Open Materials Generation (OMG): Bespoke Materials Generation with Stochastic Interpolants

We present Open Materials Generation (OMG), a state-of-the-art and open-source framework for generative modeling of atomic crystals. OMG is trained by the recently introduced stochastic interpolant (SI) framework, enabling a high degree of flexibility. Implementation of SI unifies Score-Based Diffusion Modeling (SBDM) and Conditional Flow Matching (CFM), providing the capability for SBDM in the stochastic limit, CFM in the deterministic limit, and anything in between by providing the user with maximal control over the generative flow from any arbitrary source distribution to a desired target distribution. OMG supports property-guided generation, optimal transport between base and target distributions, the incorporation of physically inductive bias into the SI framework, and the freedom to customize the functional form of the SI to maximize performance. We present results for generated atomic crystals to demonstrate the performance of OMG on various benchmarks.