- MIT has developed a new method to enhance the accuracy of climate simulators.
- The simulator can complete scenario tests in seconds, outperforming traditional models.
- Optimized emission curves provide more systemic information.
- The temperature deviation of the simplified model can reach 4.5 degrees Fahrenheit under high warming scenarios.
Researchers at the Massachusetts Institute of Technology have developed a method that trains an artificial intelligence climate simulator on several computer-generated emission scenarios. According to a report in the Proceedings of the National Academy of Sciences (PNAS) and a release from MIT, tests of simplified climate models show that this method can surpass standard training data. The climate simulator is a fast AI model that mimics Earth System Models (ESMs), which are physics-based simulations used for long-term climate forecasting. MIT states that running a scenario on an ESM can take months on a supercomputer cluster, while the simulator can complete it in seconds, allowing scientists to test thousands of scenarios.
New Method Enhances Simulator Efficiency
Simulators are typically trained on six to eight realistic emission pathways, each requiring its own ESM run. The MIT team used simple climate models to build software capable of calculating how small changes in emission inputs affect simulator errors. They then gradually adjusted the initial emission trajectories to reduce that error.
According to the release, the optimized scenarios may appear to bear little resemblance to realistic futures. The team's lead author, Christopher Whiteman, a postdoctoral researcher at ETH Zurich, stated that the team was not overly concerned with the realism of the training scenarios. Emission curves fluctuate rapidly rather than following smooth trajectories, and Whiteman noted that these fluctuations provide the simulator with more information about the entire system. In a preprint, a simulator trained on a single optimized scenario outperformed a simulator trained on six standard pathways when tested against a simple climate model.
The benchmark is scored based on its own training data, which represents its theoretical best scenario.
Testing Results of the Simplified Model
The second test used MIT's Earth System Model, a simplified ESM alternative, focusing solely on carbon dioxide scenarios. Simulators trained on one or two optimized scenarios matched or exceeded the performance of the six-scenario benchmark on other policy and idealized scenarios. The benchmark maintained its best performance on its own training scenarios. The combination of two scenarios improved the average skill on alternative policy scenarios by 32.6%. MIT stated that this training method's simulator also captured local pollution rules when one pollutant decreases while others increase, a test that traditional simulators failed to perform.
The tests utilized simpler models rather than a complete ESM. The release mentioned that users would only run an ESM once, but the preprint listed a full ESM simulation for optimized scenarios as future work, so this saving has yet to be confirmed. It assumes that scenarios optimized on simple models can transfer to the full model. The preprint found that the simple model's temperature deviation under high warming scenarios could reach 4.5 degrees Fahrenheit (approximately 2.5 degrees Celsius) when temperatures stabilize or decline. The combination of two optimized scenarios occasionally weakened the gains of a single pathway, while the simulator is fundamentally a neural network.
The Discover supercomputer from NASA is primarily used for climate forecasting.
AI forecasting has already been present in meteorological work: Google's GraphCast generates 10-day forecasts in under a minute, developers report. The team hopes to test this technology on scenarios generated by a complete ESM and apply it to other climate models and machine learning. Schmidt Science Foundation supported this work through MIT's Grand Challenges. The research was published in PNAS.
Potential Impact of Optimized Emission Scenarios
MIT's research shows that optimizing emission scenarios can significantly enhance the accuracy and efficiency of climate simulators. This method not only allows for the rapid testing of numerous scenarios but also captures complex pollution rules that traditional simulators fail to reflect. As the impacts of climate change become increasingly evident, the development of this technology is crucial for predicting future climate conditions and formulating corresponding policies. Further research may enable this technology to be applied in more complex climate models, thereby improving overall forecasting capabilities.

