Physics-Informed Deep Operator Learning for Computational Hydraulics Modeling
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.
Kai: We're looking at the paper: "Physics-Informed Deep Operator Learning for Computational Hydraulics Modeling".
Mira: Here's the gist of the paper. A physics-informed DeepONet framework is developed for computational hydraulics modeling that learns the solution operator of the 2D shallow water equations (SWEs) to create fast surrogate models because it enhances generalization capability for out-of-distribution cases while maintaining physical consistency.
Kai: Framework Architecture The DeepONet framework employs a two-branch neural network architecture designed to learn the solution operator of the 2D SWEs.
Mira: This architecture consists of a branch network and a trunk network.
Kai: The branch network processes discretized input functions that represent the boundary conditions, initial conditions, parameters, or other forcing terms of the SWEs.
Paper summary: Kai: So what does this framework actually claim? They claim that PI-SWE-DeepONet has much enhanced prediction capability when you look at scenarios that are outside the training data, which they call out-of-distribution cases.
Mira: That’s the key finding, isn't it? The physics constraint helps it generalize better to those situations compared to the purely data-driven SWE-DeepONet model.
Lev: It means if you run this on a real flood event that looks a bit different from what you trained it on, the physics part keeps the predictions more stable and prevents them from diverging too quickly.
Kai: They show that this physics-informed model exhibits slower error growth and larger breakdown distances when compared to the standard SWE-DeepONet.
Mira: But they also mention a trade-off, which is important for us to keep in mind: the simulated results have slightly higher errors for cases that are within the training data range.
Lev: That’s the cost of adding physics fidelity. It means you have to tune those loss weights carefully because you’re balancing fitting the data against satisfying those continuous equations, which is a tricky balancing act.
Kai: They set their ratio of data loss to PDE loss at two to one based on in-distribution test cases, but they acknowledge that this weighting is subjective because the scales are so different.
Conclusion: Mira: So looking at the paper "Physics-Informed Deep Operator Learning for Computational Hydraulics Modeling", what does it all boil down to in simple terms?
Kai: It boils down to this framework giving us a way to get fast predictions for hydraulic modeling, and you can choose your model based on what you need.
Lev: If you need reliable predictions for situations that are totally outside your training set, like an unexpected flood scenario, PI-SWE-DeepONet is the better choice because of its enhanced generalization.
Mira: And if your goal is just to replicate the results from a trusted traditional hydraulic solver within the data you’ve already seen, then SWE-DeepONet might be simpler and sufficient.
Kai: The implication for us is that this method gives us a tool where we can switch on or off that physics constraint, giving us flexibility in how we use it.
Lev: It provides an automated workflow through the HydroNet package for building these surrogate models, which cuts down on the computational time needed for those complex simulations.
Mira: The paper suggests that as long as you prioritize physical consistency or need to model uncertainty beyond your training data, this physics-informed operator learning approach is a worthwhile direction to explore in computational hydraulics.
Xiaofeng Liu, Yong G. Lai
Department of Civil and Environmental Engineering, The Pennsylvania State University · Institute of Computational and Data Sciences, The Pennsylvania State University · Technical Service Center, U.S. Bureau of Reclamation
physics.flu-dyn
Submitted: 2026-01-13
Updated: 2026-01-13
Code: https://github.com/psu-efd/HydroNet
License: http://creativecommons.org/licenses/by-nc-nd/4.0/
Importance score: 79/100
The gist: A physics-informed DeepONet framework is developed for computational hydraulics modeling that learns the solution operator of the 2D shallow water equations (SWEs) to create fast surrogate models…
Key concepts
- DeepONet Framework
- A neural network architecture designed to learn the solution operator of a partial differential equation (PDE). It uses two parts: a branch network for input functions (like boundary conditions) and a trunk network for spatial coordinates. These are combined via an inner product to predict the solution directly.
- Physics-Informed Mode (PI-SWE-DeepONet)
- A training method where the model learns not just from data, but also by minimizing errors related to the governing SWEs. This is achieved by adding a physics loss term to the total loss function. This forces the learned model to respect the actual laws of fluid dynamics during prediction.
