2026 Volume 7 Issue 3 Pages 595-605
To support rapid evacuation decisions during water disasters, the development of real-time and high-accuracy flood inundation prediction technology is required. Conventional inundation analysis methods involve a high computational load. Although efforts have recently been made to accelerate computation using machine learning-based prediction models, they have a limitation: model performance depends on training data, and inference results are not necessarily physically consistent. In this study, we constructed an inundation prediction model using the Physics-Informed Neural Operator (PINO), a Scientific Machine Learning (SciML) method, and examined its applicability to real-time forecasting. As a result of conducting model training considering physical constraints using inundation analysis results from a conventional physics solver and validating its accuracy, the inference speed was improved by a factor of approximately 158 under the grid resolution and computational conditions used in this study, and the inundation distribution was accurately reproduced, demonstrating high applicability.