Machine Learning-Based Structural Reliability of Freeform Thin Concrete Shells

This study presents a reliability assessment of an unreinforced prefabricated funicular concrete shell designed for membrane-dominated behaviour under uniform loading. Under non-uniform actions such as wind and snow, bending induces tensile stresses that may lead to cracking. The analyzed configuration corresponds to a triangular shell ( 15 m span, 4.8 m height) without reinforcement or prestressing, with failure defined as the exceedance of tensile resistance. To enable reliability evaluation, a surrogate-based framework was implemented. Finite element simulations were combined with an artificial neural network (ANN) trained to predict the maximum principal tensile stress under varying load combinations. Monte Carlo simulation, First-Order Reliability Method (FORM), and FORM-based importance sampling were applied to estimate failure probabilities under stochastic wind, snow, and self-weight actions. Results showed consistent reliability indices across methods. Shells with mean flexural tensile strength above 20 MPa satisfied the target reliability index 4.7, while importance sampling significantly reduced computational effort compared to Monte Carlo simulation. The proposed framework demonstrates the applicability of machine-learning-based reliability assessment for free-form concrete shells.

Keywords: Structural reliability, machine learning, concrete shells, neural network.