NeuralGCM harnesses AI to better simulate long-range global precipitation
Google Research
Google trained its NeuralGCM AI weather model directly on satellite rain data instead of messy reanalysis data. Result: it beats top forecasting and climate models at predicting rain, especially the extreme downpours everyone struggles to model.
Based on reporting by Google Research — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Rain has always been the stubborn problem child of weather modeling. Temperature behaves. Wind behaves, mostly. But precipitation depends on clouds forming and collapsing at scales as small as 100 meters, well below anything a global model can actually resolve, so forecasters have relied on rough approximations for decades. Google Research's latest update to NeuralGCM tries to fix that by changing what the model learns from in the first place.
The original NeuralGCM, released last year, blended physics with machine learning to simulate the atmosphere and did a solid job on temperature forecasts and multi-decade climate runs. But like most ML weather models, it trained on reanalysis data, essentially reconstructed history that patches real observations together with physics-based guesses. The trouble is reanalysis inherits the same blind spots as the models that built it, particularly around extreme rainfall and the timing of daily storms. So this time, the team fed NeuralGCM raw NASA satellite precipitation observations from 2001 to 2018, letting the network learn cloud behavior straight from measured reality instead of a proxy for it.
The payoff shows up clearly in testing. Against ECMWF's leading forecasting model over all of 2020, NeuralGCM beat it across most precipitation metrics for the full 15-day forecast window, including over land, where accuracy matters most for farmers and flood planners. On longer, multi-decade climate runs, NeuralGCM cut average precipitation error by roughly 40% compared with the models used in the latest IPCC report, with the biggest gains again over land. It also handled the
My take — AI-written commentary, not fact-checked reporting
I'll say the obvious thing nobody wants to hear: 280km resolution is still too coarse to help the person standing in a flooded street tomorrow, so treat this as a climate-science win, not an operational forecasting one yet. What I do like is that Google keeps open-sourcing this stuff instead of locking it behind an API, and the India monsoon pilot is exactly the kind of real-world proof that separates useful AI science from another benchmark flex.
Read more about this at: Google Research