- Competing Objectives
- The challenge in training is balancing two conflicting goals: accurately fitting the available training data and strictly minimizing violations of the governing physical equations (PDE residuals). Finding a right balance between these two objectives requires careful tuning of loss weights.
- Out-of-Distribution Performance
- This measures how well the trained model performs on situations or input data that are outside the range it was originally trained on. The physics-informed model (PI-SWE-DeepONet) shows significantly better performance and stability in these unknown cases compared to standard models.
Terminology
Summary
A physics-informed DeepONet framework is developed for computational hydraulics modeling that learns the solution operator of the 2D shallow water equations (SWEs) to create fast surrogate models because it enhances generalization capability for out-of-distribution cases while maintaining physical consistency. p. 1 manuscript submitted to
Abstract
Traditional 2D hydraulic models face significant computational challenges that limit their
applications that are time-sensitive or require many model evaluations. This study presents
a physics-informed Deep Operator Network (DeepONet) framework for computational
hydraulics modeling that learns the solution operator of the 2D shallow water equations
(SWEs) to create fast surrogate models.Where the paper says it
Framework Architecture
The DeepONet framework employs a two-branch neural network architecture designed to learn the solution operator of the 2D SWEs. This architecture consists of a branch network and a trunk network. p. 5 1 Physics-Informed DeepONet Framework
The DeepONet framework employs a two-branch neural network architecture designed to learn the solution operator of the 2D SWEs. As illustrated in Figure 1, the architecture consists of two parallel networks: a branch network and a trunk network. As illustrated in Figure 1, the architecture consists of two parallel networks: a branch network and a trunk network. The
branch network processes discretized input functions that represent the boundary conditions, initial conditions, parameters, or other forcing terms of the SWEs. These input
functions are sampled at discrete points and encoded into a high-dimensional feature representation through the branch network. The trunk network, operating in parallel to the
branch network, processes the spatiotemporal coordinates (x, y, t) at which the solution
is to be evaluated. For steady state problems in the present study, the input is only the
spatial coordinates (x, y). The outputs of the branch and trunk networks are then combined through an inner product (or more generally, a dot product) to produce the out-
–6–Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Training Signal Construction
The training of the model is carried out by minimizing the total loss function (Ltotal) consisting of four loss terms. The total loss function is expressed as a weighted sum of all loss components: Ltotal = λdataLdata + λPDEsLPDEs + λBC/IC LBC/IC + λotherLother (6). p. 8 These losses are optional and they are omitted for the examples in this work. The total loss function is then expressed as a weighted sum of all loss components:
Ltotal = λdataLdata + λPDEsLPDEs + λBC/IC LBC/IC + λotherLother (6)
where the weights λdata, λPDEs, λBC/IC , and λother are hyperparameters that measure
the relative importance of corresponding loss components. In the purely data-driven DeepONet mode, termed “SWE-DeepONet” in this work,
only the data loss function Ldata is minimized. Data loss measures the discrepancy between the predicted and the true values, which can be expressed as the mean square error (MSE):
Ldata =
1
Ndata
N
Xdata
i=1
(qˆi − qi)
2
(1)
where qˆi
is the predicted value, qi
is the true value, and Ndata is the number of data points. The true value qi can be generated by running physics-based numerical models such as
SRH-2D or HEC-RAS 2D. In the physics-informed DeepONet mode, termed “PI-SWE-DeepONet” in this work,
the flow physics governed by the SWEs is additionally incorporated into the training process by adding the physics loss function LPDEs. In the physics-informed DeepONet mode, termed “PI-SWE-DeepONet” in this work,
the flow physics governed by the SWEs is additionally incorporated into the training process by adding the physics loss function LPDEs. This loss function measures the violation of the governing equations and is computed as the summation of PDE residuals at
each PDE point.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Performance Gains and Trade-offs
The PI-SWE-DeepONet model possesses much enhanced prediction capability than SWE-DeepONet when applied to out-of-distribution scenarios. The physics constraint improves the generalization capability of the model for out-of-distribution cases and is shown to exhibit slower error growth and larger breakdown distances in comparison with SWE-DeepONet. p. 1 The physics-informed model is shown to exhibit slower error growth and larger breakdown distances in comparison with SWE-DeepONet. The
gain of the physics-informed training, however, comes with costs, chief among which are
the simulated results have slightly higher errors for in-distribution cases. The physics-informed model is shown to exhibit slower error growth and larger breakdown distances in comparison with SWE-DeepONet. The
