UNKNOWN PARAMETER ESTIMATION USING PHYSICS-INFORMED NEURAL NETWORKS WITH NOISED OBSERVATION DATA

Abstract

Due to global warming, torrential rain disasters have been occurring in various places, raising the demand for predictive simulation technology. In order to perform numerical simulations, material parameters are often determined based on observation data or empirical laws. In this context, this work presents the application of a class of neural networks, PINNs (Physics-Informed Neural Networks), to inverse problems. The characteristic of PINNs is its prediction is guaranteed by the informed physical laws, initial, or boundary conditions, by forming the loss function as a combination of predictive and physical loss. Predictive loss is the difference between PINNs prediction and the ground truth, while physical loss is defined by how much PINNs prediction satisfies the physical laws and conditions. This paper investigates the effect of each loss and the performance improvement by introducing a weighting factor, and discusses PINNs applicability to parameter estimation with noised observation data.