Environmental scientists are increasingly using huge models of artificial intelligence to anticipate changes in weather and climate, but new research conducted by researchers MIT shows that larger models are not always better.
The band shows that in some climate scenarios much simpler models based on physics can generate more accurate forecasts than the latest deep learning models.
Their analysis also reveals that a comparative technique commonly used to assess machine learning techniques for climate forecasts can be distorted by natural data changes, such as fluctuations in weather patterns. This can lead someone to the belief that the deep learning model makes more accurate forecasts when it is not.
Scientists have developed a more solid way to assess these techniques, which shows that although simple models are more accurate when estimating regional surface temperatures, deep learning approaches can be the best choice to estimate local rainfall.
They used these results to improve the simulation tool known as A climate emulatorwhich can quickly simulate the impact of human activities on the future climate.
Scientists perceive their work as a “cautionary story” about the risk of implementing large AI models in the field of climate sciences. While deep learning models showed incredible success in domains such as natural language, climate science contains a proven set of physical laws and approximations, and the challenge is to include them in AI models.
“We try to develop models that will be useful and relevant to the types of things that decision -makers need in the future, making elections regarding climate policy. Although it can be attractive to use the latest machine learning model with large images about the climate problem, the study is that backing up and really thinking about the basis of problems is important and useful,” says studies in MIT systems, society, society and society and society. (IDSS) and the Earth Department, atmosphere and planetary sciences (EAPS).
Co -authors Selin are the main author of Björn Lütjens, a former postdoc eaps, who is currently a scientist at IBM Research; Elder author of Raffaele Ferrari, professor of oceanography Cecil and Ida Green in EAPS and co -director Lorez Center; and Duncan Watson-Parris, assistant professor at the University of California in San Diego. Selin and Ferrari are also co -directions Introducing calculations to a climate challenge The project from which these research appeared. . paper appears today in Journal of Advanses in Modeling Earth Systems.
Comparison of emulators
Because the climate of the Earth is so complex, leading the most modern climate model to predict how the levels of pollution will affect environmental factors, such as temperature, may take weeks on the most powerful supercomputers in the world.
Scientists often create climate emulators, simpler approximations of the most modern climate model, which are faster and more accessible. The decision maker could use a climate emulator to see how alternative assumptions regarding greenhouse gas emissions would affect future temperatures, helping them develop recipes.
But the emulator is not very useful if it makes inaccurate forecasts regarding the local effects of climate change. While deep learning has become more and more popular for emulation, little research has been studied whether these models work better than the tried try.
MIT researchers conducted such a study. They compared a traditional technique called linear scaling of patterns (LPS) with a deep learning model using a common comparative data set to evaluate climate emulators.
Their results have shown that LPS exceeded the models of deep learning after anticipating almost all the parameters tested, including temperature and precipitation.
“Large AI methods are very attractive for scientists, but they rarely solve a completely new problem, so first the implementation of an existing solution is necessary to find out whether the complex approach to machine learning is actually improving,” says Lütjens.
Some initial results seemed to fly in the face of scientists' knowledge. The powerful model of deep learning should have been more accurate when making rainfall forecasts, because these data are not in line with the linear formula.
They discovered that the high natural variability in the course of the climate model can cause that the deep learning model will poorly perform unpredictable long -term oscillations, such as El Niño/La niña. This distorts comparative results in favor of LPS, as average these oscillations.
Constructing a new assessment
From there, scientists have constructed a new rating with a greater number of data on natural climate variability. Thanks to this new assessment, the deep learning model worked slightly better than LPS for local rainfall, but LPS was still more accurate in the case of temperature forecasts.
“It is important to use the right modeling tool, but to do this, you must also determine the problem in the right way,” says Selin.
Based on these results, scientists included LPS in the climate emulation platform to predict local temperature changes in various emission scenarios.
“We do not recommend that LPS should always be the target. It still has restrictions. For example, LPS does not provide for variability or extreme weather events,” adds Ferrari.
Rather, they hope that their results emphasize the need to develop better comparative techniques, which can provide a fuller image in which climate emulation technique is best suited for a specific situation.
“Thanks to the improved climate emulation reference points, we could use more complex machine learning methods to test problems that are currently very difficult to solve, as did the influence of aerosols or estimating extreme rainfall,” says Lütjens.
Ultimately, more accurate comparative techniques will help to provide decision -makers by making decisions based on the best available information.
Scientists hope that others are based on their analysis, perhaps by studying additional improvements of climate emulation methods and comparative tests. Such research could examine the impact -oriented indicators, such as drought indicators and the risk of fires, or new variables, such as regional wind speeds.
These studies are partly financed by Schmidt Sciences, LLC, and are part of the MIT Climate Grand Challenge team to “bring calculations to a climate challenge.”
