Building a physics-constrained, fast and stable machine learning-based radiation emulator

• DOI: 10.5194/ems2022-419
• Publication Date: 2022-06-28T10:46:43Z
• Authors: Guillaume Bertoli; Sebastian Schemm; Firat Ozdemir; Eniko Szekely; Fernando Perez Cruz
• Keywords: Parametrization (atmospheric modeling); Solver

<p>Atmospheric radiative transfer, which describes the evolution of radiation emitted by Sun, Earth's surface, clouds, and greenhouse gases, is an essential component climate weather modeling. In models, transfer approximated parameterizations. Theoretically, however, with sufficient computing power, electromagnetic equations could be solved, but in practice this out reach.  The current operational solver Icosahedral Nonhydrostatic Weather Climate Model (ICON) ecRad, which, developed at ECMWF, one most advanced available parameterizations.  It considers surface optics, gas aerosol optics cloud [1]. accurate parametrization remains computationally expensive. Therefore, usually not invoked every time step only runs on a reduced spatial grid, can affect prediction accuracy, or 1D setting without 3D transfer.</p><p>In project, we are trying to develop ecRad improved machine learning speed up computation loss accuracy. Machine learning-based parametrizations would general allow fully replace existing sub-grid scale parameterizations, once trained from data. However, such do necessarily preserve physical quantities, lead instabilities, model drifts unphysical behavior as observed [2] [3].</p><p>We present here emulation strategy, composed three steps. First, continue call significantly coarser grid predict clear-sky radiation. We thereby use regularizer while reducing costs. Then, interpolate data full using Gaussian processes. Finally, effect clouds random forests. underlying idea avoid support generalization capabilities ML method.</p><p>Our first numerical experiments aqua planet simulation promising. hope obtain valuable outcome when considering more complex datasets seasonality realistic topography. Our final goal run ICON enhanced parametrization, though online performance open. There low resolution field, computed part our expected play central role stability.</p><p>[1] <em>Hogan, R. J., & Bozzo, A. (2018). A flexible efficient scheme for ECMWF model. Journal Advances Modeling Earth Systems, 10, 1990-2008. </em></p><p>[2] <em>Brenowitz, N. D., Bretherton, C. S. Prognostic validation neural network unified physics parametrization. Geophysical Research Letters, 17, 6289–6298</em></p><p>[3] (2019). Spatially Extended Tests Neural Network Parametrization Trained Coarse‐Graining, J. Adv. Model. Syst., 11, 2728–2744</em></p>