A new review article published in Engineering outlines how the integration of artificial intelligence with advanced data assimilation techniques could enable weather forecasting at resolutions of several kilometers or even hundreds of meters—fine enough to resolve individual clouds, internal gravity waves, and potentially tornadoes.

The paper, authored by Xiaolei Zou of Nanjing University of Information Science & Technology, examines both the opportunities and challenges in modernizing atmospheric data assimilation, the mathematical process that combines observational data with numerical models to produce optimal estimates of atmospheric states. These estimates serve as initial conditions for numerical weather predictions—crucial in both meteorological services and research—while multidecadal reanalysis datasets form a backbone for weather and climate studies.

Traditional data assimilation has relied heavily on variational methods, particularly 3D-Var and 4D-Var systems. The 4D-Var approach, which assimilates observational data over a period of time using adjoint techniques, represents an optimal control method that prevents internal dynamic imbalances that can be detrimental to forecasts. However, the computational demands of these systems have historically constrained their resolution.

The review highlights two focal points for advancement: exploiting satellite-observed cloud and rainband structures in tropical cyclones for high-resolution assimilation, and re-evaluating core data assimilation techniques that may need improvement. Tropical cyclones present particular challenges because their intensity changes are affected by rapidly varying fine structural changes of clouds and precipitation within the hurricane, including the formation of secondary eyewalls that cause large oscillatory intensity changes.

Current global analyses at approximately 30 km horizontal resolution cannot adequately describe the vorticity and cloud and rainband characteristics of tropical cyclones. High-resolution data assimilation at 3 km or higher may achieve this, but requires abandoning some old practices. The paper identifies several traditional approaches needing reconsideration: the limitations of synchronizing global data assimilation with fixed coordinated universal time intervals, which creates data-void areas when polar-orbiting environmental satellites have not yet passed; mismatched grid configurations between data assimilation and forecasting models; and excessive data thinning and quality control that discards valuable small-scale information.

A single advanced technology microwave sounder scan line consists of 96 fields of view, yet typical thinning strategies retain only about 3.3% of these data. Similarly, geostationary satellite imager data are often thinned to 40–60 km resolutions, with less than about 6%–10% of original data surviving quality control. Such practices result in loss of valuable cloud-structure information that could improve forecasts.

The integration of deep learning offers pathways forward. Preliminary successes in large-scale medium-range weather forecasts have been demonstrated by deep-learning techniques using the ERA5 reanalysis dataset as training data. However, this 0.25° × 0.25° horizontal resolution dataset with 137 vertical levels limits current deep-learning models from properly forecasting extreme weather events or small-scale phenomena.

Looking ahead, the author envisions a 4D-Var AI system with AI physical parameterization models. In this framework, high-resolution numerical weather forecasting models would generate comprehensive datasets containing all scenarios of the targeted physical process for developing AI models to replace traditional physical parameterization schemes. The tangent-linear and adjoint operators of these AI models can be easily differentiated and demand much less computational cost than conventional schemes.

The review emphasizes that data assimilation and meteorological deep learning will inevitably intertwine and reinforce each other. Higher resolution data assimilation will abandon more legacy methods; frameworks will share consistent grid settings across all resolutions; all resolutions will employ deterministic dynamic cores with deep-learning physical parameterization models; and analysis fields generated over sufficiently long periods will accurately describe all relevant weather and serve as training datasets for various deep-learning models. This mutual reinforcement could ultimately enable real-time forecasts of various weather systems to reach users within a specific time limit, better meeting societal needs.

Data assimilation has always been limited by computing power, and will continue to be so in the future. Although powerful, AI’s creation involves the rapid reorganization of available information. Unconventional or even contradictory thinking, assisted with the AI technology, is more likely to give birth to truly original high-resolution data assimilation systems. Only with a thoughtful strategy and a series of incremental steps will AI data assimilation systems surpass and replace current data assimilation practices.

The paper “Traditional Data Assimilation and Its Effective Integration with Meteorological Deep Learning,” is authored by Xiaolei Zou. Full text of the open access paper: https://doi.org/10.1016/j.eng.2025.11.023. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.