Symbolic regression-aided hyperparameter relationship for developing ANN for fragility prediction

Accurate prediction of seismic fragility parameters is crucial for assessing earthquake risks and developing effective mitigation strategies. Traditional methods, such as Incremental Dynamic Analysis (IDA), impose high computational costs, limiting their practical applicability for large-scale fragility evaluations. To address this challenge, this study proposes an optimized Artificial Neural Network (ANN) architecture for predicting fragility functions of low-rise steel moment frames. Three metaheuristic optimization algorithms, including Particle Swarm Optimization (PSO), Genetic Algorithm (GA), and Bayesian Optimization (BO), were employed to optimize the number of hidden layers, the number of neurons per layer, and the learning rate of the neural network. A comparative analysis of these methods indicated that PSO outperformed the others, yielding a lower cost function value and demonstrating more excellent stability in model tuning. Additionally, the optimal learning rate in PSO was lower than in the other two methods, suggesting a slower training process but enhanced stability of the final model. Symbolic Regression (SR) was utilized to enhance prediction accuracy and derive mathematical relationships for estimating the optimal number of neurons in hidden layers using the results of optimized network architectures. As a result, based on the proposed formula, the average prediction error was reduced by approximately 23 %, demonstrating the effectiveness of the developed approach. ANN models trained based on these relationships significantly reduced computational costs while enhancing fragility prediction accuracy. Furthermore, sensitivity analysis using the Shapley Additive explanations (SHAP) algorithm was conducted to quantify the influence of input parameters on model outputs. The results indicated that structural ductility and soil type had the most significant impact on fragility estimates, whereas seismic hazard level and importance factor exhibited the least influence. These findings highlight the effectiveness of integrating ANN, metaheuristic optimization, and sensitivity analysis in developing an efficient and computationally cost-effective fragility assessment framework. The proposed methodology enhances the accuracy and efficiency of fragility models while providing a viable alternative to traditional numerical approaches. Moreover, its applicability extends to diverse structural systems and seismic vulnerability assessments. It offers a valuable tool for earthquake engineering and risk-informed decision-making in seismic-prone regions. However, as with all data-driven models, the framework's performance depends on the quality and diversity of training data, necessitating potential hyperparameter adjustments for structures with significantly different characteristics. Addressing these limitations can provide valuable insights for future research in seismic risk analysis.