A Large Language Model Framework for Causal Reasoning and Performance Prediction in Multimodal Time-Series Data

Understanding the dynamic evolution of electrocatalysts under operando conditions is critical for advancing sustainable energy conversion. However, interpreting complex multimodal time-series data remains challenging. In this work, we presents Multimode Operando GPT (MOGPT), a large language model-based framework for causal reasoning and performance prediction in electrocatalysis. MOGPT integrates multimodal data processing with a Temporal Causal Discovery Module, a Catalytic Evolution Knowledge Graph, and a Causal Consistency Loss to identify temporal and causal relationships in catalyst behavior. A large-scale dataset of causal question–answer pairs across various catalyst systems is constructed for benchmarking. Experimental results show that MOGPT achieves superior performance in spatio-temporal reasoning, causal inference, and performance prediction, while maintaining strong robustness and generalization. This approach highlights the potential of large language models for interpretable and data-driven discovery in electrocatalysis.