Machine learning-based surrogate models for predicting the optimized weight and prestress level of double-curvature cable domes

Abstract

A major challenge in designing cable domes is the high sensitivity of these lightweight flexible structures to geometry change. Including shape parameters with size of all elements and prestress level as discrete design variables in optimization algorithm may lead it to be computationally prohibitive. In this regard, this study employs and compares a multi-linear regression (MLR) model, a multiplicative non-linear regression (MNLR) model, and an artificial neural network (ANN) model with the aim to estimate the optimized prestress level at the minimum weight for different geometries of double-curvature cable domes. The models are trained using a large dataset of dome geometries developed and optimized numerically. The models show high accuracy with a slight advantage for ANN over MLR and MNLR. However, the prediction functions using MNLR model are much simpler and concise with only 7 coefficients when compared to a complete quadratic function of 28 terms needed in the MLR model and to the complexity of ANN equation. Moreover, unlike MLR and ANN, the MNLR model can synthetically generate a larger dataset of optimized values quite accurately beyond the trained range. The developed functions facilitate the prediction of the optimized prestress level and structural weight of double-curvature cable domes when compared to the traditional optimization algorithms. They can be utilized as a design support tool when implemented as a built-in library in design software. They also save time and effort when employed as an objective function for a shape optimization problem. Moreover, they fulfill designers’ preferences for the dome stiffness and load-bearing capacity. Finally, the impact of considered geometrical parameters on the optimized values is explored using the trained dataset, providing recommendations for designers.

Topics & Concepts