First World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022): Machine Learning/Deep Learning in Materials and Manufacturing V
Program Organizers: Taylor Sparks, University of Utah; Michael Dawson-Haggerty, Kerfed, Inc.; Elizabeth Holm, University of Michigan; Jin Kocsis, Purdue University; Adam Kopper, Mercury Marine; Benji Maruyama; James Warren, National Institute of Standards and Technology
Wednesday 9:00 AM
April 6, 2022
Room: Riverboat
Location: Omni William Penn Hotel
Session Chair: Elizabeth Holm, University of Michigan
9:00 AM Invited
Addressing Annotated Data Scarcity and Materials Diversity with Advanced Deep Learning Architectures: Ali Riza Durmaz1; Aurčle Goetz1; Edward Kreutzarek1; Martin Müller2; Chris Eberl1; 1Fraunhofer Iwm; 2Saarland University
Materials' microstructures are the footprint of the entire processing route rendering them very diverse. As materials discovery is accelerated and materials become increasingly intricate, deep learning (DL) gains in significance for many tasks. Especially when domain knowledge is incomplete, DL frequently outperforms classical techniques by a large margin. However, its poor data-efficiency (1) and inter-domain generalizability (2) are inherently conflicting with the expense of data annotation and materials diversity, respectively. To alleviate both issues, we propose to apply two advanced computer vision architectures that are trained in partially supervised fashion, i.e., additionally exploiting unannotated data. Specifically, we utilize (1) semi-supervised learning using dynamic mutual training and (2) unsupervised domain adaptation (transfer across materials domains) to address both challenges. Both frameworks are investigated on a complex phase steel segmentation task utilizing micrographs with different surface etchings and imaging modalities. These state-of-the-art frameworks outperform conventional supervised training strategies and pre-trained networks.
9:30 AM
Prediction of Anisotropic Plastic Flow from Indentation Responses Via Neural Networks Combined with Finite Element Analysis: Kyeongjae Jeong1; Kyungyul Lee1; Siwhan Lee1; Jinwook Jung1; Hyukjae Lee1; Nojun Kwak1; Dongil Kwon1; Heung Nam Han1; 1Seoul National University
In this study, anisotropic plastic behavior is predicted from spherical indentation responses using neural networks (NNs) trained with information collected by finite element (FE) simulations. We present a robust FE–NN modeling approach that inversely solves the problem of leading the mechanical properties of anisotropic materials from the load-depth curve, pile-up/sink-in, and residual in-plane displacement field. The predictive performance of the FE–NN model is evaluated by comparison with uniaxial tensile curves measured in the rolling, diagonal, and vertical directions. Furthermore, we propose that the Knoop indenter enables accurate inverse analysis without additional indentation data other than the load-depth curve. By paying attention to the characteristics of the Knoop indenter whose load-depth curve varies depending on the rotation angle about the indentation axis, as opposed to the spherical indenter, we discuss the potential capability of the Knoop indenter in the inverse analysis.
9:50 AM
Combining Organic and Inorganic Descriptors for Predictions of Solubility and Volatility Across Vast Chemical Space: Anand Chandrasekaran1; Simon D. Elliot1; Asela Chandrasinghe1; Yuling An1; H. Shaun Kwak1; Mathew D. Halls1; 1Schrodinger Inc
Traditional QSPR approaches use machine learning to map the 2D structures of molecules to a particular label/property. These 2D structures are usually numerically encoded as bit-based fingerprints that capture the substructures around atoms in a molecule. However, such approaches do not perform well for inorganic or metal-containing compounds. More recently, materials informatics descriptors have been developed that span the entire periodic table, capturing both chemical and stoichiometric information. In this work we utilize a combination of traditional QSPR fingerprints and inorganic descriptors to train ML models for solubility and volatility across a huge chemical space of organic, organometallic, and inorganic materials. For solubility, we use the AqSolDB dataset and for volatility we have carefully curated an organometallic dataset with more than 3000 pressure vs temperature measurements. We benchmark a number of different machine learning approaches and show that stacking estimators perform significantly better in comparison to using a single algorithm/model.
10:10 AM Break
10:40 AM
Nanoindentation Mapping Defects Filtration for Heterogeneous Materials Using Generative Adversarial Networks (GAN’s): Giuseppe Bianco Atria1; Ambreen Nisar1; Cheng Zhang1; Benjamin Boesl1; Arvind Agarwal1; 1Florida International University
Advanced materials with multiple phases and heterogeneous microstructure necessitate mapping their mechanical properties such as elastic modulus to develop constitutive relations. However, structural heterogeneities such as surface roughness, porosity, and secondary phases result in anomalous underrepresentation of the true materials’ properties that conventional experimental strategies cannot amend. This paper establishes a novel deep learning-based strategy to rectify incorrect experimental spatial measurements acquired during nanoindentation modulus mapping. The integrated bicubic interpolation and generative adversarial networks (GANs) model was trained using 14 ceramic and 18 metallic data sets, each comprising 65,536 measurements. The developed algorithm was validated against experimental measurements on 4 unknown specimens. The standard deviation in measured elastic modulus reduces by ~50% in ceramics and ~72% on metals. This computational framework thus constitutes a significant advancement in applying deep learning algorithms for developing advanced materials and their correlation with manufacturing.
11:00 AM
Supervised Machine Learning for Collision Weld Process Optimization: Blake Barnett1; Glenn Daehn1; 1Ohio State University
Statistical process control (SPC) and artificial intelligence/machine learning (AI/ML) have enjoyed widespread successful use in fusion and solid-state welding process control. However, collision welding techniques such as explosive welding (EXW), magnetic pulse welding (MPW), laser impact welding (LIW), and vaporizing foil actuator welding (VFAW) cannot be subject to in-process control due to the microsecond timescales over which acceleration and impact occur. Machine learning techniques offer accessible pathways to understanding collision welding results and identifying key process variable-weld performance correlations across multiple welding technologies. The feasibility of this concept was demonstrated through the application of supervised learning algorithms to perform classification of simulated collision welds as inside or outside of classical analytic welding windows with prediction accuracy above 80%. Top performing model-trained algorithms were also tested against literature data to evaluate real-world performance.
11:20 AM
Interpretation of Convolutional Neural Networks for Predicting Volume Requirements in Studies of Microstructurally Small Cracks: Karen Demille1; Ashley Spear1; 1University Of Utah
Previously, representative volume elements for microstructurally small cracks (RVEMSC) were established using a finite-element (FE) framework. The process of establishing RVEMSC values—which indicate the minimum volume requirements for studies involving microstructurally small cracks (MSCs)—required thousands of 3D FE simulations. In this work, convolutional neural networks (CNNs) are deployed in response to the prohibitive computational expense of determining RVEMSC values strictly via FE simulations. CNNs are trained to predict RVEMSC values given 3D descriptions of microstructural and geometrical features near points along the crack front. In addition to providing RVEMSC predictions, the CNNs provide insight into the influence of microstructure on RVEMSC values: CNN input sensitivity studies compare the relative importance of microstructural and geometrical features on RVEMSC predictions, and saliency maps highlight regions of microstructure most important to RVEMSC value predictions. Through interpretation of CNN models, the link between RVEMSC requirements and microstructural neighborhoods is clarified.