Can weather forecasts and artificial intelligence coexist? In what way could artificial intelligence (a human and scientific component) be connected to a meteorological and natural aspect?
When looking at a weather application, for example on a phone, it may seem that predicting the weather simply means understanding the atmospheric conditions of the day in question or how the week is expected to develop; whether it will rain, be sunny, cloudy, and so on.
In reality, behind a weather forecast there is a much more complex process, which describes the evolution of the atmosphere starting from an enormous amount of information.
Temperature, humidity, wind, atmospheric pressure and many other parameters are collected through observations from weather stations, satellites, radiosondes and other detection instruments. These data are used by traditional numerical weather prediction models, which use physical equations to simulate how the state of the atmosphere should change over the course of hours and as the days go by. This is a highly sophisticated method that requires considerable computing power.
In recent years, however, research has begun to explore a different possibility: using AI to learn from past meteorological data in order to produce new forecasts.
The principle is different from that of a standard physical model. In fact, it is based on an artificial intelligence system that is trained using large quantities of data describing the state of the atmosphere at different moments. For example, data relating to variables such as pressure, humidity, temperature and wind are used, collected in different areas of the planet and, for some variables, at different levels of the atmosphere.
These data could therefore be imagined as a sort of “snapshot” of the state of the atmosphere at a given moment. The model receives the information and has to learn to predict what the atmosphere will be like afterwards.
During training, the AI is compared with the values present in the reference data, and the model progressively modifies its parameters in an attempt to reduce the error. By repeating this process on a large scale of examples, the neural network learns the relationships present in the data and how different atmospheric conditions tend to evolve over time.
Once training is complete, the model can receive the atmospheric state at a given moment and can therefore produce a new estimate of its future state. The forecasts can then be used as a starting point for calculating subsequent ones, allowing the system to reach longer time intervals as well.
One of the best-known examples is GraphCast, developed by a group of researchers from Google DeepMind and published in the journal Science in 2023. The system was trained on decades of meteorological data and designed to produce global forecasts for up to 10 days. In the evaluation described in the study, GraphCast achieved better results than the European Centre for Medium-Range Weather Forecasts (ECMWF) deterministic HRES system for most of the variables and time intervals analysed.
A particularly interesting aspect, however, concerns speed. Once trained, GraphCast can produce a ten-day forecast in less than a minute using a dedicated computing unit. Traditional weather models, on the other hand, require high-performance computing systems to carry out their simulations. This does not mean that artificial intelligence has simply “replaced” traditional weather models, but it shows how fast a forecast based on machine learning can be.
GraphCast is not, however, the only system developed in this field. Around the same period, a group of researchers presented Pangu-Weather, published in Nature in 2023. This model also uses artificial intelligence to produce global weather forecasts and was trained using 39 years of data. In the study, its performance was compared with that of the ECMWF operational system, showing better results on the reanalysis data used in the evaluation for all the variables tested, as well as a much higher computational speed.
Another field in which these systems can be useful is that of extreme weather events. Predicting the evolution of a cyclone with greater advance notice, for example, can be important for emergency management and for organising activities that may be affected by an intense atmospheric event. In the case of Pangu-Weather, the researchers also studied the model’s ability to track the trajectory of tropical cyclones. In the analysed sample, consisting of 88 cyclones from 2018, the model achieved lower average position errors than ECMWF-HRES in three- and five-day forecasts.
The interest in these systems does not mean, however, that traditional weather forecasts have become useless. On the contrary, numerical models continue to be a fundamental component of operational meteorology. Artificial intelligence therefore represents a new tool that can work alongside existing methods, offering in some cases very fast forecasts with a high level of accuracy.
An important point is also to understand that artificial intelligence does not “guess” the weather. The model does not possess an intuitive understanding of the atmosphere: it processes data and recognises, through training, relationships present in meteorological evolution. For this reason, it is necessary to verify its forecasts through systematic comparisons with reference data and with weather systems already in use.
Research is also moving towards an increasing integration of different approaches. Models based on artificial intelligence can be extremely fast, while traditional systems directly incorporate the physical laws that describe the atmosphere. The possibility of combining data, physical knowledge and computing power therefore represents one of the most interesting directions in meteorological research.
Artificial intelligence in this field therefore does not mean having a machine capable of knowing with certainty what the weather will be like in ten days. Rather, it means developing new tools capable of analysing enormous quantities of data and rapidly producing forecasts that can be verified. Studies on GraphCast and Pangu-Weather show that this technology has already achieved scientifically measurable results and is opening up new possibilities for making weather forecasts increasingly fast and efficient.
Author: Ms Maria Lagani, psychology graduate- Master’s student. Research Team for JUMP staff (Italy)
Source: Lam, R. et al. (2023). “GraphCast: Learned Global Weather Forecasting”. Science.
Bi, K. et al. (2023). “Accurate medium-range global weather forecasting with 3D neural networks”. Nature, 619, 533–538.
European Centre for Medium-Range Weather Forecasts (ECMWF). Documentazione sui modelli meteorologici basati sul machine learning, inclusi GraphCast e Pangu-Weather