Physics knowledge-based transfer learning between buildings for seismic response prediction

The recent advance in deep learning has attracted considerable interest for employing the state-of-the-art methods to solve engineering problems. However, the applicability of machine learning based models is hindered by the high cost of big data acquisition and task-specific difficulties. This paper presents a framework of physics knowledge-based transfer learning (Phy-KTL) neural networks that integrates the powerful learning capacity of physics-informed neural networks (PINNs) and the flexible transferability of model-based transfer learning technique to enhance structural seismic response prediction in the context of limited labelled datasets. The leverage of physics knowledge (represented by Runge-Kutta solver) allows the neural networks to better capture the structural nonlinear pattern. The use of model-based transfer learning improves the model generality by transferring features extracted from the source building to target buildings. The effectiveness of Phy-KTL in predicting seismic responses between target buildings is numerically validated as compared with Data-driven neural networks, PINNs, and Data-based transfer learning (Data-KTL). A practical application, which uses Phy-KTL to transfer features extracted from the numerical model to the physical building tested on the shaking table, validates that Phy-KTL is robust and effective to improve seismic response prediction of target buildings with limited labelled data.