MicroscopyGPT: Generating Atomic-Structure Captions from Microscopy Images of 2D Materials with Vision-Language Transformers
Kamal Choudhary
Abstract
Determining complete atomic structures directly from microscopy images remains a long-standing challenge in materials science. MicroscopyGPT is a vision-language model (VLM) that leverages multimodal generative pretrained transformers to predict full atomic configurations, including lattice parameters, element types, and atomic coordinates, from scanning transmission electron microscopy (STEM) images. The model is trained on a chemically and structurally diverse data set of simulated STEM images generated using the AtomVision tool and the JARVIS-DFT as well as the C2DB two-dimensional (2D) materials databases. The training set for fine-tuning comprises approximately 5000 2D materials, enabling the model to learn complex mappings from image features to crystallographic representations. I fine-tune the 11-billion-parameter LLaMA model, allowing efficient training on resource-constrained hardware. The rise of VLMs and the growth of materials data sets offer a major opportunity for microscopy-based analysis. This work highlights the potential of automated structure reconstruction from microscopy, with broad implications for materials discovery, nanotechnology, and catalysis.