Process-Structure Linkages Using a Data Science Approach: Application to Simulated Additive Manufacturing Data
- 2.1k Downloads
A novel data science workflow is developed and demonstrated to extract process-structure linkages (i.e., reduced-order model) for microstructure evolution problems when the final microstructure depends on (simulation or experimental) processing parameters. This workflow consists of four main steps: data pre-processing, microstructure quantification, dimensionality reduction, and extraction/validation of process-structure linkages. Methods that can be employed within each step vary based on the type and amount of available data. In this paper, this data-driven workflow is applied to a set of synthetic additive manufacturing microstructures obtained using the Potts-kinetic Monte Carlo (kMC) approach. Additive manufacturing techniques inherently produce complex microstructures that can vary significantly with processing conditions. Using the developed workflow, a low-dimensional data-driven model was established to correlate process parameters with the predicted final microstructure. Additionally, the modular workflows developed and presented in this work facilitate easy dissemination and curation by the broader community.
KeywordsPSP linkages Workflows Microstructure quantification Additive manufacturing Monte Carlo simulation
Acceleration in the rate of material development and deployment has been the focus of several recent efforts in current literature (e.g., [1, 2, 3, 4, 5, 6]). In this regard, multiscale modeling and simulation has been identified as a key enabler [7, 8, 9, 10, 11], because of its potential to dramatically reduce time and effort expended in experimentation. However, there is now an increasing recognition that this alone cannot bring about the desired acceleration in material development. There is a critical need for the development and deployment of a suitable supporting data infrastructure that efficiently integrates closed-loop iterations between experimental and multiscale modeling/simulation efforts. This need is being addressed by a new cross-disciplinary field known as materials data science and informatics [1, 3, 12, 13, 14, 15, 16, 17, 18, 19, 20].
A central impediment in the implementation of the approach described in Fig. 1 comes from a lack of validated and broadly adopted frameworks for the rigorous quantification of hierarchical material structures or microstructure. Microstructure plays a central role in the formulation of PSP linkages and is often an important input and/or output. Furthermore, microstructure can often require a higher dimensional representation compared to other variables involved in the PSP linkages. From a practical viewpoint, it becomes essential to seek suitable reduced-order representations of material structure and use them in formulating PSP linkages. Traditionally, this dimensionality reduction has been performed by materials scientists based on intuition or insight of the materials phenomena studied. As a specific example, one might quantify polycrystalline microstructures using grain size or shape distributions, and possibly orientation and misorientation distributions, when studying their plastic response. However, such approaches have not yet identified a common set of low-dimensional measures that can be universally applied across diverse material systems for identification of a majority of material response characteristics. This, however, is a key element in the formulation of re-usable, high-value, material knowledge systems.
Emerging toolsets in materials data science and informatics have demonstrated tremendous promise in addressing some of the key challenges described above. It is now possible to generate a large ensemble of datasets (inputs and outputs) from a simulation toolset and publicly share these with the broader scientific community in an open-access data repository . Once this is accomplished, it is possible to engage the broader scientific community in the extraction of the embedded knowledge of these datasets. If this activity is guided in a suitable framework for PSP linkages, it could lead to accelerated and robust curation of the knowledge, while simultaneously ensuring the highest levels of access, sharing, and dissemination for re-use.
The main goal of this work is to explore the viability of the concepts and philosophies described above with an example demonstrator focused on process-structure (P-S) linkages with a view toward additive manufacturing. Additive manufacturing (AM) is a rapidly growing field of advanced materials processing [26, 27]. Process improvements in recent years have enabled the creation of near-fully dense parts with sophisticated geometries that are unobtainable using traditional manufacturing techniques . While AM has seen significant adoption as a prototyping and small-batch production tool, the science behind AM part creation is complex and only partially understood. Variations in factors such as powder composition, processing technique, and component shape can result in dramatically different microstructures and material properties. Additionally, microstructure can vary significantly even within a single as-built part. The interplay between the length scales of AM builds and those of processing (e.g., localized melt pool size and shape) presents new challenges in the analysis and prediction of microstructure-sensitive performance characteristics. Furthermore, irregular component geometries and material anisotropies create compounded difficulties for traditional analysis methods .
