MaskTerial: a foundation model for automated 2D material flake detection.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Automated detection and classification of 2D materials are crucial for efficient sample fabrication and large-scale data collection.
• Existing computer vision algorithms struggle with low-contrast materials and require extensive training data.

Purpose of the Study:

• To develop a deep learning model for reliable identification of 2D material flakes.
• To overcome limitations of existing methods in detecting low-contrast materials and reduce training data requirements.

Main Methods:

• Implementation of an instance segmentation network (MaskTerial) for 2D material flake identification.
• Extensive pre-training using a synthetic data generator for realistic microscopy image creation.
• Integration of an uncertainty estimation model for classification based on optical contrast.

Main Results:

• MaskTerial reliably identifies 2D material flakes, including challenging low-contrast materials like hexagonal boron nitride.
• The model adapts quickly to new materials with as few as 5-10 training images.
• Significant improvements demonstrated over existing techniques across eight datasets and five 2D materials.

Conclusions:

• The MaskTerial model offers an accurate, efficient, and data-efficient solution for 2D material detection and classification.
• This approach enhances objectivity and enables large-scale data collection in materials science research.
• The synthetic data generation and uncertainty estimation provide a robust framework for adapting to diverse materials.