Harnessing large language models for data-scarce learning of polymer properties

Area of Science:

• Materials Science
• Artificial Intelligence
• Computational Modeling

Background:

• Large language models (LLMs) show potential for material modeling but require extensive data for accuracy.
• Acquiring sufficient experimental data for LLM fine-tuning is often limited and costly.
• Data scarcity poses a significant challenge for developing accurate LLM-based material models.

Purpose of the Study:

• To develop a physics-based training pipeline to overcome data scarcity in LLM material modeling.
• To enable accurate LLM training and fine-tuning even with limited experimental datasets.
• To improve the reliability and applicability of LLMs in material evaluation, analysis, and design.

Main Methods:

• A physics-based modeling framework generates abundant synthetic data for initial LLM alignment.
• A two-phase training strategy is employed: supervised pretraining with synthetic data, followed by fine-tuning with experimental data.
• The pipeline was validated using polymer flammability metrics, a domain with sparse experimental data.

Main Results:

• Supervised pretraining with synthetic data is crucial for achieving accurate fine-tuned LLMs.
• The physics-based pipeline effectively mitigates the negative impact of data scarcity.
• The approach demonstrated success in learning polymer flammability metrics.

Conclusions:

• Physics-based data generation is a viable strategy to address data scarcity in LLM material modeling.
• The proposed two-phase training pipeline enhances LLM accuracy and reliability.
• This methodology offers a pathway for more efficient and effective material design and analysis using LLMs.