Date:
Thu, 11/05/2023 - 11:00 to 12:00
Location:
Los Angeles Bld., Jerusalem, Israel
Abstract
Molecular dynamics are a key tool for the study of a wide variety of solid-state and liquid systems. However, theapplication of molecular dynamics (MD) simulations to the interpretation of Raman scattering spectra is hindered byinability of atomistic simulations to account for the dynamic evolution of electronic polarizability, requiring the use ofeither ab initio method or parameterization of machine learning models. More broadly, the dynamic evolution ofelectronic-structure-derived properties cannot be treated by the current atomistic models. Furthermore, theconstruction of accurate atomistic models is often difficult even for modeling of interatomic dynamics due to theneed to manually choose the relevant energy terms. While machine learning potentials have offered an automatedand flexible approach for construction of potential energy and polarizability models for MD simulations, they aretypically at least an order of magnitude less computationally efficient than traditional analytic atomistic potentials. Inour recent work, we have developed a simple, physically-based atomistic model with few adjustable parameters thatcan accurately represent the changes in the electronic polarizability tensor for molecules and solid-state systems. Dueto its compactness, the model can be applied for simulations of Raman spectra of large (~1,000,000-atom) systemswith modest computational cost. To demonstrate its accuracy, the model was applied to the CO2 molecule, waterclusters, and BaTiO3 and CsPbBr3 perovskites and showed good agreement with ab-initio-derived and experimentalpolarizability tensor and Raman data. We have also showed the moments theorem can be used to derive featuresthat enable parameterization of small but accurate neural networks for electronic structure and atomic forceevolution of solid state systems. Finally, our investigation of machine learning model reliability revealed that a simpledistance-based criterion can estimate the accuracy of prediction of machine learning methods such as SVR andrandom forest.
Molecular dynamics are a key tool for the study of a wide variety of solid-state and liquid systems. However, theapplication of molecular dynamics (MD) simulations to the interpretation of Raman scattering spectra is hindered byinability of atomistic simulations to account for the dynamic evolution of electronic polarizability, requiring the use ofeither ab initio method or parameterization of machine learning models. More broadly, the dynamic evolution ofelectronic-structure-derived properties cannot be treated by the current atomistic models. Furthermore, theconstruction of accurate atomistic models is often difficult even for modeling of interatomic dynamics due to theneed to manually choose the relevant energy terms. While machine learning potentials have offered an automatedand flexible approach for construction of potential energy and polarizability models for MD simulations, they aretypically at least an order of magnitude less computationally efficient than traditional analytic atomistic potentials. Inour recent work, we have developed a simple, physically-based atomistic model with few adjustable parameters thatcan accurately represent the changes in the electronic polarizability tensor for molecules and solid-state systems. Dueto its compactness, the model can be applied for simulations of Raman spectra of large (~1,000,000-atom) systemswith modest computational cost. To demonstrate its accuracy, the model was applied to the CO2 molecule, waterclusters, and BaTiO3 and CsPbBr3 perovskites and showed good agreement with ab-initio-derived and experimentalpolarizability tensor and Raman data. We have also showed the moments theorem can be used to derive featuresthat enable parameterization of small but accurate neural networks for electronic structure and atomic forceevolution of solid state systems. Finally, our investigation of machine learning model reliability revealed that a simpledistance-based criterion can estimate the accuracy of prediction of machine learning methods such as SVR andrandom forest.