Modeling process–structure–property relationships in metal additive manufacturing: a review on physics-driven versus data-driven approaches

Modeling the process-structure (p-s) relationship in MAM involves computational techniques that were established to model traditionally manufactured metals, welding processes that translate to MAM, and, in some cases, specific MAM processes. In process modeling, the physical variables are predicted as a function of space and time during the MAM process. Due to the fact that the MAM process is a multi-scale process, process models can be subcategorized into mesoscale models and macroscale models [7].

Mesoscale models resolve the geometry of individual particles and predict the temperature, pressure, and fluid flow (velocity) field when the source interacts with the particles [8–11]. The calculation of these physical variables is achieved using computational fluid dynamics (CFD) simulations, which numerically solve a set of partial differential equations (PDEs). In order to distinguish the geometry of particles and molten pool, the interface between the metal phase and the gaseous phase needs to be captured. That has been achieved by a variety of interface capturing techniques for multi-phase CFD simulations. Thermo-fluid physics in the building process, such as surface tension, Marangoni effects, laser absorption, and metal evaporation, can be incorporated into the model through special treatment of the boundary conditions and source terms when solving the PDEs. These multiphysics considerations enable the model to predict key physical events, e.g., balling [9–11], keyholing [12, 13] and spattering [14], that lead to mesoscale defects such as lack-of-fusion and porosity [15]. Therefore, mesoscale process models can play a crucial role in quantitatively understanding the formation of these defects.

Macroscale models implement simplifying assumptions of the physics to relieve the computational costs and allow to simulate the building process of an entire part. A major simplification is to treat the metal phase as a continuum without resolving individual particles or the fluid flow in the molten pool [7]. The temperature and stress/strain fields in the part during its building process can be predicted with a thermo-mechanical approach [16]. The macroscale models have shown to predict with considerable accuracy the residual thermal stress/strain after the part is finished [17, 18]. With this predictive power, the macroscale models have been used to optimize process parameters [19]. Compared with mesoscale models, macroscale models are more relevant to solve practical problems but have reduced physical fidelity. The hybrid macroscale-mesoscale modeling approach is a prominent way to balance physical fidelity and practicality [20].

Efforts of structure modeling have been primarily focused on two length scales, mesoscale and microscale. Mesoscale modeling predicts the grain structure resulting from the solidification of the metal, while microscale modeling predicts sub-grain features such as dendrite morphology and phase distribution.

Mesoscale models are used to predict the grain structure in MAM based on either the cellular automata (CA) [21–23] or the Monte Carlo method [24, 25]. The temperature field is a necessary input for the model, and the temperature reduction during solidification is the thermodynamic driving force for nucleation and grain growth to occur. The CA and Monte Carlo methods use different algorithms to predict grain structure evolution based on the temperature information. Each method has its own advantages and disadvantages, and they appear to have comparable popularity in literature. The mesoscale models can reproduce typical grain structures observed in experiments and have shown to capture several key mechanisms of grain structure formation, including the epitaxial nucleation and growth, bulk nucleation, competitive growth, and columnar-to-equiaxed transition. This modeling capability may potentially facilitate the process design to achieve site-specific control of grain structures [26].

Microscale models primarily simulate the evolution of sub-grain microstructure morphologies (e.g., dendrite) and/or phase transformation during solidification based on given temperature history. Phase field is arguably the most widely used method to simulate the sub-grain microstructure evolution for AM processes [27–31]. The method is composed of a set of thermodynamically based PDEs. It is applicable to fairly complex metal alloy systems if the thermodynamic descriptions of the systems are available, usually from the CALPHAD (calculation of phase diagram) database. The computational costs for the phase field are rather demanding, which limits the simulations to relatively small spatial and temporal domains.

Structure-property (s-p) modeling is based on the use of solid-mechanics models to link structure to properties in MAM. In this case, constitutive equations from continuum mechanics are used to describe the evolution of plasticity and, in some cases, damage. The constitutive equations are solved over the simulation domain using numerical methods like the finite-element method (FEM) or fast Fourier transform (FFT) method. In MAM, efforts have been broadly divided into three types of predictions. In the first one, conventional continuum-based modeling is used to predict mechanical properties of AM structures, disregarding any types of AM defect structures. In the second modeling approach, defect structures (e.g., gas porosity, lack-of-fusion) in MAM are incorporated. In the third modeling approach, the underlying crystal structure of the MAM part is incorporated.

Many researchers have performed s-p predictions for AM parts by neglecting the process modeling and excluding any defects and crystal structure resulting from the AM process, thereby treating the part as a conventional, continuum domain absent voids. In an application to lattice components, a number of investigations simulated the mechanical response of AM lattices [32–34] by modeling the geometry of the component, treating each strut as a material continuum. Specifically, Hedayati et al [33] investigated the fatigue behavior of an open-cell lattice component with AM Ti-6Al-4V mechanical properties using FEM. Similarly, work by Andani et al [34] compared experimentally measured and simulated elastic moduli of NiTi open-cell lattice components with a user material subroutine in Abaqus. Another recent effort in which researchers modeled the mechanical response of an AM component treated as a continuum is highlighted in the Third Sandia Fracture Challenge [35–44]. In the challenge scenario, participants were tasked with predicting ductile fracture in an AM 316L part geometry containing multiple internal features, including angled channels, a hole, and a large void. Most of the participants modeled the AM material as a continuum, neglecting to explicitly include details like pore structure, surface roughness, and grain structure, which were provided. Because most of the mechanical response of the AM part was governed by the large-scale geometrical features of the part, many of the teams were able to perform reasonably well in the blind challenge, despite neglecting to model the p-s-p relationships explicitly.

In the second type of modeling approach to investigate the s-p link, multiple efforts have incorporated defect structures in the MAM structure. A number of studies simulate the stress distribution over synthetic grain and/or pore structures under certain loading conditions and predict the effective mechanical properties of the simulated materials [45–47]. In the context of the Sandia Fracture Challenge, one team [44] accounted for AM-induced pore structure in their solid–mechanics simulations. The team leveraged x-ray computed tomography data (which was provided to all participants) to initialize an element-wise value of a damage parameter based on the local pore volume to element volume by juxtaposing the tomography data with a finite-element mesh as shown in figure 4. Another effort [42] adopted a similar constitutive model to incorporate the effects of AM-induced void defects in their predictions with more accurate prediction results. In another effort, Erickson et al [48] established a new, unique descriptor of pore structures that shows a promising relationship with a variety of mechanical properties. The new descriptor is based on a radial distribution function that takes into account the sizes, clustering, and locations of pores within a specific part. The maximum value of the radial distribution function (or the location of the maximum value) was shown to have a stronger correlation with ultimate tensile strength, elongation, toughness modulus, and fracture location than any of the pore metrics previously reported in the literature.

Zoom In Zoom Out Reset image size

Finally, there have been recent efforts in physics-driven, s-p modeling to incorporate the crystal structure of the AM material into predictions of mechanical properties. Work by Ozturk and Rollett [49] provides an example of modeling the s-p relationships using statistically representative volumes of AM Ti-6Al-4V microstructures. The microstructural instantiations were created in DREAM.3D [50] using grain-scale statistics for AM Ti-6Al-4V from the literature. Subsequently, an elasto-viscoplastic FFT code was used to simulate the full-field micromechanical response and corresponding homogenized stress–strain curve of each volume. In another approach, Ahmadi et al [45] simulated an AM microstructure by repeating a melt-pool geometry with a grain structure generated via Voronoi tessellation. They predicted the effective stress–strain response of the polycrystalline domain using finite-element modeling with cohesive elem-ents along the melt-pool interfaces and compared the results to experimental observations. Likewise, Andani et al [46] used a similar grain-structure model but investigated the response of representative volume elements of interest within the simulated microstructure. While the above works consider a simulated AM microstructure in the mechanical simulation, none used physics-based process modeling to predict the microstructure. The accuracy of all s-p predictions described in this section hinged upon rigorous selection and calibration of the material constitutive model(s) used to represent the mechanical behavior of the AM material, which is further discussed later in the manuscript.