Among the many processing variables of interest, beam power density and scan pattern stand out as relatively dominant factors. Power density is directly controlled by beam parameters (spot size, power, scan rate, etc.), but is also indirectly influenced by the scan pattern used to construct the build. Together, power density and scan pattern greatly influence both the overall microstructure and the local microstructural variations [30, 31, 32]. Although a number of experimental and simulation studies are underway [27, 33, 34, 35] to quantify the P-S relationships in AM, the opportunity for advanced data analysis has also been recognized [27, 35, 36, 37, 38]. The multiscale heterogeneity present throughout a solidified AM build would suggest that a rigorous, quantitative, and statistical analysis is essential to achieve high-fidelity success in the realm of qualification for significant industrial or high-consequence applications .
A Monte Carlo Potts model has been employed successfully to simulate grain growth , recrystallization , electron beam welding , and AM processing , and has demonstrated a remarkable qualitative agreement with experimental data. The simulation method yields predictions of three-dimensional (3-D) polycrystalline microstructures under a variety of scenarios and has even been demonstrated to couple effectively with additional models for the inclusion of additional physics . With recent advances in computational infrastructure, it is now possible to conduct a large number of simulations to generate an aggregate dataset composed of thousands of individual simulations, where input parameters are systematically varied to cover specific ranges of interest. While extracting re-usable P-S linkages in the form of low-cost surrogate models from these datasets is a non-trivial task, this paper will address this task using emerging toolsets of materials data science and informatics.
Additive Manufacturing Simulation Dataset
The approach described here allows for rapid exploration of varying simulation conditions and the use of relatively large simulation domains (300 × 300 × 200 elements) at low computational costs. The kMC simulations are non-dimensional but include an implied length scale resulting from the shape of the molten zone and the height of the remelt layers as both determine significant amounts of the resulting microstructure’s arrangement. In the simulations presented, a molten zone of 60, 70, 80, or 90 sites corresponds to physical dimension of 0.3, 0.35, 0.4, or 0.45 mm, respectively. These layer-by-layer structures with limited remelting of prior layers are consistent with the low-power experimental validation comparisons of  presented in . A total of 1799 microstructures (each corresponding to a different combination of process parameters) were generated on a Linux-computing cluster to comprise the ensemble dataset for this study. In comparison, state-of-the-art thermofluid, multiphysics, simulations of AM processes are generally capable of simulating only a single pass under a similar computational cost .
The range of simulation conditions used in the study
(X/XY) Scan pattern
Parallel (X) or cross hatch (XY)
(W) Molten zone width (lattice sites)
60, 70, 80, 90
(V) Velocity (sites/Monte Carlo step)
2.5, 5, 7.5, 10, 15
(D) Molten zone depth (sites)
(L) Molten zone tail length (sites)
50, 60, 70
(HAZ) Heat-affected-zone width (sites)
5, 20, 35
(T) Tail heat-affected-zone length (sites)
5, 20, 35
Data Science Workflow for Extracting Process-Structure Linkages
The first step in the workflow is a pre-processing step aimed at ensuring quality and consistency of the dataset. While the identification of the phases, boundaries, or other features of interest in simulated data is trivial in most cases, experimental data often requires segmentation of images to properly identify a given feature of interest. As needed, one might set a criterion to eliminate spurious or questionable data (e.g., the data that does not conform to known physics). In this step, the inputs (process parameters) are also clearly associated with the outputs (microstructure data).
In the second step, microstructures are quantified to obtain salient statistical measures of microstructures. In a data science approach, it is desirable to capture a very large set of measures at this stage. Consequently, it is preferable to adopt a microstructure quantification framework that allows one to increase systematically the numbers of potential features included in the analyses. In this regard, the framework of n-point spatial correlations [12, 53, 54] offers tremendous promise because of its scalability (ability to define an infinite number of microstructural features), organization (value of n can start with one and increase systematically), and available access to efficient computational toolsets [55, 56]. Another option for this step includes lineal path functions  or chord-length distributions [58, 59] that provide information about shape and size distribution of a specific feature of interest.
