ADDA: A MODULAR FRAMEWORK FOR REPRESENTING, SIMULATING AND ASSIMILATING DYNAMICS WITH END-TO-END DIFFERENTIABILITY

How can researchers overcome the high implementation friction, incompatibility between simulation and assimilation codes, and lack of automatic differentiation that currently make it difficult to develop, compare, and apply data assimilation methods across diverse dynamical systems in the geosciences?

The authors introduce ADDA (Automatic Differentiation for Data Assimilation), a flexible Python and PyTorch software framework designed to seamlessly connect simulations, observation operators, and data assimilation algorithms. ADDA abstracts system states to support diverse geometries including regular grids, staggered grids, unstructured meshes, and Lagrangian variables. It provides native first-class support for parallel batched processing on CPUs and GPUs, automatic differentiation for gradient-based variational methods, and interoperability with JAX-based dynamics. To prove its versatility, the framework includes ready-to-use, differentiable implementations of ten distinct dynamical systems spanning various scales and dimensionalities, alongside classical assimilation algorithms such as Kalman filters, smoothers, ensemble methods, and 4D-Var.

ADDA successfully unifies both variational and sequential data assimilation workflows under a single clean interface, removing the traditional bottleneck of manually deriving and maintaining adjoint models for 4D-Var. By leveraging modern automatic differentiation, the framework allows researchers to execute strong- and weak-constraint 4D-Var optimizations out of the box using backpropagation. Furthermore, its modular object-oriented design cleanly separates state representations from observation schemes and solver algorithms, letting users mix and match components without rewriting underlying physical simulators.

Data assimilation is foundational for accurate forecasting in geosciences, hydrology, and climate modeling, but methodological progress has historically been slowed by rigid, application-specific software silos. By providing an open-source, auto-diff-native framework that bridges classical variational techniques, ensemble methods, and modern machine learning approaches, ADDA drastically reduces the engineering overhead required to evaluate new algorithms and conduct fair cross-system comparisons.