Regression mode collapse in deep learning surrogates for stochastic structural response prediction

Deep-learning surrogates can greatly reduce the cost of the Stochastic Finite Element Method (SFEM), but standard accuracy metrics do not show whether stochastic response variability is preserved. This study examines Regression Mode Collapse (RMC), in which different random-field realizations lead to nearly identical predicted response curves. We treat RMC as an ensemble-level loss of response diversity. Conditional averaging is a possible mechanism, underdispersion is its main statistical manifestation, and over-smoothing may appear in the shape of individual curves. Two nonlinear problems are studied: phase-field fracture of concrete under uniaxial tension and the cyclic response of a U-shaped reinforced concrete shear wall. CNN-RNN, CNN-LSTM, and ViT achieve high R 2 values and low MAE and RMSE values, yet they lose most of the response variability. The Standard Deviation Ratio (SDR) and Wasserstein Distance Ratio (WDR) are used to measure the preservation of ensemble dispersion and inter-sample response differences. Across the tested learning rates, batch sizes, training-set sizes, and loss functions, the lost diversity does not clearly recover. CNN-KAN retains much more response variability and does not show severe RMC in either case. These results suggest that deep-learning surrogates for stochastic analysis should be judged by both pointwise accuracy and their ability to preserve response diversity.