Machine Learning-Aided Performance-Based Optimal Design of Steel Moment-Resisting Frames Using NSGA-II

This paper introduces a novel and rigorous probabilistic methodology for the performance-based design of steel moment-resisting frame buildings. The proposed methodology is consistent with the modern performance-based earthquake engineering framework and aims to apply this framework directly in the process of designing new buildings. To this end, a multiobjective optimization problem has been defined considering the cross section of structural elements as design variables. The probability of exceeding a targeted performance level during the building’s lifetime under earthquake hazards, serves as the safety criterion to be minimized. This, along with the minimization of the total structural weight as the cost criterion, constitutes the considered objective functions. A multiobjective optimization framework using the nondominated sorting genetic algorithm version II has been employed to solve the optimization problems and determine the set of Pareto optimal solutions. A machine learning (ML) model is developed to predict the performance parameters for each potential building. The outputs of this ML model are employed in the inner loop of the optimization to reduce computation efforts, enabling direct performance-based design of the building. The effectiveness of the proposed methodology is illustrated through a numerical example of optimal section design of a 6-story moment-resisting frame (MRF). Application of the proposed method has resulted in several reliable MRF scenarios with optimum weights, which are assessed in terms of fragility and reliability. Moreover, the results of sensitivity analysis shed light on the importance of element cross sections affecting the reliability of the structure.