Leveraging Uncertainty from Deep Learning for Trustworthy Material Discovery Workflows.

Area of Science:

• Materials Science
• Machine Learning
• Deep Learning

Background:

• Machine learning (ML) applications in materials science often face challenges with data requirements and model reliability.
• Deep neural networks (DNNs) can provide predictions but lack clear indicators of their confidence in these predictions.
• Understanding the uncertainty of DNNs is crucial for dependable material application workflows.

Purpose of the Study:

• To demonstrate how predictive uncertainty in DNNs can address key challenges in ML-based material science.
• To enable users to determine optimal training dataset sizes for desired classification accuracy.
• To introduce methods for detecting and handling ambiguous or out-of-distribution data samples.

Main Methods:

• Leveraging predictive uncertainty quantification from deep neural networks.
• Developing an uncertainty-guided decision referral system for sample classification.
• Utilizing scanning electron microscope (SEM) image-derived microstructure data as a use case.

Main Results:

• Predictive uncertainty effectively guides the determination of necessary training data size for specific classification accuracies.
• The uncertainty-guided referral system successfully identifies and flags confusing samples, preventing erroneous decisions.
• Predictive uncertainty accurately detects out-of-distribution test samples, including those with altered image acquisition or synthesis conditions.

Conclusions:

• Uncertainty-aware deep learning significantly enhances the performance and dependability of classification models in materials science.
• Predictive uncertainty provides a robust mechanism for managing data variability and improving ML model robustness.
• This approach offers practical solutions for material scientists using ML in their workflows.