Open Materials Generation using Stochastic Interpolant to Discover New Superconductors

Generative models have made significant strides in various domains, including images, text, and video. Recently, there has been considerable progress in materials generation using methods such as diffusion and flow matching, through models such as diffCSP, MatterGen, and FlowMM. In this study, we leverage Open Materials Generation (OMG), where the training is guided by the property of interest. OMG is based on stochastic interpolants and provides flexible pathways for material generation, allowing both diffusion and flow matching approaches. The core idea is to map a base distribution to a target distribution, facilitating the discovery of new materials with the desired attribute. We fine-tune our model on a dataset of 7,000 superconductors, targeting the strength of electron-phonon interaction. Generated materials undergo a rigorous screening process to identify stable, unique, and novel candidates. These are further screened for superconductivity using the BETE-NET [arXiv:2401.16611] framework and refined through density functional theory calculations.

This work was supported by National Science Foundation under award number 2311632.