LLM-Fusion introduces a novel approach combining multiple data modalities through large language models to accelerate materials discovery and accurately predict properties.

• Integrates chemical structure, compositional, and property data using transformer-based architecture
• Achieves superior performance in material property prediction across multiple benchmark datasets
• Provides explainable predictions through attention mechanisms that highlight influential features
• Creates richer representations than traditional fusion methods by leveraging contextual relationships

This engineering breakthrough has significant implications for materials science, potentially reducing development time and costs while enabling the discovery of new materials with tailored properties for specific applications.

LLM-Fusion: A Novel Multimodal Fusion Model for Accelerated Material Discovery