CatalyticMLLM: A Graph-Text Multimodal Large Language Model for Catalytic Materials

CatalyticMLLM presents a unified graph-text multimodal large language model that integrates property prediction and inverse structural design for catalytic materials within a single framework. This approach overcomes limitations of traditional decoupled systems by eliminating representation space inconsistencies and evaluator bias, enabling more stable closed-loop optimization workflows for materials discovery.

CatalyticMLLM addresses a fundamental inefficiency in computational materials science by unifying two traditionally separate workflows—property prediction and inverse design—into a single multimodal model. The research demonstrates that fragmented approaches create data distribution shifts and evaluator bias when generation models and prediction models operate independently, degrading the stability of optimization cycles. The unified framework leverages both three-dimensional structural information and textual data, allowing the model to predict catalytic properties accurately while simultaneously generating and screening physically feasible candidate structures based on desired properties.

This advancement reflects broader trends in AI toward more integrated, end-to-end solutions rather than pipelined systems. The materials discovery domain has traditionally struggled with the costly iterative cycle of generating structures, evaluating them separately, and redesigning based on misaligned evaluation criteria. By consolidating these processes within shared representation spaces, CatalyticMLLM reduces computational overhead and improves consistency across the entire optimization workflow.

For the materials science and catalysis industries, this approach could accelerate the discovery of novel catalytic materials for applications ranging from battery development to chemical synthesis and carbon capture. The closed-loop optimization paradigm—inverse design, prediction, screening, and redesign—represents a more efficient path to functional materials than traditional trial-and-error methods. The research validates that joint modeling outperforms decoupled baselines on both relaxed-energy prediction and inverse design tasks, suggesting meaningful practical improvements.

The work positions multimodal LLMs as increasingly valuable tools for domain-specific scientific problems, potentially inspiring similar unified frameworks in other materials science subfields and opening opportunities for interdisciplinary AI-materials research collaborations.

• →CatalyticMLLM unifies property prediction and inverse design in one model, eliminating representation space inconsistencies between separate systems.
• →The unified framework enables closed-loop optimization with improved stability through consistent evaluation criteria across design cycles.
• →Joint modeling of structure generation and property prediction outperforms traditional decoupled baseline approaches on catalytic materials tasks.
• →Integration of 3D structural and textual information within a multimodal model improves both prediction accuracy and design feasibility.
• →The approach could accelerate discovery of novel catalytic materials for batteries, chemical synthesis, and carbon capture applications.