The Woodruff School

SUMMARY

The behavior of material systems is governed by complex physical phenomena taking place over a hierarchy of length scales, making it computationally expensive to model with physics-driven simulation methods alone. Atomic systems at the first-principles level are typically considered the smallest relevant length scale in materials modeling, and thus form the foundation of the multiscale physics governing the material response. This work introduces a new Voxelized Atomic Structure (VASt) framework for formulating physics-based, high-fidelity reduced-order material structure-property relationships using density functional theory (DFT) computations.
We first use DFT with the Boltzmann transport equation to compute the thermal conductivity of the novel electride Ca2N and to study the role of defects and sample size on thermal transport in Ga2O3. These studies demonstrate the complex physics that govern the relationship between material structure and property and highlight the significant computational cost of modeling the underlying linkages with physics-driven simulations alone. We then present VASt machine learning-based interatomic potentials, in which the physical volume around the atom of interest is mapped to a voxelized three-dimensional domain and utilized directly as the input to the convolutional neural network for implicit feature engineering. We demonstrate that VASt potentials are capable of rapidly and accurately modeling the relationship between hydrostatic strain and thermal conductivity in silicon. Finally, we develop the VASt homogenization framework, in which atomic structure is quantified by two-point spatial correlations of the charge density field, projected to a salient low-dimensional feature-space via principal component analysis, and correlated to physical properties by Gaussian process regression (GPR). The uncertainty quantification inherently provided by GPR is utilized to implement an active learning strategy based on Bayesian experiment design, which minimizes the number of computationally expensive physics simulations required to achieve highly accurate VASt structure-property relationships. We utilize this framework to elucidate physical insight into the relationship between chemical composition and thermo-mechanical properties in high entropy alloys.

SUBJECT:	Ph.D. Proposal Presentation

BY:	Matthew Barry

TIME:	Thursday, December 8, 2022, 11:00 a.m.

PLACE:	shorturl.at/cCMO6, Virtual

TITLE:	Voxelized Atomic Structure Framework for Atomistic Modeling of Multifunctional Materials

COMMITTEE:	Dr. Satish Kumar, Co-Chair (ME) Dr. Surya R. Kalidindi, Co-Chair (ME) Dr. David L. McDowell (ME) Dr. Andrew J. Medford (CHBE) Dr. Michael E. Chandross (Sandia National Laboratories)

SUMMARY The behavior of material systems is governed by complex physical phenomena taking place over a hierarchy of length scales, making it computationally expensive to model with physics-driven simulation methods alone. Atomic systems at the first-principles level are typically considered the smallest relevant length scale in materials modeling, and thus form the foundation of the multiscale physics governing the material response. This work introduces a new Voxelized Atomic Structure (VASt) framework for formulating physics-based, high-fidelity reduced-order material structure-property relationships using density functional theory (DFT) computations. We first use DFT with the Boltzmann transport equation to compute the thermal conductivity of the novel electride Ca2N and to study the role of defects and sample size on thermal transport in Ga2O3. These studies demonstrate the complex physics that govern the relationship between material structure and property and highlight the significant computational cost of modeling the underlying linkages with physics-driven simulations alone. We then present VASt machine learning-based interatomic potentials, in which the physical volume around the atom of interest is mapped to a voxelized three-dimensional domain and utilized directly as the input to the convolutional neural network for implicit feature engineering. We demonstrate that VASt potentials are capable of rapidly and accurately modeling the relationship between hydrostatic strain and thermal conductivity in silicon. Finally, we develop the VASt homogenization framework, in which atomic structure is quantified by two-point spatial correlations of the charge density field, projected to a salient low-dimensional feature-space via principal component analysis, and correlated to physical properties by Gaussian process regression (GPR). The uncertainty quantification inherently provided by GPR is utilized to implement an active learning strategy based on Bayesian experiment design, which minimizes the number of computationally expensive physics simulations required to achieve highly accurate VASt structure-property relationships. We utilize this framework to elucidate physical insight into the relationship between chemical composition and thermo-mechanical properties in high entropy alloys.