Bridging physics and statistical learning methodologies for the accurate modeling of the radiative properties of non-uniform atmospheric paths.

• A fast though accurate modeling strategy is presented to treat non-uniform atmospheres. • A functional form, based on physical and statistical arguments, is derived to handle path non-uniformities. • Evaluation of the unknown coefficients of the model (Lévy–Khintchine representation) relies on regression on LBL data. • The proposed method is shown to provide highly accurate results at a very low CPU cost. The objective of the present work is to describe a technique to approximate atmospheric path transmissivities using a recurrent structure, following a method proposed recently but limited to date to high temperature applications. The physical model together with its underlying statistical assumptions is detailed. It is found to involve a rather simple analytical formula that applies both to two-layers systems and to more general multi-layers non-uniform configurations. This treatment of path non-uniformities uses several unknown parameters that are first trained on LBL reference data in two-layers configurations to illustrate the relevance of the proposed approximate model. Then, in a second time, model's parameters are trained on non-uniform path transmission curves representative of multi-layers atmospheres. The corresponding recurrent formulation is shown to provide accurate estimates of transmissivities of non-uniform atmospheric paths (maximum relative errors are below 0.35 % in all the considered test cases) at a very low CPU cost. [ABSTRACT FROM AUTHOR]