Predicting Stress-Strain Curve with Confidence: Balance Between Data Minimization and Uncertainty Quantification by a...

Area of Science:

• Materials Science
• Computational Materials Science
• Polymer Science

Background:

• Machine learning (ML) shows promise for predicting polymer stress-strain curves.
• Challenges include data scarcity, inherent processing-property variability (aleatoric uncertainty), and model confidence (epistemic uncertainty).

Purpose of the Study:

• To develop an uncertainty-aware framework for accurate stress-strain curve prediction.
• To differentiate between aleatoric and epistemic uncertainty in polymer processing.
• To minimize data requirements while ensuring prediction reliability.

Main Methods:

• Utilized a dual Bayesian model integrating Bayesian neural networks (BNNs).
• Employed a Taguchi array for efficient data sampling (27 samples in 4D space).
• Incorporated hidden and output-distribution layers for uncertainty quantification.

Main Results:

• The dual Bayesian model accurately predicted stress-strain curves.
• Successfully distinguished between aleatoric and epistemic uncertainties.
• Provided a 95% confidence interval for predictions with modest uncertainty.

Conclusions:

• The developed framework balances data minimization and robust uncertainty quantification.
• Enables reliable material property prediction even with limited data.
• Offers a novel approach for uncertainty-aware machine learning in polymer science.