Original scientific paper
https://doi.org/10.17794/rgn.2026.5.14
PHYSICS-INFORMED NEURAL NETWORKS FOR OFFSHORE TSUNAMI SIMULATION OF THE 2006 SOUTHERN JAVA EARTHQUAKE
Sorja Koesuma
; Sebelas Maret University, Department of Physics, Surakarta, Indonesia
*
Arnida Lailatul Latifah
; National Research and Innovation Agency (BRIN), Research Center for Computing, Cibinong, Indonesia
Angger Zufan Hanggara
; Sebelas Maret University, Department of Physics, Surakarta, Indonesia
Ahmad Fauzi Pohan
; Andalas University, Department of Physics, Padang, Indonesia
Sismanto Sismanto
; Gadjah Mada University, Department of Physics, Yogyakarta, Indonesia
Budi Eka Nurcahya
; Gadjah Mada University, Department of Physics, Yogyakarta, Indonesia
Sugeng Purwo Saputro
; National Research and Innovation Agency (BRIN), Research Center for Geological Resources, Jakarta, Indonesia
Rian Amukti
; National Research and Innovation Agency (BRIN), Research Center of Geological Disaster, Jakarta, Indonesia
* Corresponding author.
Abstract
This study presents a comparative analysis between a conventional numerical model, COMCOT, and Physics-Informed Neural Networks (PINNs) for simulating tsunami waves generated by the July 17, 2006 Java earthquake (Mw 7.7). The mesh-free PINNs framework integrates the nonlinear shallow water equations directly into its loss function, enabling simulations without spatial discretization. The model was trained to forecast wave elevation fields using initial conditions derived from the Okada fault model. Results demonstrate that the PINNs model accurately replicates the COM COT benchmark, successfully capturing the initial tsunami formation and subsequent propagation dynamics. Key quantitative metrics, including correlation coefficients (remaining above 0.98 for most of the simulation) and spatial error distribution (mean absolute percentage error <5% in the open ocean), confirm the model’s high accuracy. Critically, once trained, the PINN model generates a complete wave elevation field in under one second, offering a substantial operational advantage over conventional grid based solvers. Despite slight smoothing effects in high-gradient coastal regions, the study concludes that PINNs offer a reliable and accurate, mesh-free alternative to traditional methods, showing significant promise for future application in rapid tsunami hazard assessment and data-integration for early warning systems, with full real-time deployment identified as a key direction for future research.
Keywords
modelling; COMCOT; PINNs; non-linear shallow water equation; machine learning
Hrčak ID:
352893
URI
Publication date:
27.10.2026.
Visits: 0 *