In MAM, the investigation of the p-s-p relationships is not limited to the three-way relationship paradigm. A number of efforts bypassed the (micro)structure prediction, and the physics-based model was used to predict, for example, the thermal history during the process followed immediately by a prediction of the residual stress field and part distortions [51–54]. In the work by Johnson et al [52], the transient temperature field was first solved over a finite-element mesh for a 304L stainless steel tube manufactured using laser engineered net shaping (LENS). Output from the thermal model was then used to predict the thermo-mechanical strain as a function of temperature-dependent material properties and, finally, the residual-stress state was predicted using an assumed constitutive model.

While the above modeling efforts focus on modeling either the first link, last link, or bypass the structure link in the p-s-p lifecycle, there are modeling approaches that capture the complete lifecycle from process to mechanical properties for MAM. In this area, Yan et al [47] implemented a framework that, given the processing parameters, predicts using FEM a number of mechanical properties (e.g., yield strength and fatigue) of the simulated microstructure. The process (powder bed flow, heat transfer, and fluid flow) and structure (grains and voids) are modeled and incorporated into the framework. Rodgers et al [6], used direct numerical simulations to simulate tensile loading on multiple hollow cylindrical structures with synthetic microstructures ranging from equiaxed to columnar grains (by changing the processing parameters, namely the scan speed). The 3D, synthetic voxelized microstructures were generated using the kinetic Monte Carlo method. The resultant microstructures were mapped onto a highly refined conformal FE mesh of a part geometry. A grain-scale anisotropic crystal elasticity model was then used to represent the constitutive response of each grain. In a similar framework, detailed in figure 5, Herriott et al [55] implemented a parallelized elasto-viscoplastic FFT model to predict the spatial variability of the yield strength and elastic modulus throughout different MAM build domains, exemplified for SS316L DED build volumes. To achieve that, four microstructural volumes, ranging from fully equiaxed to columnar, were predicted using process modeling (viz., the CA method) to mimic different combinations of processing parameters in the DED process. Then, the full-field mechanical response of microstructural subvolumes extracted from multiple layers of interest within the simulated build volumes was predicted using EVPFFT. Later, the full-field response of each subvolume was homogenized, and stress–strain curves of each subvolume were extracted and analyzed for values of the elastic modulus and yield strength. Finally, the homogenized mechanical properties were plotted using contour maps to investigate their location-specific variability with respect to the laser scan tracks.

Zoom In Zoom Out Reset image size

More recently, Shi et al [56] established a physics-driven framework to establish the p-s-p relationships and tested it on a Ti-6Al-4V single-track LPBF system. The framework consists of a high-fidelity, powder-scale, 3D simulation of transient heat transfer and melt flow dynamics (figure 6(B)), a CA simulation of solidification grain structure and texture, phase field modeling of precipitation and dissolution of second-phase precipitate during repeated thermal cycles as presented in figure 6(A), and microstructure-based micro- and mesoscopic elastic response calculation (figure 6(D)). The framework was capable of simulating concurrent nucleation, growth, and coarsening of the equilibrium $\alpha$ phase in a bicrystalline $\beta$ matrix (figure 6(C)).

Zoom In Zoom Out Reset image size

Some other approaches have combined both physics-driven methods combined with experimental support, which are not discussed in the current review. An extensive review of those efforts are detailed by Smith et al and can be found in [57].

Physics-driven approaches provide a number of advantages and disadvantages that are discussed in detail in section 4. Another approach to model the p-s-p relationship in MAM is data-driven models, which predict the p-s-p relationships of MAM by correlating the expected properties to build conditions (e.g., laser processing parameters, location of a part on the build plate, gas flow in the printer, etc.) that uniquely define the manufacturing process. Such models utilize specific data-driven techniques such as machine learning or data-mining algorithms, for example, to establish the p-s-p relationships and are capable of predicting the behavior of metal components that are manufactured using specific processing parameters without the need to explicitly identify the constitutive model governing its behavior.

