Learning Model Discrepancy in Physics-guided Neural Networks

Abstract

Recent machine learning models such as deep learning become complicated and difficult to understand the meaning of learned weights. And also, there is a possibility of obtaining output ignoring the prior knowledge because machine learning model is learned from the observed data including the noise outlier. Especially in the natural science field exploring the principle, non interpretable model cannot be a useful model unless the model has descriptiveness even if model could perform well with high accuracy. On the other hand, numerical simulation using physical model is difficult to predict long-term due to the model discrepancy. In order to solve such disadvantages, we focused on the method that integrate machine learning model and physical model. This paper proposes the algorithm that can predict two components, namely outputs based on the law of physics and their model discrepancy. As an example, we used on predicting winds in the upper troposphere from thermal wind equations.