LLM-Prop: Leveraging Large Language Models to Predict Crystalline Material Properties

The LLM-Prop framework utilizes a fine-tuned T5 model encoder to predict the physical and electronic properties of crystal materials based on text descriptions of their structures. Unlike traditional Graph Neural Networks (GNNs), it leverages text-based data, preprocessing crystal descriptions by tokenizing and enhancing them with special tokens like [NUM] and [ANG] for bond distances and angles. This approach reduces complexity and improves predictive performance by incorporating more context. The framework was tested on various crystal properties, achieving state-of-the-art results with fewer training samples than GNN models. Results indicated that stripping numerical details can sometimes enhance model performance, as LLMs tend to process language more effectively than numeric data. LLM-Prop surpassed baselines like MatBERT in key metrics while also demonstrating robust transfer learning capabilities across different properties. This research highlights the advantages of text-driven methods for crystal property predictions and emphasizes the need for quality datasets in materials science.

Source link