The prediction of the p-s link in MAM using data-driven approaches is still at its infancy due to the complexity of the problem and the multiphase, multiscale processes involved. Data-driven methods have been used to uncover the p-s relationship by predicting phase formation in the microstructures of traditionally manufactured metals. In fact, phase formation and microstructures are predicted by mapping thermal processing using continuous cooling transformation (CCT) and time–temperature–transformation (TTT) diagrams. Primarily, artificial intelligence techniques have been employed to predict CCT curves [70–72]. Initial efforts by Wang et al [72] trained an ANN to using 151 CCT diagrams to predict the diagrams of class Fe-xC-0.4Si-0.8Mn-1.0-Cr-0.003P-0.002 S (x within 0.1 through 0.6) steels. Similarly, Dobrzański et al [71] trained a neural network using 400 CCT diagrams made for structural and machinable steels. The model takes as inputs the chemical composition and austenitising temperature and predicted the temperature of the beginning and the end of transformation as a function of cooling rate, the chemical composition, and the hardness of steel cooled from austenitising temperature with a fixed rate which then allow to predict the CCT diagram of the material. Using a similar approach, Trzaska et al [70] trained a neural network and combined it with regression methods to predict the CCT curve for steel with a known chemical composition. The method allowed the analysis of the influence of particular elements on the characteristic points and transformation curves of the supercooled austenite using a set of 300 experimental data sets that are available in the literature. The CCT and TTT diagram of AM metals, however, have been shown to be different from their traditionally manufactured counterparts. Moving forward, the aforementioned approaches could be adapted to MAM processes and used in the prediction of the p-s relationship in MAM.

In the context of p-s linkage for MAM, Popova et al [73] presented a workflow, reproduced in figure 9(A), that can be applied to augment computational materials science using data science approaches. The authors also suggested a general workflow template that could be used to establish the p-s link, as shown in figure 9(B). The blue boxes describe the general steps to achieve that (data pre-processing, microstructure quantification, dimensionality reduction, and extraction/validation of the p-s relationship), while the white boxes present examples of methods that can be used to achieve each. In a specific application of the framework to MAM, the authors established a microstructural descriptor (chord length) on a synthetic microstructure and established the p-s link using distribution calculation, principal component analysis, and multivariate polynomial regression methods. In a different approach, Wang et al [74] successfully trained a two-level surrogate model based on 280 thermal simulations and 150 grain growth simulations to predict temperature profile and materials microstructures. The authors quantified the sensitivity of the material microstructure to specific heat capacity and grain growth activation energies and its uncertainty quantification and presented optimal combination of process parameters for EBM Ti-6Al-4V.

Zoom In Zoom Out Reset image size

Data-driven approaches have been also used to identify microstructural features and defects in MAM structures using in situ data. Combining acoustic emission (recorded using a fiber Bragg grating sensor) and training a neural network, Shevchik et al [75, 76] performed in situ quality monitoring in MAM to identify the formation of porosity during the build, as presented in figure 10. The inputs to the neural network were the relative energies of the narrow frequency bands of the wavelet packet transform, and the outputs of the models provided a classification of the build quality of the layer of interest with accuracy around 83%–89%. In a different approach for the classification of in situ defect formation, other studies focused on the use of visual data and image processing techniques. For example, Scime and Beuth [77, 78] and Caggiano et al [79] employed classification techniques in identifying defects such as keyhole-induced pores and balling from in situ process images. Both investigations implemented convolutional neural networks to identify and classify the types of defects in the printed layer. More details of the employed techniques can be found in the flowcharts in figures 11(A) [79] and (B) [78]. Other classification efforts focused on the identification of microstructural features [80] and identifying feedstock powder source and type for AM metals [81] based on images of the microstructure.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

After printing, MAM parts have been shown to exhibit geometrical deviations post solidification compared to the intended design (CAD model). Such deviations in the geometry from the initial design may limit the certification and qualification of MAM parts, especially in applications where qualification requirements are stringent. To control the geometrical fidelity of MAM parts, Koeppe et al [84] trained a neural network based on FEM simulation results of MAM cell structures to predict load-displacement curves and maximum stress, which were validated using experimental results. The neural network was then used to reproduce the loading history of the FEM simulation and later as a substitute for the physics-driven simulations.

