AtomGPT for Inverse Materials Design

Large language models (LLMs) have demonstrated transformative potential across various domains, yet their application in materials science remains underexplored. We present AtomGPT and DiffractGPT, two transformer-based models designed to accelerate materials characterization and design through advanced machine learning techniques. AtomGPT specializes in atomistic property prediction and structure generation, integrating chemical composition and structural descriptors to achieve predictive accuracies comparable to state-of-the-art graph neural networks for properties such as formation energies, electronic bandgaps, and superconducting transition temperatures. DiffractGPT focuses on multi-value datasets, including X-ray diffraction (XRD) pattern interpretation and structural refinement, leveraging transformers to decode complex diffraction patterns for rapid phase identification and crystal structure determination. Together, AtomGPT and DiffractGPT create a synergistic framework that bridges computational and experimental materials science, offering an efficient, comprehensive platform for forward and inverse materials design. AtomGPT is publicly available at https://github.com/usnistgov/atomgpt.