Weather forecasting is a continuous race against time, one that has been ongoing since its inception. At present scientists are engaged in the collection of data using satellites, weather stations, ocean sensors, and aircrafts to feed them into complex models from which an estimation of the expected behaviour of the atmosphere is carried out. With the introduction of artificial intelligence into the weather forecasting systems, the process of forecasting is expected to be changed significantly. The new machine learning technology will be used not only to replace the traditional forecasting techniques but also to come up with new methods of identifying and analyzing atmospheric changes, forecasting atmospheric conditions faster, and improving the quality of all the information regarding storm situations.
Traditionally, meteorology has been based upon numerical weather prediction whereby equations can be used to describe air moisture, pressure, and temperature moves through the atmosphere. Such systems are still very effective when it comes to highly accurate forecasting in cases when the importance of physical processes is very high. However, they are very expensive in terms of calculations. Machine learning models suggest a different approach to the problem since instead of using equations, they learn from available historical data.
Meteorology Enters a New Forecasting Era
The increasing role of AI in the field of meteorology can be observed notably in the activities of the European Centre for Medium-Range Weather Forecasts. In February 2025, the ECMWF introduced its Artificial Intelligence Forecasting System – AIFS – which will be used alongside the world-famous Integrated Forecasting System based on physics principles. ECMWF representatives stated that AI would save about 1,000 times more energy while completing forecast tasks and receiving better results in many aspects, including the ability to forecast the development of cyclones.
This doesn’t mean that the classical model has lost its relevance overnight. On the contrary, the ECMWF deliberately keeps utilizing both models to understand the prospects of forecasting. The future of forecasting suggests a cooperation between physics and AI, where machine learning will take its part in solving some questions while physics continues to provide the capabilities that AI hasn’t yet delivered.
The resolution of models is one of the points worth discussing. For instance, when talking about the AIFS, ECMWF stated its model operates on the resolution equal to 31 kilometers, while its physics-based model has the resolution of 9 kilometers.
Why Speed Matters When Storms Are Moving
The most remarkable benefit of forecasts created by artificial intelligence may not be that a machine makes predictions but that machine-learning systems can provide a vast number of forecasts in a limited period of time.
This is critical for meteorologists as the nature of the atmosphere changes every second. Based on this system, ECMWF introduced AIFS in July 2025. This consists of 51 forecasts that differ due to the parameters of the weather. Its purpose is the creation of the forecasts that contain the element of uncertainty thus providing meteorologists with more potential results instead of errors caused by misleading deterministic forecasts.
The significance of uncertainty becomes apparent while investigating hurricanes and other dangerous phenomena. The weather can be changed noticeably even by a slight fluctuation in the direction of the storm but this can still mean that large areas are threatened by strong winds or they will manage to avoid the bad weather. Thanks to the high speed of such simulation systems every day meteorologists can provide several options and alter the risks they give to society.
The Problem With Letting AI Guess Too Freely
Despite the excitement about the role of AI in meteorology, one must tread carefully. The truth is that machine learning models can be very accurate without understanding the atmosphere like a trained meteorologist. These models identify statistics from the historical data. This is powerful but creates some issues due to unorthodox situations, data constraints, and the physical consistency that exists.
The second problem relates to earlier machine learning models which often produce oversmoothed atmosphere fields, eliminating small variability. The European Centre for Medium-Range Weather Forecasts has pointed this drawback out many times regarding the first generation of machine-learning models.
This is when the significance of probabilistic forecasts emerges. One should realize that weather forecasting should not merely say, “This is what will happen.” Often the more helpful question is, “What are the possible outcomes and what is their probability?” GenCast attempts to provide an answer by developing portfolios of forecasting instead of producing the only one deterministic prediction.
Meteorologists Are Still Central to the Process
The emergence of AI does not imply that human forecasting has lost its relevance. Rather, it seems to enhance the importance of human expertise. A machine-learning model can generate thousands of possibilities in no time, but humans must analyse those results and validate if they make sense, relate them to observations, and communicate them to the public.
This human role is especially critical in emergencies. A forecast becomes useless if it has just been made with minimal errors. People want to know if they need to evacuate, if any outdoor event should be cancelled, or if they should ppotect their harvests or move their machinery.
The shift in meteorology thus appears less to be about the emergence of automation replacing human capabilities, and more like the role of scientists changing from that of forecasters to that of trusted partners who are capable of performing advanced analyses on an enormous scale.
What Comes Next
The subsequent phase of meteorology which is guided by AI seems to be that models are supposed to be more advanced, more probabilistic and connected to the existing infrastructure of traditional meteorology. The development of AIFS as well as Anemoi framework by ECMWF is revealing how fast this industry is moving from labs into real forecasting.
There is also a growing interest in forecasts that go beyond conventional medium-term forecasts. The use of AI Weather Quest that was introduced by ECMWF was meant to study machine learning approaches in dealing with sub-seasonal forecasts, a very challenging issue that might be useful for agriculture and energy management.
Most probably, the most important change will be philosophical. In meteorology, computations have been associated with sophisticated computing power that has based on very advanced physical simulations. Now, having the use of AI can offer another approach of learning the functions of the atmosphere by means of massive computations.
Nonetheless, that does not mean that physics has lost its relevance. It means that forecasting toolbox has become more versatile. Now that extreme weather is becoming one of the leading issues all around the world, that approach might be of great value.
