It was reported that Google research’s team NeuralGCM is going to develop hybrid weather and climate models that will combine scientific computing features with AI. Moreover, recently the team has working on publishing Nature, they introduced an atmospheric model that blends a differentiable numerical solver for large-scale atmospheric circulations with neural networks to represent complex processes, according the Research Engineer Dmitrii Kochkov.
Also, he added, this research aims to fuse traditional physics-based models and data-driven approaches. However, by using numerical methods for well-understood phenomena and ML for hard-to-parameterize processes, they aim to develop models that not only excel at short-term forecasting but also maintain physical consistency at much longer timescales. Additionally, the weather forecasts have trained on 3 to 5 days, and their models demonstrate skills comparable to existing physics-based and ML approaches in medium-range weather forecasting.
Meanwhile, it was noted that their models produce realistic forecasts at much longer lead times extending from months to decades. In year-long simulations with prescribed sea surface temperatures, we observed emergent phenomena such as accurate tropical cyclone behavior and seasonal cycles.
Importantly, Dmitrii Kochkov explained, that their models have achieved a level of performance with a fraction of the computational cost of traditional high-resolution models. It was believed that the hybrid models have the potential to improve weather forecasting and understanding of the impact of changing climate. Moreover, the high level of accurate medium range in forecasts could lead to better decision-making and reduced losses due to weather-related disruption.
Accurately, the multiyear simulations could enhance our understanding of weather variability in the current climate, aiding in climate adaptation strategies. Further, he described their next step is to apply the approach to even more complex Earth system models. This will allow them to tackle longer-term simulation challenges like sub-seasonal forecasting and provide a valuable understanding of weather variability in the changing climate.
The words shared by the research team of Google, NeuralGCM, with an official video update, showcased their process and upcoming development of hybrid AI and differential approaches in solving atmospheric modeling.