Compound Parameterization for a Quality Control of Outliers and Larger Errors in Neural Network Emulations of Model Physics

Development of neural network (NN) emulations depends significantly on our ability to generate a representative training set. Because of high dimensionality of the input domain that is in 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. In this situation the emulating NN may be forced to extrapolate, which is 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 and climate change predictions and other applications.