Ceramics and Glasses Modeling by Simulations and Machine Learning: Session II
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
Wednesday 2:00 PM
October 20, 2021
Room: B231
Location: Greater Columbus Convention Center
Session Chair: Mathieu Bauchy, University of California, Los Angeles; Peter Kroll, University of Texas at Arlington; N. M. Anoop Krishnan, Indian Institute of Technology Delhi
2:00 PM
Ceramics from Polymers –– Results of Ab Initio Molecular Dynamic Simulations: Peter Kroll1; 1University of Texas at Arlington
Chemical reactions during thermal processing of hybrid organic-inorganic polymers transform the precursor into a ceramic. Here we present ab-initio molecular dynamic (aiMD) simulations of principal chemical reactions of this process. We detail simulations of polysiloxanes and polysilazanes with different carbon-bearing side groups yielding SiCO and SiCN ceramics. Models are evolved at elevated temperatures for simulation times between 0.1 and 1 ns. We detail intra-chain and inter-chain coupling, cross-linking, and elimination reactions. Of particular interest are mechanisms that incorporate carbon into the -Si-O- and -Si-N- polymer backbone.Beyond those principal chemical reactions, we can follow transformations of polydimethylsiloxane (PDMS) and polyvinylsilazane (PSZ20) into systems with mixed SiCnO4-n and SiCnN4-n tetrahedra. After removal of gaseous species, the systems evolve into SiCO and SiCN ceramics, respectively.
2:20 PM
Fusing Experimental and Simulation Datasets in Machine Learning for Predicting Glass Properties: Mathieu Bauchy1; 1University of California, Los Angeles
Machine learning models require as a prerequisite the existence of data that are available, complete, consistent, accurate, and numerous. Although experimental data are usually accurate, they are often not numerous enough to enable meaningful deep learning approaches. As an alternative path, synthetic data generated by high-throughput molecular dynamics simulations can offer large, consistent datasets. However, their limited accuracy does not always yield a perfect agreement with experiments—which makes it challenging to directly combine experimental and simulation data within universal, unifying datasets. Here, we present a new “data fusion” approach that can simultaneously leverage the advantages of experimental and simulation data—wherein experimental and simulation data mutually inform, augment, and advance each other. We demonstrate that our fused model systematically outperforms models that are solely trained based on experimental (or simulation) data.
2:40 PM
Bayesian Optimization of Silicon Nitride Empirical Potentials: Tobias Kroll1; Peter Kroll1; 1University of Texas at Arlington
Engineering classical empirical potentials or force fields to closely match experiment or available quantum-chemically data requires optimization of parameters. Depending on the complexity of the force field, the target properties to be matched, and the available data, this process attains potentially extreme dimensionality associated with large computational costs. Here we apply Bayesian Optimization, a so-called machine learning technique, to optimize empirical potentials used in modeling silicon nitride.The optimization process uses static and dynamic data generated using Density-Functional-Theory calculations. It offers several hyperparameter options that allow generalization to a variety of empirical potentials.
3:00 PM
Development of a Reactive Force Field (ReaxFF) for Simulation of Polymer-derived Ceramics: Shariq Haseen1; Peter Kroll1; 1University of Texas at Arlington
Polymer-derived ceramics (PDCs) are materials that exhibit many desirable properties such as enhanced mechanical properties at high-temperatures, oxidation resistance, and creep resistance. We develop a reactive force field (ReaxFF) for large-scale simulation of the polymer-to-ceramic transformation of select Si-based PDCs with density functional theory-like (DFT) accuracy. Using our extensive library of hypothetical crystalline and amorphous structures, we first aim to reproduce DFT energy differences within ReaxFF. We further optimize parameters by matching ReaxFF molecular dynamics energies and forces with those of DFT ab initio molecular dynamics. In order to evaluate and improve our parameters, we generate melt-quench (MQ) ReaxFF models and optimize these models within DFT in a “self-learning” process that is used to augment the training set. We use our final set of ReaxFF parameters to investigate the thermal conversion of polymers to PDCs.
3:20 PM Panel Discussion: Challenges and Opportunities in Machine Learning for Materials