ChatProp: Integrating Large Language Models and Physical Chemistry Tools for Enhanced Material Property Prediction

ACL ARR 2025 February Submission1740 Authors

Abstract:

Material property prediction is essential for optimizing physical processes and developing novel materials in physical chemistry and materials science.
Large language models (LLMs) have emerged as powerful tools for this task but encounter challenges in physics-related applications due to limited access to specialized external knowledge.
To overcome these limitations, we present ChatProp, an intelligent agent that integrates first-principles (FP) calculations with machine learning-driven potential energy surface (PES) models to enhance the accuracy and efficiency of material property prediction.
Leveraging LLMs such as GPT-4, ChatProp extracts critical information from textual inputs and generates appropriate responses, thereby eliminating the need for rigid, structured queries.
The system forms a robust pipeline for tasks such as data retrieval and property prediction. In evaluations using GPT-4, ChatProp achieves accuracy rates of 96.8\% for property prediction.
As the first agent to integrate FP and machine learning PES models for material property prediction, ChatProp demonstrates the potential of combining LLMs with databases and machine learning in physical chemistry, showcasing transformative capabilities for future scientific advancements.

Paper Type: Long

Research Area: Question Answering

Research Area Keywords: reasoning

Contribution Types: NLP engineering experiment

Languages Studied: English

Submission Number: 1740