ABSTRACT
Development of neural network (NN) emulations for fast calculations of physical processes in numerical climate and weather prediction models depends significantly on our ability to generate a representative training set. Owing to the high dimensionality of the NN input vector which is of the order of several hundreds or more, it is rather difficult to cover the entire domain, especially its "far corners" associated with rare events, even when we use model simulated data for the NN training. Moreover the domain may evolve (e.g., due to climate change). In this situation the emulating NN may be forced to extrapolate beyond its generalization ability and may lead to larger errors in NN outputs. A new technique, a compound parameterization, has been developed to address this problem and to make the NN emulation approach more suitable for long-term climate prediction and climate change projections and other numerical modeling applications. Two different designs of the compound parameterization are presented and discussed.