The third step in the workflow focuses on reducing the dimensionality of microstructure representation using data science approaches. Some of the established dimensionality reduction techniques include principal component analysis , factor analysis , projection pursuit , and independent component analysis , among others. These methods are designed to reduce dataset dimensions, while losing only the smallest amounts of information. The use of dimensionality reduction leads to savings in both computational time and storage, and leads to identification of salient features that can be used to establish models. For example, in prior work , PCA has proven to be remarkably efficient in producing high-value, low-order, representations of microstructures that are ideally suited to establishing PSP linkages in a broad variety of material systems.
After obtaining a data-driven model, errors are checked, and if they do not satisfy the error criteria, a new iteration in model building is launched (see Fig. 4). It is, however, important to identify which step contributed most to the unreliable model. If this insight is available, suitable modifications can be implemented in any step of the workflow in the next iteration. For instance, one might select a different model learning algorithm or identify new features using a different dimensionality reduction technique. The modular nature of the workflow shown in Fig. 4 allows one to explore a very large number of potential models in highly computationally efficient toolkits [67, 55] before settling on the best model for the phenomena studied.
Suitable error criteria for an acceptable model should be defined or set by the user for any practical implementation of the workflow shown in Fig. 4. These criteria are likely to be highly dependent on the intended purpose of the reduced-order model extracted using this workflow. In most MGI or ICME applications, a materials designer is likely to use the reduced-order models for rapid screening of a large design space under consideration. Therefore, the requirements for accuracy should be based on obtaining reliable guidance for meaningful down-selection of the design choices. Note also that all data sources (in the present case, simulation codes employed to generate an ensemble of microstructures) inherently exhibit certain (often non-negligible) uncertainty (or inaccuracy) that can be attributed to the numerous approximations and idealizations employed. Therefore, it would be unwise to establish an error criterion that exceeds the inherent uncertainty in the data source.
Case Study: Application to Additive Manufacturing Datasets
The workflow discussed in Fig. 4 provides a generalized template to extract a P-S linkage from a collection of data points, where each data point includes both the final microstructure (measured or simulated) and the process parameters associated with it. In this section, we demonstrate the application of this workflow to analyses of the additive manufacturing simulation dataset described in “Additive Manufacturing Simulation Dataset.”.
The first step in the workflow is a data check to ensure that the data points are reliable and consistent. The additive manufacturing dataset described in “Additive Manufacturing Simulation Dataset” has been made publicly accessible  and consists of 1799 individual synthetic microstructures derived from simulations performed with varying AM processing parameters. A check of the data revealed some of the downloaded data to be corrupt (could not be opened), and a small number of microstructures showed unusually large grains that typically extended in length over the entire domain (in one direction). These instances were considered as outliers and eliminated from the analyses presented here. The total data for analyses reported in this paper consisted of 1599 structures.
The second most apparent variation between CLDs shown in Fig. 6 is in regard to the value of the highest frequency (corresponding to the most populous chords), excluding chord lengths of the size of one voxel. In general, it is seen that the chord length corresponding to the highest frequency is around 5 voxel lengths, but the frequency varies from 15% to 7% for the four cases shown. It should be noted that higher frequencies for the peak of the distribution generally correspond to narrower distributions (as each distribution is normalized such that the sum of the frequencies adds to one), implying that the grains within the microstructure are more similar to one another in both size and shape. Additionally, a slightly larger variation between the CLDs resolved in all three directions was observed for cases 1 and 2, in comparison to cases 3 and 4. This can be attributed to the fact that cases 3 and 4 implemented a crosshatching scan pattern, which is expected to produce more isotropic grain structures. Most interestingly, the tails of the distributions (capturing the decay in the distributions) vary significantly for the different microstructures, and are likely influenced by the changes in the size of the molten zone. In general, the parallel build pattern exhibits a sharper decay (narrower distribution of grain sizes) compared to the cross-hatching build pattern.
While some, but not all, experimental AM processing conditions can produce columnar grains that extend over several build layers, builds of this type would certainly produce heavily skewed Z-direction CLDs in comparison to those of the X and Y directions. However, in the simulations presented here, a maximum of no more than a 20% sublayer remelting was imposed. This was done to reduce the propensity for overwhelmingly biased Z-directional CLDs and produce microstructures which are in effect more reminiscent of powder-fed processes; e.g., directed energy deposition (DED) or laser-engineered net shaping (LENS) AM techniques. These processes often create builds with larger layer heights and significantly less remelting of prior layers than those of powder-bed systems .
