Accelerating Calculations Based on Many-body Perturbation Theory with Machine Learning
Abstract: Machine learning (ML) is transforming how we study and design materials by enabling predictions that would otherwise require expensive first-principles simulations. However, applying ML to the most accurate quantum mechanical methods is challenging because they rely on large intermediate quantities that are costly to compute and store.
In many-body perturbation theory, particularly GW calculations, this bottleneck is the frequency-dependent dielectric matrix, whose size scales unfavorably with system size. The Without Empty State (WEST) code (https://west-code.org/) represents the dielectric matrix through Projective Dielectric Eigenpotentials (PDEPs), which capture the essential screening information in a more compact form. However, even this representation remains high-dimensional for ML applications.
In this talk, I will present ML strategies to address these challenges. The focus will be on using 3D variational autoencoders (VAEs) to learn compressed representations of PDEPs, achieving 1000× dimensionality reduction while preserving physical accuracy.
I will discuss the VAE architecture, training methodology, and integration of decoded representations back into WEST, demonstrating mean absolute errors below 0.2 eV on standard molecular benchmarks. This compressed latent space provides a tractable target for structure-to-property models, circumventing the curse of dimensionality in predicting high-dimensional electronic structure quantities directly. I will also briefly discuss physics-based approximations for reducing computational costs at heterogeneous interfaces.
Bio: Joseph Frimpong is a postdoctoral researcher at the Center for Nanoscale Materials at Argonne National Laboratory and the Midwest Integrated Center for Computational Materials. He received his Ph.D. in Chemistry from Wayne State University. His research focuses on many-body perturbation theory calculations of electronic and optical properties, ML for electronic structure and the development of computational workflows to accelerate materials discovery for renewable energy and optoelectronic applications.
See all upcoming talks at https://www.anl.gov/mcs/lans-seminars.