Cloud range and ephemeral nature pose challenges
To simulate precipitation, we have to go to the clouds, that are the sources of precipitation. Clouds can exist on scales lower than 100 meters, the dimensions of a taking part in area. That is far under the kilometer-scale decision of worldwide local weather fashions and the tens of kilometer-scale decision of worldwide local weather fashions. Clouds are available many differing kinds and alter quickly. Advanced bodily phenomena that happen on even smaller scales can generate water droplets and ice crystals. It’s inconceivable to resolve or calculate all this complexity in large-scale fashions.
To account for the affect of small-scale atmospheric processes, resembling cloud formation, on the local weather, fashions use approximations known as parameterizations based mostly on different variables. Somewhat than counting on these parameterizations, NeuralGCM makes use of neural networks to be taught the consequences of such small-scale occasions immediately from current climate knowledge.
We improved the illustration of precipitation on this model of the mannequin by coaching the ML a part of NeuralGCM immediately on satellite-based precipitation observations. NeuralGCM’s first product, like most ML climate fashions, was skilled on a recreation of earlier atmospheric situations, a reanalysis that mixes physically-based fashions and observations to fill in gaps in noticed knowledge. Nevertheless, the physics of clouds is extraordinarily complicated, so even reanalysis is tough to precisely decide precipitation quantities. Coaching on the reanalysis output means reproducing weaknesses resembling excessive precipitation or diurnal cycles.
As an alternative, we skilled the precipitation portion of NeuralGCM immediately on NASA satellite-based precipitation observations spanning 2001 to 2018. NeuralGCM’s differential dynamical core infrastructure allowed us to coach on satellite tv for pc observations. Earlier hybrid fashions that mixed physics and AI might solely use output from high-fidelity simulations or reanalysis knowledge. By coaching NeuralGCM’s AI element immediately on high-quality satellite tv for pc observations, somewhat than counting on reanalysis, we will successfully discover higher machine studying parameterizations for precipitation.
