Autoencoder-based dimensionality reduction with multi-layer perceptron surrogates for global sensitivity analysis of high-dimensional nonlinear responses

Global sensitivity analysis (GSA) of models producing high-dimensional distributed responses remains challenging, as classical Sobol’-based methods become computationally inefficient to apply directly to complex response fields. This paper proposes an autoencoder (AE)-based dimensionality reduction framework combined with multi-layer perceptron (MLP) surrogates for efficient Sobol’-based GSA of high-dimensional nonlinear responses. To accommodate different response characteristics, multiple AE architectures are incorporated to enable effective compression of sequential, spatially correlated, and uncertainty-driven model outputs, while preserving the dominant variance structure of the original responses in a reduced latent space. An MLP surrogate is constructed to learn the nonlinear mapping from model parameters to latent variables. To ensure consistency of variance-related characteristics between latent-space and original-space sensitivity measures, a decoder-based pull-back strategy is introduced to propagate variance contributions back to the physical response domain, facilitating Sobol’-based GSA in the original response space. The proposed autoencoder-assisted multi-layer perceptron (AE-MLP) framework is capable of addressing sensitivity analysis problems associated with highly complex systems and extremely high-dimensional distributed responses, that are difficult to handle without using dimensionality reduction techniques. Numerical benchmark studies and engineering applications demonstrate that the AE-MLP framework achieves accurate sensitivity ranking with significantly reduced computational cost.