Fewer efforts have been invested in the exploration of the s-p relationship in MAM using data-driven approaches as compared to the physics-driven approaches. In an effort to comprehend the s-p link, Gupta et al [85] proposed a general data-driven framework to investigate the aforementioned branch as detailed in figure 12. The framework consists of three steps that start with the generation of a calibration dataset consisting of representative microstructures (either from measurements or generated synthetically) and their subsequent mechanical response (from an established physics-based numerical model, or experimental results). The second step consists of extracting objective, reduced-order (based on principal component analysis), quantitative measure(s) of the microstructure. In the third and final step, the s-p linkages are established using regression methods and then validated using leave-one-out-cross-validation. The authors evaluated the framework on non-metallic inclusions/a steel composite system and compared it to results from conventional linkage techniques. The results demonstrated high accuracy of the surrogate model that was based on the data-driven framework compared to the conventional approaches. Moving forward, the presented framework could also be applied to s-p data that is available in the MAM literature to establish those linkages.

Zoom In Zoom Out Reset image size

In a different approach, recent efforts applied machine learning techniques to predict the mechanical properties of given metallic microstructures [86–88]. Even though these efforts have been presented for non-additively manufactured materials, the approaches present powerful tools for potential applications in MAM. In the first effort, Jung et al [86] employed the GPR method to predict the full-field response of two-phase microstructures generated with periodic boundary conditions using DREAM.3D. Over 1100 different synthetic microstructures were used to predict the uniform elongation, ultimate tensile strength, and strain localization index defined in the soft phase of each microstructure and later validated using another 1000 microstructural instantiations. The accuracy of the surrogate model was calculated as relative error from finite element simulation results and showed that as little as 40 training points were enough to decrease the error between the surrogate and finite element models to less than 1%. The authors also employed the Efficient Global Optimization method, a Bayesian optimization technique that utilizes an expected improvement algorithm to maximize an unknown black-box objective function to predict the optimal microstructure in the dual phase steel to result in the targeted material properties. The results were shown to predict with high accuracy using a relatively low number of inputs. In a different approach, Mangal and Holm predicted the formation of stress hotspots in face centered cubic materials [87] and hexagonal close packed materials [88] under uniaxial tensile deformation by integrating full-field crystal plasticity-based deformation models and machine learning techniques as summarized in figure 13. In both studies, synthetic microstructural data were generated using DREAM.3D [50]. The local micro-mechanical fields under uniaxial tensile deformation were then simulated using a 3D full-field, image-based, FFT technique, and stress hotspots were defined as grains having stress values above the 90^th percentile of the stress distribution. Then, hotspots were characterized using metrics that reflect local crystallography, geometry, and connectivity. The data were used to create input feature vectors to train a random forest learning algorithm, which predicted the grains that will become stress hotspots. Other work, tailored specifically towards MAM, have also investigated hotspots in build structures with respect to the processing parameters.

Zoom In Zoom Out Reset image size

In a similar approach tailored towards MAM, Herriott and Spear [89] utilized the results from their p-s-p framework discussed in section 2.3 to establish a data-driven model that predicts the s-p link in four AM SS316L build domains exhibiting vastly different microstructures. The main purpose of the data-driven method is to omit the high cost and computational time associated with the crystal-plasticity physics-driven method. To predict the s-p relationship, machine-learning (Ridge regression and XGBoost) and deep-learning (convolutional neural networks) models were trained using microstructural inputs as features and the effective mechanical properties (calculated from the full-field response resulting from the crystal plasticity FFT method) as targets. Output predictions of the effective yield strength in a certain loading direction were then tested for build layers that were omitted from the training set of different MAM build domains. The results were compared to a reference property map (generated using the physics-driven model). An example of the output from the machine- and deep-learning models are presented in figure 14. Among all of the data-driven models tested, the deep-learning models that use crystal orientation as the primary input significantly outperform all other models considered, achieving an R² value upwards of 0.86 in blind predictions of the holdout layers for two different MAM build domains. Furthermore, the deep-learning models offer the benefit of requiring very little pre-processing and feature extraction, instead allowing the model, itself, to extract relevant higher-level features from the microstructural image data. In all cases, the data-driven approach was capable of predicting the properties in a matter of seconds once trained.

