The goal of the Model Parameter Optimization(MPO) framework presented here is to use machine-learning to map the parameter space of one resolution of a model configuration to another higher resolution. This talk presents a brief introduction to neural networks, details of the technical implementation of the MPO, and a demonstration of the MPO using the double gyre test case within MOM6. We show that a neural network trained on output from a 1/2 degree version of the test case run with a range of Laplacian viscosities ~(500 m^2/s to 50000 m^2/s) generates a nonlinear mapping to the equivalent viscosity in a 1/4 degree resolution case. Given a particular value of viscosity at one resolution, this mapping can thus predict the value of viscosity needed to maximize the behavior of the model to the other, higher resolution. Additionally, this talk examines various subsets of diagnostics (mass transports, streamfunction, and dye tracers injected at various points in the domain) to determine which are most useful to train the neural network. Lastly, the MPO framework is shown to be extensible to more complex simulations and parameter spaces.

# GFDL Informal Seminar

Tue, Dec 4, 2018, 10:30 am to 12:00 pm

Location:

Smagorinsky Seminar Room 209

Speaker(s):

Sam Partee, Haverford College, Department of Computer Science

Trans-resolution model parameter optimization using an artificial intelligence approach