FURI | Spring 2020

Machine Learning and Dimensionality Reduction to Accelerate DFT Simulations of Material Properties and Enable Computational Material Design

Data icon, disabled. Four grey bars arranged like a vertical bar chart.

This research is focused on applying machine learning to create an alternative to computationally expensive DFT and MD simulations of material properties. We specifically investigate the utility of graph neural network in learning Koopman operators for converting nonlinear (atomic) dynamical models into (approximately) linear ones. For n-body dynamics (MD and DFT), we hypothesize that GNNs will help learn Koopman operators from small scale systems where high-fidelity simulations are affordable and generalize the operators to high cost large scale systems. If successful, the learned GNN model will drastically accelerate material property prediction and enable computational material design.

Student researcher

John Haas Daly

Mechanical engineering

Hometown: Albuquerque, New Mexico, United States

Graduation date: Spring 2021