Physics-Anchored Sequence Learning for Predicting Hysteresis of RC Shear Walls

We propose a physics-anchored sequence model for predicting force–displac ement hysteresis of reinforced concrete shear walls under reversed cyclic loading. A degrading Bouc–Wen model is first identified for each specimen using robust fitting and displacement sub-stepping, providing interpretable priors for stiffness, strength scale, and degradation. A recurrent network then predicts the latent hysteretic state and reconstructs restoring force, trained with a data loss and discrete Bouc–Wen constraints, and further regularized by tangent-stiffness and per-cycle energy consistency. We evaluate the method on 21 open-access shear-wall tests with heterogeneous loading protocols. On seven held-out specimens, the predictor achieves a mean \(\:{R}^{2}\) realizes around 0.97 (median around 0.99); errors concentrate near reversals and at the largest drift cycles. The learned latent state and identified parameters enable direct inspection of stiffness deterioration and strength loss, offering both predictive accuracy and interpretable diagnostics.