First World Congress on Artificial Intelligence in Materials and Manufacturing (AIM 2022): Machine Learning/Deep Learning in Materials and Manufacturing II
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
Monday 1:30 PM
April 4, 2022
Room: Riverboat
Location: Omni William Penn Hotel
Session Chair: Francesca Tavazza, National Institute of Standards and Technology
1:30 PM Invited
Interpretable Machine Learning-Assisted Phase Classification for High Entropy Alloys: Kyungtae Lee1; Mukil Ayyasamy1; Paige Delsa2; Timothy Hartnett1; Prasanna Balachandran1; 1University of Virginia; 2Louisiana School for Math, Science, and the Arts
High entropy alloys (HEAs) are multicomponent materials with nearly equal amount of four or more principal elements. HEAs exhibit excellent mechanical, thermal, and electrochemical properties with immense potential for materials design due to a vast compositional space. In this work, we develop a machine learning (ML) approach with post hoc interpretability capability for HEA phase prediction. The ML methods establish a quantitative relationship between chemical composition and experimentally determined phase information. The interpretability methods offered unprecedented insights into the local behavior (i.e., each composition) of the trained black-box models. We developed a novel algorithm that combined the data from local interpretability analysis with clustering to identify similar compositions. This analysis uncovered previously unknown phase-specific correlations between key features and the HEA phases. We are currently developing novel methods to explore local interpretability of ML models that are trained to predict the mechanical properties of HEAs.
2:00 PM
Convolutional Neural Networks for Image Classification in Metal Selective Laser Melting Additive Manufacturing: Rodolfo Ledesma1; Andy Ramlatchan; 1NASA Langley Research Center
Selective laser melting (SLM) is a metal additive manufacturing process that has several advantages such as the large range of metal materials that can be accommodated, 3D printing of complex shape components, the ability to adjust material properties, and cost reduction as expensive production equipment may not be required. Therefore, process monitoring is crucial in different stages of the component building. In this work, convolutional neural networks (CNNs) are investigated as a suitable technique for post-inspection of builds. The monitoring of manufactured parts was conducted by collecting computed tomography (CT) images and identifying defects. Five CNN models were implemented and tested for the classification of the CT images. The models were based on NASNetMobile and DenseNet121, and a custom-built CNN model. The results of this work show that CNNs can be feasible and reliable for rapid monitoring and classification of defects in CT images from build fabrication using SLM.
2:20 PM
Microstructural Classification of Bainite Subclasses in Low-Carbon Multi-Phase Steels Using Supervised Machine Learning: Martin Mueller1; Dominik Britz2; Thorsten Staudt3; Frank Muecklich1; 1Saarland University; 2Material Engineering Center Saarland; 3Aktiengesellschaft der Dillinger Hüttenwerke
Artificial intelligence (AI) and machine learning (ML) have now made their way into materials science, too and are omnipresent. Especially for microstructure segmentation and classification, for which simple, traditional methods like thresholding or manual judgment by human experts are still common, AI and ML offer new potentials and promise decisive improvements. The segmentation, classification and following quantification of the microstructure are the foundation for establishing process-microstructure-property correlations, which in turn are the basis for developing or optimizing materials. This work deals with the automated, objective, and reproducible classification of the carbon-rich second phase objects in multi-phase steels. In total, seven microstructure classes, including bainite subclasses, are considered which can all be present simultaneously in one micrograph. Based on SEM images, textural features and morphological parameters are calculated and classified with a support vector machine with an accuracy of 89.2 % regarding the area of second phase objects.
2:40 PM Break
3:10 PM
Predictive Synthesis of Quantum Materials with Offline Reinforcement Learning: Pankaj Rajak1; Aravind Krishnamoorthy; Aiichiro Nakano1; Rajiv Kalia1; Priya Vashistha1; 1University of Southern California
Finding strategies to synthesize novel functional materials and structures is one of the central endeavors of chemical and materials science. Optimal material design involves a high dimensional parameter space search and time dependent sequential decision-making process, which requires development of new machine learning techniques that are capable of implicit decision-making over long period of times with minimal human supervision. Here, we have developed an autonomous agent using offline reinforcement learning for the predictive synthesis of a prototypical quantum material, monolayer MoS2, via chemical vapor deposition. After training, the RL agent successfully learns optimal synthesis policies in terms of threshold temperatures and chemical potentials for the onset of chemical reactions and provides mechanistic insight to predict new synthesis schedules for well-sulfidized, crystalline and phase pure MoS2 in minimum time, which is validated by molecular dynamics and experimental synthesis.
