MOFGPT: Generative Design of Metal-Organic Frameworks using Language Models.

Area of Science:

• Materials Chemistry
• Computational Materials Science
• Artificial Intelligence in Chemistry

Background:

• Discovering Metal-Organic Frameworks (MOFs) with tailored properties is challenging due to their vast structural complexity.
• Traditional computational methods like DFT are accurate but too slow for large-scale screening.
• Machine learning (ML) offers a data-driven alternative to accelerate materials discovery.

Purpose of the Study:

• To develop a novel, scalable framework for the de novo design of MOFs using advanced AI techniques.
• To address the challenges posed by MOF complexity in generative modeling.
• To accelerate the discovery of synthesizable MOFs with desired functional attributes.

Main Methods:

• A reinforcement learning (RL)-enhanced, transformer-based generative framework was developed.
• A chemically informed string representation, MOFid, was used to encode MOF connectivity and topology.
• The pipeline integrates a GPT-based generator, a transformer property predictor (MOFormer), and an RL optimization module.

Main Results:

• The framework successfully generates topologically valid and synthesizable MOFs.
• Property-guided reward functions in the RL module optimize generated MOF candidates.
• The approach demonstrates efficient inverse design of MOFs with specific properties.

Conclusions:

• Large language models coupled with reinforcement learning can significantly accelerate inverse design in reticular chemistry.
• This AI-driven approach unlocks new possibilities for computational MOF discovery.
• The MOFid representation enables scalable and effective generative modeling for complex materials.