Published on:

28 October 2023

Primary Category:

Machine Learning

Paper Authors:

Phillip Pope,

David Jacobs

•

Generated large dataset of charge densities for catalysts using DFT

•

Trained graph neural network model on simpler catalysts

•

Model generalizes to unseen combinations of elements

•

Learned densities speed up DFT convergence by 13% on average

•

Shows progress towards combinatorial generalization

Machine learning for predicting charge densities

This paper investigates using machine learning to predict the charge density field for catalyst materials. Charge density prediction can enable faster convergence in density functional theory calculations. The authors generated a large dataset of charge densities for unary, binary, and ternary catalysts. They trained a graph neural network model on simpler unary and binary cases, and tested if it could generalize to more complex binary and ternary combinations not seen during training. The results showed the learned densities led to faster convergence versus standard baselines in over 80% of test cases, reducing iterations needed for convergence by 13% on average. This demonstrates a step towards combinatorial generalization, an important property for catalyst discovery applications.

Machine learning improves molecular force fields

Machine learning improves density functional theory

Predicting cluster size in particle systems

Machine learning predicts thermal conductivity in layered materials

Machine learning for molecular dynamics simulations

Machine learning prediction of thermal conductivity

No comments yet, be the first to start the conversation...

Sign up to comment on this paper