A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics

Peter Rivera-Casillas, Sourav Dutta, Shukai Cai, Mark Loveland, Kamaljyoti Nath, Khemraj Shukla, Corey Trahan, Jonghyun Lee, Matthew Farthing, Clint Dawson · 2025

Read A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics by Peter Rivera-Casillas, Sourav Dutta, Shukai Cai, Mark Loveland, Kamaljyoti Nath, Khemraj Shukla, Corey Trahan, Jonghyun Lee, Matthew Farthing, Clint Dawson (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-14782.txt.

Introduction

Peter Rivera-Casillas and co-authors build a neural emulator for shallow water flow on coasts and rivers. Storms and floods threaten these zones, and fast forecasts guide planning. Full numeric models are too slow for live use, while cheap ones fail on new conditions. Their network learns in a squeezed latent space and steps forward on its own.

Two test beds were used: a tidal inlet in New York and a river reach in Louisiana. Inputs span start states, shifting edge flows, and bottom friction values. Skill scores stayed high, with match values above 0.9 across runs. Speed gains ran from one hundred to over a thousand times. Long rollouts of 175 days held up from random starts.

Catalog record

CREATORPeter Rivera-Casillas, Sourav Dutta, Shukai Cai, Mark Loveland, Kamaljyoti Nath, Khemraj Shukla, Corey Trahan, Jonghyun Lee, Matthew Farthing, Clint Dawson
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE105 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-14782