gain of the physics-informed training, however, comes with costs, chief among which are
the simulated results have slightly higher errors for in-distribution cases. It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution.Where the paper says it
Where the paper says it
Where the paper says it
Model Selection Guidelines
Guidelines are developed for selecting the appropriate approach based on a real-world case. PI-SWE-DeepONet is recommended for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority. p. 1 It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution. It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution.Where the paper says it
Where the paper says it
Competing Objectives
Adding PDE residuals into DeepONet loss changes the learning target fundamentally. The network is additionally required to satisfy the continuous shallow water equations, which the numerical model (e. p. 1 This study presents
a physics-informed Deep Operator Network (DeepONet) framework for computational
hydraulics modeling that learns the solution operator of the 2D shallow water equations
(SWEs) to create fast surrogate models. The framework can operate in two modes: a
purely data-driven SWE-DeepONet that learns from numerical solver such as SRH-2D,
and a physics-informed PI-SWE-DeepONet that additionally incorporates the continuous SWEs as constraints during training. 1 Introduction
Two-dimensional (2D) depth-averaged hydraulic models solve the shallow water
equations (SWEs), which are derived from depth-averaging the Navier-Stokes equations
under the hydrostatic pressure assumption. They have been widely used in recently years
due to the availability of models such as SRH-2D (Sedimentation and River Hydraulics
2D) (Y. It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution.Where the paper says it
Where the paper says it
Where the paper says it
Conclusion
The study presents a physics-informed DeepONet framework for computational hydraulics modeling. The developed framework can switch on and off the SWEs physics constraint, resulting in PISWE-DeepONet and SWE-DeepONet models. p. 1 This study presents
a physics-informed Deep Operator Network (DeepONet) framework for computational
hydraulics modeling that learns the solution operator of the 2D shallow water equations
(SWEs) to create fast surrogate models. The framework can operate in two modes: a
purely data-driven SWE-DeepONet that learns from numerical solver such as SRH-2D,
and a physics-informed PI-SWE-DeepONet that additionally incorporates the continuous SWEs as constraints during training. Based on a real-world case, steady flows in
a reach of the Sacramento River in California, it is demonstrated that PI-SWE-DeepONet
possesses much enhanced prediction capability than SWE-DeepONet when applied to
out-of-distribution scenarios. The physics-informed model is shown to exhibit slower error growth and larger breakdown distances in comparison with SWE-DeepONet. It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution. This flexibility makes the framework adaptable to diverse
hydraulic modeling scenarios, from constant discharges (steady-state) to complex unsteady
open channel flows with multiple inflow and exit boundaries. In this study, the algorithms and workflow presented are implemented in the opensource Python package HydroNet, which uses PyTorch for deep learning model development and training.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Open Research Section
All code and data that support the findings of this study are available from the HydroNet GitHub repository. The framework can switch on and off the SWEs physics constraint, resulting in PISWE-DeepONet and SWE-DeepONet models. p. 1 This study presents
a physics-informed Deep Operator Network (DeepONet) framework for computational
hydraulics modeling that learns the solution operator of the 2D shallow water equations
(SWEs) to create fast surrogate models. The framework can operate in two modes: a
purely data-driven SWE-DeepONet that learns from numerical solver such as SRH-2D,
and a physics-informed PI-SWE-DeepONet that additionally incorporates the continuous SWEs as constraints during training. This flexibility makes the framework adaptable to diverse
hydraulic modeling scenarios, from constant discharges (steady-state) to complex unsteady
open channel flows with multiple inflow and exit boundaries. In this study, the algorithms and workflow presented are implemented in the opensource Python package HydroNet, which uses PyTorch for deep learning model development and training. The physics-informed training process allows the model to
learn from both available simulation data and the underlying flow physics, creating a more
robust and reliable surrogate for applications. This approach has been shown to reduce
the need for a large set of training data and potential to improve the predictive capability to untrained scenarios (Wang et al. The physics-informed training process allows the model to
learn from both available simulation data and the underlying flow physics, creating a more
robust and reliable surrogate for applications. This approach has been shown to reduce
the need for a large set of training data and potential to improve the predictive capability to untrained scenarios (Wang et al. It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Acknowledgments
We acknowledge the time, effort, and insights provided by the reviewers and the editor.
