#C04 Solar-J: Improved Solar-Heating

Abstract

An alternative module for solar heating is being developed and implemented into ACME  by UC Irvine (Earth System Science and Information & Computer Science) and LLNL.  Based on the recently published Solar-J code (Geosci.Model Dev., doi:10.5194/gmd-2017-27), it represents a high-accuracy solar heating module alternative to RRTMG-SW, providing also a photolysis module for aerosol and gas-phase chemistry.  The increased accuracy comes with increased computational costs (15x), and thus a core part of this proposal is the computer science effort to optimize the backbone code for multi-stream scattering on the new DOE HPC systems.  We have identified several opportunities for 3x to 10x speedups using GPUs and even reduced precision arithmetic.  In parallel, we are pursuing alternative radiation-science optimizations with fewer visible-infrared wavelength bins but lower fidelity (possible 5x speedup). 

Solar-J takes RRTMG-SW's spectral data as the gold standard for solar heating at visible-infrared wavelengths (0.7-12 µm), but it keeps the UCI Fast-J cross sections for O₂ and O₃ at uv-visible wavelengths (0.18-0.7 μm) because UCI has a long history of optimizing radiative transfer in the stratosphere.  For clear sky, overhead sun, both models agree to within 1-2% on tropospheric heating rates, as they should since we copied the RRTMG code for water vapor and other trace gas absorption.  We found errors in RRTMG heating rates for large solar zenith angles (>84⁰), becoming very large near the terminator.  For the stratus cloud, RRTMG has a 3% low bias in planetary albedo across most solar zenith angles due to its 2-stream scattering vs. 8-stream scattering of Solar-J. Discrepancies with the cirrus cloud using any of RRTMG’s three different parameterizations are larger than for stratus, less systematic. Averaging over cloud structures within a grid cell with RRTMG is by Monte Carlo (MC) selection of independent column atmospheres (ICAs) for each wavelength bin (RRTM MCICA paper describes rms errors >10% per column), while Solar-J uses cloud quadrature over those same ICAs using the same ICAs for all wavelengths (errors ~ 1%). 

When ACME has solar-heating models that can be run with a range of accuracies, we can assess how such uncertainty maps onto climate simulations.  Solar-J’s resolved scattered light field at the surface can improve the fidelity of land/ocean biogeochemistry.  Long-term goals are a single solar module providing a more coherent linkage with the biogeochemistry and surface heating.