cfuse 0.4.1

Differentiable FUSE hydrological model with Enzyme AD

cFUSE - Differentiable FUSE

A differentiable implementation of the FUSE hydrological model framework with Enzyme automatic differentiation.

Note cFUSE is in active development

Features

• Differentiable physics: Full gradient computation via Enzyme AD
• PyTorch integration: Custom autograd function for gradient-based calibration

Quick Start

1. Install Dependencies

# Python dependencies
pip install numpy torch netCDF4 tqdm matplotlib

# macOS: Install LLVM 19 for Enzyme
brew install llvm@19

2. Build Enzyme (one-time setup)

git clone https://github.com/EnzymeAD/Enzyme.git
cd Enzyme && mkdir build && cd build
cmake ../enzyme -DLLVM_DIR=$(brew --prefix llvm@19)/lib/cmake/llvm
make -j
sudo cp Enzyme/ClangEnzyme-19.dylib /opt/homebrew/lib/

3. Build cFUSE

cd cFUSE
mkdir build && cd build
cmake .. -DCMAKE_BUILD_TYPE=Release \
    -DDFUSE_BUILD_PYTHON=ON \
    -DDFUSE_USE_NETCDF=ON \
    -DDFUSE_USE_ENZYME=ON \
    -DCMAKE_CXX_COMPILER=$(brew --prefix llvm@19)/bin/clang++
make -j
cp cfuse_core*.so ../python/

3b. Build Python package with Enzyme + NetCDF (optional)

# Use local Enzyme plugin and enable NetCDF in the build
CMAKE_ARGS="-DDFUSE_USE_ENZYME=ON -DDFUSE_USE_NETCDF=ON -DENZYME_PLUGIN=/path/to/ClangEnzyme-19.dylib" \
    python -m build --no-isolation --skip-dependency-check

4. Run Example Optimization

cd path/to/cFUSE/python
python optimize_basin.py

Command Line Options

python optimize_basin.py --help

# Examples:
python optimize_basin.py --iterations 700          # More iterations
python optimize_basin.py --lr 0.05                 # Lower learning rate
python optimize_basin.py --loss nse                # Optimize NSE instead of KGE
python optimize_basin.py --spinup-days 730         # 2-year spinup

Test Data

Example data is distributed as a release asset starting with v0.4.0. Download the Bow-at-Banff dataset from https://github.com/DarriEy/cFUSE/releases/tag/v0.4.0 and extract it so the path is data/domain_Bow_at_Banff_lumped_era5/ before running examples.

• Basin: Bow River at Banff (2210 km²)
• Forcing: ERA5 reanalysis (precipitation, temperature, PET)
• Observations: Streamflow for validation

Project Structure

cFUSE/
├── data/                    # Downloaded example data (release asset)
├── include/cfuse/          # C++ headers
├── python/
│   ├── optimize_basin.py   # Example optimization script
│   └── cfuse/              # Python package
├── build/                  # Build output
└── README.md

License

MIT License

Acknowledgments

• Original FUSE: Clark et al. (2008), Water Resources Research
• Enzyme AD: Moses & Churavy (2020), NeurIPS