Developing a physics-informed and physics-penalized neural network model for preliminary design of multi-stage friction pendulum bearings

Abstract

Over the last few decades, the field of base isolation systems has made significant strides forward by developing new systems to improve the behavior of isolated structures under moderate and severe seismic excitations. One of the most efficient systems is the multi-stage friction pendulum that can provide a wide range of effective pendula with various regimes to reach high energy dissipation capability. The difficulty in designing such bearing at the preliminary stage comes from the relatively long process of trials to select the parameters of each sliding surface in order to ensure that the required effective period, effective damping, and displacement capacities are met. Thus, the following study proposes a direct design approach relying on a physics-informed and physics-penalized neural network model to overcome this issue. Within the context of the analysis, a large dataset composed of over 35000 isolators that covers a wide range of properties of the recently developed generation of friction pendulum “Quintuple Friction Pendulum” was generated following the mechanics-driven approach and then utilized as the case study to test the reliability of suggested design strategy and capability of the proposed multi-output neural network method. Thereafter, the performance of the physics-informed and physics-penalized model was compared to a purely data-driven approach and a physics-informed one. Generally, the results have shown that the proposed model has considerable accuracy with a maximum MAPE of 4.89% for achieving the required QFP properties. Additionally, they showed that the presented method provides an effective way for rapidly designing a quintuple friction pendulum isolator.