MetaIMNet: A physics-informed neural network architecture for surrogate response and fragility modeling of structures subjected to time-varying hazard loads

Extreme events such as earthquakes and hurricanes cause widespread damage and disruption to infrastructure assets such as buildings and bridges. Catastrophe modeling and accurate extreme event risk and resilience assessment require portfolio-level fragility functions of these assets, which involve the establishment of functional relationships between a relevant peak response quantity, also known as the engineering demand parameter (EDP), and select features characterizing the hazard. Given the computational demands of analyzing several statistical combinations of hazard and structural features, while running nonlinear time history analyses for each combination, surrogate demand models relating peak EDP to relevant intensity measures (IMs) of the input time history are popular. Although traditional IMs such as peak accelerations and velocities, average velocities, and peak spectral accelerations determined a priori have been traditionally found to be effective predictors of response and damage, their use in surrogate models in fragility model development introduces additional model uncertainties. In a bid to enable more robust and accurate surrogate modeling, we propose MetaIMNet; a physics-informed framework based on a neural network that simultaneously extracts key features from the time history of the load and leverages these features for structure specific response prediction. The framework is illustrated through a case study which shows that it outperforms traditional surrogate modeling strategies at a nominal added computational cost associated with model training, and can be used as an effective surrogate model for developing fragility functions for a wide range of hazards and structures.