Dual-backend (JAX + NumPy) implementation of the Sacramento Soil Moisture Accounting model (SAC-SMA) with optional Snow-17 coupling — usable standalone or as a SYMFLUENCE plugin.
Part of the SYMFLUENCE JAX-native model family — self-contained packages that run standalone (NumPy fallback, no JAX required) and register automatically with SYMFLUENCE when installed alongside it.
- Differentiable: automatic differentiation through the full simulation (JAX)
- Fast: JIT compilation via
lax.scan;vmapfor ensembles; GPU-capable - Dependency-light: pure-NumPy fallback when JAX is not installed
- Plugin architecture: auto-registers with SYMFLUENCE via entry points
pip install jsacsma # NumPy backend
pip install 'jsacsma[jax]' # with JAX (differentiable, JIT)from jsacsma.model import simulate
# coupled Snow-17 + SAC-SMA (temperature drives the snow module)
flow, state = simulate(precip, temp, pet, day_of_year=doy, latitude=51.17)
# standalone SAC-SMA (no snow module)
flow, state = simulate(precip, temp, pet, snow_module="none")The JAX backend makes the full simulation differentiable end-to-end, so model parameters can be calibrated with gradient descent:
import jax
from jsacsma.losses import kge_loss, get_kge_gradient_fn
grad_fn = get_kge_gradient_fn(precip, temp, pet, observed)
value, grads = grad_fn(params) # dKGE/dparam for every parameternse_loss / kge_loss and their gradient factories are JIT-compatible and work
with any optax optimizer. Within SYMFLUENCE the same interface powers the
ADAM and L-BFGS calibration options.
jsacsma registers with SYMFLUENCE
through the symfluence.plugins entry point — installation is the integration:
pip install symfluence jsacsma# config.yaml (excerpt)
model:
hydrological_model: SACSMASYMFLUENCE then handles forcing preparation, calibration, evaluation, and benchmarking for the model with no further wiring.
SAC-SMA (Burnash, 1995) partitions the soil into upper- and lower-zone tension and free water storages, generating direct runoff, surface runoff, interflow, and primary/supplemental baseflow. The optional Snow-17 coupling (Anderson, 2006) converts precipitation to rain-plus-melt before the soil accounting.
16 calibration parameters (jsacsma.PARAM_BOUNDS); snow parameters are handled by
the embedded Snow-17 component.
pip install -e '.[dev]'
pytestIf you use jSACSMA in your research, please cite the SYMFLUENCE companion papers, which describe the design of the JAX-native model family (registry integration, differentiability, and the calibration experiments they enable):
Eythorsson, D., et al. (2026). The registry as social contract: Architectural patterns for community hydrological modeling. Water Resources Research (submitted).
Eythorsson, D., et al. (2026). From configuration to prediction: Multi-model, multi-basin experiments with SYMFLUENCE. Water Resources Research (submitted).
Citation metadata for this package is provided in CITATION.cff;
a version-specific DOI is minted via Zenodo for each GitHub release.
- Burnash, R. J. C. (1995). The NWS River Forecast System — catchment modeling. In V. P. Singh (Ed.), Computer Models of Watershed Hydrology (pp. 311–366). Water Resources Publications.
- Anderson, E. A. (2006). Snow Accumulation and Ablation Model — SNOW-17. NOAA Technical Report NWS HYDRO-17.
Apache-2.0. See LICENSE.