A machine-learning method for deriving state-dependent fragility curves of existing steel moment frames with masonry infills

Seismic assessment of existing buildings is usually a building-specific task that relies on refined finite element models. Such a task may require considerable computational demand, especially when predicting the seismic fragility of existing buildings under the framework of performance-based earthquake engineering. However, the computational cost can be significantly reduced by replacing the finite element model with a well-trained machine learning-based model, for example, an artificial neural network model. This paper presents the application of feedforward neural networks to derive the state-dependent fragility curves of existing steel moment frames, taking into account the effects of masonry infills. The network models can be trained to predict explicitly whether a structure exceeds the target limit state based on representative intensity measures of ground motions, which is in nature a binary classification problem. The number of non-linear time-history analysis required to generate the training data for the network models tends to be significantly lower compared to the case of conventional incremental dynamic analysis, particularly when a great number of ground motions are adopted aiming at higher accuracy of the fragility curves.