AI/Data Informatics: Computational Model Development, Validation, and Uncertainty Quantification: On-Demand Oral Presentations
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Saurabh Puri, VulcanForms Inc; Dennis Dimiduk, BlueQuartz Software LLC; Darren Pagan, Pennsylvania State University; Anthony Rollett, Carnegie Mellon University; Francesca Tavazza, National Institute of Standards and Technology; Christopher Woodward, Air Force Research Laboratory
Monday 8:00 AM
March 14, 2022
Room: Materials Design
Location: On-Demand Room
Extracting and Making Use of Materials Data from Millions of Journal Articles via Natural Language Processing Techniques: Anubhav Jain1; 1Lawrence Berkeley National Laboratory
Historically, both data and knowledge in the materials domain has been recorded mainly as text, figures, or tables in journal articles. Such data is critical to developing, training, and validating machine learning models. In this talk, I will describe some of our efforts to extract information from the research literature automatically based on natural language processing techniques. For example, data on the dopability of materials is difficult to simulate and not part of a standard database, but is present either implicitly or explicitly as part of many published research studies. Similarly, data on materials synthesis can be difficult to simulate and compile but can be extracted from the historical research literature. The talk will summarize our most recent progress towards extracting both individual data items as well as "knowledge" in various areas.
Orchestrating Multi-task Material Design Campaigns with Artificial Intelligence: Logan Ward1; 1Argonne National Laboratory
Materials design requires being judicious about how to use resources. Careful thought and analysis on how new data should inform the next experiment or computation is the key to success. However, the continual arrival of increasingly faster ways of gathering data (e.g., exascale supercomputers, robotic laboratories) leaves much shorter times for engineers to be circumspect. In this talk, we discuss how artificial intelligence systems can augment the ability of humans to quickly identify promising leads and develop better materials through illustrative examples including the design of battery electrolytes and conductive polymer films. We will focus on the software and machine learning algorithm which can enable such techniques to be used broadly through materials engineering.
Variational System Identification of the Partial Differential Equations Governing Microstructure Evolution in Materials: Inference over Sparse and Spatially Unrelated Data: Krishna Garikipati1; Xun Huan1; Zhenlin Wang1; 1University of Michigan
Pattern formation in materials is mechanism-specific, and encoded by partial differential equations (PDEs). With the aim of discovering hidden physics, we have developed a variational approach to identifying such systems of PDEs in the face of noisy data at varying fidelities (Comp Meth App Mech Eng, 353:201, 2019 and 377:113706, 2021). We apply our methods to image data on microstructures in materials physics. PDEs are posed as initial and boundary value problems over combinations of time intervals and spatial domains whose evolution is either fixed or can be tracked. However, micrographs of pattern evolution in materials are over domains that are unrelated at different time instants, and come from different physical specimens. The temporal resolution can rarely capture the fastest time scales, and noise abounds. We exploit the variational framework to choose weighting functions and identify PDE operators from such dynamics. The framework is demonstrated on synthetic and real data.
Automatic Segmentation and Quantification of Microscopy Data Using Transfer Learning from a Large Microscopy Database: Joshua Stuckner1; 1NASA Glenn Research Center
A transfer learning approach for the automatic segmentation of microscopy data is presented. Many encoder architectures, including VGG, Inception, ResNet, and others, were trained on ~100,000 microscopy images from 54 material classes to create pre-trained models that learn representations that are more relevant to downstream microscopy analysis tasks than models pre-trained on natural images. The pre-trained encoders were embedded into segmentation architectures including U-Net and DeepLabV3+ to evaluate the performance of models pre-trained on a large microscopy dataset. Each encoder/decoder pair was evaluated on several benchmark datasets. Our testing shows that models pre-trained on a large microscopy dataset generalize better to out-of-distribution data (micrographs taken under different imaging or sample conditions) and are more accurate when training data is limited than models pre-trained on ImageNet.
Multi-fidelity Surrogate Modeling of Epistemic Uncertainty Arising from Microstructure Reconstruction: Arulmurugan Senthilnathan1; Pinar Acar1; 1Virginia Tech
Microstructure reconstruction enables multi-scale analysis using small-scale experimental data and, thus, eliminates the infeasibilities arising from the cost and time requirements of experiments in large domains. Markov Random Field is one efficient method that produces statistically-equivalent synthetic representations to small-scale test samples while introducing epistemic uncertainty on microstructural features. These uncertainties can alter material properties by propagating over multiple scales. The present work addresses multi-scale modeling for grain topology of polycrystalline microstructures under the effects of the microstructural uncertainties. The special focus is on the Titanium-7wt\%-Aluminum alloy (Ti-7Al), which is a candidate material for many aerospace systems owing to its outstanding mechanical performance in elevated temperatures. The shape moment invariants are used to quantify the grain topology of polycrystalline microstructures and a surrogate model based on Gaussian Process Regression (GPR) is developed as a function of shape moment invariants by utilizing the experiments and crystal plasticity simulations for training data.
