Learning accurate electronic interactions in materials from cheap models

Electronic interactions govern a wide range of transport and nonequilibrium physics in materials. Computing these interactions accurately – typically, using from first principles techniques – is currently an important research direction. As these calculations are computationally demanding, one could attack this problem using data-driven or machine-learning (ML) methods, which can accelerate calculations in various fields. However, learning electronic interactions requires a large training set with diverse data for thousands to millions of compounds. Generating a database of this size for electronic interactions is a seemingly insurmountable task.

In this talk, we show an effective solution to this problem. Focusing on electron-phonon (e-ph) interactions, we propose using synthetic data based on flexible tight-binding electronic models and interatomic lattice potentials, as opposed to using first-principles data. Using these cheaper models allows us to rapidly generate millions of e-ph coupling matrices and build a large data set for ML of e-ph interactions. This data generation process that can be easily controlled and fine-tuned. Our ML model trained on this data set employs a convolutional autoencoder architecture. We show that this model can accurately reconstruct e-ph matrices for real materials, while also achieving a high compression ratio for the e-ph matrices in the latent representation. An analysis of latent representation will also be discussed.