BatGPT-Chem: A Foundation Large Model for Chemical Engineering.

Abstract

Large language models (LLMs) have showcased remarkable capabilities in the realm of AI for Science, and chemistry has greatly benefited from the advancement of AI tools. With a strong capacity for learning sequential data like natural language, LLMs offer immense potential. Despite this promise, the application of LLMs in chemistry remains limited, with few models specifically designed for chemical data and tasks. Hence, we propose leveraging LLMs to comprehensively model both chemical sequences and natural language sequences, aiming to tackle diverse chemical tasks. We introduce BatGPT-Chem, a general foundation large-scale model with 15 billion parameters tailored for chemical engineering. Built on a corpus of over 100 million chemical instances, BatGPT-Chem specializes in 5 core tasks: retrosynthesis prediction, molecule design, molecule description, product inference, and yield prediction. BatGPT-Chem comprehensively models the information flow between chemical language and natural language, enabling full-spectrum prediction across chemical tasks. It is one of the largest bilingual chemistry-specific LLMs, supporting both English and Chinese for input and output. BatGPT-Chem is also the first automated retrosynthesis tool capable of explicitly predicting reaction conditions, a critical but often overlooked aspect in previous models. Through rigorous zero-shot evaluations, BatGPT-Chem demonstrates state-of-the-art performance, surpassing both existing chemical LLMs and general-purpose models in accuracy and validity across a diverse range of tasks. Notably, it demonstrates superior ability in predicting both reactants and reaction conditions, as well as strong generalization in low-data settings. These results suggest that BatGPT-Chem is among the most advanced and practical chemical LLMs, with strong potential to support real-world applications in synthesis planning, drug discovery, and materials design.