References
Baydin, A. G., Pearlmutter, B. A., Radul, A. A., & Siskind, J. M. (2018). Automatic differentiation in machine learning: a survey. p. 7 In this study, the non-conservative form of the 2D SWEs is adopted, versus the conservative form, because it produces simpler automatic differentiation graphs and reduces
residual imbalance during training within the physics-informed DeepONet framework. Deep-learning surrogate models generally do not enforce conservation laws on a control
volume by computing fluxes across control-volume boundaries. Instead, they enforce conservation through pointwise PDE residuals. This choice of using non-conservative form
is also consistent with prior physics-informed SWE studies (Bihlo & Popovych, 2022)
and is physically valid because the supervised SRH-2D data inherently enforce conservation at the solution level. Major operator learning architectures include
DeepONet (Lu et al. , 2021, 2022), Fourier Neural Operator (FNO) (Li, Kovachki, Azizzadenesheli, liu, et al. The physics-informed DeepONet framework for open channel flows governed by the
2D shallow water equations. The training of the model (weights and biases in the branch and trunk networks)
is carried out by minimizing the total loss function (Ltotal) consisting of four loss terms:
data loss (Ldata), physics loss (LPDEs), boundary conditions and initial conditions loss
(LBC/IC ), and other losses (Lother); details are explained below. , 2021; Li, Kovachki, Azizzadenesheli, Liu, et
al. , 2021; Boule & Alex, 2024; Chen et al. , 2021; Kovachki et al. , 2024; Li, Kovachki, Azizzadenesheli, liu, et al. manuscript submitted to
Physics-Informed Deep Operator Learning for
Computational Hydraulics Modeling
Xiaofeng Liu1,2and Yong G. Lai3
1Department of Civil and Environmental Engineering, The Pennsylvania State University, University
Park, PA, USA
2
Institute of Computational and Data Sciences, The Pennsylvania State University, University Park, PA,
USA
3Technical Service Center, U. They have been widely used in recently years
due to the availability of models such as SRH-2D (Sedimentation and River Hydraulics
2D) (Y. Lai, 2010) and HEC-RAS 2D (Hydrologic Engineering Center’s River Analysis System) (Brunner, 2016). They have been widely used in recently years
due to the availability of models such as SRH-2D (Sedimentation and River Hydraulics
2D) (Y. Lai, 2010) and HEC-RAS 2D (Hydrologic Engineering Center’s River Analysis System) (Brunner, 2016). manuscript submitted to
Physics-Informed Deep Operator Learning for
Computational Hydraulics Modeling
Xiaofeng Liu1,2and Yong G. Lai3
1Department of Civil and Environmental Engineering, The Pennsylvania State University, University
Park, PA, USA
2
Institute of Computational and Data Sciences, The Pennsylvania State University, University Park, PA,
USA
3Technical Service Center, U. , 2021; Wang et al. , 2021; Goswami
et al. Lai3
1Department of Civil and Environmental Engineering, The Pennsylvania State University, University
Park, PA, USA
2
Institute of Computational and Data Sciences, The Pennsylvania State University, University Park, PA,
USA
3Technical Service Center, U. Bureau of Reclamation, Denver, CO, USA
Key Points:
• Physics-informed DeepONet framework is developed for computational hydraulics
modeling based on the 2D shallow water equations
• Physics constraint comes with the cost of higher errors for in-distribution cases
because of the inconsistency between discrete and continuous governing equations
• Physics constraint improves the generalization capability of the model for out-ofdistribution cases
Corresponding author: Xiaofeng Liu, xzl123@psu. Such models have been explored using the convolutional neural network (CNN) for spatial variation, long short-term memory (LSTM) network for temporal evolution, and fully connected neural network for parameterto-output mapping (Song et al. , 2023; Haces-Garcia et al. Surrogate models learn the input-output
relationship of the original model from a limited set of training simulations, then provide rapid predictions for new inputs without having to solve the full partial differential equation (PDE) system. It is widely accepted that the most promising surrogate models are based on deep machine learning techniques. , 2024; Rivera-Casillas et al. , 2025;
Keppler et al. , 2024; Rivera-Casillas et al. , 2025;
Keppler et al. The physics-informed DeepONet framework for open channel flows governed by the
2D shallow water equations. The training of the model (weights and biases in the branch and trunk networks)
is carried out by minimizing the total loss function (Ltotal) consisting of four loss terms:
data loss (Ldata), physics loss (LPDEs), boundary conditions and initial conditions loss
