2023 Volume 79 Issue 15 Article ID: 22-15011
With the rapid increase in torrential rainfall disasters and associated landslides, the demand for predictive numerical simulations has been growing. Due to computational limits, one needs to introduce approximations, however, the parameters to link detailed and approximated simulations (e.g. drag / bed friction coefficients) are determined empirically, and their applicability remains vague. In this context, this paper presents the application of a deep learning model, PINN (Physics-Informed Neural Network) to inverse analysis. This work assumes a scenario where one has an access to limited data (which is the case for real-site observation), and proposes utilizing data’s spatial features extracted from POD (Proper Orthogonal Decomposition) instead of conventional random number-based method. We found that proposed method supports PINN for faster training convergence and efficient parameter identification.
