A physics-encoded recurrent neural network inspired by Bouc–Wen class models for hysteretic behavior identification and prediction of flexural-failure reinforced concrete columns

A physics-encoded recurrent neural network inspired by Bouc–Wen class models is proposed for identifying nonlinear hysteretic behavior. Essential physical properties, such as memory effects, bounded-input bounded-output stability, rate-independence, elastoplasticity, and energy dissipation, are explicitly encoded into the network architecture, significantly improving interpretability and generalization compared to purely data-driven methods. Experimental validation using reinforced concrete column datasets demonstrates that the proposed network achieves superior identification accuracy compared to the conventional particle swarm optimization method under complex loading conditions. The model robustly captures degradation and pinching effects observed in hysteresis curves, without requiring empirically designed formulas or parameters as in traditional Bouc–Wen models. By avoiding manual formulation, the model reduces subjective bias and improves consistency in nonlinear hysteretic behavior identification. Furthermore, as system characteristics are flexibly incorporated into the neural network inputs, the hysteretic behavior of previously unseen columns can also be effectively predicted.