Zoom In Zoom Out Reset image size

In another application that is based on the use of structure data, Kantzos et al [90] trained a convolutional neural network to predict stress concentration of rough surfaces, which are detrimental to fatigue life of MAM parts. The model was trained using the elasto-viscoplastic FFT stress field results of a database of 512 synthetically generated surface height maps (figures 15(A) and (B)) that mimic the surface roughness of AM metals and predicted the strain energy metric for each surface. The images were augmented to 1536 data points by performing transformations on the data (rotations and mirroring), and the results showed that the convolutional neural network was able to achieve predictions with an R² value around 0.75 over surfaces that were unknown to the trained model (figure 15(A)). One interesting contribution of this work is the integration of a 'viewpoint' within the convolutional neural network for better interpretability by looking at the activation of the second convolution layer, as shown in figures 15(B)(d).

Zoom In Zoom Out Reset image size

In a review of the efforts undertaken by GE, Aggour et al [91] detailed GE's current efforts to implement artificial intelligence and machine learning methods to manufacturing techniques (both traditional and additive) to support design, processing, and inspection. In terms of MAM, GE presented a framework utilizing probabilistic machine learning and intelligent sampling and optimization protocols, coupled with high-throughput printing, testing, and characterization tools to accelerate process parameter development for LPBF MAM [92]. The process optimization framework was tested on the crack-susceptible Nickel-based alloys used in high temperature locations of aircraft engines. The protocol is based on a Bayesian hybrid model [93, 94], which utilizes a Gaussian process to develop a predictive, data-driven modeling technique, and the GE intelligent design and analysis of computer experiments, as shown in figure 16. The initial model is built with key inputs that are dependent on the processing parameters in use (e.g., laser power, scan speed, hatch spacing, and beam spot size), along with initial constraints and experimental results. In addition to predicting mean values of the model response, the data-driven model also provides uncertainty in the model response. The uncertainty data is then used to iteratively identify new process parameter combinations for test coupons and builds. The framework has been used to minimize the model uncertainty or optimize quantities of interest such as porosity and mechanical properties and was shown to improve with increasing number of iterations. Using these tools, LPBF AM process parameter development cycle time has been reduced by about four times, as shown in figure 16(B) (from [95]), leading to significant savings in material cost, labor, and machine resources compared to traditional design-of-experiment methods. In a different application of part repair using AM, GE used machine learning to define toolpaths (physical path through which the spray nozzle moves when applying material) during a cold-spray process to improve yield and reduce manufacturing time and cost. To perform the repair, GE incorporated artificial intelligence to guide the 12 degree-of-freedom robotic arm movements and allow the robotic arm to adjust in situ to deliver high quality builds of parts such as aviation airfoils. Furthermore, GE shares their experience where the cold-spray deposition process may be improved by incorporating in situ 3D scanning of the deposition surface to intelligently modify the toolpaths on the fly, completing the deposition process more quickly using less material.

Zoom In Zoom Out Reset image size

To establish the p-s-p relationships for MAM, Yan et al [96] proposed a vision for the use of a data-driven multi-scale multi-physics framework, detailed in figure 17. The framework aims at quantitatively predicting the working performance of the final MAM products from given manufacturing parameters for a range of AM technologies (EBM, LPBF, and LENS) by combining process modeling, mechanical modeling, and data mining. The inputs to the model are derived from physics-driven models for all three links (process, structure, and property). Data mining is then used in a cycle of design–predict–optimize, and the preliminary results were validated experimentally. In the future, real-time monitoring is envisioned by Yan et al as a means to validate the numerical models.