(LBC/IC ), and other losses (Lother); details are explained below. This study presents
a physics-informed Deep Operator Network (DeepONet) framework for computational
hydraulics modeling that learns the solution operator of the 2D shallow water equations
(SWEs) to create fast surrogate models. The framework can operate in two modes: a
purely data-driven SWE-DeepONet that learns from numerical solver such as SRH-2D,
and a physics-informed PI-SWE-DeepONet that additionally incorporates the continuous SWEs as constraints during training. manuscript submitted to
Physics-Informed Deep Operator Learning for
Computational Hydraulics Modeling
Xiaofeng Liu1,2and Yong G. Lai3
1Department of Civil and Environmental Engineering, The Pennsylvania State University, University
Park, PA, USA
2
Institute of Computational and Data Sciences, The Pennsylvania State University, University Park, PA,
USA
3Technical Service Center, U. Despite their widespread adoption and proven reliability, traditional 2D hydraulic
models face computational challenges that have placed a limit on their applicability range. High-fidelity simulations of large-scale flood events or complex river systems can require
hours to days of computational time, even on modern high-performance computing systems. , 2017, 2019). The framework also leverages pyHMT2D, a Python interface for the
SRH-2D hydraulic model, to process simulation results, extract flow field data from result files, and generate collocation points for physics-informed training. Other challenges are also discussed, such
as the loss weighting approach. 1 Introduction
Two-dimensional (2D) depth-averaged hydraulic models solve the shallow water
equations (SWEs), which are derived from depth-averaging the Navier-Stokes equations
under the hydrostatic pressure assumption. The Sacramento River is one of the major rivers in California, USA, and has historically been subject to complex flow dynamics due to its confluence with the San Joaquin River and the intricate channel geometry in the delta region. In this study, a section of the river near the Fremont Weir is simulated under a constant-discharge flow condition. Operator learning represents a transformative approach in scientific machine learning (SciML) that addresses the fundamental challenge of learning solution operators for
parametric partial differential equations (Lu et al. , 2021; Kovachki et al. Operator learning represents a transformative approach in scientific machine learning (SciML) that addresses the fundamental challenge of learning solution operators for
parametric partial differential equations (Lu et al. , 2021; Kovachki et al. Operator learning represents a transformative approach in scientific machine learning (SciML) that addresses the fundamental challenge of learning solution operators for
parametric partial differential equations (Lu et al. , 2021; Kovachki et al. Operator learning represents a transformative approach in scientific machine learning (SciML) that addresses the fundamental challenge of learning solution operators for
parametric partial differential equations (Lu et al. , 2021; Kovachki et al. , 2021; Kovachki et al. , 2024; Li, Kovachki, Azizzadenesheli, liu, et al. , 2021; Kovachki et al. , 2024; Li, Kovachki, Azizzadenesheli, liu, et al. , 2021; Kovachki et al. , 2024; Li, Kovachki, Azizzadenesheli, liu, et al. manuscript submitted to
mappings between infinite-dimensional function spaces rather than finite-dimensional vectors. Operator learning represents a transformative approach in scientific machine learning (SciML) that addresses the fundamental challenge of learning solution operators for
parametric partial differential equations (Lu et al. manuscript submitted to
mappings between infinite-dimensional function spaces rather than finite-dimensional vectors. Operator learning represents a transformative approach in scientific machine learning (SciML) that addresses the fundamental challenge of learning solution operators for
parametric partial differential equations (Lu et al. Despite their widespread adoption and proven reliability, traditional 2D hydraulic
models face computational challenges that have placed a limit on their applicability range. High-fidelity simulations of large-scale flood events or complex river systems can require
hours to days of computational time, even on modern high-performance computing systems. , 2021; Li, Kovachki, Azizzadenesheli, Liu, et al. , 2021), PhysicsInformed Neural Operators (PINO) (Li, Zheng, et al. manuscript submitted to
mappings between infinite-dimensional function spaces rather than finite-dimensional vectors. Operator learning represents a transformative approach in scientific machine learning (SciML) that addresses the fundamental challenge of learning solution operators for
parametric partial differential equations (Lu et al. (2021), was among
the first to demonstrate the feasibility of learning solution operators for PDEs (Lu et al. DeepONet employs a two-branch architecture inspired by the universal approximation theorem for operators: a branch network processes discretized input functions
(e. , 2021; Wang et al. , 2021; Goswami
et al. The framework can operate in two modes: a
purely data-driven SWE-DeepONet that learns from numerical solver such as SRH-2D,
and a physics-informed PI-SWE-DeepONet that additionally incorporates the continuous SWEs as constraints during training. Based on a real-world case, steady flows in
a reach of the Sacramento River in California, it is demonstrated that PI-SWE-DeepONet
possesses much enhanced prediction capability than SWE-DeepONet when applied to
out-of-distribution scenarios. In this study, the algorithms and workflow presented are implemented in the opensource Python package HydroNet, which uses PyTorch for deep learning model development and training. The neural network architectures, including both the branch and
trunk networks of DeepONet, are built using PyTorch’s neural network modules, enabling
automatic differentiation for computing PDE residuals in physics-informed training (Paszke
et al. This flexibility makes the framework adaptable to diverse
hydraulic modeling scenarios, from constant discharges (steady-state) to complex unsteady
open channel flows with multiple inflow and exit boundaries. In this study, the algorithms and workflow presented are implemented in the opensource Python package HydroNet, which uses PyTorch for deep learning model development and training. , 2021; Kovachki et al. , 2024; Li, Kovachki, Azizzadenesheli, liu, et al. , 2021; Li, Kovachki, Azizzadenesheli, Liu, et al. , 2021), PhysicsInformed Neural Operators (PINO) (Li, Zheng, et al. , 2024; Kramer et al. , 2024; Rivera-Casillas et al. Lai3
1Department of Civil and Environmental Engineering, The Pennsylvania State University, University
Park, PA, USA
2
Institute of Computational and Data Sciences, The Pennsylvania State University, University Park, PA,
USA
3Technical Service Center, U. Bureau of Reclamation, Denver, CO, USA
Key Points:
• Physics-informed DeepONet framework is developed for computational hydraulics
modeling based on the 2D shallow water equations
• Physics constraint comes with the cost of higher errors for in-distribution cases
because of the inconsistency between discrete and continuous governing equations
• Physics constraint improves the generalization capability of the model for out-ofdistribution cases
Corresponding author: Xiaofeng Liu, xzl123@psu. manuscript submitted to
Abstract
Traditional 2D hydraulic models face significant computational challenges that limit their
applications that are time-sensitive or require many model evaluations. This study presents
a physics-informed Deep Operator Network (DeepONet) framework for computational
hydraulics modeling that learns the solution operator of the 2D shallow water equations
(SWEs) to create fast surrogate models. , 2021; Li, Kovachki, Azizzadenesheli, Liu, et
al. , 2021; Boule & Alex, 2024; Chen et al. It is widely accepted that the most promising surrogate models are based on deep machine learning techniques. Such models have been explored using the convolutional neural network (CNN) for spatial variation, long short-term memory (LSTM) network for temporal evolution, and fully connected neural network for parameterto-output mapping (Song et al. In recent years, researchers and practitioners have increasingly turned to surrogate modeling approaches that can approximate the behavior of traditional hydraulic
models at a fraction of the computational cost. Surrogate models learn the input-output
relationship of the original model from a limited set of training simulations, then provide rapid predictions for new inputs without having to solve the full partial differential equation (PDE) system. manuscript submitted to
mappings between infinite-dimensional function spaces rather than finite-dimensional vectors. Operator learning represents a transformative approach in scientific machine learning (SciML) that addresses the fundamental challenge of learning solution operators for
parametric partial differential equations (Lu et al. Major operator learning architectures include
DeepONet (Lu et al. , 2021, 2022), Fourier Neural Operator (FNO) (Li, Kovachki, Azizzadenesheli, liu, et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Out-of-Distribution Performance
