For decades, numerical models have been the backbone of weather forecasting operations worldwide. These models run on high-performance computers and use complex physical equations to represent the behavior of the atmosphere. In recent years, however, a significant shift has taken place, as artificial intelligence has entered the field strongly, offering fast and efficient alternatives in certain areas. This development has raised an important question: will numerical models remain the foundation of forecasting, or are we moving toward a stage where artificial intelligence becomes dominant?
The beginning of the shift: how did AI enter weather forecasting?
The atmospheric field is rich in data: satellites, radars, surface stations, weather balloons, in addition to the vast outputs of numerical models. This enormous volume of data has made the atmosphere an ideal environment for deep learning techniques, which typically perform best when large and diverse datasets are available.
AI-based systems have evolved gradually, starting with improving satellite image quality, moving to correcting statistical errors in numerical models, and now reaching the stage where AI models can generate full atmospheric forecasts across multiple layers, at speeds far exceeding traditional models.
Advantages of AI compared to numerical models
1. High speed
Running a numerical model may take hours, especially for global models, while an AI model can generate a full forecast map within seconds.
2. Handling imperfect and incomplete data
Artificial intelligence can effectively work with imperfect or noisy data, such as inaccurate observations or regions lacking monitoring stations. By learning from large datasets, it identifies general patterns, filters out random errors, and can reasonably estimate missing values without significantly affecting forecast accuracy.
3. Efficiency in resource usage
AI systems do not require expensive supercomputers. Many can run on significantly smaller servers.
4. Competitive accuracy
Some AI-based models have reached levels of accuracy comparable to leading global models for short- and medium-range forecasts.
Why numerical models remain essential
Despite the rapid advancement of artificial intelligence, several factors ensure that numerical models remain fundamental:
1. Lack of explicit physical interpretation
Numerical models directly simulate the laws of physics, whereas AI predicts patterns without explicit physical understanding. Under unusual weather conditions, relying entirely on AI can be risky.
2. Dependence on reference data
AI does not learn from nothing; it depends on data—often generated by numerical models—for training. Any decline in the quality of numerical outputs will directly affect AI performance.
3. Weaker performance in long-range forecasting
Numerical models remain more stable when forecasting periods extend to two weeks or more, where physical equations are more reliable than statistical patterns.
4. Importance in early warning systems
Many global emergency systems rely officially on numerical model outputs, especially for severe storms and cyclones. There is still caution regarding full reliance on AI in such cases.
Can the two complement each other instead of competing?
The current global trend is toward integration rather than replacement.
Artificial intelligence has become a supporting layer that enhances the accuracy of numerical models, reduces their errors, and accelerates the use of their outputs. At the same time, numerical models remain the physical source that provides the core data used to train AI systems.
This integration is already evident in some international centers, where numerical models are run in their original form, and their outputs are then passed to AI systems to improve accuracy or enhance fine-scale details.
How is the role of numerical models evolving?
Based on the current situation and operational experience worldwide, the disappearance of numerical models is not expected in the foreseeable future. What is actually observed is a shift in roles between numerical models and artificial intelligence.
Numerical models will continue to represent the physical foundation for understanding atmospheric behavior, while artificial intelligence acts as a supporting tool that accelerates forecast production and improves accuracy. In other words, AI is not replacing numerical weather prediction, but reshaping how it is used, making forecasting more efficient, faster, and more cost-effective.
References
- Bauer, P., Thorpe, A., & Brunet, G. (2015). The quiet revolution of numerical weather prediction. Nature, 525(7567), 47–55.
- Schultz, M. G. et al. (2021). Can deep learning beat numerical weather prediction? Philosophical Transactions of the Royal Society A, 379.
- Dueben, P. D., & Bauer, P. (2018). Challenges and design choices for global weather and climate models based on machine learning. Geoscientific Model Development.
- Weyn, J. A., Durran, D. R., & Caruana, R. (2019). Can machine learning replace numerical weather prediction? Journal of Advances in Modeling Earth Systems.