Bayesian neural networks for uncertainty quantification in data-driven materials modeling

Modern machine learning (ML) techniques, in conjunction with simulation-based methods, present remarkable potential for various scientific and engineering applications. Within the materials science field, these data-based methods can be used to build efficient structure-property linkages that can be further integrated within multi-scale simulations, or guide experiments in a materials discovery setting. However, a critical shortcoming of state-of-the-art ML techniques is their lack of reliable uncertainty/error estimates, which severely limits their use for materials or other engineering applications where data is often scarce and uncertainties are substantial. This paper presents methods for Bayesian learning of neural networks (NN) that allow consideration of both aleatoric uncertainties that account for the inherent stochasticity of the data-generating process, and epistemic uncertainties, which arise from consideration of limited amounts of data. In particular, algorithms based on approximate variational inference and (pseudo-)Bayesian model averaging achieve an appropriate trade-off between accuracy of the uncertainty estimates and accessible computational cost. The performance of these algorithms is first presented on simple 1D examples to illustrate their behavior in both extrapolation and interpolation settings. The approach is then applied for the prediction of homogenized and localized properties of a composite material. In this setting, data is generated from a finite element model, which permits a study of the behavior of the probabilistic learning algorithms under various amounts of aleatoric and epistemic uncertainties. (C) 2021 Elsevier B.V. All rights reserved.

Automated summary

Modern machine learning techniques, in conjunction with simulation-based methods, show great potential for scientific and engineering applications. However, a critical shortcoming is the lack of reliable uncertainty estimates. This paper presents methods for Bayesian learning of neural networks, achieving a balance between accuracy of uncertainty estimates and computational cost.