3:30 PM
Data-Driven Reduced-Order Multiscale Materials Modeling Under Inhomogeneous Porosity Distributions: Shiguang Deng1; Carl Soderhjelm1; Diran Apelian1; Ramin Bostanabad1; 1University of California Irvine
Cast aluminum alloys often contain non-uniformly distributed pores of complex morphologies. Since such porosity defects have significant influence on material behaviors and affect the usage in high-performance applications, it is vital to understand the cross-scale impact of microscale porosity characteristics on the cast component’s macro-mechanical properties. In this talk, we will introduce a computationally efficient data-driven multiscale framework to simulate the behavior of metallic components containing process-induced porosity distributions. Major components of our approach include: (1) a porosity-oriented 3D microstructure reconstruction algorithm which mimics the material’s local heterogeneity with reconstructed pores from tomography characterization, (2) a novel reduced-order model which significantly reduces computational costs by projecting solution variables into a lower dimensional space where the material’s elasto-plastic behaviors are approximated, and (3) a machine learning-based metamodel which correlates material responses with pore morphology and deformation history. We will compare our approach against direct numerical simulations to demonstrate performance and versatility.
3:50 PM
Using Uncertainty in Machine Learning to Inform Decision Making on Structural Characterization of Materials: Austin Mcdannald1; Brian DeCost1; A. Gilad Kusne1; 1National Institute of Standards and Technology
We show how quantifying and propagating uncertainty in the machine learning analysis of structural characterization data can inform the decision making about those measurements. This allows us to answer questions about what can be inferred from the measurements, and about what conditions or locations to investigate next. We present two case studies on this topic. In neutron powder diffraction for the identification of phase transitions there is a need to choose the temperature to measure at that quickly spans the range of interest without missing pertinent information. Since the transitions can be abrupt and hysteretic the experiment temperature is limited to only increase. We show how the Autonomous Neutron Diffraction Explorer uses uncertainty to efficiently navigate this space. In a second case study we show how uncertainty information can be used to inform the acquisition of orientation data from the Electron Backscatter Diffraction Measurement maps of grain structure. We encode the relevant physical constraints such as the requiring contiguous grains, and abrupt changes in orientation at grain boundaries, as well as the crystal symmetry in the algorithms and their uncertainty. This is key to providing an accurate understanding of the uncertainty landscape and allows us to choose the locations for new measurements.
4:10 PM
Data-Driven Modelling of Graphene Synthesis: Aagam Shah1; Mitisha Surana1; Jad Yaacoub1; Elif Ertekin1; Sameh Tawfick1; 1University of Illinois at Urbana-Champaign
Despite that graphene synthesis via chemical vapor deposition (CVD) was first demonstrated in 2009, today economically viable manufacturing of continuous single crystals of graphene is still elusive. Hundreds of coupled synthesis parameters and complex kinetic pathways impede the development of recipes that are reproducible, environmentally sustainable, and compatible with electronics fabrication. To overcome these challenges, we present Gr-ResQ (Graphene Recipes for Synthesis of High-Quality Materials) - a platform enabling the sharing and use of graphene synthesis data. At its core is a crowd-sourced database of CVD recipes and characterization with a suite of tools to enable quick analysis of Raman spectra and microscopy images. We then propose a Bayesian optimization technique that uses the experimental results in the database to automate the search for an optimal synthesis recipe. This technique can guide the selection of the next experiment and iterate through the process till a predictive model is prepared.
4:30 PM
Deep Learning Surrogate Models for Multiscale Simulation of Advanced Materials: Cornwell Cornwell1; Flavio Souza1; 1Siemens Digital Industries Software
Multiscale modeling has proven to be the most accurate approach for characterizing advanced materials in finite element analysis (FEA). Fully-coupled multiscale analysis utilizes FEA at the material level to represent micromechanical attributes which provides high levels of accuracy compared to analytical constitutive material laws. However, multiscale analysis can be computationally expensive if many microstructural models are run concurrently. Using machine learning, specifically deep-learning, these microstructure models can be replaced with neural networks to predict the constitutive properties of advanced materials. In this paper, many network configurations demonstrate high levels of accuracy and flexibility in predicting multiple microstructural models. Accuracy was evaluated by comparing stress-strain curves of the models compared to results computed through FEA.