Torsion design of CFRP-CFST columns using a data-driven optimization approach

Abstract

A challenging issue of utilizing the merit of the machine learning model to the multi-objective optimization (MOO) problem is that sufficient physical experiments data are hard to get. With the limited training data, overfitting of the surrogate model is inevitable, which may mislead the search engine. A data-driven model based on simulations commonly has a better performance for addressing overfitting problems. However, there is a gap between the numerical model and the real structure/physical experiments. In this paper, a framework called data-driven optimization is proposed for structural performance optimization. First, the transfer learning algorithm (two-stage TrAdaBoost) is designed to reweight the simulation data points that have more significant residuals predicted by a base learner (i.e., ensemble machine learning model), which goal is to select the “accuracy” of simulation data points as supplements to the real structure/physical experiments data points, instead of trying to reduce the gap by model-updating methods, as the traditional methods do. In this way, simulation data points relevant to the real structure/physical experiments will be assigned a large weight value. Then, the generated two-stage TrAdaBoost model incorporated with nondominated sorting genetic algorithm II (NSGA-II) is used to optimize structure design. Finally, this paper used the proposed framework to conduct feature impact analysis, fast predict, and optimize torsion design for the concrete-filled steel tube (CFST) column subjected to combined compression-bending-torsion. The results showed that the two-stage TrAdaBoost algorithm performs better and outperforms the baseline model, an extreme gradient boosting (XGBoost). By typical examples, the proposed framework can be a viable tool for the preliminary design of the CFST column. The Pareto fronts of the two objectives (ultimate torsion strength and cost) are successfully obtained.