As mentioned previously, CLDs are computed in each orthogonal direction (X, Y, and Z) and are then concatenated one after the other in a specific sequence to produce a large feature vector for each microstructure. The largest possible chord could, in theory, be equal to the dimensions of a microstructure in a given direction producing 300, 300, and 200 chord length statistics (in the X, Y, and Z directions, respectively). However, the maximum chord lengths in the ensemble of 1599 microstructures studied were identified to be 210, 203, and 90 voxels in X, Y, and Z directions, respectively. The CLDs in each direction were therefore truncated at these levels for all microstructures studied. The three CLDs for each microstructure were then concatenated to produce one large feature vector of 503 chord length statistics (the sum of 210, 203, and 90 chord length statistics obtained for each microstructure). It is unwieldy to utilize such high-dimensional representations in the practical extraction of P-S linkages. Therefore, a dimensionality reduction is performed as a next step of the workflow using principal component analysis (PCA).
Once the reduced-order representations are established, the next step of the workflow is to build a model using machine learning methods. The reduction of dimensionality from 503 to 4 has significantly reduced the difficulty associated with this step. The P-S linkage of interest in the present case study was extracted using a regression technique. Regression typically consists of four primary steps : (1) defining dependent (output) and independent (input) variables, (2) identifying the form of the function (linear, parabolic, exponential, etc.), (3) computing the regression function, and (4) performing error analysis.
The models explored in this work were evaluated for accuracy using both a data splitting approach  and a leave-one-out cross validation (LOOCV). Data splitting allows an unbiased evaluation of the model for new inputs that were not utilized in the model development. For this purpose, the dataset is divided into non-overlapping calibration (training) and validation (test) sets. (Note that “calibration set” and “training set” as well as “validation set” and “test set” are used interchangeably in this work.) More specifically, the data points corresponding to values of variables V = 7.5 and W = 70, comprising a total of 684 synthetic structures, were selected as the validation set. The remaining dataset of 915 structures composed the calibration set used to build the models for each α j . Note that the validation set was excluded even in the dimensionality reduction step. Thus, the validation conducted here is a validation of the entire workflow, including all the choices made for microstructure quantification (CLDs), dimensionality reduction (PCA), and model forms (multivariate polynomials). Although one can implement a number of other strategies to obtain the split between calibration and validation datasets, the above strategy was preferred in this study due to its ability to evaluate critically the model predictions for new inputs not included in the model building effort.
Error metric values of the acceptable models for PC1, PC2, PC3, and PC4
Error metric values of the validation of the models for PC1, PC2, PC3, and PC4
Test = R 2
The truncation level of PCs in the model as well as the degree of polynomial were varied to arrive at an optimized data-driven model. After numerous iterations between the steps of the workflow, it was identified that the first four principal components provided the best balance between the accuracy of the model and the number of features used. It is somewhat remarkable that specific values of acceptable error in the present study were not pre-determined. Rather, the specific models that exhibited the lowest errors (computed using the measures defined earlier) for the validation set, were identified and selected.
Resulting acceptable third-order polynomial models consisted of over 70 terms and coefficients. Although all 70+ terms of the model were used in this work, the authors acknowledge that optimization can be performed on the model to arrive at a more compact form with a smaller number of terms. Removing polynomial terms based on their coefficient decimals (e.g., smaller coefficients) resulted in increased mean error of predicting PC1 scores for test set from 0.0032 to 0.0132. Simply eliminating one term at a time also did not improve the results. Therefore, better optimization techniques are needed if one would like to obtain a model with fewer number of terms. However, since the computational cost of the reduced-order model produced in this work is minimal, there is no significant benefit to such pruning of the model terms.
The P-S linkage for the simulated additive manufacturing microstructures presented here consists of a large table of coefficients of polynomials, as well as the basis functions and the mean value A 0. These tables are not presented here due to their size; however, the authors are willing to share the results upon request.
