Edit on GitHub

CalibrateEDMF.jl

CalibrateEDMF is a julia package that enables the calibration of atmospheric turbulence and convection parameterizations using gradient-free ensemble Kalman methods. It provides a user-friendly framework to train parameterizations implemented in TurbulenceConvection.jl, using the ensemble-based optimization methods implemented in EnsembleKalmanProcesses.jl.

Some of the options enabled by the package are:

  • Automatic regularization of parameter learning as an inverse problem,
  • Minibatch training using EnsembleKalmanProcesses.jl,
  • Kalman inversion with isotropic or anisotropic regularization,
  • Tracking of validation diagnostics, given a user-specified validation dataset.

The rationale behind the calibration framework implemented in this package is thoroughly described in our paper:

Lopez-Gomez, I., Christopoulos, C., Langeland Ervik, H. L., Dunbar, O. R. A., Cohen, Y., Schneider, T. (2022) Training physics-based machine-learning parameterizations with gradient-free ensemble Kalman methods, Journal of Advances in Modeling Earth Systems, 14(8), e2022MS003105. doi

If you use this package for your own research, or find any of the ideas presented useful, please cite our work. The article also includes results for an extended eddy-diffusivity mass-flux (EDMF) closure of turbulence and convection trained using this package. The EDMF scheme is implemented in TurbulenceConvection.jl, and described in the following papers:

Cohen, Y., Lopez-Gomez, I., Jaruga, A., He, J., Kaul, C., and Schneider, T. (2020) Unified entrainment and detrainment closures for extended eddy-diffusivity mass-flux schemes. Journal of Advances in Modeling Earth Systems, 12, e2020MS002162. doi

Lopez-Gomez, I., Cohen, Y., He, J., Jaruga, A., Schneider, T. (2020) A generalized mixing length closure for eddy-diffusivity mass-flux schemes of turbulence and convection. Journal of Advances in Modeling Earth Systems, 12, e2020MS002161. doi

Authors

CalibrateEDMF.jl is being developed by the Climate Modeling Alliance. The main developers are Ignacio Lopez-Gomez (lead), Haakon Ludvig Langeland Ervik, Charles Kawczynski, Costa Christopoulos and Yair Cohen.