Knowledge-Enhanced Deep Learning for Wind-Induced Nonlinear Structural Dynamic Analysis

Recent advancements of performance-based wind design of tall buildings have placed increasing importance on effectively modeling of the nonlinear structural dynamic response under extreme storms. However, the numerical estimation of wind-induced nonlinear structural response based on the high-fidelity finite element model is computationally intensive due to its small time-step size and long simulation duration. To this end, the reduced-order modeling methodology using either physics-based analytical models or data-driven metamodels is widely applied to the simulation of nonlinear structural dynamics. With the rapid developments of machine learning techniques, the deep neural networks have recently become a popular data-driven approach for the efficient and accurate estimation of nonlinear structural responses. Due to a high demand on the quality and quantity of data, training the deep neural networks can become intractable. In this study, a knowledge-enhanced deep learning (KEDL) algorithm is proposed to simulate the wind-induced linear/nonlinear structural dynamic response. More specifically, the machine-readable knowledge in terms of both physics-based equations and/or semiempirical formulas is leveraged to enhance regularization mechanism during training of deep networks for structural dynamics. The KEDL methodology is data-efficient and robust to noise by effectively utilizing both the available input-output data and the prior knowledge on the structure of interest. In addition, the KEDL methodology is coupled with the wavelet-domain projection to simplify the input-output relationship, and hence to accelerate the training process. The data-efficient and noise-resistant characteristics of the KEDL methodology have been comprehensively investigated based on a single-degree-of-freedom (SDOF) system. Finally, it is clearly demonstrated that the trained knowledge-enhanced deep neural network presents both high simulation accuracy and computational efficiency in estimating the nonlinear dynamic response of a multidegree-of-freedom (MDOF) system under wind excitations.