ChemST-LLM: A Multi-Modal Spatiotemporal Question-Answering System for Dynamic Defect-Performance Synergy in Catalysts

The development of high-performance electrocatalysts for reactions such as the oxygen evolution reaction (OER) is limited by the complex, dynamic evolution of defects under operando conditions. Existing methods and large language models (LLMs) face challenges in integrating and interpreting the diverse spatiotemporal data generated by advanced characterization techniques. To overcome this, we propose ChemST-LLM, a multi-modal spatiotemporal question-answering framework designed to understand and explain the relationship between defect evolution and OER performance in high-entropy layered double hydroxides (LDHs). ChemST-LLM features specialized temporal and graph encoders, a gated cross-modal fusion module, and an instruction-tuned LLM core with retrieval-augmented generation. We construct ChemVac-STQA, a large-scale dataset of experimental trajectories and spatiotemporal QA pairs, to support training and evaluation. Experimental results show that ChemST-LLM outperforms strong baselines across spatiotemporal QA tasks, achieving higher accuracy and robustness, particularly in challenging out-of-distribution scenarios. This work demonstrates a new paradigm for data-driven understanding and accelerated discovery of advanced electrocatalysts.