{"id":5193,"date":"2024-02-08T12:20:35","date_gmt":"2024-02-08T11:20:35","guid":{"rendered":"https:\/\/ingenius.ecoledesponts.fr\/?p=5193"},"modified":"2025-07-29T16:04:06","modified_gmt":"2025-07-29T14:04:06","slug":"artificial-intelligence-and-meteorology-we-are-in-the-middle-of-a-revolution","status":"publish","type":"post","link":"https:\/\/ingenius.ecoledesponts.fr\/en\/articles\/artificial-intelligence-and-meteorology-we-are-in-the-middle-of-a-revolution\/","title":{"rendered":"Artificial Intelligence and Meteorology: \u201cWe are in the middle of a revolution.\u201d"},"content":{"rendered":"\n\n\n
\"\"
Ph: Adobe Stock<\/figcaption><\/figure>\n\n\n\n

Since 2020, you have been working with the European Centre for Medium-Range Weather Forecasts (ECMWF). What role do you play in this?<\/strong><\/p>\n\n\n\n

A meteorological center has two main tasks: to produce operational forecasts and to carry out research to improve the accuracy of future forecasts. To do this, we need to develop efficient and accurate algorithms. My research involves refining these algorithms using the data we have available. Since 2019, our teams have been working on combining data assimilation and machine learning 1<\/a><\/sup> to train neural networks to correct forecasting models. In other words, instead of replacing \u2014 as many in the private sector are doing \u2014 a weather forecasting model based on decades of experience, we are proposing to add a statistical model (for example, a neural network) to improve forecast quality.<\/p>\n\n\n\n

Where do you start?<\/strong><\/p>\n\n\n\n

We digitize what exists, discretize it, and translate it mathematically. For example, to map the Earth, or one of its regions, we first define a grid of points, and the number of points determines how precise the discretization will be. Next, we assign a value to each grid point for each variable in our model. To make a forecast, we need to propagate the system in time. We generally differentiate between two types of contribution: large-scale dynamics, such as wind transportation, and small-scale physics, such as cloud formation. The contribution from physics is largely based on \u201csub-mesh parameterizations,\u201d which represent what happens at a smaller scale than the grid\u2019s mesh. In the past, these parameterizations were built and calibrated by meteorologists, based on their expertise.
Despite the best efforts of meteorologists, these parameterizations are still marred by errors. This is where machine learning comes in, by providing techniques for calibrating neural networks to correct model error.<\/p>\n\n\n\n

\"\"

@ Alban Farchi, 2024<\/figcaption><\/figure>\n\n\n\n

Why is correcting a model so complex?<\/strong><\/p>\n\n\n\n

Think about it: an image of the entire Earth with all its variables (temperature, pressure, etc.) takes up several gigabytes of storage space. It is technically very ambitious to calibrate a correction on this basis. It means ingesting a huge amount of data and then making sense of it. Most methods use a 40-year historical database. Multiply that by 365 days and then by the number of images per day, and you have a huge amount of data! Training neural networks is therefore very time-consuming. As researchers, we have to be very realistic about how we can make the most of our limited technological resources.<\/p>\n\n\n\n

How do you go about it?<\/strong><\/p>\n\n\n\n

One solution is to reduce the resolution of the mapping we were talking about, and therefore of our grid. We usually base our grid on something like 32 points in latitude and 64 points in longitude. This gives us a total of just 2,048 points on the Earth. This is a very rough approximation, but as we are working on correcting errors in the model, which is often on a large scale, we manage to depict the targeted phenomena with varying degrees of success.<\/p>\n\n\n\n

Weather forecasts are extremely important. They impact our agriculture, our armed forces, our daily lives, and our economy. Like Google (GraphCast), Microsoft (Climax), Huawei (Pangu-Weather), and NVIDIA (FourCastNet) before them, major multinationals are now investing in this field, relying solely on artificial intelligence. And they claim that their models are at least as accurate as \u2014 and faster than \u2014 the best physics-based models. How true is this?<\/strong><\/p>\n\n\n\n