The performance on out-of-distribution cases is now presented to show the generalizability of trained surrogate models when scenarios are beyond the parameter range of the training data. PI-SWE-DeepONet shows a superior generalization capability in comparison with SWE-DeepONet and increased advantages are observed with increasing Wasserstein distance. p. 1 Based on a real-world case, steady flows in
a reach of the Sacramento River in California, it is demonstrated that PI-SWE-DeepONet
possesses much enhanced prediction capability than SWE-DeepONet when applied to
out-of-distribution scenarios. The physics-informed model is shown to exhibit slower error growth and larger breakdown distances in comparison with SWE-DeepONet. Based on a real-world case, steady flows in
a reach of the Sacramento River in California, it is demonstrated that PI-SWE-DeepONet
possesses much enhanced prediction capability than SWE-DeepONet when applied to
out-of-distribution scenarios. The physics-informed model is shown to exhibit slower error growth and larger breakdown distances in comparison with SWE-DeepONet. Other challenges are also discussed, such
as the loss weighting approach. 1 Introduction
Two-dimensional (2D) depth-averaged hydraulic models solve the shallow water
equations (SWEs), which are derived from depth-averaging the Navier-Stokes equations
under the hydrostatic pressure assumption. Lai3
1Department of Civil and Environmental Engineering, The Pennsylvania State University, University
Park, PA, USA
2
Institute of Computational and Data Sciences, The Pennsylvania State University, University Park, PA,
USA
3Technical Service Center, U. Bureau of Reclamation, Denver, CO, USA
Key Points:
• Physics-informed DeepONet framework is developed for computational hydraulics
modeling based on the 2D shallow water equations
• Physics constraint comes with the cost of higher errors for in-distribution cases
because of the inconsistency between discrete and continuous governing equations
• Physics constraint improves the generalization capability of the model for out-ofdistribution cases
Corresponding author: Xiaofeng Liu, xzl123@psu. Based on a real-world case, steady flows in
a reach of the Sacramento River in California, it is demonstrated that PI-SWE-DeepONet
possesses much enhanced prediction capability than SWE-DeepONet when applied to
out-of-distribution scenarios. The physics-informed model is shown to exhibit slower error growth and larger breakdown distances in comparison with SWE-DeepONet. Lai3
1Department of Civil and Environmental Engineering, The Pennsylvania State University, University
Park, PA, USA
2
Institute of Computational and Data Sciences, The Pennsylvania State University, University Park, PA,
USA
3Technical Service Center, U. Bureau of Reclamation, Denver, CO, USA
Key Points:
• Physics-informed DeepONet framework is developed for computational hydraulics
modeling based on the 2D shallow water equations
• Physics constraint comes with the cost of higher errors for in-distribution cases
because of the inconsistency between discrete and continuous governing equations
• Physics constraint improves the generalization capability of the model for out-ofdistribution cases
Corresponding author: Xiaofeng Liu, xzl123@psu.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Discussion on Objectives
Adding PDE residuals into DeepONet loss changes the learning target fundamentally. The competing objectives are fit the training data versus minimize PDE residuals. p. 1 It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution. The
gain of the physics-informed training, however, comes with costs, chief among which are
the simulated results have slightly higher errors for in-distribution cases. It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution. Other challenges are also discussed, such
as the loss weighting approach. These losses are optional and they are omitted for the examples in this work. The total loss function is then expressed as a weighted sum of all loss components:
Ltotal = λdataLdata + λPDEsLPDEs + λBC/IC LBC/IC + λotherLother (6)
where the weights λdata, λPDEs, λBC/IC , and λother are hyperparameters that measure
the relative importance of corresponding loss components.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The choice between SWE-DeepONet and PI-SWE-DeepONet models should be guided by the application requirements. PI-SWE-DeepONet is preferred when predictions are needed for out-of-distribution cases, training data contains relatively large uncertainties, or physical consistency is the priority. p. 1 It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution. It reflects the
existence of a tension between the two competing training objectives: fitting the results
