AI Helps to Produce Breakthrough in Weather and Climate Forecasting

 

AI Helps to Produce Breakthrough in Weather and Climate Forecasting

A team of scientists, led by Google, has made significant advancements in weather and climate forecasting by integrating artificial intelligence (AI) with conventional atmospheric physics models. Their innovative model, called NeuralGCM, successfully combines machine learning with established forecasting techniques, enabling it to track long-term climate trends and extreme weather events more accurately and efficiently.

Combining AI with Physics-based Models

NeuralGCM, which leverages both AI and traditional physics-based models, demonstrates how this hybrid approach can enhance the accuracy and speed of climate simulations. According to Stephan Hoyer, a senior staff engineer at Google Research and co-author of the study published in Nature, “NeuralGCM shows that when we combine AI with physics-based models, we can dramatically improve the accuracy and speed of atmospheric climate simulations.”

Superior Performance and Efficiency

In rigorous tests, NeuralGCM outperformed traditional forecasting models. It proved faster and more accurate, while also using less computing power compared to X-SHiELD, a current forecasting model developed by the US National Oceanic and Atmospheric Administration (NOAA). For example, NeuralGCM identified nearly the same number of tropical cyclones as conventional trackers and twice as many as X-SHiELD. In temperature and humidity level predictions for 2020, NeuralGCM’s error rate was significantly lower, ranging between 15% and 50% less.

Additionally, NeuralGCM’s efficiency is remarkable. It was able to generate 70,000 simulation days within 24 hours using Google’s custom AI tensor processing units, whereas X-SHiELD required 13,824 computer units to produce only 19 simulation days.

Collaborative Efforts and Open Access

The development of NeuralGCM was a collaborative effort between Google and the European Centre for Medium-Range Weather Forecasts (ECMWF). The ECMWF made its model publicly available in June, and Google has released the NeuralGCM code as open access, utilizing 80 years of ECMWF observational data and reanalysis for machine learning.

Google’s DeepMind unit had previously unveiled an AI-only weather forecasting model, GraphCast, which outperformed conventional methods for short-term predictions up to 10 days ahead. However, combining AI with physics-based models, as seen with NeuralGCM, provides a more robust and reliable approach for longer-term forecasts.

Expert Opinions and Future Work

Peter Dueben, head of ECMWF’s earth system modelling and a co-author of the paper, emphasized the advantage of this hybrid approach. He noted that while AI-only models often face skepticism for not being based on physics-derived equations, combining them with physics-based models offers the best of both worlds. This approach is seen as a significant step forward in climate modelling with machine learning.

Despite the impressive advancements, there is still more work to be done. For instance, NeuralGCM needs to improve its ability to estimate the impact of CO₂ increases on global surface temperatures and to simulate unprecedented climate conditions.

Cédric M. John, head of data science for the environment and sustainability at Queen Mary University of London, who was not involved in the study, noted the compelling evidence of NeuralGCM’s superior accuracy and efficiency. He highlighted the model’s capacity to capture an ensemble of predictions, allowing for an estimate of the uncertainty of the predictions, which is crucial for practical applications.

Broader Implications

Google’s involvement in environmental surveillance initiatives extends beyond NeuralGCM. The company supports satellite missions to track methane emissions and partners with NASA to help local governments monitor air quality. These efforts underscore the potential of AI in enhancing our understanding and response to environmental challenges.

Conclusion

The successful integration of AI with traditional climate models in NeuralGCM marks a significant breakthrough in weather and climate forecasting. This hybrid approach not only improves accuracy and efficiency but also provides a promising template for using AI in other scientific fields, from materials discovery to engineering design. As climate challenges grow more complex, such innovative solutions are crucial for better predicting and mitigating the impacts of climate change.