Knowing whether it will rain in three days is now within the reach of any smartphone, but trying to estimate the weather in a month or a month and a half has always been a bit like flipping a coin. For almost fifty years, modern meteorology has collided with a real “wall”: the chaos theory. Today, a study published on Nature by MIT researchers Judah Cohen and Lester Mackey demonstrates that a new approach based on iartificial intelligence and learning from past errors (post-processing) is capable of double the accuracy forecasts in the sub-seasonal range, also known as the trend or projection period (2 to 6 weeks). An extraordinary step forward for science, which opens up completely new scenarios for global modeling.
How AI makes weather forecasts more reliable: the study on post-processing
To understand the significance of the discovery we must take a small step back to the 1970s. He was the mathematician and meteorologist at MIT Edward Lorenz to develop the concept of chaos theory applied to the atmosphere. Lorenz demonstrated that the atmosphere is a non-linear and non-stationary dynamic system: very small variations in the initial conditions can amplify enormously over time (the famous “butterfly effect”).
Lorenz identified a theoretical barrier placed around the 14 days: Beyond two weeks, the background “noise” of the atmosphere becomes so loud that it overwhelms any initial signals. Entering the range between 2 and 6 weeks, the so-called “desert of predictability” or sub-seasonal scale, meant coming across slow but extremely complex variables (such as El Niño fluctuations or stratospheric oscillations) that caused the precision of traditional models to collapse.
Cohen and Mackey’s intuition is based on a fascinating principle: the atmosphere can also generate an infinite number of configurations, but it naturally tends to “get attached” to certain patterns and repeat them cyclically over time.
Instead of simply running complex physics simulations from scratch, the AI model developed by MIT researchers uses the technique of post-processing:
- Error analysis: the algorithm constantly analyzes the inconsistencies between what physical models predicted in the past and what actually happened in climate history.
- Dynamic Correction: By learning from the systematic errors made by traditional simulations, AI applies real-time corrections to future projections.
Tested in European weather center competitions ECMWF (the world benchmark for modeling), the system has proven to double the accuracy of 2-6 week trends compared to historical climatological averages. A result that offers a very useful margin of notice for agricultural activities, energy resources and medium-long term planning.
Because the human being remains irreplaceable
The evolution of algorithms does not eliminate the figure of the meteorologist, but supports it by enhancing its analytical capacity. Algorithms and calculators are unbeatable in crunching enormous amounts of data and in drawing probabilistic scenarios on a planetary scale, but when the prediction falls on the real level, and in particular in the short term, the the human factor remains the irreplaceable element.
Let’s think about theItalya real puzzle of microclimates and orographic roughness:
- One pnarrow and elongated island immersed in the Mediterranean, between marine basins with different thermal responses.
- Impressive mountain barriers: the Alps to the north and the Apennine chain generate local dynamics that no high-resolution model can fully decipher without careful critical reading.
In the short term (1-3 days), when the real impact of a disturbance on the territory must be established, numerical processing alone is not enough. We need the experience of those who study meteorology and experience it in the field, capable of:
- Master the complex local geography: understand how a specific valley, gulf or mountainside responds to the arrival of certain wind flows.
- Interpret and correct models: the human brain knows how to identify any “mistakes” of the algorithm, contextualizing the dry data on the basis of local historical experience.
- Take care of risk communication: translating numbers, maps and percentages into a clear, calibrated and alarm-free message for citizens and Civil Protection is a scientific and empathetic task that belongs only to people.
Artificial intelligence therefore gives us increasingly sophisticated calculations and models to look far into the future, but the final decision and the real reading of what will happen above our heads will always remain entrusted to the competence of the human being.
