Artificial neural networks in environmental sciences. II. NNs for fast parameterization of physics in numerical models

Abstract

A new generic approach, based on the neural networks (NN) technique, to improve computational efficiency of parameterizations in numerical environmental models is formulated. Such parameterizations generally require computations involving complex mathematical expressions, including differential and integral equations, rules, restrictions and highly nonlinear empirical relations bused on physical or statistical models. From a mathematical point of view, such parameterizations can usually be considered as continuous mappings (continuous dependencies between true vectors). NNs are a generic tool for fast and accurate approximation of continuous mappings and, therefore, they can be used to replace primary parameterization algorithms. In addition to fast and accurate approximation to the primary parameterization, NN also provides the entire Jacobian for very little computation cost.