Learning accurate electronic interactions from cheaper models and synthetic materials 

The interactions of electrons and atomic vibrations (phonons) control a wide range of material properties, including electronic and thermal transport, superconductivity, optical response, spin decoherence, and polaron formation. Although electron-phonon (e-ph) interactions can be accurately computed using established first-principles methods, such calculations are computationally demanding—particularly for materials with large unit cells—and require material-specific workflows, which hinders automation. The absence of extensive databases for first-principles e-ph interactions further limits the application of machine learning (ML) methods in this field.

In this talk, we present an efficient strategy for generating a large and diverse database of e–ph interactions by combining tight-binding electronic structures with interatomic potentials for lattice dynamics. Instead of relying on costly ab initiodata, our approach leverages synthetic data obtained from model calculations, enabling scalable data generation and systematic exploration. This dataset is then used to train a convolutional autoencoder neural network. We demonstrate that the model, trained entirely on synthetic data, can successfully encode and reconstruct first-principles e–ph interaction data for real materials. Several proof-of-concept applications—such as e–ph interaction compression, materials classification, and super-resolution in momentum space—illustrate the power and flexibility of our approach. Our results highlight a promising route toward learning complex first-principles interactions using computationally inexpensive surrogate models and synthetic material datasets.