Adaptable Method for Crystal Design across Diverse Constraints and Objectives with Pretrained Property Predictors
Researchers have developed a new, data-efficient method for crystal design that can satisfy multiple properties and physical constraints simultaneously. This approach utilizes predictor-guided gradient optimization, combining off-the-shelf property predictors with site-wise element masks and task-specific losses. The method demonstrated superior performance in perovskite design compared to generative and Bayesian baselines, achieving competitive band-gap targeting with significantly less training data. This adaptable framework also successfully supported half-metal design, offering a modular solution for optimizing candidate crystals with minimal computational cost. AI