Generative Pretrained Transformer for Heterogeneous Catalysts.

Area of Science:

• Chemistry
• Material Science
• Artificial Intelligence

Background:

• Discovering new materials is crucial for chemistry and material science.
• Traditional methods like trial-and-error and machine learning are used for material design.
• Transformer-based language models show promise for generating novel material structures.

Purpose of the Study:

• Introduce the catalyst generative pretrained transformer (CatGPT) for inorganic catalyst discovery.
• Demonstrate CatGPT's capability to generate valid and accurate catalyst structures.
• Showcase CatGPT's flexibility for generating specific catalyst types through text-conditioning and fine-tuning.

Main Methods:

• Trained a transformer-based language model (CatGPT) on a large dataset of inorganic catalyst structures.
• Utilized text-conditioning and fine-tuning to guide catalyst generation.
• Applied fine-tuning to a binary alloy catalyst dataset for oxygen reduction reaction (ORR) screening.

Main Results:

• CatGPT successfully generates valid and accurate inorganic catalyst structures.
• The model demonstrates high performance in expanding the chemical space for materials.
• Fine-tuning enabled the generation of catalysts specialized for the two-electron oxygen reduction reaction (2e-ORR).

Conclusions:

• Generative language models, like CatGPT, are powerful tools for accelerating catalyst discovery.
• CatGPT offers a foundation model for exploring diverse catalyst structures.
• This approach has significant potential for discovering new materials with desired properties.