Accelerating domain-aware electron microscopy analysis using deep learning models with synthetic data and image-wide...

Area of Science:

• Materials Science
• Computational Biology
• Image Analysis

Background:

• Machine learning (ML) models improve microscopy feature detection but rely on limited, flawed, manually labeled datasets.
• Lack of domain awareness in current ML models hinders their practical application in scientific research.

Purpose of the Study:

• To develop a physics-based synthetic data generator for ML in microscopy.
• To create an ML model trained on synthetic data with comparable performance to models trained on human-labeled data.
• To enhance ML model reliability and domain awareness through confidence metrics.

Main Methods:

• Created a physics-based synthetic image and data generator.
• Trained an ML model on the generated synthetic data.
• Developed an image-wide confidence metric using feature prediction scores.
• Applied thresholding to filter ambiguous or out-of-domain images.

Main Results:

• The synthetic data-trained ML model achieved precision (0.86), recall (0.63), F1 scores (0.71), and engineering property predictions (R ² = 0.82), comparable to human-labeled data.
• The confidence metric improved performance by 5-30% with a 25% filtering rate.
• Demonstrated the efficacy of synthetic data in eliminating reliance on human labeling.

Conclusions:

• Physics-based synthetic data generation is a viable alternative to manual labeling for ML in microscopy.
• Synthetic data enables domain awareness and improves the reliability of ML models for feature detection.
• This approach reduces human dependency and enhances the scalability of ML applications in scientific imaging.