The diversity and fleeting nature of clouds poses challenges
To simulate precipitation, we must go to its source: the clouds. Clouds can exist on scales smaller than 100 metres, the size of an athletic field – much below. kilometer-scale resolution of global weather models, or tens of kilometers-Scale resolution of global climate models. Clouds come in a variety of types, change rapidly, and have complex physics occurring at small scales that can generate water droplets or ice crystals. It is impossible for large-scale models to solve or calculate all this complexity.
To take into account the influence of small-scale atmospheric processes, such as cloud formation, on climate, models use approximations, called parameterizationWhich are based on other variables. Instead of relying on these parameters, NeuralGCM uses a neural network to learn the impacts of such small-scale events directly from existing weather data.
We improved the representation of precipitation in this version of our model by training the ML portion of NeuralGCM directly on satellite-based precipitation observations. NeuralGCM’s initial offering, like most ML weather models, was trained on recreations of past atmospheric conditions, i.e., reanalyzesWhich combines physics-based models with observations to fill gaps in observational data. But the physics of clouds is so complex that even reanalyses struggle to get the precipitation right. Training on the outputs from reanalysis means reproducing their vulnerabilities, for example, on rainfall extremes and the diurnal cycle.
Instead, we directly trained the precipitation part of NeuralGCM nasa satellite based Rainfall observations from 2001 to 2018. neuralgcm differential dynamic core The infrastructure allowed us to train it on satellite observations. Previous hybrid models that combine physics and AI could only use output from high-fidelity simulations or reanalysis data. By training the AI component of NeuralGCM directly on high-quality satellite observations rather than relying on reanalysis, we are effectively finding a better, machine-learned parameterization for precipitation.
