Unified Graph-Text Model for Catalytic Material Property Prediction and Inverse Design
A new multimodal large language model, QE-Catalytic-V2, unifies property prediction and inverse structural design of catalytic materials within a single framework. Traditionally, these tasks are decoupled, leading to data distribution shifts and evaluator bias that limit closed-loop optimization. QE-Catalytic-V2 integrates both tasks in a shared graph-text representation space, enabling reliable property prediction and candidate structure generation. The model addresses inconsistencies between generative and predictive models, aiming to stabilize the optimization workflow. The research is published on arXiv as preprint 2605.17254.
Key facts
- QE-Catalytic-V2 is a unified graph-text multimodal large language model for catalytic materials.
- It integrates property prediction and inverse design within the same model.
- Traditional decoupled paradigm causes data distribution shifts and evaluator bias.
- The model uses a shared representation space for both tasks.
- It enables reliable property prediction and candidate structure generation.
- The research is published on arXiv with identifier 2605.17254.
- The model aims to stabilize closed-loop optimization.
- The work addresses inconsistency between generative and predictive models.
Entities
—