An adaptive weighted extrapolation neural network for dispersive shock waves generated by square-root initial condition of the KdV equation
Abstract
Simulating dispersive shock waves (DSWs) using physics-informed neural networks (PINNs) presents significant numerical difficulties. In particular, the square-root initial condition exhibits singularity near the origin, making it difficult to fit accurately. This initial fitting error is amplified through nonlinear evolution and gradually accumulates during extrapolation, imposing high demands on the long-term extrapolation capability of numerical methods. Accordingly, this paper proposes an improved physics-guided multistage neural network (IPgMSNN) framework. IPgMSNN is built upon the original PgMSNN framework, with its core innovation focused on the third-stage training strategy. Specifically, online fine-tuning and a weighted target mechanism are introduced, which effectively suppress error accumulation during long-term extrapolation. Numerical experiments systematically evaluate IPgMSNN from three perspectives: forward problem solving, model stability, and parameter inversion. Experimental results demonstrate that IPgMSNN effectively captures high-frequency wavefront details, maintains long-term extrapolation stability, and achieves high-precision parameter inversion, providing an efficient solution for deep learning-based simulation of dispersive nonlinear systems.
Related
- New matrix-weighted norm inequalities and extrapolation
- Adaptive spline-based weighting functions for blended survival curve extrapolation
- Gradient estimates and volume doubling for locally finite weighted graphs with $CDψ(n,-K)$ condition
- Weighted Two-Parameter Quantum Divergences Generated by Matrix Power-Mean Gaps
- Fine-Tuning on Self-Generated and Reward-Weighted Data: Learning Dynamics, Convergence Rates, and Benefits of Off-Policyness
BibTeX
@misc{wang2026adaptive,
title = {{An adaptive weighted extrapolation neural network for dispersive shock waves generated by square-root initial condition of the KdV equation}},
author = {Xuan-Jie Wang and Rui Guo and An-Yao Jin and Hua-Ying Ren},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.27835},
eprint = {2609.27835},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.27835},
}
RIS
TY - UNPB
TI - An adaptive weighted extrapolation neural network for dispersive shock waves generated by square-root initial condition of the KdV equation
AU - Xuan-Jie Wang
AU - Rui Guo
AU - An-Yao Jin
AU - Hua-Ying Ren
PY - 2026
DA - 2026/09/01
PB - arXiv
DO - 10.48550/arxiv.2609.27835
UR - https://arxiv.org/abs/2609.27835
AB - Simulating dispersive shock waves (DSWs) using physics-informed neural networks (PINNs) presents significant numerical difficulties. In particular, the square-root initial condition exhibits singularity near the origin, making it difficult to fit accurately. This initial fitting error is amplified through nonlinear evolution and gradually accumulates during extrapolation, imposing high demands on the long-term extrapolation capability of numerical methods. Accordingly, this paper proposes an improved physics-guided multistage neural network (IPgMSNN) framework. IPgMSNN is built upon the original PgMSNN framework, with its core innovation focused on the third-stage training strategy. Specifically, online fine-tuning and a weighted target mechanism are introduced, which effectively suppress error accumulation during long-term extrapolation. Numerical experiments systematically evaluate IPgMSNN from three perspectives: forward problem solving, model stability, and parameter inversion. Experimental results demonstrate that IPgMSNN effectively captures high-frequency wavefront details, maintains long-term extrapolation stability, and achieves high-precision parameter inversion, providing an efficient solution for deep learning-based simulation of dispersive nonlinear systems.
ER -