The Diversity and Ephemeral Nature of Clouds Pose Challenges
To simulate precipitation, you have to go back to its source: the clouds. Clouds can exist at scales smaller than
100 meters, the size of a sports field – well below the kilometer-scale resolution of global weather models, or the
tens of kilometers scale of global climate models. Clouds come in different types, change rapidly, and the complex
physics that occur at even smaller scales can generate water droplets or ice crystals. All this complexity is
impossible to solve or calculate for large-scale models.
To account for the effect of small-scale atmospheric processes like cloud formation on climate, models use
approximations, called parameterizations, based on other variables. Rather than relying on these settings,
NeuralGCM uses a neural network to learn the effects of these small-scale events directly from existing weather
data.
We improved the precipitation representation in this version of our model by training the ML part of NeuralGCM
directly on satellite precipitation observations. NeuralGCM’s initial offering was, like most ML weather models,
trained on recreations of previous atmospheric conditions, i.e. reanalyses, which combine physics-based models with
observations to fill in gaps in observational data. But cloud physics is so complex that even reanalyses struggle to
correctly determine precipitation. Training on the results of reanalyses amounts to reproducing their weaknesses,
for example on extreme precipitation and the daily cycle.
Instead, we trained the precipitation portion of NeuralGCM directly on NASA satellite precipitation observations
spanning from 2001 to 2018. NeuralGCM’s differential dynamics core infrastructure allowed us to train it on
satellite observations. Previous hybrid models combining physics and AI could only use results from high-fidelity
simulations or reanalysis data. By training the AI component of NeuralGCM directly on high-quality satellite
observations instead of relying on reanalyses, we actually find better, machine-learned precipitation
parameterization.
“`
