Domain-Trained Language Model for Inverse Design and Synthesis of High-Performance Hydrogen Storage MOFs.

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Developing advanced materials like metal-organic frameworks (MOFs) is crucial for efficient hydrogen storage.
• Traditional methods for MOF discovery are often time-consuming and resource-intensive.
• Improving structure-property relationships is key to designing MOFs with enhanced performance.

Purpose of the Study:

• To develop a domain-specific large language model (MOFs-LLM) for accelerating the inverse design and synthesis of MOFs.
• To enhance the model's ability to reason about structure-property relationships in MOFs.
• To enable the design of MOFs with optimized hydrogen storage capacity and synthetic accessibility.

Main Methods:

• Trained MOFs-LLM on a large dataset of 210 million tokens from MOF publications and crystal structures.
• Integrated chemical knowledge and structural features into the language model.
• Validated the model's performance by comparing its structure-property reasoning against baseline methods.
• Utilized MOFs-LLM for inverse design of candidate MOF structures.
• Synthesized a novel MOF guided by the model's predictions.

Main Results:

• MOFs-LLM demonstrated a 46.7% enhancement in capturing structure-property relationships compared to baseline methods.
• Successfully designed 60 candidate MOF frameworks optimized for hydrogen storage and synthesis.
• Synthesized a novel MOF, Cu-LLMs-1, in three experimental iterations.
• The synthesized MOF exhibited a hydrogen uptake of 1.33 wt% at room temperature, ranking among the top pure MOFs.
• The study successfully bridged virtual screening with experimental realization in materials discovery.

Conclusions:

• Domain-trained language models like MOFs-LLM can significantly accelerate materials discovery.
• MOFs-LLM shows great potential for the inverse design and synthesis of functional materials.
• The successful synthesis of Cu-LLMs-1 validates the model's predictive capabilities and practical applicability.