In this work, our focus was exclusively on building a reduced-order model for the process-structure linkage. In prior work, we have demonstrated the viability of employing the same overall strategy for structure-property linkages . Because of the use of a consistent framework for microstructure quantification and its low-dimensional representation in both classes of linkages, it should now be possible to establish interoperable process-structure and structure-property linkages. These reduced-order PSP linkages are central to the realization of the ambitious goals set forth in the MGI and which is implicitly necessary in ICME frameworks. This is because the reduced-order PSP linkages are the only practical way forward for conducting a rapid screening of extremely large design spaces (i.e., strategies for inverse solutions scanning large spaces). Keeping in mind that the main requirement in such efforts is objective (data-driven) guidance in down-selection of the design space, the authors offer reduced-order PSP linkages are the only practical way forward. Of course, one must keep in mind the limitations on the expected accuracy of these models, and develop and implement strategies to continuously refine and improve the reduced-order models with new data (both new simulations and new experiments). Indeed, the reduced-order models can serve as a natural bridge between the modeling and experimental efforts identifying not only new opportunities with high potential payoff (e.g., improved properties or performance) but also providing objective guidance on where (and how much) effort should be expended (e.g., improving fidelity mainly in the input ranges that lead to the desired changes in the microstructure).
A novel workflow template is presented to extract process-structure linkages in microstructure evolution problems through the utilization of advanced data science techniques. The presented workflow is scalable and expandable and can be applied to a broad variety of microstructure evolution datasets. This workflow consists of four modular steps: (1) data pre-processing, (2) microstructure quantification, (3) dimensionality reduction, and (4) extraction and validation of process-structure linkages. Each step of the workflow allows selection and utilization of readily accessible codes from a large library of repositories.
The application of this template to quantify and predict synthetic additive manufacturing microstructures has been demonstrated. A publicly available set of simulated additive manufacturing microstructures has been created and shared to support exploration of AM processing parameters and the resultant grain-scale microstructural arrangements. The dataset consisted of 1599 unique microstructures and would have been extremely difficult to analyze effectively and comprehensively with conventional materials science approaches. Using the data-science approach presented here, chord length distribution calculations, principal component analysis, and multivariate polynomial regression were combined to produce a reliable reduced-order model, which was also cross-validated.
Although the process-structure linkage obtained here using a data science approach showed excellent results, the goal of this work was to establish a generic workflow to extract process-structure linkage for microstructure evolution problems. While the methods used in this case study are specific for the datasets presented, they can be altered to suit a variety of investigations and data types. Additionally, this workflow can be fully automated. This test case has demonstrated that exploration of process-structure linkages can be conducted most efficiently by exploiting modern data science-based workflows, the central feature of which is their automated consideration of a very large number of regression fits leading to a selection of surrogate models that meet the defined error and validation criteria.
Evdokia Popova and Surya R. Kalidindi would like to acknowledge support from NIST grant 70NANB14H191. Xinyi Gong acknowledges support from NSF award 1435237. Ahmet Cecen acknowledges support from AFOSR award FA9550-12-1-0458. Sandia National Laboratories is a multi-mission laboratory managed and operated by Sandia Corporation, a wholly owned subsidiary of Lockheed Martin Corporation, for the U.S. Department of Energy’s National Nuclear Security Administration under contract DE-AC04-94AL85000.
- 2.Drosback M (2014) Materials Genome Initiative: Advances and Initiatives. JOM. 66: 334–335Google Scholar
- 5.Holdren JP (2011) Materials genome initiative for global competitiveness. National Science and Technology Council OSTP. Washington, USAGoogle Scholar
- 8.Pollock TM, Allison JE, Backman DG et al (2008) Integrated computational materials engineering: a transformational discipline for improved competitiveness and national security. Washington DC, The National Acamedies PressGoogle Scholar
- 10.Spanos G, Allison J, Cowles B, Deloach J, Pollock T (2013) Integrated Computational Materials Engineering (ICME): implementing ICME in the aerospace, automotive, and maritime industries, Tech. rep., The Minerals, Metals & Materials Society (TMS)Google Scholar
- 11.Voorhees P and G Spanos (2015) Modeling across scales: a roadmapping study for connecting materials models and simulations across length and time scales. Tech. rep., The Minerals, Metals & Materials Society (TMS)Google Scholar
- 12.Kalidindi SR (2015) Hierarchical materials informatics: Novel analytics for materials data. ElsevierGoogle Scholar
- 13.Krein MP, Natarajan B, Schadler LS et al (2012) Development of materials informatics tools and infrastructure to enable high throughput materials design. MRS Online Proceedings Library. 1425: doi: 10.1557/opl.2012.57.
