Repository navigation
Home
We build and apply differentiable Earth system models — models whose gradients we can take with respect to parameters, inputs, and internal structure.
The organizing question is what scientific questions become newly askable because a model is differentiable?
Three things become possible that were not before:
Learn from data at scale. Parameters, structural choices, and ML closures can all be fit against observations by gradient descent, rather than by hand-tuning or by sampling schemes that scale badly with dimension.
Embed machine learning inside physics. Learned components can be trained while coupled to the physical model, so they see the states the physics actually produces instead of offline snapshots. Online training is the point; offline training is what we already know how to do.
Ask new types of questions. Sensitivity, attribution, and causal structure become directly computable rather than approximated by finite-difference ensembles.
Turn differentiable components into instruments for gradient-based science.
- Iterative solvers
- GPU/TPU performance
- Handling stiffness and discontinuities without destroying the gradient
- Parameter optimization and calibration
- Coupling ML parameterizations into physical models
- Data assimilation
- Sensitivity and attribution analyses
Demonstrate the value of these tools to the broader community.
Concrete target: a differentiable testbed for land–atmosphere interactions — a single-column atmosphere coupled to a multilayer canopy land model.
Small enough to iterate on quickly, complete enough that the coupled problem is real. It gives us a shared object to develop methods against and a credible demonstration for people outside the group.
Test examples and educational tools for the Earth system modeling community.