Authors:
Agnimitra Dasgupta | University of Southern California | United States
Qian Fang | University of Southern California | United States
Erik Johnson | University of Southern California | United States
Steven Wojtkiewicz | Clarkson University | United States
Hideo Fujitani | Kobe University | Japan
Yoichi Mukai | Kobe University | Japan
Eiji Sato | National Institute for Earth Science and Disaster Prevention, E-Defense | Japan
Machine learning (ML) models, particularly deep learning (DL) models, are increasingly being augmented with information from physics-based (PB) simulation models to develop surrogates that reduce the computational cost of multiple forward model runs for uncertainty quantification, optimization, control, etc. Traditionally, physics-informed ML/DL models are trained by augmenting the loss/cost function with the PB model’s governing equation. However, a priori knowledge of PB model that can accurately capture the underlying physics is usually unavailable.
This study presents a strategy to develop physics-informed ML/DL models in the absence of accurate prior PB models, to infer from the data a ‘sufficiently accurate’ PB model. The PB model accuracy is evaluated using model falsification (De et al. 2018), and the ML/DL model is trained using a modified loss function that accounts for the inferred PB model. The approach is used to develop neural network (NN) surrogate models for a magnetorheological (MR) fluid damper tested in different excitations while recording displacement, velocity and commanded electromagnet current. A Bingham-viscoplastic damper model is inferred from some data. The physics-informed NN models are trained and validated using the remaining data, and compared against pure PB models. The developed surrogate models will be used in hybrid simulation studies, implementing various control strategies, on a full-scale structure over the E-defense shake table.