These new private-sector players are developing their own forecasting models and claiming to compete with the major traditional forecasting centers. While it is true that, in pure performance terms, these models are more accurate than physics-based models, they tend to produce forecasts that are not always consistent with physics, and are therefore generally unstable over the long term. However, there is no denying that these new models are several orders of magnitude faster. Whereas up to now, we have used CPUs (central processing units), these companies are using an architecture based on graphical supercomputers such as GPUs (graphics processing units), and these are capable of much faster matrix operations. But the method chosen can also explain the difference in computation time. The statistician\u2019s method is designed around the advantages from using GPUs, while the physicist\u2019s is based on model equations and numerical schemes, which generally involve more elementary operations.<\/p>\n\n\n\n

How long have you been wondering what AI could bring to meteorology?<\/strong><\/p>\n\n\n\n

Five years ago, most of the scientific community was frankly skeptical about the benefits of AI, as was the ECMWF, for example, which published a comparison between physics-based and statistical model<\/a>s. At that time, there was no argument: statistical model implementations were still relatively unsophisticated and a long way from matching the performance of existing meteorological models. But it was not forecasting specialists who were leading these early \u201cmachine learning\u201d initiatives, and international companies subsequently recruited meteorologists to help them design suitable architectures that performed efficiently.
Then, at the beginning of 2022, NVIDIA launched FourCastNet (short for Fourier Forecasting Neural Network) and, for the first time, a private-sector team published a model that attracted a great deal of attention. Despite all the reservations, here was a large-scale model that produced more or less realistic images with very low computation times. It was a turning point.<\/p>\n\n\n\n

In light of these new developments, should researchers and forecasting centers adopt a more proactive strategy towards artificial intelligence techniques?<\/strong><\/p>\n\n\n\n

To give you an example, the ECMWF has significantly overhauled its organization in recent months, and two approaches are currently being explored. The first involves adopting the methods developed by the big players in the private sector, reproducing and even improving their models to provide an objective comparison with the traditional model. The second approach, the one we are working on, aims to create a hybrid model, combining AI and traditional models \u2014 in other words, statistics and physics \u2014 to achieve greater accuracy. In our view, this is the most promising approach, as systems based purely on neural networks also have their weaknesses. For example, they tend to smooth forecasts over longer time frames, providing less resolution than they do over shorter time frames. The physics-based model, on the other hand, maintains its resolution whatever the time frame, because it has been calibrated to do so.<\/p>\n\n\n\n

How is the engineering profession being affected?<\/strong><\/p>\n\n\n\n

Its center of gravity is shifting. It is no longer purely physics-based. It also needs to be able to build a neural network and have it interact with the physics-based model. But we should not forget that, ideally, it is still meteorologists, with their expertise, who have the legitimacy to decide which statistical correction to apply. With hybrid methods, they will know how to move the slider to balance the model, and with purely statistical methods, they will know how to explain what the neural network should produce.<\/p>\n\n\n\n

And what does tomorrow have in store for weather forecasting?<\/strong><\/p>\n\n\n\n

In 2030, will the work we have done still be in use, or will it be considered totally obsolete? It is hard to say. Currently, a transition is taking place and two camps are battling it out: hybrid models versus purely statistical models. And who knows whether the latter will dominate in the years to come. But be aware that, until now, we have only been talking about the forecasting aspect. But to make a forecast, you first need to assimilate the observations collected to calibrate the model. For the time being, neural network models are not designed for this task. This means that hybrid models currently have an advantage. But things are moving fast, and we are in the middle of a revolution. There are already teams working on the use of neural networks for data assimilation. Observations are, by their very nature, \u201cunstructured,\u201d and making them readable by a network is not an easy task. But five years ago, we were just as skeptical about forecasting. At any rate, this is a subject that our teams will no doubt be working on very soon.<\/p>\n\n\n\n

Interview by St\u00e9phan Lemonsu.<\/p>\n\n\n

  1. In the field of meteorology, artificial intelligence primarily takes the form of machine learning (or statistical learning), and in particular deep learning, in other words, neural networks. A statistical model is a model that can be trained by machine learning, and is typically a neural network. \u21a9\ufe0e<\/a><\/li><\/ol>\n\n\n
    \n