Mutual information for explainable deep learning of multiscale systems

JOURNAL OF COMPUTATIONAL PHYSICS

卷:444

Article

Taverniers, Soren¹  Hall, Eric J.²  Katsoulakis, Markos A.³  Tartakovsky, Daniel M.¹ 

[1] Stanford Univ, Dept Energy Resources Engn, Stanford, CA 94305 USA

[2] Univ Dundee, Div Math, Dundee DD1 4HN, Scotland

[3] Univ Massachusetts, Dept Math & Stat, Amherst, MA 01003 USA

关键词: Surrogate model;    Mutual information;    Global sensitivity analysis;    Black box;    Probabilistic graphical model;    Electrical double-layer capacitor;   

DOI  :  10.1016/j.jcp.2021.110551

来源: Elsevier

【 摘 要 】

Timely completion of design cycles for complex systems ranging from consumer electronics to hypersonic vehicles relies on rapid simulation-based prototyping. The latter typically involves high-dimensional spaces of possibly correlated control variables (CVs) and quantities of interest (QoIs) with non-Gaussian and possibly multimodal distributions. We develop a model-agnostic, moment-independent global sensitivity analysis (GSA) that relies on differential mutual information to rank the effects of CVs on QoIs. The data requirements of this information-theoretic approach to GSA are met by replacing computationally intensive components of the physics-based model with a deep neural network surrogate. Subsequently, the GSA is used to explain the surrogate predictions, and the surrogate-driven GSA is deployed as an uncertainty quantification emulator to close design loops. Viewed as an uncertainty quantification method for interrogating the surrogate, this framework is compatible with a wide variety of black-box models. We demonstrate that the surrogate-driven mutual information GSA provides useful and distinguishable rankings via a validation step for applications of interest in energy storage. Consequently, our information-theoretic GSA provides an outer loop for accelerated product design by identifying the most and least sensitive input directions and performing subsequent optimization over appropriately reduced parameter subspaces. (C) 2021 Elsevier Inc. All rights reserved.

【 授权许可】

Free