AtomVision: A Machine Vision Library for Atomistic Images.

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Computer vision offers significant potential for advancing materials design.
• Microscopy images are crucial for understanding material structures.
• Existing tools may lack integration for comprehensive image analysis and dataset generation.

Purpose of the Study:

• Introduce AtomVision, an integrated library for materials design.
• Enable generation and curation of microscopy image datasets.
• Apply diverse machine learning techniques to materials imaging data.

Main Methods:

• Established a diverse atomistic image dataset of ~10,000 materials.
• Developed and compared convolutional and graph neural networks for Bravais lattice classification.
• Utilized U-Net for pixelwise atom/background classification and GANs for super-resolution.
• Applied Natural Language Processing for dataset curation from arXiv.
• Integrated the framework with experimental microscopy data (Rh, Fe3O4, SnS).

Main Results:

• Demonstrated the capability to generate and curate large, diverse materials imaging datasets.
• Achieved accurate classification of Bravais lattices using neural networks.
• Successfully applied deep learning for atom identification and image super-resolution.
• Showcased seamless integration with experimental microscopy data.

Conclusions:

• The AtomVision library provides a versatile platform for materials design using computer vision.
• It facilitates data generation, curation, and advanced machine learning analysis of microscopy images.
• This integrated approach accelerates the discovery and design of novel materials.