Efficient few-shot machine learning for classification of EBSD patterns.

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Deep learning is crucial for materials science, but large labeled datasets are difficult to obtain.
• Transfer learning, using pre-trained models, accelerates training and improves performance, especially with limited data (few-shot learning).

Purpose of the Study:

• To evaluate the effectiveness of few-shot transfer learning for classifying electron backscatter diffraction (EBSD) patterns.
• To compare this approach against training a model from scratch using EBSD data.
• To enhance model interpretability through visualization techniques.

Main Methods:

• Utilized a few-shot transfer learning approach for EBSD pattern classification.
• Compared performance metrics and training history with a model trained from scratch.
• Employed filter, activation map, and Shapley value visualizations for explainability.

Main Results:

• Few-shot transfer learning demonstrated effectiveness in classifying EBSD patterns into specific space groups.
• The transfer learning model showed comparable or improved performance over the model trained from scratch.
• Visualizations provided insights into the deep learning model's decision-making process.

Conclusions:

• Few-shot transfer learning is a viable and efficient method for EBSD pattern classification in materials science.
• This technique facilitates accurate phase identification, even with limited data.
• Explainability methods enhance trust and understanding of the deep learning models used.