Climate-PINN: Physics-Informed Neural Networks for Climate Modeling

Project Overview

This repository implements a Physics-Informed Neural Network (PINN) approach to climate modeling, combining deep learning with physical constraints from fluid dynamics. The model is trained on ERA5 climate reanalysis data and predicts near-surface temperature and wind patterns while respecting fundamental physics principles derived from the Navier-Stokes equations.

By incorporating physics-based constraints, the model maintains physical consistency while delivering accurate predictions, even in complex climate scenarios.

Key Features

Physics-Informed Architecture

Neural network incorporating Navier-Stokes equations as physics constraints to ensure physically consistent predictions that respect fluid dynamics principles.

Multiple Model Variants

Several model architectures with progressive improvements, from basic implementation to advanced neural network-based Reynolds number estimation.

ERA5 Dataset Integration

Training on comprehensive climate reanalysis data with multiple variables including geopotential, temperature, wind patterns, and geographical constraints.

Advanced Visualization Tools

Generate animated temperature and wind field maps with prediction analysis, error visualization, and uncertainty quantification.

Model Architecture

The Climate-PINN consists of several key components:

MeteoEncoder: Processes meteorological input variables (geopotential and temperature)
MaskEncoder: Handles geographical masks (orography, land-sea mask, soil type)
CoordProcessor: Processes spatial (lat/lon) and temporal coordinates
Physics Constraints: Incorporates Navier-Stokes equations with learnable Reynolds number

The repository includes several model variants with progressive improvements:

model_0: Original baseline model
model_0_Re: Enhanced with clipped gradient and momentum on the Reynolds number
model_1: Model with modified dropout placement
model_2: Model with correctly placed dropout and improved Reynolds number handling
model_3: Advanced model with a neural network for Reynolds number estimation

Learn more about model variants

Getting Started

Installation

git clone https://github.com/enzolvd/PINN_Climate.git
cd PINN_Climate
conda create -n climate_pinn python=3.12
conda activate climate_pinn
pip install -r requirements.txt

Training a Model

python train.py \
    --model=model_2 \
    --experiment_name=climate_run_1 \
    --hidden_dim=64 \
    --initial_re=100.0 \
    --nb_years=10 \
    --train_val_split=0.8 \
    --physics_weight=0.5 \
    --data_weight=1.0

Generating Visualizations

python video_gen.py   # For animated visualizations
python image_gen.py   # For static comparison images

View full documentation

Citation

If you use this code in your research, please cite:

@misc{louvard2025climatepinn,
  author = {Louvard, Enzo},
  title = {Climate-PINN: Physics-Informed Neural Networks for Climate Modeling},
  year = {2025},
  publisher = {GitHub},
  journal = {GitHub Repository},
  howpublished = {\url{https://github.com/enzolvd/PINN_Climate}}
}