LLMs Show Potential for Materials Optimization

Discovering materials with specific desirable properties often involves searching through vast candidate spaces, where experimental or computational evaluations are costly and time-consuming. Active learning strategies address this by using previous observations to intelligently select the next candidate for evaluation, typically relying on probabilistic surrogate models. This research explores whether open-weight Large Language Models (LLMs) can function as effective standalone acquisition policies in this context. The study evaluated five different LLMs across four retrospective finite-pool materials optimization tasks, testing various candidate-presentation strategies. Their performance was benchmarked against both random selection and conventional Gaussian-process methods. The findings indicate that LLM policies generally reach the global optimum in fewer iterations than random selection, suggesting they provide a useful signal for acquisition without requiring task-specific training. However, the performance of LLMs relative to Gaussian-process methods was mixed. While conventional acquisition methods performed better on most tasks, LLMs occasionally matched or even outperformed them in certain settings. The study also highlighted significant variability in performance across different tasks, LLM models, initializations, and candidate presentation formats, indicating that no single LLM approach consistently performed best across all scenarios. Overall, open-weight LLMs demonstrate potential as acquisition policies for finite-pool materials search, though their reliability remains sensitive to task specifics and presentation methods.