Using explainability to design physics-aware CNNs for solving subsurface inverse problems

Abstract

We present a novel method of using explainability techniques to design physics-aware convolutional neural networks (CNNs). We demonstrate our approach by developing a CNN for solving an inverse problem for shallow subsurface imaging. Although CNNs have gained popularity, the development of CNNs remains an art, as there are no clear guidelines regarding the selection of hyperparameters that provide the best network. While optimization algorithms can select hyperparameters automatically, these methods develop networks with high predictive accuracy while disregarding model explainability (descriptive accuracy). The field of Explainable Artificial Intelligence (XAI) addresses the absence of model explainability by providing tools to evaluate the internal logic of networks. In this study, we use the explainability methods Score-CAM and Deep SHAP to select hyperparameters (e.g., kernel size and network depth) to develop a physics-aware CNN for shallow subsurface imaging. We begin with an Encoder-Decoder network, which uses surface wave dispersion images to generate 2D shear wave velocity images. Through model explanations, we find that a shallow CNN using two convolutional layers with an atypical kernel size of 3 × 1 yields comparable predictive accuracy but increased descriptive accuracy. We believe this method can be used to develop networks with high predictive accuracy while providing inherent explainability.