Strategies for Assessing and Improving the Diversity and Utility for Generative Models for Material Structure

Generative machine learning (ML) models can use computational and experimental materials databases to create large quantities of novel material structures. Via additional property-prediction models, generated materials can be assessed for suitability in design tasks. Identified materials can be studied in detail, synthesized, and characterized. Here, we assess how one state-of-the-art generative model, the physics-guided crystal generation model (PGCGM), can be used as part of the inverse design process. We show that the PGCGM's learned latent space is not smooth with respect to variation in model inputs and output material properties, making material optimization difficult and limited. We also demonstrate that most of its generated structures are predicted to be thermodynamically unstable by a separate property-prediction model, partially due to out-of-domain data challenges in stability-prediction. Our findings suggest strategies and mechanisms for improving the effectiveness of generative models for inverse design. In particular, we implement and evaluate some of these suggested strategies, such as developing smoother latent spaces and more generalizable property-prediction models.