PNNL

The Science 

Predicting floods that arise from the combined effects of heavy rain, tides, and storm surge, known as compound flooding, is difficult for large-scale Earth system models (ESMs). Researchers developed and tested two types of machine learning approaches to address this: physics-informed neural networks (PINNs) that follow physical laws and data-driven models that learn directly from past events. The study introduces a finite-difference PINN that accelerates conventional PINNs’ training by ~6.5 times while improving accuracy and evaluates hybrid convolutional-recurrent models (e.g., CNN-LSTM) that capture complex flood dynamics with high fidelity. Together, these approaches show potential for integration into ESMs to represent flood behavior at river-ocean interfaces where river flow interacts with tides and storm surge.

The Impact 

This research introduces a new machine learning method that makes PINNs over six-times faster, removing a key barrier to their practical use in ESMs. By systematically comparing physics-informed and data-driven neural networks, the study identifies where each excels and how they can be combined in a hybrid, AI-enhanced modeling framework. The work provides a foundation for integrating machine learning into large-scale ESMs to better capture local flood dynamics, improve prediction efficiency at higher resolutions, and accelerate scientific discovery in complex Earth system processes. This work demonstrates Pacific Northwest National Laboratory’s leadership in combining process-aware diagnostics with machine learning surrogates to target difficult coupling zones.

Summary 

Simulating compound flooding caused by the combined effects of river discharge, tides, and storm surges remains challenging for ESMs because these processes interact across different scales, while AI-based models have shown great potential to capture complex dynamics. This study compares two advanced machine learning strategies to improve how ESMs capture local compound flooding. The first, a PINN, learns by enforcing physical laws, while the second uses data-driven architectures that learn directly from historical floods. Researchers developed finite-difference PINN, a new PINN variant that runs about six-times faster than conventional PINNs while achieving improved accuracy. They also introduced a new data-generation method that samples historical flood and surge events to train robust data-driven models for predicting extreme conditions. Using a river-ocean test case based on the Delaware River and Hurricane Irene, the study shows that finite-difference PINN offers efficient, physically consistent results and that hybrid data-driven models, when combining convolutional and recurrent neural networks, best balance accuracy and computational cost. Together, these results outline a practical, hybrid pathway to embed machine learning surrogates at river-ocean interfaces in ESMs.

Contact 

Renu R. Joseph, Earth System Model Development program area, renu.joseph@science.doe.gov

Zeli Tan, Pacific Northwest National Laboratory, zeli.tan@pnnl.gov

Funding 

This work was supported by the Earth System Model Development program area within the Department of Energy, Office of Science, Biological and Environmental Research program as part of the multiprogram, collaborative Integrated Coastal Modeling project.