Modern weather prediction relies on powerful models that simulate the atmosphere. To make these forecasts more accurate, data assimilation techniques like 4D-Var are used to combine model output with real-world observations. A key part of this process is building a “tangent linear” or “adjoint” version of the forecast model, which helps understand how small changes in the initial conditions affect the future state of the atmosphere. However, creating these linearized models is extremely difficult—especially for parts of the forecast that involve complex physical processes like clouds or radiation.

This research explores a new approach called Hybrid Tangent Linear Models (HTLMs), which use information from ensembles—many slightly different forecasts run in parallel—to better capture the influence of nonlinear physical processes. HTLMs can reduce errors caused by oversimplifying the model physics and offer a more flexible and scalable alternative to traditional adjoint models.

We apply a new HTLM framework, developed by the Joint Center for Satellite Data Assimilation, to the U.S. weather model (FV3/GFS) using the JEDI software. The results show improved sensitivity estimates and better representation of uncertainty. This approach has the potential to improve operational weather forecasts and extend data assimilation techniques to longer-term climate predictions.