CatPath‐GPT: A Mixture of Experts System for Computational Catalyst Design

2026/04/01 by Xingyu Wang, Zihao Jiao, Ziyun Wang · 1 voice
Chemical Engineering · Energy · Materials Science · #CO2 Reduction Techniques and Catalysts #Catalysis and Oxidation Reactions #Machine Learning in Materials Science

paper · doi:10.1002/smm2.70080

openalex publication_date 2026/04/01 · openalex created_date 2026/04/28 · openalex updated_date 2026/06/18

Abstract

ABSTRACT Despite advances in computational catalysis, the complexity of theoretical calculations and specialised expertise requirements limit the broader adoption of catalyst design tools. This work introduces CatPath‐GPT, a mixture‐of‐experts framework that democratizes computational catalyst design by integrating three AI specialists: product prediction (77.2% accuracy), computational planning, and automated code generation through a unified BERT‐based router. Experimental validation through two case studies demonstrates practical impact: systematic screening of Cu x Zn 1 − x catalysts identifies optimal compositions for selective CO 2 RR (Cu 75 Zn 25 for ethanol), while high‐throughput metal oxide screening reproduces Nørskov's classical scaling relationships and identifies high‐activity materials. Benchmarking against GPT‐4 and Mistral‐7B demonstrates superior performance across catalyst‐related tasks, particularly in modeling complex surface reactions. The open‐source framework enables researchers without computational expertise to perform advanced catalyst design, potentially transforming how catalytic materials are discovered and optimized.

Citations

• The surface energy of metals
• Large language models generate functional protein sequences across diverse families
• Evolutionary-scale prediction of atomic-level protein structure with a language model
• SWIFT:A Scalable lightWeight Infrastructure for Fine-Tuning
• Mixtral of Experts
• Mistral 7B
• Llama 2: Open Foundation and Fine-Tuned Chat Models
• Crystal Structure Generation with Autoregressive Large Language Modeling
• ChemCrow: Augmenting large-language models with chemistry tools
• GPT-4 Technical Report
• Understanding and Improving Knowledge Distillation for Quantization-Aware Training of Large Transformer Encoders
• LoRA: Low-Rank Adaptation of Large Language Models
• Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity
• High-Throughput Computational-Experimental Screening Protocol for the Discovery of Bimetallic Catalysts
• Language Models are Few-Shot Learners
• Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
• DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter
• RoBERTa: A Robustly Optimized BERT Pretraining Approach
• EDA: Easy Data Augmentation Techniques for Boosting Performance on Text Classification Tasks
• Attention Is All You Need
• Self-Consistent Equations Including Exchange and Correlation Effects
• Contextual feature extraction hierarchies converge in large language models and the brain

Discussions

• CatPath-GPT: A Mixture of Experts System for Computational Catalyst Design @OpenAI doi.org/10.1002/smm2... [bsky, 5 points, 1 comments]

Related

• AI-Empowered Catalyst Discovery: A Survey from Classical Machine Learning Approaches to Large Language Models
• Catalyst Diffusion Transformer: Generative Inverse Design of Heterogeneous Catalysts
• AceWGS: An LLM-Aided Framework to Accelerate Catalyst Design for Water-Gas Shift Reactions
• Artificial Intelligence in Energy Catalysis: From Catalyst Screening to Decision-Making
• Catalysis in the digital age: Unlocking the power of data with machine learning