Zoom In Zoom Out Reset image size

In another effort, Wang et al [97] employed a physics-informed data-driven surrogate model applicable for MAM EBM technique. Surrogate models for each of the process, structure, and property links were trained using data from physics-driven results. Namely, the meltpool was simulated using FEM (280 thermal simulations), grain growth was modeled using the phase-field method (150 grain growth simulations), and the properties were predicted using the FFT crystal plasticity method (150 elasto-viscoplastic FFT simulations) as shown in figure 18(A). The authors then trained a Gaussian process model to serve as a surrogate model for each of the levels (process, structure, property) and tested the model quality using leave-one-out cross-validation, figure 18(B). Even though the authors communicated the difficulty of generating the thermal surrogate model that is a high-dimensional response, results were shown to capture the temperature fields with reasonable accuracy (figure 18(C)). The final step of Wang's framework [97] consisted of uncertainty quantification using a brute-force Monte Carlo approach, from process to property estimates. Sensitivity analysis was applied to a test case of EBM Ti-6Al-4V alloy, as summarized in figure 18(D). More details about the uncertainty quantification can be found in [97].

Zoom In Zoom Out Reset image size

As it is clear from the previously presented works, the scientific literature provides a wealth of modeling, simulation, and experimental data that captures the processing, structure, and properties of materials. Mining this information from the literature requires the development of a number of tools [98]. Using these tools, data-driven techniques provide a powerful means to learn from 'prior experience'. Extracting information from the literature, Liu et al [99] built a GPR model to correlate process conditions (i.e., laser power, laser speed, hatch spacing, layer thickness, and powder size) across seven laser powder bed machines to density and microhardness of the finished part. Then, using this model, they predicted the build conditions necessary to achieve target densities and microhardness on a build platform that was not part of the original training data. This is one of many studies that leverage existing information in the scientific literature to map process–property relationships directly [99–101].

It is worth noting that another category of data-driven models, models not captured in this review, and not feasible to implement using physics-based models but worth mentioning due to their implementation using data-driven approaches, are property–property models. That is, models that use measurable properties, such as yield strength or springback, to model and predict qualitative properties (e.g., formability and rollability). An example of these efforts, which have not yet been implemented for AM materials, have looked at formability or rollability [102–105].

Both physics-driven and data-driven modeling approaches result in the generation of big/rich data. With the wealth of data in the scientific community comes a growing challenge of data storage, management, and sharing, especially for data relevant to AM. Materials informatics, which accelerate materials, products, and manufacturing innovations, is the way forward in facilitating materials innovation, as detailed in the discussion by Ramakrishna et al [110]. With the availability of the large datasets, a number of databases (such as Citrination [111], the AM Materials Database developed by NIST [112], and the Materials Commons repository by PRISMS [113]) are available to the scientific community to store and share scientific materials data. This list is not exhaustive but is intended to offer examples of how the community is handling data storage and curation. More efforts are needed in unifying and sharing data via established materials databases.

A new paradigm of theory-guided data science was introduced by Karpatne et al [114], in which physics-driven knowledge and data science are integrated to leverage the potential of data science in the advancement of materials discovery. The strength of ICME is the ability to incorporate physics-based modeling data into data-driven models [67, 115–119]. In the specific application of predicting p-s-p relationships for MAM design, output from physics-driven model predictions for MAM digital twins can be used as training data for a data-driven model; in which case, the latter could be trained to predict the relevant output, despite not having explicit access to the governing equations. In turn, the forward predictions from trained data-driven models could be used to inform subsequent iterations of physics-driven models. In this way, data-driven and physics-driven models are not mutually exclusive but can be used to inform each other.

Combining both approaches will allow the physics-driven approaches to support the interpretability of results from the data-driven techniques, revealing the contents of the black box or even making it slightly less opaque. Outputs from the physics-driven models used as inputs to the data-driven models may provide larger sets of input data for training based on governing physics. In turn, the interpretability of the data-driven modeling results may support in formulating more sophisticated governing equations to be used in the physics-driven models in support of the digital twin of MAM parts. The understanding of the p-s-p relationships in the digital twins of MAM parts is ultimately indispensable to the exploration of the MAM design space, which is not feasible solely through experimentation. The exploration of the process design space of MAM will enable the process optimization and the production of parts with specific material properties for a given application to enable qualification and certification of MAM parts.