Ceramics and Glasses Modeling by Simulations and Machine Learning: Machine Learning Modeling of Glass and Ceramics
Sponsored by: ACerS Glass & Optical Materials Division
Program Organizers: Mathieu Bauchy, University of California, Los Angeles; Peter Kroll, University of Texas at Arlington; N. M. Anoop Krishnan, Indian Institute of Technology Delhi
Monday 8:00 AM
October 10, 2022
Room: 408
Location: David L. Lawrence Convention Center
Session Chair: Mathieu Bauchy, UCLA; Peter Kroll, The University of Texas at Arlington; Anoop Krishnan, IIT Delhi
8:00 AM Invited
Machine Learning Defect Properties of Semiconductors: Arun Kumar Mannodi Kanakkithodi1; 1Purdue University
Defects and impurities in semiconductors heavily influence their performance in optoelectronic applications. Quick predictions of defect properties are desired in technologically important semiconductors, but complicated by difficulties in assigning measured levels to specific defects and by the expense of large-supercell first principles computations that involve charge corrections and advanced functionals. We address this issue by combining high-throughput density functional theory (DFT) with machine learning (ML) to develop predictive models for defect formation energies and charge transition levels, for three distinct defect datasets: (a) substitutional and interstitial impurities in zincblende semiconductors, (b) Pb-site doping in A(Pb)X3 hybrid perovskites, and (c) A/X vacancies in complex halide perovskite alloys. ML models combine unique encoding of the defect atom’s elemental properties, coordination environment, and unit cell defect data with rigorous training using random forests, Gaussian processes, and neural networks. DFT-ML datasets and models are made available as online tools for easy prediction and screening.
8:30 AM
Natural Language Processing Aided Understanding of Material Science Literature: Mohd Zaki1; Tanishq Gupta1; N. M. Anoop Krishnan1; Mausam Mausam1; 1Indian Institute of Technology Delhi
Material science literature has been an indispensable and reliable source of information for designing materials for targeted applications. Many research papers are now available, which can be referred by researchers to come up with novel materials for answering industrial and societal needs. However, it is humanly impossible to go through and understand all the published research literature. In this work, we use a natural language processing based solution by training a language model, namely MatSciBERT, on materials science literature. The model’s capability to understand the material science domain by evaluating its performance on downstream tasks of named entity recognition, abstract classification, and relation classification is evident in the achieved state of the art results on these tasks. We have made all the resources publicly available for the scientific community to use and accelerate material discovery.
8:50 AM
Machine Learning-Derived Atomistic Potentials for Y2Si2O7 and Yb2Si2O7: Cameron Bodenschatz1; Wissam Saidi2; Jamesa Stokes1; 1NASA Glenn Research Center; 2University of Pittsburgh
Incorporation of SiC/SiC ceramic matrix composite (CMC) hot section components into aircraft engines promises to increase efficiency and safety. However, SiC/SiC CMCs are subject to water vapor-induced oxidation and recession at the high temperatures of engine operation, and thus environmental barrier coatings (EBCs) are required to reduce this degradation and enable their widespread adoption. An understanding of EBCs failure mechanisms, including thermochemical and thermomechanical mechanisms, is essential as coating degradation leads to reduced CMC component service life. Computational modeling approaches can provide insight into EBC material properties important for coating design. However, density functional theory (DFT) is computationally expensive and atomistic potentials are lacking for materials of interest. In this work, we utilize a machine learning approach and DFT training data to parameterize atomistic potentials for two candidate EBC materials, Y2Si2O7 and Yb2Si2O7. These potentials enable near DFT-accurate calculations of thermodynamic and thermomechanical properties essential to EBC design.
9:10 AM
Using Machine Learning Empirical Potentials to Investigate Interdiffusion at Metal-Chalcogenide Alloy Interfaces: Siddarth Achar1; Derek Stewart2; Julian Schneider3; 1University of Pittsburgh; 2Western Digital Technologies; 3Synopsys Inc.
Chalcogenide alloys for selector and memory elements for next generation non-volatile memory cells may suffer from interdiffusion at interfaces due to Joule heating and high applied fields. This interdiffusion can degrade device performance over time. While first principles atomistic simulation can provide insight into the electronic structure and local atomic bonding configurations, this approach is limited to small atomic systems and time scales. To explore this problem on a broader scale, we developed machine-learning empirical potentials that can be used in molecular dynamic simulations of Ge-Se alloys interfaced with Ti electrodes. We used the moment-tensor approach with active learning (as implemented in QuantumATK) to construct these potentials, drawing on a dataset of ab-initio calculations for relevant Ge, Se, and Ti systems. We will discuss the evolution of the composition profile over time and the impact interdiffusion will have on the electronic properties of these device structures using our potentials.
9:30 AM
Quantifying the Local Structure of Metallic Glass as a Function of Composition and Atomic Size: Thomas Hardin1; Michael Chandross1; Murray Daw2; 1Sandia National Laboratories; 2Clemson University
The structural complexity and large compositional design space of metallic glass are an enduring challenge for those who seek performance gains beyond the crystalline/amorphous binary. We report a series of simulations using the EAM-X interatomic potential (a recently-developed formalism that loosely captures the behavior of a wide range of metallic alloys with a few easy-to-change parameters) to sample the design space of binary metallic glasses, specifically focusing on variation in composition and atomic size ratio. We used data mining techniques (the Gaussian Integral Inner Product Distance with agglomerative clustering and diffusion maps) to map out the local structural states of the glass as a function of these variables. This analysis enables the development of a science basis for rule-of-thumb relationships between composition and atomic size, and local structure. SNL is managed and operated by NTESS under DOE NNSA contract DE-NA0003525 (SAND2022-4502 A).
9:50 AM
Data-driven Prediction of Room Temperature Density of Multicomponent Silicate-based Glasses: Kai Gong1; Elsa Olivetti1; 1Massachusetts Institute of Technology
Density is one of the most commonly measured/estimated materials properties, especially for glasses and melts that are of significant interest to many fields, including metallurgy, geology, materials science and sustainable cements. Here, two types of machine learning models (i.e., random forest (RF) and artificial neural network (ANN)) have been developed to predict the room-temperature density of glasses in the compositional space of CaO-MgO-Al2O3-SiO2-TiO2-FeO-Fe2O3-Na2O-K2O-MnO, based on ~2100 data points mined from literature. The results show that both RF and ANN models exhibit accurate density predictions with R2 value of ~0.96-0.98 and MAPE of ~0.59-0.79% for the 15% testing set, better than empirical density models based on ionic packing ratio (R2 values and MAPE of ~0.28-0.91 and ~1.40-4.61%, respectively). Analysis of the predicted density-composition relationships from these models suggests that the ANN model exhibits a certain level of transferability and captures known features, including the mixed alkaline earth effects for (CaO-MgO)0.5-(Al2O3-SiO2)0.5 glasses.
10:10 AM Break
10:30 AM Invited
A Physics Informed Machine Learning Approach to Predict Glass Forming Ability: Collin Wilkinson1; Cory Trivelpiece2; Rebecca Welch1; John Mauro1; 1Pennsylvania State University; 2Savannah River National Laboratory
Predicting the liquid compositions that will vitrify at experimentally accessible quench rates remains one of the grand challenges in the field of condensed matter physics. This glass-forming ability can be quantified as the critical quench rate needed to suppress crystallization. Knowledge of this critical quench rate also informs which glass composition could be used for new applications. There have been several physical and empirical models presented in the literature to predict the critical quench rate/glass-forming ability. These models range from those theoretically derived to those quantified only through experimental characterization. In this work, we instead propose a new method to calculate the critical quench rate using the recently developed physical models combined with machine learning. The results are then compared to traditional glass-forming ability metrics.
11:00 AM
Data Driven Design and Enhancement of Machinable Glass Ceramics
: Prachi Garg1; Scott Broderick1; Baishakhi Mazumder1; 1University at Buffalo
Machinable glass ceramics (MGC) are composed of a borosilicate glassy matrix with a crystalline phase, and with excellent machinability and high temperature working conditions. Machinability arises due to the crack propagation in the individual crystals, making crystallization a critical point of consideration. However, a significant challenge is that there are limited hardness values available along with a nearly infinite chemical search space. To address this challenge, we developed an ensemble data mining approach to model the hardness of over 250 new MGC compositions with high accuracy and robustness. Thus, our model guides future experiments based on the precious limited data available. Designing an experiment is challenging due to limited hardness values and infinite chemical search base. Our model was instead trained on the extracted data from different sources to guide future experiments. Coupling these results with microstructural studies provides a unique feedback on the fabrication of state-of-the-art machinable glass ceramics.
11:20 AM
Predicting and Accessing Metastable Phases: Vancho Kocevski1; James Valdez1; Benjamin Derby1; Ghanshyam Pilania1; Blas Uberuaga1; 1Los Alamos National Laboratory
Metastable phases can have distinct properties compared to the ground state, increasing the usability of the materials in technological applications. The applicability of lanthanide sesquioxides (Ln2O3) in solid-oxide fuel cells and as irradiation resistant materials have been attributed to the relative ease with which they transform to different polymorphs. An efficient method to estimate the amount of stored energy that different metastable phases require by irradiation for their practical realization can aid in understanding and rationalizing their irradiation response. Here, we calculate metastable phase diagrams, from which we extract the metastability threshold – the excess energy stored in the metastable phase relative to the ground state. We demonstrate how metastable phase diagrams provide new insight into the synthesis and irradiation behavior of Ln2O3. We successfully predict the sequence of metastable phase formation of Lu2O3 irradiated at cryogenic temperature, forming three metastable phases with increasing irradiation fluence, displaying unique irradiation behavior.
11:40 AM
Predicting the Dynamics of Atoms in Glass-Forming Liquids by a Surrogate Machine-Learned Simulator: Mathieu Bauchy1; 1University of California, Los Angeles
Molecular dynamics (MD) is a workhorse of computational material science. However, the inner-loop algorithm of MD is computationally expensive. This is a key limitation since, as a result, MD simulations of glasses are typically limited to very fast cooling rates. Here, we introduce a surrogate machine learning simulator that is able to predict the dynamics of liquid glass-forming systems with no prior knowledge of the interatomic potential or nature of the Newton’s law of motion. The surrogate model consists of a graph neural network (GNN) engine that is trained by observing existing MD-generated trajectories. We demonstrate that the surrogate simulator properly predicts the dynamics of a variety of systems featuring very different interatomic interactions. The development of machine-learned surrogate simulators that can effectively replace costly MD simulations could expand the range of space and time scales that are typically accessible to MD simulations.