Thermodynamic Analysis for the Design of High-strength Aluminum Alloys at High Temperatures: Takeshi Kaneshita1; Shimpei Takemoto1; Yoshishige Okuno1; Kenji Nagata2; Junya Inoue3; Manabu Enoki3; 1Showa Denko K.K.; 2National Institute for Materials Science; 3The University of Tokyo
We discuss the design of 2000 series high-strength aluminum alloys at high temperatures using Bayesian learning for neural networks and thermodynamic analysis. It is known that the strength of aluminum alloys decreases rapidly above 150°C, so improving the strength at high temperatures is essential for industrial applications. In order to design high-strength alloys, it is necessary to optimize the additive element compositions and the heat treatment conditions such as temperature and time for homogenization, solution processing, and aging. A data science approach using neural networks is suitable for handling such multi-dimensional problems and exploring the optimal process conditions from the vast design space. This study focuses on the thermodynamic calculations, including the CALPHAD method, the Langer-Schwartz-Kampmann-Wagner approach, and the phase-field method for analyzing the strengthening mechanism of the alloy designs suggested by the neural network.
Using Machine Learning to Improve Melt Pool Prediction in Additive Manufacturing: Data Denoising and Predictive Modeling
: Yaohong Xiao1; Zhuo Wang1; Lei Chen1; 1University of Michigan-Dearborn
Accurate modeling of melt pool is vital to achieving process optimization and quality control in additive manufacturing. In this research, a data-driven melt pool modeling framework based on machine learning (ML) is established, further fueled by massive experimental data from National Institute of Standards and Technology (NIST). First, a convolutional neural network is used to pre-process as-received melt pool images, enabling removal of spattering noises and thus extraction of high-quality melt pool data. Then, a novel melt pool prediction model using multi-layer perceptron is trained by incorporating raw, long scanning history as input features, which best accounts for effects (e.g., heating) of printing history on melt pool development. Testing results under various manufacturing conditions show that the average relative error of predicting melt pool area drops to 2.8 %, compared to 14.8 % of prior art–the Neighboring Effect Modeling Method, representing a significant step towards reliable melt-pool-guided quality control.
Digital Image Correlation Based Machine Learning Predictions for Grain-boundary Strain Accumulation in a Polycrystalline Metal: Renato Vieira1; John Lambros2; 1Pontifícia Universidade Católica do Rio de Janeiro; 2University of Illinois Urbana Champaign
Non-uniform accumulation of plastic strains at the grain scale in polycrystalline metals is a precursor to damage and eventual crack nucleation during fatigue. Here we investigate the accumulation of plastic strains at the microscale using a high-resolution digital image correlation (HiDIC) technique in conjunction with electron backscatter diffraction (EBSD). With the large datasets resulting from the HiDIC experiments, we develop machine-learning algorithms capable of predicting the accumulation of strains at the grain scale from microstructural and load inputs. As a proof of concept, a neural-network has been trained to correlate grain boundary orientation with the resulting local strains surrounding boundary mantle regions. The obtained neural-network predictions show good correlation with the DIC-measured strain fields for most measured cases. These results showed that the local geometrical angle between a grain boundary and the loading axes is in many cases a good predictor for the accumulation of strains around that boundary.
