SWIN-DeepONet: A Swin Transformer Enhanced DeepONet for Learning Wave Dynamics

Parth Sethi, Miguel Beneitez, Tianning Tang, Anirbit Mukherjee

Keywords: scientific-ML, operator learning

TL;DR: We introduce the SWIN-DeepONet architecture for predicting rough wave prediction.

Abstract:

Accurate short-horizon prediction of nonlinear wave dynamics is critical for the safe
operation of offshore infrastructure, marine vessels, and coastal systems. While neural
operator frameworks such as DeepONet offer a principled approach to learning solution
operators for parametric PDEs, their standard MLP branch networks struggle to encode
the coherent, multi-scale spatial structure characteristic of dispersive wave envelopes.
We introduce *SWIN-DeepONet*, which replaces the MLP branch of DeepONet with a
Swin Transformer encoder. By lifting the 1-D input profile into a 2-D token grid and
applying hierarchical shifted-window self-attention, the Swin branch captures
local-to-global spatial dependencies at linear computational cost. We evaluate both models on single-step and fully autoregressive rollout tasks using wave envelope data governed by the modified nonlinear Schrodinger (MNLS) equation. SWIN-DeepONet reduces
average rollout MSE by $24.3\%$ and final relative $L^2$ error by $16.0\%$ over hand-crafted
Fourier-feature using DeepONet baselines, while exhibiting substantially better generalisation,
with no observable train--test divergence. We also show that erstwhile strong baselines like the DINO entirely fail on this task. Thus our SWIN enhanced DeepONet opens the door to more physically faithful surrogate models for nonlinear wave dynamics.

Paper Type: Long Paper

Email Sharing: We authorize the sharing of all author emails with Program Chairs.

Data Release: We authorize the release of our submission and author names to the public in the event of acceptance.

Submission Number: 5