AtomGPT: Atomistic Generative Pretrained Transformer for Forward and Inverse Materials Design

Area of Science:

• Materials Science
• Artificial Intelligence
• Computational Chemistry

Background:

• Large language models (LLMs) show commercial promise but are underexplored for materials design.
• Existing methods for materials design can be computationally intensive.

Purpose of the Study:

• Introduce AtomGPT, a transformer-based LLM tailored for materials design.
• Evaluate AtomGPT's capability in atomistic property prediction and structure generation.
• Demonstrate the potential of LLMs for both forward and inverse materials design.

Main Methods:

• Developed AtomGPT using transformer architectures.
• Utilized a combination of chemical and structural text descriptions for property prediction.
• Employed density functional theory (DFT) calculations for validation of generated structures.

Main Results:

• AtomGPT accurately predicts material properties like formation energies, bandgaps, and superconducting transition temperatures, comparable to graph neural networks.
• The model successfully generates atomic structures for designing new superconductors.
• DFT calculations validated the predictive accuracy and generative capabilities of AtomGPT.

Conclusions:

• AtomGPT offers an efficient and powerful approach for materials discovery and optimization.
• LLMs can be effectively leveraged for both predicting properties and generating novel material structures.
• This work opens new avenues for applying AI in materials science.