Title: Integrating Large Language Models with Physics-Based Simulations to Generate Inorganic Materials

Abstract: While contemporary generative machine learning (ML) architectures excel at designing novel inorganic crystals with specific desired attributes, the subsequent challenge of planning their synthesis remains significant. This difficulty stems from the intricate nature of the underlying physical processes and the scarcity of robust computational resources. To address this gap, we present a new hybrid framework designed to assess the capabilities of Large Language Models (LLMs) in the realm of inorganic synthesis planning. This approach merges thermodynamic databases with simplified kinetic models to simulate realistic synthesis environments.

Our case study centers on the niobium-oxygen system, selected for its array of industrially important oxide phases and comprehensive data availability. Through computational experiments, we juxtapose synthesis pathways generated by LLMs against those produced by classical path-planning algorithms. The results indicate that the implicit priors embedded within LLMs facilitate the derivation of more feasible strategies. Notably, in our evaluation context, traditional search algorithms functioned mainly as a baseline for comparison rather than as direct rivals. This outcome underscores the inherent complexity of the synthesis problem and demonstrates the distinct advantages provided by the implicit knowledge structures of LLMs.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC