A flexible framework for radiation parameterizations traceable to benchmarks

Radiative transfer is unique ... because the solution to fully specified problems is known to great accuracy the representation of radiation is one of the most pure exercises in parameterization in atmospheric models. But fundamental understanding does not mean that the representation of clearsky is a solved problem; every assessment identifies significant parameterization errors in the treatment of gaseous absorption and scattering. Some errors arise from different choices between accuracy and computational cost, but the real barrier is the difficulty of creating and updating a parameterization for absorption by gases. Users of externally-developed parameterizations have no control over either cost-accuracy tradeoffs nor the range of applicability for which the parameterization is developed. Accuracy and applicability are intertwined: as one example, weather models using a general-purpose parameterization pay the computational cost for accurate calculations with greenhouse gas concentrations quite different than today's even though that capability will never be exercised.

We propose to build a flexible framework that allows modelers to build the parameterization of absorption by gases that best suits their needs, allowing them to control the atmospheric conditions under which the model is to be trained and the aspects of the errors to be minimized. We will provide practical methods for building parameterizations of absorption by gases and allow users to explore cost-accuracy tradeoffs. We will use this framework to exploring promising new ideas for replacing k-distributions with more robust and efficient techniques for treating the spectral variation of absorption by gases. In broad outline our tasks will include a) work to establish baseline spectroscopy, subjecting widely-used collections such as HITRAN to rigorous validation with observations; b) construct and/or adapt existing tools for fast and flexible benchmark line-by-line calculation of absorption coefficients and spectrally-integrated fluxes that results from these; c) develop methods to generate custom k-distributions in which users specify training data, spectral discretization, and error metrics, and a framework that allows users to choose one or more cost-accuracy tradeoffs; d) exercise the system by developing a range of focused parameterizations aimed at particular applications, including sub-seasonal to seasonal forecasting with models at GFDL and idealized modeling at NCAR, as well as parameterizations for very unusual conditions including deep-time paleoclimate simulations; e) explore alternatives to the development of k-distributions including the use of very spectrally-sparse line-by-line calculations and the use of machine learning approaches to parameterization.

The project proposes to build a flexible framework of clear-sky radiative gas absorption and scattering parameterizations for use in atmospheric models. This will allow climate model developers/users to construct parameterizations that meet their needs for balancing cost and accuracy under conditions where models require creating and updating parameterizations due to absorption under changing gas concentrations. The goal is to make it easier for climate model developers/users to select and implement appropriate absorption and scattering parameterizations in radiative transfer codes based on cost-accuracy metrics.