Unified Hysteresis Modeling via Physics‐Based Deep Learning and Data Augmentation

Deep learning‐based models have recently emerged as alternatives to traditional form‐constrained hysteresis models, including Bouc‐Wen class models, offering significant potential for unified hysteresis modeling to capture complex nonlinearities and diverse response patterns exhibited under stochastic excitations such as ground motions. This paper proposes a unified hysteresis modeling framework based on deep learning, leveraging (1) physics‐encoded deep learning through a custom architecture that emulates the solution process of traditional models, (2) physics‐informed deep learning with an efficient loss function to enforce non‐negative energy dissipation, and (3) data augmentation via resampling hysteresis data to enhance the training dataset. The proposed unified hysteresis model can be trained on a relatively small amount of force–displacement data obtained under seismic excitations and enabling efficient and accurate time history analysis. The proposed model can account for complex stiffness and strength degradations and pinching effects. Tests across various traditional hysteresis models demonstrate that the proposed deep learning‐based unified hysteresis model can effectively reproduce diverse hysteresis behaviors. The proposed model is also validated against experimental hysteresis data from modular yielding links, confirming its capability to accurately represent real‐world hysteresis behavior. The source code and accompanying data can be accessed online for reproducibility at https://github.com/JaehwanJeon/Testing_torch .