Old Tools Outshine AI in Climate Forecasting Race todayheadline

A simple truth is rattling the climate science world. Sometimes the old models work better than the new.

MIT researchers report that simple physics-based models can outperform massive deep-learning systems at predicting future regional temperature shifts. The finding, published August 26 in the Journal of Advances in Modeling Earth Systems, challenges assumptions that artificial intelligence will always be the sharper tool. The team also showed that the natural swings of climate—like El Niño or random weather noise—can make AI systems look better than they really are. In reality, linear models were often more accurate for temperature, while deep learning had an edge in some rainfall forecasts.

Why Simpler Can Be Better

Climate prediction is a monster of a problem. Running a full-scale Earth system model to see how pollution affects temperature can take weeks on the world’s fastest supercomputers. To make this manageable, researchers often use “climate emulators,” stripped-down models that approximate the big simulations. These are what policymakers depend on when drafting emission targets or climate regulations.

But what if those emulators misfire? The MIT team set out to test. They compared linear pattern scaling (LPS), a decades-old statistical method, with a cutting-edge deep-learning model. Using a benchmark dataset, they found LPS beat the AI in predicting most parameters—especially temperature.

“Large AI methods are very appealing to scientists, but they rarely solve a completely new problem,” said lead author Björn Lütjens of IBM Research.

At first glance, the results were puzzling. Precipitation is messy and nonlinear, so deep learning should have excelled. But the researchers discovered that natural variability—the chaotic ups and downs of climate—was tricking the AI into overfitting. Essentially, it was memorizing the noise instead of capturing the signal.

Building a Fairer Test

To correct this, the team built a new evaluation system using many more climate simulations. When they did, the deep-learning model improved, edging out LPS in precipitation forecasts. Still, LPS remained more accurate for surface temperature.

The lesson? Benchmarking methods matter. “It is important to use the modeling tool that is right for the problem, but in order to do that you also have to set up the problem the right way in the first place,” said MIT professor Noelle Selin, a co-author and director of the Center for Sustainability Science and Strategy.

“We are trying to develop models that are going to be useful and relevant for decision-makers,” Selin said.

Implications for Policy and Science

The study is not a dismissal of AI but a cautionary tale. Climate science already has strong physical laws to lean on. The challenge is blending those laws with machine learning in a way that avoids overfitting and misuse.

Key findings include:

• Linear regression models outperformed deep learning on 3 of 4 key climate variables, including temperature.
• Deep learning struggled with internal variability, often overfitting to climate noise like El Niño/La Niña cycles.
• With more data, deep learning gained an edge for local precipitation forecasts but not for temperature.

For policymakers, this means caution is warranted before trusting the biggest AI models to drive climate planning. A simpler emulator may give a truer picture of how emissions shift local temperatures. At the same time, deep learning remains promising for thornier questions such as extreme rainfall or aerosol impacts.

Takeaway

This MIT study shows that simpler, physics-based models can outperform large AI systems in predicting local temperature changes, while deep learning may be more effective for rainfall. The results highlight the importance of better benchmarks to guide which modeling approach to use.

The next step, the authors argue, is developing improved climate emulation tools that combine the strengths of both approaches—helping decision-makers face a future where the stakes could not be higher.

Journal: Journal of Advances in Modeling Earth Systems
DOI: 10.1029/2024MS004619