Artificial Intelligence and Data Science
Online ISSN : 2435-9262
Surrogate modeling of Boussinesq-type wave transformation using FNO/PINO and identifiability of an internal variable
Kazuki MASUDA, Yoji TANAKA, Takenori SHIMOZONO, Masahiko ISOBE
Author information
JOURNAL OPEN ACCESS

2026 Volume 7 Issue 3 Pages 180-198

Details
Abstract

This study investigates surrogate modeling of Boussinesq-type wave transformation using FNO/PINO and the inverse estimation of the breaking-induced eddy viscosity coefficient, which is treated as an internal variable. Training data were generated by extracting the centerline section from planar two-dimensional simulations performed with a physics-based model and treating it as a one-dimensional spatiotemporal wave field. PINO, FNO, and PINN were compared under identical data conditions. FNO/PINO reproduced the water surface elevation and line discharge more accurately than PINN. The effect of physics-based constraints depended on the inference conditions, and FNO without such constraints sometimes achieved higher accuracy under conditions close to the training data. Performance degraded when extrapolating to high-wave conditions, whereas FNO/PINO maintained a certain level of reproducibility for bathymetric extrapolation when the response statistics remained close to the training range. Without supervised data for the internal variable, the estimated variable tended to act as a degree of freedom that compensated for forward-prediction errors and inconsistencies with the physics loss, making stable inverse estimation difficult. In contrast, a small amount of supervised data stabilized the estimated scale of the internal variable.

Content from these authors
© 2026 Japan Society of Civil Engineers
Previous article Next article
feedback
Top