JAX-based differentiable hydrological modeling framework implementing FUSE

A fully differentiable JAX implementation of the Framework for Understanding Structural Errors (FUSE) hydrological model from Clark et al. (2008), with Muskingum-Cunge routing.

Note: jFUSE is in active development

Features

• Fully differentiable: Automatic differentiation through the entire model using JAX
• JIT-compiled: Fast execution with XLA compilation
• FUSE decision file compatible: Read standard FUSE decision files to configure model structure
• Muskingum-Cunge routing: Network-based streamflow routing with adaptive parameters
• Gradient-based calibration: Built-in calibration with optax optimizers
• GPU-ready: Seamless scaling to GPU with JAX

Requirements

• Python >= 3.11
• JAX >= 0.4.0
• equinox >= 0.11.0
• optax >= 0.1.7
• numpy
• xarray (for NetCDF I/O)

Installation

# Clone or download the package
git clone https://github.com/symfluence-org/jFUSE.git
cd jfuse

# Install in development mode
pip install -e .

# Or install with all dependencies
pip install -e ".[dev]"

# Or install with PyPi
pip install jfuse          # CPU
pip install jfuse[gpu]     # CUDA

Command Line Interface

jFUSE provides a CLI compatible with FUSE file manager format:

# Run simulation
jfuse run fm_catch.txt bow_at_banff

# Run gradient-based calibration
jfuse run fm_catch.txt bow_at_banff --mode=calib --method=gradient

# Show file manager configuration
jfuse info fm_catch.txt

Download Example Data

We provide a ready-to-use example dataset for jFUSE (lumped and distributed setups).

Download the ZIP release asset and unzip it into a data/ folder:

wget https://github.com/symfluence-org/jFUSE/releases/latest/download/jfuse-example-data.zip
unzip jfuse-example-data.zip
mv jfuse-example-data data

Lumped configuration example

This example calibrates a lumped jFUSE model using ERA5 forcing.

jfuse run \
  data/domain_Bow_at_Banff_lumped_era5/settings/FUSE/fm_catch.txt \
  Bow_at_Banff_lumped_era5 \
  --mode=calib \
  --method=gradient \
  --loss=kge \
  --lr=0.01 \
  --epochs=500 \
  --plot

Distributed configuration example

This example calibrates a distributed jFUSE model using a Muskingum–Cunge routing network.

jfuse run \
  data/domain_Bow_at_Banff_distributed/settings/FUSE/fm_catch.txt \
  Bow_at_Banff_distributed \
  --mode=calib \
  --method=gradient \
  --network=data/domain_Bow_at_Banff_distributed/settings/mizuRoute/topology.nc \
  --obs-file=data/domain_Bow_at_Banff_distributed/observations/streamflow/preprocessed/Bow_at_Banff_distributed_streamflow_processed.csv \
  --loss=kge,nse \
  --lr=0.01 \
  --epochs=1000 \
  --plot

License

GNU General Public License v3.0 (GPL-3.0-only) - see LICENSE file for details.

jFUSE implements the FUSE model of Clark et al. (2008); the original FUSE (NCAR/FUSE) is licensed GPL-3.0, and jFUSE is distributed under the same license.

References

• Clark, M. P., et al. (2008). Framework for Understanding Structural Errors (FUSE). Water Resources Research, 44, W00B02.
• Cunge, J. A. (1969). On the subject of a flood propagation computation method (Muskingum method). Journal of Hydraulic Research, 7(2), 205-230.