from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on
a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet
is recommended if the modeling objective is to replicate faithfully the traditional hydraulic
model results within the training distribution. This flexibility makes the framework adaptable to diverse
hydraulic modeling scenarios, from constant discharges (steady-state) to complex unsteady
open channel flows with multiple inflow and exit boundaries. In this study, the algorithms and workflow presented are implemented in the opensource Python package HydroNet, which uses PyTorch for deep learning model development and training. A 2D mesh is generated for use by SRH-2D and shown in Figure 3(a); it has 5,036
mixed triangle-quad cells. For the DeepONet surrogate models, collocation points are
additionally needed on which the PDE residuals and loss functions due to boundary conditions are computed. manuscript submitted to
Abstract
Traditional 2D hydraulic models face significant computational challenges that limit their
applications that are time-sensitive or require many model evaluations. This study presents
a physics-informed Deep Operator Network (DeepONet) framework for computational
hydraulics modeling that learns the solution operator of the 2D shallow water equations
(SWEs) to create fast surrogate models.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Physics informed DeepONet framework can be designed to facilitate practical adoption in real-world engineering applications by developing an automated process for all four stages of modeling: generating training data, designing the operator learning architecture, training the model, and testing performance.
-
The improved AI system can perform rapid scenario analysis, uncertainty quantification, risk analyses, and real-time forecasting while maintaining physical consistency required for engineering decision-making.
-
The PI-SWE-DeepONet model can be used for out of distribution predictions or when physical consistency is a priority, as it
exhibit[s] slower error growth and larger breakdown distances in comparison with SWE-DeepONet.
Abstract
Traditional 2D hydraulic models face significant computational challenges that limit their applications that are time-sensitive or require many model evaluations. This study presents a physics-informed Deep Operator Network (DeepONet) framework for computational hydraulics modeling that learns the solution operator of the 2D shallow water equations (SWEs) to create fast surrogate models. The framework can operate in two modes: a purely data-driven SWE-DeepONet that learns from numerical solver such as SRH-2D, and a physics-informed PI-SWE-DeepONet that additionally incorporates the continuous SWEs as constraints during training. Based on a real-world case, steady flows in a reach of the Sacramento River in California, it is demonstrated that PI-SWE-DeepONet possesses much enhanced prediction capability than SWE-DeepONet when applied to out-of-distribution scenarios. The physics-informed model is shown to exhibit slower error growth and larger breakdown distances in comparison with SWE-DeepONet. The gain of the physics-informed training, however, comes with costs, chief among which are the simulated results have slightly higher errors for in-distribution cases. It reflects the existence of a tension between the two competing training objectives: fitting the results from the traditional hydraulic model and satisfying the continuous governing equations. In this study, guidelines are developed for selecting the appropriate approach based on a real-world case: PI-SWE-DeepONet is preferred for out-of-distribution predictions, uncertain training data, or when physical consistency is a priority, while SWE-DeepONet is recommended if the modeling objective is to replicate faithfully the traditional hydraulic model results within the training distribution. Other challenges are also discussed, such as the loss weighting approach.
Sources
- Physics-Informed Neural Operator for Learning Partial Differential Equations
- U-NO: U-shaped Neural Operators
- A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics
Related papers
- Emergent Transfer of a Physics Foundation Model from Simulation to Laboratory Turbulence
- Early onset of secondary shear instability in Kelvin-Helmholtz braids at high Reynolds number
- Rotational covariance, continuous spectrum and non-modal growth of three-dimensional planar weak viscous shock layers
- Deep Learning-based Algebraic Reynolds Stress Closures for RANS Simulations of Turbulent Flows
- Neptuna: A Comprehensive Machine Learning Framework for Benchmarking Complex Multiphase Flows
- Size and shape of condensing steam jets and plumes