Surrogate Modeling of Stochastic Hysteresis Response in Irregular RC Shear Walls Using CNN-KAN and Manifold Random Fields

Abstract

Abstract Accurately predicting the hysteretic response of irregular reinforced concrete (RC) shear walls is essential for seismic performance assessment. However, spatial variability in material properties and geometric complexity introduce significant stochasticity, making conventional stochastic finite element (SFEM) methods computationally inefficient due to difficulties in random field assignment and the high cost of nonlinear analyses. This study proposes an efficient deep learning framework that uses a convolutional neural network-Kolmogorov–Arnold network (CNN-KAN) model to predict stochastic hysteresis responses of irregular shear walls with manifold random fields. The random field is constructed using the isometric mapping (isomap) method to accurately represent the spatial variability of material properties in SFEM. Then, hysteresis curves obtained from SFEM analysis are used to train the CNN-KAN model, the CNN-long short-term memory (LSTM) model, and the CNN-multilayer perceptron (MLP) model. While all three models achieve high prediction accuracy, only CNN-KAN successfully preserves the stochastic diversity of structural responses. The traditional CNN-MLP and CNN-LSTM models suffer from an averaging effect, failing to capture the inherent randomness. The trained CNN-KAN model is then used as a surrogate model for Monte Carlo simulation in large-scale sample analysis. The results demonstrated that it can efficiently generate hysteresis curves and energy dissipation distributions to study the influence of spatial correlation on structure stochastic responses. The proposed model significantly enhances computational efficiency, reducing the computing time to only 1/300 of the conventional SFEM analysis. Further, the proposed framework enables efficient uncertainty quantification and is extensible to other complex structural systems involving spatial randomness.

The authors' abstract, as published at the source. Journal of Computing in Civil Engineering, 2026 · DOI ↗