Wednesday, 9 January 2019: 9:00 AM
North 124B (Phoenix Convention Center - West and North Buildings)
This work analyzes whether machine learning emulation is superior to traditional emulation (regression, curve fits) used in predictive geophysical weather and climate models. The impact of atmospheric aerosol particles on clouds is the largest uncertainty in the prediction of anthropogenic forcing of climate. This process is described in bulk and empirical form in most large scale climate simulations through changes in the rain formation process (called ‘autoconversion’ of cloud condensate to rain). Here we replace the bulk treatment of autoconversion with a computationally expensive detailed treatment of the stochastic collection process, and explore emulating both treatments using machine learning methods. To build a training dataset for the machine learning emulator, a global General Circulation Model (GCM), the Community Atmosphere Model (CAM) run with detailed inputs and outputs of the process. A diverse set of machine learning models, including random forests, deep dense neural networks, and generative adversarial networks, are trained to estimate the autoconversion tendencies. A random hyperparameter search provides information to optimally tune each machine learning model, and then each scheme is validated with a temporally independent test set. Evaluations of each machine learning model examine their ability both to minimize approximation error and reproduce the correct distributions of each tendency.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner