1359 The Use of a Deep Neural Network to Represent Radiation Transfer Calculations in the E3SM

Wednesday, 15 January 2020
Hall B (Boston Convention and Exhibition Center)
Linsey Passarella, ORNL, Oak Ridge, TN; and A. Pal, S. Mahajan, and M. R. Norman

We plan to develop a Deep Neural Network (DNN) to replace the full radiation model in the Department of Energy’s (DOE)Energy Exascale Earth System Model (E3SM). This study will build upon research that demonstrated the ability of a DNN to imitate just the shortwave and longwave radiative transfer calculations in a super-parameterized version of E3SM with an accuracy of 90-95% while also proving to be qualitatively similar to the original parametrizations in year-long simulations (Pal et al. 2019, Geophysical Research Letters). As a preliminary effort to train a DNN, we have run the E3SM at an ultra-low resolution (800 km) for seven years while saving the input and output variables of the radiation module for each grid box at each time step. We are using this data to train a dense, fully-connected, feed-forward DNN. We plan to validate the impact of using the DNN on the model climate in a decades long run of the ultra-low resolution model. A successful implementation of the DNN to emulate the full radiation transfer calculations will allow for a significant decrease in computational expense, while remaining highly accurate.
- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner