A data-driven surrogate model and its test validation for mechanical response analysis of air-supported membrane structures

Air-supported membrane structures exhibit significant geometric nonlinearity and complex contact coupling effects due to the interaction between membrane surfaces and cable nets. This interaction makes multi-parameter nonlinear finite element analyses computationally expensive. This study presents a data-driven surrogate model that integrates proper orthogonal decomposition (POD) and artificial neural networks (ANN). This model efficiently represents and predicts membrane stress and cable net tension fields. First, a parametric finite element database is established. Then, unstructured response fields from different geometries are transformed into a unified reference domain. Then, POD is employed to extract the dominant spatial modes, and parallel multilayer perceptron (MLP) are constructed to capture the nonlinear mapping between the design parameters and the modal coefficients. The proposed framework was validated using additional numerical cases and a small-scale physical model test. Results showed that the first three POD modes preserved 99.0% and 97.4% of the cumulative energy for the membrane stress and cable net tension fields, respectively. For ten additional numerical cases, the mean relative errors (MRE) are 1.46%-3.43% and 2.94%-4.55%, respectively. Further test validation demonstrates MRE of 3.70% and 6.26% compared to finite element results. The surrogate model accurately reconstructs full-field mechanical responses and provides an efficient computational approach for the rapid analysis of forms and the parametric design of large-span, air-supported membrane structures.