Uncertainty quantification in multivariable regression for material property prediction with Bayesian neural...

Area of Science:

• Materials Science
• Computational Materials Science
• Machine Learning

Background:

• Data-driven methods and machine learning are increasingly vital in materials science.
• Reliable uncertainty quantification (UQ) is crucial for informed decision-making in material property prediction.
• Challenges in UQ for materials include multi-scale physics, complex interactions, and limited data.

Purpose of the Study:

• To introduce a novel physics-informed Bayesian Neural Networks (BNNs) approach for UQ in material science.
• To evaluate the effectiveness of this BNNs approach for predicting creep rupture life in steel alloys.
• To assess the performance of UQ in active learning scenarios for materials prediction.

Main Methods:

• Developed a physics-informed Bayesian Neural Networks (BNNs) framework integrating material governing laws.
• Applied the BNNs approach to predict creep rupture life using three experimental steel alloy datasets.
• Compared BNNs performance against Gaussian Process Regression and other neural network variants.

Main Results:

• The physics-informed BNNs approach demonstrated competitive or superior point predictions and uncertainty estimations compared to conventional UQ methods.
• The BNNs framework showed promising performance in active learning scenarios for materials prediction.
• Bayesian Neural Networks utilizing Markov Chain Monte Carlo (MCMC) approximation yielded more reliable results than variational inference approximations.

Conclusions:

• Physics-informed BNNs provide a robust framework for uncertainty quantification in material property prediction.
• The MCMC-based BNNs approach is particularly effective for accurate and reliable creep life prediction.
• This methodology enhances the utility of machine learning in materials science by providing trustworthy predictions and uncertainty estimates.