Physics-informed neural networks as surrogate models of hydrodynamic simulators

Area of Science:

• Environmental Science
• Computational Fluid Dynamics
• Machine Learning

Background:

• Climate change is increasing flood risk, necessitating accurate and rapid flood prediction models.
• Current high-resolution flood simulations are computationally intensive, limiting their practical application.
• Many scientific problems, including flood modeling, face challenges with sparse data, requiring 'small-data' solutions.

Purpose of the Study:

• To develop an efficient machine learning surrogate model for hydrodynamic simulators.
• To address the need for rapid and precise flood prediction in the context of climate change.
• To create a model that performs effectively in 'small-data' scenarios.

Main Methods:

• A novel Physics-Informed Neural Network (PINN) based surrogate model was developed for Shallow Water Equations.
• Physics-based prior information, specifically conservation of mass, was integrated into the neural network architecture.
• The model was demonstrated on high-resolution inland flood and large-scale regional tidal simulations.

Main Results:

• The proposed PINN-based surrogate model demonstrated superior performance compared to existing data-driven methods.
• The model achieved up to a 25% improvement over state-of-the-art data-driven approaches.
• The method effectively incorporated physics-based constraints without requiring continuous derivative calculations in the loss function.

Conclusions:

• Physics-informed approaches offer significant benefits and robustness for surrogate modeling in flood and hydroclimatic studies.
• The developed model provides a computationally efficient and accurate alternative for flood prediction.
• This research highlights the potential of PINNs for advancing climate change adaptation strategies through improved flood risk assessment.