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics Peter Rivera-Casillas, Sourav Dutta, Shukai Cai, Mark Loveland, Kamaljyoti Nath, Khemraj Shukla, Corey Trahan, Jonghyun Lee, Matthew Farthing, Clint Dawson Abstract Coastal regions and river floodplains are particularly vulnerable to the impacts of extreme weather events. Accurate real-time forecasting of hydrodynamic processes in these areas is essential for infrastructure planning and climate adaptation. Yet high-fidelity numerical models are often too computationally expensive for real-time use, and lower-cost approaches, such as traditional model order reduction algorithms or conventional neural networks, typically struggle to generalize to out-of-distribution conditions. In this study, we present the Multiple-Input Temporal Operator Network (MITONet), a novel autoregressive neural emulator that employs latent-space operator learning to efficiently approximate high-dimensional numerical solvers for complex, nonlinear problems that are governed by time-dependent, parameterized partial differential equations. We showcase MITONet’s predictive capabilities by forecasting regional tide-driven dynamics in the Shinnecock Inlet in New York and riverine flow in a section of the Red River in Louisiana, both described by the two-dimensional shallow-water equations (2D SWE), while incorporating initial conditions, time-varying boundary conditions, and domain parameters such as the bottom friction coefficient. Despite the distinct flow regimes, the complex geometries and meshes, and the wide range of bottom friction coefficients studied, MITONet displays consistently high predictive skill, with anomaly correlation coefficients above 0.9, a maximum normalized root mean square error of 0.011, and computational speedups between 100 ​ x − 1,250 ​ x 100\text{x}-1,250\text{x} , even for 175 175 days of autoregressive rollout forecast from random initial conditions and with unseen parameter values. † † journal: Journal of Geophysical Research: Machine Learning and Computation † † corresponding: Peter Rivera-Casillas, [email protected] keypoints We present MITONet, an autoregressive neural operator framework that emulates shallow-water dynamics in coastal and riverine regions. MITONet shows superior predictive skill and stability in comparison to other well-known neural operator models. MITONet shows strong performance on two realistic shallow-water flow problems and has the potential to extend to other applications. Plain Language Summary Coastal and riverine hydrodynamics can be modeled accurately through numerical simulations. However, traditional numerical solvers are computationally intensive and require domain expertise. In this study, we introduce a machine learning framework, MITONet, which learns the governing physics from simulation data. Once trained, MITONet can quickly generate accurate predictions under similar or changing conditions, offering a faster, more accessible tool for planning and decision-making. 1 Introduction Coastal regions are densely populated and host critical infrastructure vulnerable to the impacts of rising sea levels and increased frequency of extreme weather conditions Oppenheimer et al. (2019) . River floodplains, which provide essential resources for society, are likewise susceptible to extreme weather, and human-induced disturbances Arnell and Gosling (2016) ; Alifu et al. (2022) . High-fidelity numerical models of the two-dimensional (2D) depth-averaged shallow water equations (SWE) are typically employed to capture coastal circulation patterns and riverine hydrodynamics in order to aid in crucial decision-making and planning efforts Pachev et al. (2023a) ; Pevey et al. (2020) . Despite recent technological and algorithmic advances, these methods can still be computationally expensive, time consuming, and technically challenging. Machine learning (ML) approaches have been successful in addressing some of these challenges due to their flexibility and computational efficiency Abouhalima et al. (2024) ; Giaremis et al. (2024) ; Karim et al. (2023) ; Pachev et al. (2023b) ; Song et al. (2023) ; Liu et al. (2024) ; Dutta et al. (2021b) . Predicting behaviors of highly nonlinear systems remains a persistent challenge in computational modeling. Although the universal function approximation property of neural networks has allowed deep learning to make substantial progress in the development of surrogate models Brunton et al. (2019) ; Karniadakis et al. (2021) ; Dutta et al. (2021a) ; Dutta et al. (2021c) , achieving consistent generalizability remains a challenge. To this end, operator approximation has recently emerged as an increasingly popular field of study in deep learning Azizzadenesheli et al. (2024) . The earliest attempts included efforts like nonlinear integro-differential operator regression Patel and Desjardins (2018) , where deep neural networks (DNNs) and Fourier pseudospectral methods were employed to recover complex operators from data, and physics-informed neural networks (PINNs), where DNNs were used to regress partial differential equation (PDE) models by exploiting physics knowledge Raissi et al. (2019) . While useful in certain applications, these and other earlier approaches had limitations in scalability and generalizability across varying domains of physics, highlighting the challenges of learning high-dimensional systems and multiscale features. The introduction of the Deep Operator Network (DeepONet) Lu et al. (2021) marked a new era in operator approximation. DeepONet addressed many of the issues from earlier approaches by leveraging the universal approximation theorem for operators Chen and Chen (1995) , enabling the network to efficiently approximate maps between infinite-dimensional function spaces. Subsequent developments in neural operators include the Fourier neural operator (FNO) Li et al. (2020a) , graph neural operators Li et al. (2020b) , transformer neural operators Cao (2021) , and the variationally mimetic operator network Patel et al. (2024) , among others, many of these with multiple extensions. Another critical challenge in data-driven modeling of nonlinear, real-world physical systems that has been an active area of research is the ability to forecast over long lead times Benidis et al. (2022) ; Bodnar et al. (2024) ; Lam et al. (2022) ; Lim and Zohren (2021) . Recent advances, including the Diffusion-inspired Temporal Transformer Operator (DiTTO) Ovadia et al. (2023) , the Universal Physics Transformer Alkin et al. (2024) , and OceanNet Chattopadhyay et al. (2024) , demonstrate the effectiveness of neural operator-based autoregressive models in overcoming some of these challenges by addressing the spectral bias. DiTTO uses temporal bundling to jointly predict multiple future steps, which reduces recursive error accumulation and helps correct the small-scale spectral bias that typically appears in autoregressive rollouts. The Universal Physics Transformer conditions its prediction blocks through feature modulation, allowing the network to adapt across temporal and dynamical regimes and better represent multiscale behavior. OceanNet couples an FNO backbone with a PEC-style convergent integration scheme and an explicit spectral regularizer to mitigate autoregressive error growth and the small-scale spectral bias. This research focuses on the DeepONet architecture and subsequent extensions Jin et al. (2022) ; Kontolati et al. (2023) ; Ovadia et al. (2023) ; Wang et al. (2022) that demonstrate improved performance. We specifically focus on the applications to modeling real-world coastal ocean circulation and river dynamics. An application to ocean dynamics was previously explored in Rajagopal et al. (2023) , where latent DeepONet and FNO are used for one-shot forecasting of ocean dynamics. Similarly, Choi et al. (2024) compared FNO variations and explored a recursive approach for one-shot forecasting of ocean dynamics given an initial condition (IC). Both efforts demonstrate the potential of neural operators to predict ocean dynamics; however, their usefulness is limited by their structure and data requirements, relying solely on initial ocean states without incorporating external forcings as boundary conditions (BC), which are crucial for real-world scenarios. Additionally, neither handles varying domain parameters, limiting their ability to generalize in dynamic environments. Another recent study developed an FNO-based digital twin model that autoregressively predicts the sea-surface height of the northwest Atlantic Ocean’s western boundary

Read the complete text