Machine learning-aided cloud analysis for seismic fragility assessment of multi-span bridges

The present study explores the implementation of machine learning (ML) algorithms in the context of the fragility analysis of existing multi-span reinforced concrete bridges composed of simply supported girders, accounting for the record-to-record variability and the uncertainty related to structural and mechanical parameters. The study compares a conventional cloud analysis based on nonlinear time history analysis (NLTHA) and a proposed ML-aided cloud method. In this latter, random forest and neural network models are trained to surrogate NLTHA in predicting the seismic demand of the piers. The ML-aided approach implements a novel strategy to optimise the choice of input features (seismic intensity measures and uncertain structural parameters) to be used for enhancing the model training. The accuracy of the method is discussed by comparing probabilistic demand models and fragility curves conditioned to variable uncertain parameters. In the case study application, the ML-aided cloud analysis, implemented by reducing the number of NLTHA to approximately one-third compared to the conventional approach, achieves high accuracy when using the random forest algorithm. This latter outperforms neural network leading to a lower number of required input features and improved performance metrics. In conclusion, the outcomes are used to propose a novel ML-aided cloud analysis approach for future applications of bridge risk assessment.