From LLM to Agent: A large-language-model-driven machine learning framework for catalyst design of MgH₂ dehydrogenation

Magnesium hydride (MgH₂), a promising high-capacity hydrogen storage material, is hindered by slow dehydrogenation kinetics. AI-driven catalyst discovery to address this is often hampered by the laborious extraction of data from unstructured literature. To overcome this, we introduce a transformative “LLM to Agent” framework that synergistically integrates Large Language Models (LLMs) for automated data curation with Machine Learning (ML) for predictive design. We automatically constructed a comprehensive database of 809 MgH₂ catalysts (6555 data rows) with high fidelity and an ~40-fold acceleration over manual methods. The resulting ML models achieved high accuracy (average R² > 0.91) in predicting dehydrogenation temperature and activation energy, subsequently guiding a Genetic Algorithm (GA) in an exploratory inverse design that autonomously uncovered key design principles for high-performance catalysts. Encouragingly, a strong alignment was found between these AI-discovered principles and the design strategies of recently reported, state-of-the-art experimental systems, providing substantial evidence for the validity of our approach. The framework culminates in Cat-Advisor, a novel, domain-adapted multi-agent system. Cat-Advisor translates ML predictions and retrieval-augmented knowledge into actionable design guidance, demonstrating capabilities that surpass those of general-purpose LLMs in this specialized domain. This work delivers a practical AI toolkit for accelerated materials discovery and advances the emerging Agent-based paradigm for designing next-generation energy technologies.