Weather forecasting has two halves: estimating the current state (data assimilation) and projecting it forward (prediction). Traditional systems handle these separately — a physics-based assimilation system ingests observations, then a numerical model integrates the equations forward. The assimilation is the harder half, requiring the inversion of a chaotic system from sparse, noisy observations.
FuXiWeather2 unifies both halves in a single neural network. The same architecture that forecasts also assimilates — it learns to produce the atmospheric state estimate that, when forecast forward, best matches future observations. Training aligns the objectives: the analysis is good insofar as it produces good forecasts, not insofar as it matches a reanalysis dataset.
Note (2026-11-08): the abstract (Xu, Sun, Zhu, Zhong and Huang) says FuXiWeather2 aligns its training objectives with a combination of real-world observations and reanalysis data, so that it can rectify errors in reanalysis products, uses recursive unrolling to handle the distribution shift between training backgrounds and self-generated backgrounds, and trains on raw and simulated observations; its analyses surpass NCEP-GFS across most variables and ERA5 and ECMWF-HRES in lower-tropospheric and surface variables, and its forecasts exceed HRES in 91 percent of evaluated metrics. The essay says the analysis is good insofar as it produces good forecasts, 'not insofar as it matches a reanalysis dataset', and that assimilation and prediction share a loss function; the abstract describes the objective as observations plus reanalysis, and does not say the loss is forecast skill alone.
The recursive unrolling method bridges a gap that has plagued learned weather models: the training inputs (reanalysis fields) differ from the operational inputs (self-generated analyses). By training the system on its own outputs iteratively, the model learns to work with the imperfect analyses it will actually encounter in production.
The results: 0.25° global analysis and 10-day forecasts in minutes, surpassing NCEP-GFS analysis and exceeding ECMWF HRES forecasts in 91% of evaluated metrics. Strong typhoon track prediction demonstrates that the learned system captures extreme-event dynamics, not just climatological averages.
The unification is the point. When assimilation and prediction share a loss function, the analysis optimizes for what matters (forecast skill) rather than for a proxy (matching reanalysis). The two halves constrain each other.