Uncertainty Prediction for a Variety of Material Properties Modelled via Machine Learning: Francesca Tavazza1; Kamal Choudhary1; Brian DeCost1; 1National Institute of Standards and Technology
Uncertainty quantification in AI-based predictions of material properties is of immense importance for the success and reliability of AI applications in material science. While confidence intervals are commonly reported for machine learning (ML) models, prediction intervals, i.e. the evaluation of the uncertainty on each prediction, are seldom available. In this work we compare 3 different approaches to obtain such individual uncertainty, testing them on 12 ML-physical properties. Specifically, we investigated using the Quantile loss function, machine learning the prediction intervals directly, and using Gaussian Processes. We identify each approach’s advantages and disadvantages and compare their results. All data for training and testing were taken from the publicly available JARVIS-DFT database, and the codes developed for computing the prediction intervals are available through JARVIS-tools
Accelerated Alloy Design by Batch Constrained Multiobjective Optimization Using Surrogate Models: Gérard Ramstein1; Franck Tancret1; 1Université de Nantes
Recent alloy design strategies use constrained multiobjective optimization by genetic algorithms. However, due to the high number of evaluations they require, they can rely only on fast predictive tools like machine learning or parsimonious results from the calculation of phase diagrams (Calphad / Thermo-Calc). The present work addresses multiobjective optimization when objectives and/or constraints require computer-intensive calculations (e.g. solidification, diffusion, precipitation, phase field, ab initio simulation…). It is inspired from batch Bayesian optimization; it uses an evolutionary optimizer relying on the predictions of Kriging surrogate models (Gaussian processes), and a selector of the best solutions to be evaluated in sequential batches to improve both the global Pareto-optimal solutions and the surrogate models accuracy. The algorithm is exemplified and benchmarked in the case of the design of nickel-based superalloys. Massive reductions in computing times are achieved, with negligible losses on the attainment of the Pareto front.
A Statistical-physical Framework for the Analysis of Uncertainties due to Material Parameters in Multi-physics Modelling: Amanda Giam1; Jiaxiang Cai1; Fan Chen1; Zhisheng Ye1; Wentao Yan1; 1National University of Singapore
A bottleneck in Laser-Powder Bed Fusion (L-PBF) metal additive manufacturing (AM) is the quality inconsistency of products. To avoid costly experimentation, computational multi-physics modelling is being used to tackle this issue, but its’ effectiveness is limited by modelling parameter uncertainties. Therefore, a statistical-physical framework is utilized to characterise uncertainty in multi-physics models. Data is gleaned from a high-fidelity thermal-fluid model with a two-level full factorial design for five selected material parameters. Statistical techniques including the analysis of variance, scatter plots and linear regression are employed for sensitivity analyses of input factors on the response melt pool dimensions. To account for physics in the L-PBF process, crucial physical phenomena are thoroughly analysed. This complements statistical findings with domain knowledge, yielding a validated joint evaluation. Uncertainty propagation is investigated via graphical relations of input-output standard deviations. Consistent results from the sensitivity and uncertainty analyses can provide practical guidance for simulations and experiments.
Austenitic Parent Grain Reconstruction in Martensitic Steel Using Deep Learning: Patxi Fernandez-Zelai1; Andrés Márquez Rossy1; Quinn Campbell1; Andrzej Nycz1; Michael Kirka1; 1Oak Ridge National Laboratory
Phase transformations take place in many structural materials following solidification. Phase reconstruction algorithms are commonly used to infer the underlying parent phase crystal structure from spatial-orientation patterns present in electron backscatter diffraction micrographs. In the past decade machine learning methods have been shown to perform exceptionally well in a number of vision tasks. In this work we develop a deep convolutional architecture for estimating prior austenite micrographs from observed martensite electron backscatter diffraction micrographs. Efficient learning of the orientation relationships within the network is facilitated by a novel data augmentation strategy. Training is performed using only four micrographs by exploiting the arbitrariness of the reference sample coordinate system. Model inference is much faster than algorithmic approaches and generalizes well when applied to micrographs of a completely different material. This work demonstrates that modern computer vision models, trained with only a few micrographs, are well suited for analyzing orientation imaging micrographs.
Semi-mechanistic Gaussian Process Model for Disentangling Structural and Chemical Influences on Material Properties: Brian DeCost1; Howie Joress1; Jason Hattrick-Simpers2; 1National Institute of Standards and Technology; 2University of Toronto
Current AI systems used in the sciences make and test predictions, but lack the mechanistic modeling components needed for formulating and evaluating scientific hypotheses that explain these predictions in terms of generalizable physical principles. Furthermore, the need to reconcile multiple structure and property data streams make it difficult to comprehensively explore complex and multiscale structures that determine the performance of engineering materials.We present exploratory research incorporating mechanistic models (such as Hall-Petch grain size effect models) into the Gaussian Process modeling framework. By using mechanistic modeling components for microstructure-driven effects and Bayesian nonparametric modeling components for chemistry-driven effects, we can gain insight into the role of both chemistry and microstructure even without robust physical models for chemical effects. We use this approach to analyze the dependence of the corrosion behavior of multicomponent alloys by fusing high throughput electrochemical assays with data from structural characterization methods, such as x-ray diffraction.