Artificial intelligence models suggest that the true boundaries of the “butterfly effect” remain unknown.
The two weeks limit, based on Chaos Theory and in the notions of “Butterfly Effect” of the 1960s, has been transmitted from generation to generation, says Peter Duebenhead of modeling of the terrestrial system of the European Center for Medium term meteorological forecasting.
“It’s basically a divine rule“But even the Gods can make mistakes.
According to, through a meteorological model of artificial intelligence (IA) Developed by Google, atmospheric scientists have found that predictions of one month or more in the future may be possible. “We didn’t find a limit to how much you can go,” he says Trent VonichDoctoral student at the University of Washington (UW) who led the work last month at Arxiv. “First we stayed without memory.”
With powerful computational models, researchers have already achieved significant predictions for about 10 daysapproaching the limit of two weeks. Show that this limit can, in principle, be exceeded “means that AI will be able to do this someday, which is really exciting“, it says Amy McGoverncomputer scientist and meteorologist at the University of Oklahoma.
The article has caveats. On the one hand, it does not make real forecasts beyond two weeks, points out Tobias Selzatmospheric scientist at Ludwig Maximilian University of Munich. So far, UW researchers have tested their AI predictions only with reconstructed instantaneous time.
Moreover, as Selz and his colleagues demonstrated in a 2023 study in the magazine Geophysical Research Lettersthe models of I would ignore the atmospheric processes On a small scale – effects as small as beating a butterfly wings – which are believed to accumulate and drive the predictability limit. “I am very reluctant to use these models to make statements about atmospheric predictability.”
The notion of an intrinsic forecast limit dates back to Edward Lorenzthe famous mathematician and meteorologist at the Massachusetts Institute of Technology (MIT).
In an article of 1963, he pointed out that even a slogan In the representation of the initial state of the atmosphere or a similar chaotic system would eventually cause significant differences in predictions. So, in an article of 1969, he suggested that even if these initial conditions were known almost perfectly, the system would still have a predictability limit driven by the rapid growth of error in small scales.
Lorenz, however, did not really specify a two -week limit. According to a recent study led by Bo-wen ShenMathematician at San Diego State University, Lorenz presented a variety of possible boundariesbut never decided for one.
The number of two weeks came, instead of Jule Charneyof the MIT, and other pioneers who were evaluating the capabilities of the world’s early numerical timely models in the world at the same time.
Shen also notes that Lorenz’s modeling exercise in 1969 was based on Highly sensitive equations Input data, which led Shen to question whether the butterfly effect is an artifact. Anyway, there is no reason to think that the two -week limit is a rule, he says. “It is not a law based on physics. It is a empirical assumption“.
In his new work, Vonich and Hakim were based on Google, a model of AI trained with 40 years of “reanalysis data” – High resolution of the planet climate based on observations and forecasts of short -term models. The pair wanted to see how the graphcast would work if they could radically increase the accuracy of the initial conditions, the initial instantaneous.
They did this by comparing the final state of the atmosphere from the reanalysis data with the graphcast forecasts. Deficiencies in a forecast could then be used to Adjust the initial conditions of reanalysis data which the model used to start its forecast, potentially bringing them closer to the actual state of the atmosphere.
Operational weather models can also be adjusted retroactively in this way as subsequent observations are accumulated.
But the calculations needed to look more than 12 hours in time quickly become overwhelming. The structure of the graphcast, on the other hand, makes these analyzes Easy to execute thousands of times and further back in timeallowing the model to focus on an initial instantaneous almost perfect For the atmosphere, says Hakim. “Basically, they were delivering them to a silver tray.”
With the initial conditions trained, the accuracy of the graphcast for its 10 -day forecast improved at 86% on average – “Absolutely huge” in meteorological terms, says Vonich. Even more surprising, the model showed the ability to predict the time more than 33 days in the future.
At first it was difficult for Hakim to believe, given what he had learned. “It’s almost like a disconnection of reality,” he says. “However, here are the results. It is possible to repeat this calculation.”
The pair also analyzed how the model was changing the initial conditions, fearing that it was doing something unrealistic. They found that the model was making minor adjustments in parameters such as the temperature at large scales. It also seemed to be reinforcing certain wind patterns that traditional meteorological models often mitigate.
Which only shows that there are ways for AI, if you have enough data, overcome the approximations and errors incorporated into traditional models, says Animashree anandkumarComputer Scientist at the California Institute of Technology. “When you discard everything, you have the opportunity to rethink things.”
Selz, however, says there is no evidence that the adjusted initial conditions are really closer to the reality observed in the atmosphere. Adjustments may simply be creating a Ideal starting point for graphcast forecastsin a kind of self-realizable prophecy. If this perfect version is disturbed, Selz suspects that the prolonged forecast window may close again.
“And that’s exactly what you tell us.” Regardless of this, the work is raise many questions about the acquired knowledgesays Dueben, who has always been a little skeptical about the applicability of the butterfly effect to the climate. “It is probably a very limited view to say that they are only small scales that move up and destroy the limits of the forecast,” he says.
This opinion is shared by James DoyleMeteorologist researcher at the Naval Research Laboratory. Lorenz was not wrong to say that small mistakes can proliferate, he says. “But maybe it’s not so critical.”
For now, a A month’s forecast is still an aspirationas it would require a much more refined view of the atmosphere than it is currently possible with satellites and weather balloons. But if the new forecast horizon continues to attract, Doyle says it is not the time to retreat in meteorological research. “This tells us that there is more to gain by taking the models further.”
Teresa Oliveira Campos, Zap //
