Numerical modeling in hydrology and climate science often requires immense computational resources—especially for high-resolution, long-term simulations. As the demand for fast, accurate, and scalable forecasting grows, AI-based surrogate models offer a promising alternative.

This project explores the use of Neural Operators, particularly the Adaptive Fourier Neural Operator (AFNO) and Deep Operator Network (DON), as efficient surrogates for hydrological modeling. These architectures go beyond conventional deep learning by learning mappings between functional spaces, enabling faster and more generalizable solutions to spatiotemporal problems.

Project Highlights

  • ⚡ Surrogate Modeling for Hydrology: Replaces traditional physics-based models with fast, data-driven approximations.
  • 🧠 Advanced Neural Operators: Implements and compares AFNO and DON with classical LSTM and CNN-based models.
  • 🔬 Physics-Informed Learning: Infuses models with hydrological constraints for improved accuracy and interpretability.
  • 🌍 Spatiotemporal Forecasting: Predicts variables like groundwater head across both time and space.

Tech Stack

  • Frameworks: PyTorch, DeepXDE, Keras, JAX, nvidia-modulus
  • Architectures: LSTM, CNN, AFNO, DON
  • Metrics: NSE, RMSE, spatial distribution error

Key Insights

  • Neural Operators achieved significant speed-ups over numerical models with comparable or better accuracy.
  • AFNO and DON showed strong generalization on unseen regions and hydrological conditions.
  • Physics-informed constraints helped maintain physical realism while reducing overfitting.