CatalyticMLLM: A Unified Multimodal Large Model Enabling Closed-Loop Optimization for Catalytic Material Property Prediction and Inverse Design

This article introduces QE-Catalytic-V2, a unified graph-text multimodal large language model that integrates catalytic material property prediction and inverse design through a shared representation space. It addresses issues such as inconsistent representations and distribution shifts in traditional decoupled paradigms, enables closed-loop optimization, and outperforms decoupled baselines, providing a tool for the intelligent transformation of materials science.

In catalytic material research, property prediction and inverse design have traditionally been separated. While this facilitates the 'generation-evaluation-screening' process, inconsistencies in representation spaces and training objectives lead to data distribution shifts and evaluator biases, limiting the stability of closed-loop optimization. Core issues include: inconsistent representations, distribution shifts, evaluation biases, and unstable optimization.

QE-Catalytic-V2 adopts a graph-text multimodal architecture, processing 3D structures (CIF format) and textual information with a shared representation space. It has bidirectional capabilities: forward prediction (structure → performance) and inverse generation (performance → structure). It implements a closed-loop optimization process: inverse design → property prediction → intelligent screening → iterative optimization. Technical highlights: guarantees of physical feasibility (chemical rationality, crystallographic constraints, energy stability) and end-to-end training (simultaneously optimizing prediction and generation).

Experiments show: In terms of prediction performance, the unified paradigm significantly reduces prediction errors for catalytic relaxation energy; in terms of design quality, generated structures meet performance requirements with better diversity and rationality; in terms of closed-loop stability, the accumulation of errors in iterative optimization is reduced, outperforming decoupled methods.

This work can accelerate material discovery (rapid screening, guiding experiments, discovering new materials), transfer across domains such as batteries, optoelectronic materials, and drug molecules, promote the development of AI4Science, and provide a new methodological reference.

QE-Catalytic-V2 addresses the problem of separating property prediction and inverse design, verifying the advantages of the unified paradigm. Future directions: expanding material categories and performance indicators, introducing experimental feedback, integrating molecular dynamics simulations, and developing interactive interfaces.