Abstract: Stable production of high-strength, ultra-thin cold-rolled strips places extremely high demands on rolling mill performance.Addressing the challenge of frequent and unpredictable strip warpage defects in S6-high cold rolling mill during the production of such products, this paper proposes a strip warpage prediction model that integrates physics-based simulation with data-driven approaches.First, the kinematic mechanism of self-excited crossing and misalignment of work rolls, induced by complex contact-friction coupling within the unique floating work roll configuration of the S6-high cold rolling mill, is thoroughly analyzed.Second, a roll-strip coupled finite element model is constructed, incorporating roll system elastic deformation, strip elastoplastic flow, and contact nonlinearity.The reliability of the model is rigorously validated by comparing simulated results with industrial measurements of post-rolling strip profiles and thickness reduction.To alleviate the scarcity of onsite samples for abnormal instability conditions, the Latin Hypercube Sampling (LHS) method is employed to design experiments for key disturbance variables, such as work roll misalignment and crossing angles, thereby constructing a high-fidelity warpage behavior database covering non-self-stabilizing conditions.Subsequently, a nonlinear warpage prediction model is established using the CatBoost ensemble learning algorithm with an ordered boosting mechanism.The results demonstrate that the model achieves a coefficient of determination R² of 0.921 and a root mean square error (RMSE) of 3.81×10^-5 mm on the test set, significantly outperforming traditional algorithms such as XGBoost and SVR.Finally, an interpretability analysis based on SHAP values is conducted to quantitatively identify the key sensitive features leading to warpage in high-strength thin strips.This study provides a novel approach to controlling warpage behavior under non-self-stabilizing states in S6-high cold rolling mill.