variables using ML. Still, they used only surface information instead of microstructure information to train the model because of the difficulty and high cost of microstructural data generation.

The generation of a large experimental AM dataset is time consuming and costly. Different equipment, such as universal testing machine, hardness testing machine, IR camera, SEM, and EBSD microscopes, are used to obtain the data to train the ML models. Recently, AE sensors have also become popular in identifying defects in AM. AE sensors and micro-CT can be useful to generate AM defect-related data for predicting properties and microstructures in AM. Data generation from numerical simulations of AM processes is a good option, but it has limitations, such as time consumption and requirement of high processing computational facilities. Adaptation of data augmentation and transfer learning techniques may help to train ML models better, even with small datasets. Also, new ML techniques like few-shot and one-shot learning may be useful.

After observing the trend of research, it can be pointed out that the research on the application of ML models for the analysis and prediction of AM microstructures is still in its infancy. The reviewed applications of ML in AM for microstructure analysis are mainly focused on the prediction of microstructural features, the acceleration of numerical simulations for microstructure analysis, and the optimization of microstructural features to enhance the properties and quality of AM-fabricated materials. The applications of the ML techniques for microstructure prediction and analysis are limited to materials, namely, steel 316L, Ti-6Al-4V, Inconel 718, PETG, and copper. Data from directional reflectance microscopy and microstructural images generated from SEM and EBSD are predominantly used in the ML input dataset for microstructure prediction and analysis. Neural networks are the most popular ML models, followed by SVR, GPR, and POLY. Efforts have been made to predict microstructure corresponding to different process parameters and improve the microstructure of materials using ML methods. In the future, we can expect research on the ML-based development of new microstructures in AM that provide superior qualities and properties to the AM parts.

There are no conflicts of interest.

The datasets generated and supporting the findings of this article are obtainable from the corresponding author upon reasonable request.