2024 Volume 80 Issue 17 Article ID: 24-17058
Machine learning is effective in terms of computational cost and accuracy, but the black box nature of its computation poses challenges to its reliability. To address this issue, Physics-Informed Neural Networks (PINNs), which incorporate physical laws into learning, are gaining attention. This study examines the validity of PINNs under complex conditions in real marine environments using the case of the tsunami caused by the 2024 Noto Peninsula earthquake. It was found that the use of Fourier network-based learning methods improves accuracy, and compared to difference methods, PINNs can capture the general behavior of water level propagation and reduce computation time. However, significant errors were noted in discontinuous boundary areas, revealing challenges in application under complex conditions. Future prospects include the potential for improved accuracy through data assimilation with observations and learning using sparse numerical solutions.