- 22.Olson GB (2000) Pathways of discovery designing a new material world. Science 228(12):933–998Google Scholar
- 25.McDowell DL, Panchal J, Choi HJ, Seepersad C, Allen J, et al (2009). Integrated design of multiscale, multifunctional materials and products. Butterworth-HeinemannGoogle Scholar
- 28.Brackett, D., I. Ashcroft, and R. Hague (2011) Topology optimization for additive manufacturing. In Proceedings of the Solid Freeform Fabrication Symposium. Austin, TXGoogle Scholar
- 37.Regli W, Rossignac J, Shapiro V & Srinivasan V (2016) The new frontiers in computational modeling of material structures. Comput Aided Des 77:73–85Google Scholar
- 41.Tikare V, Hernandez-Rivera E, Madison JD, Holm EA, Patterson BR, & Homer ER (2013) Hybrid models for the simulation of microstructural evolution influenced by coupled, multiple physical processes, Brigham Young University, Provo, UT; Sandia National Laboratories (SNL-NM), Albuquerque, NM (United States)Google Scholar
- 42.Plimpton S, Battaile C, Chandross M, Holm L, Thompson A, Tikare V, & Slepoy A (2009) Crossing the mesoscale no-man’s land via parallel kinetic Monte Carlo. Sandia National LaboratoryGoogle Scholar
- 45.Rodgers TM, J Madison, and V Tikare (2016) Simulation of metal additive manufacturing microstructures using kinetic Monte Carlo. Computational Materials Science - submitted for reviewGoogle Scholar
- 54.Adams BL, Kalidindi S, Fullwood DT (2013) Microstructure-sensitive design for performance optimization. Butterworth-HeinemannGoogle Scholar
- 55.Wheeler D, Brough D, Fast T, Kalidindi S, & Reid A (2014) PyMKS: Materials Knowledge System in Python (Figshare, 2014). doi: 10.6084/m9.figshare.1015761
- 56.Agrawal A, Deshpande PD, Cecen A, Basavarsu GP, Choudhary AN & Kalidindi SR (2014) Exploration of data science techniques to predict fatigue strength of steel from composition and processing parameters. Integr Mater Manuf Innov 3(8):1–19Google Scholar
- 60.Mardia KV, Kent JT, Bibby JM (1980) Multivariate analysis (probability and mathematical statistics). Academic Press, LondonGoogle Scholar
- 61.Fodor IK (2002) A survey of dimension reduction techniques. Center for Applied Scientific Computing, Lawrence Livermore National Laboratory 9:1–18Google Scholar
- 62.Hyvärinen, A. (1999) Survey on independent component analysis. Neural Computing Surveys 2(4):94–128Google Scholar
- 63.Quinlan JR (1992) Learning with continuous classes. In 5th Australian joint conference on artificial intelligence. SingaporeGoogle Scholar
- 67.Pedregosa F, Varoquaux G, Gramfort et al. (2011) Scikit-learn: machine learning in Python. J Mach Learn Res 12(Oct):2825–2830Google Scholar
- 68.Rodgers T (2015) Exploration of process-structure linkages in simulated additive manufacturing microstructures. Harvard Dataverse V1. doi: 10.7910/DVN/KJMK9Z
- 71.Team RC (2013) R: a language and environment for statistical computing 2013 (Global Biodiversity Information Facility, Copenhagen, Denmark)Google Scholar
- 72.Berthold MR, Cebron N, Dill F, Gabriel TR et al. (2009) KNIME—the Konstanz information miner: version 2.0 and beyond. AcM SIGKDD Explor Newsl 11(1):26–31Google Scholar
- 75.Sinha P (2013) Multivariate polynomial regression in data mining: methodology, problems and solutions. Int J Sci Eng Res 4(12):962–965Google Scholar
- 76.Jones E, Oliphant T, Peterson P (2015) SciPy: Open source scientific tools for Python, 2001. URL http://www.scipy.org . 73: p. 86
- 77.Plimpton S, Thompson A, Slepoy A (2012) SPPARKS kinetic Monte Carlo simulator. http://spparks.sandia.gov/