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

Area of Science:

• Artificial Intelligence in Chemistry
• Large Language Models (LLMs)
• Chemical Engineering

Background:

• Large language models (LLMs) show promise for AI in science, particularly chemistry, due to their ability to process sequential data.
• Current applications of LLMs in chemistry are limited, with few models specifically designed for chemical data and tasks.
• There is a need for advanced AI tools to model chemical sequences and natural language for diverse chemical applications.

Purpose of the Study:

• To leverage LLMs for comprehensive modeling of chemical and natural language sequences to address various chemical engineering tasks.
• To introduce BatGPT-Chem, a large-scale, bilingual foundation model tailored for chemical applications.
• To enable full-spectrum prediction across chemical tasks by modeling information flow between chemical and natural language.

Main Methods:

• Developed BatGPT-Chem, a 15-billion-parameter foundation model trained on over 100 million chemical instances.
• Specialized BatGPT-Chem for 5 core tasks: retrosynthesis prediction, molecule design, molecule description, product inference, and yield prediction.
• Enabled bilingual (English/Chinese) input/output and integrated explicit prediction of reaction conditions for retrosynthesis.

Main Results:

• BatGPT-Chem demonstrated state-of-the-art performance in zero-shot evaluations, outperforming existing chemical LLMs and general models.
• Achieved superior accuracy and validity across diverse tasks, particularly in predicting reactants and reaction conditions.
• Showcased strong generalization capabilities, even in low-data settings, highlighting its practical utility.

Conclusions:

• BatGPT-Chem represents a significant advancement in chemistry-specific LLMs, offering practical solutions for real-world applications.
• The model's ability to predict reaction conditions explicitly addresses a critical gap in automated retrosynthesis.
• BatGPT-Chem has strong potential to support and accelerate synthesis planning, drug discovery, and materials design.