Introduction

Date	Time	Title	Speaker
Dec. 5th	15:40 - 16:10	Machine learning functionals and force-fields with multipole features	Prof. Andrew J Medford

Abstract

Machine-learning has broadly impacted the field of atomistic simulations and computational chemistry with the promise of models that can maintain the accuracy of high-level physical and quantum-mechanical simulations at far lower computational costs. A key aspect of many machine-learning models is how atomic or electronic structure of a chemical system is converted into a “feature vector”. Most existing methods are specialized to specific types of chemical systems (e.g. molecules/crystals/atoms/electrons), scale poorly to data sets with many (>5) chemical elements, or require large amounts of training data to “learn” features (e.g. deep learning models). In this talk the concept of “multipole features” is introduced, and we show that this framework is sufficiently flexible to be adapted to both electrons and atoms, enabling the development of exchange-correlation functionals and machine-learned force-fields that are based on multipole features. The multipole features have a fixed length regardless of the number of chemical elements present, and can even extrapolate to new elements, making them ideally suited for problems involving many elements. The features also have a physical interpretation based on the multipole expansion, potentially opening the door to new types of hybrid physics/machine-learning models.

Biography

AJ Medford is an Assistant Professor in the School of Chemical & Biomolecular Engineering. He attended North Carolina State University as an undergraduate, and subsequently spent a year as a Fulbright fellow at the Technical University of Denmark before attending Stanford University where he received his Ph.D in Chemical Engineering in the group of Prof. Jens Nørskov. His research spans computational chemistry, machine learning, and heterogeneous catalysis. Specific projects include application of density functional theory to understand the mechanism of photocatalytic nitrogen fixation and solar fertilizer production, development of machine-learning schemes for electronic structure theory and atomistic simulations, utilizing neural networks and data analytics to extract intrinsic rate constants from transient kinetic data, and applying machine learning tools to understand complex mixtures of chemical compounds for separations and biomass catalysis.