A deep learning approach to calculate elementary effects of morris sensitivity analysis

Abstract

The shape optimization of the multicomponent hydraulic turbomachinery is complex and computationally expensive due to the high number of computational fluid dynamics simulations. It is essential to identify the most influential parameters for which sensitivity analysis is needed to reduce the number of simulations. Morris sensitivity analysis provides a cost‐effective alternative for global sensitivity analysis that screens the essential parameters, requiring only a few computations to identify the most influential parameters from many parameters. This method is based on the elementary effects (EEs), which calculates the derivatives using the finite difference method. A deep learning (DL) approach is proposed to estimate the Morris method's EE. Two DL methods are proposed: the first utilizes the backpropagation of deep neural networks to calculate the partial derivatives of outputs to inputs; the second method relies on an artificial neural network‐based surrogate model which is trained using the optimization run dataset of hydraulic machinery with 30 parameters. The experimental results showed that the surrogate model trained with at least 7000 samples computes similar EEs as the classical Morris method with 310 samples. However, the backpropagation approach on Morris samples was observed to be less effective compared to a surrogate modeling approach.