Efficient seismic fragility analysis of structures based on probabilistic machine learning method

Abstract

Reliable and rapid estimate of the performance of civil infrastructure against seismic hazards is one of the most essential tasks in civil engineering. Machine learning (ML) technology have been widely adopted in engineering practices recently by establishing the relationship between seismic intensities and the corresponding structural demands based on very limited experimental or numerical simulation databases. Nevertheless, its predictions are essentially deterministic, which largely have ignored the uncertainty inherent within the structural system.

In this study, a new ML algorithm, that is natural gradient boosting (NGBoost), is applied to directly evaluate the conditional probability distribution P(y|x) instead of producing a point estimate E[y|x] for each value of x. This type of probability prediction can directly obtain the failure probability of the structure under an intensity measure, while avoiding the extra input of uncertainties from structure properties in traditional ML methods. Therefore, the evaluation of structural seismic fragility is especially simple and efficient.

Finally, seismic fragility analysis of an inelastic single degree of freedom (SDOF) system and a multi-span concrete bridge are carried out respectively to illustrate and demonstrate NGBoost model. The results indicate that the average accuracy of NGBoost agrees with conventional ML algorithms reasonably well in seismic fragility analysis, while without re-input for each additional set of structural material properties and geometric parameters.