DiffGA: Steering Frozen Diffusion Models for Materials Design via Evolutionary Search
DiffGA is a retraining-free method that guides frozen diffusion models for materials design toward user-defined objectives, even when those objectives are black-box and non-differentiable. It injects evolutionary search into the model's sampling path using intermediate states and an on-the-fly Bayesian surrogate to replace expensive oracle calls, significantly reducing computational cost while maintaining fidelity.