LeMat-GenBench: Bridging the gap between crystal generation and materials discovery

Generative machine learning models hold great promise for accelerating materials discovery, particularly through the inverse design of inorganic crystals---enabling an unprecedented exploration of chemical space. Yet, the lack of standardized benchmarks makes it difficult to evaluate, compare and further develop these ML models meaningfully. In this benchmark paper, we introduce LeMat-GenBench, a unified framework for assessing generative models of crystalline materials. In particular, we propose a set of evaluation metrics alongside a set of tasks (unconditional, conditional, and limited-budget crystal generation), designed to better inform model developers as well as downstream, practical applications. To support it, we release an open-source evaluation suite and a public leaderboard on Hugging Face with verified submissions. Altogether, LeMat-GenBench aims to guide model development and bridge the gap between generative modeling and practical materials discovery.