<p>Accurate prediction of water surface profiles (WSP) is essential for reliable flood hazard assessment in compound channels with converging floodplains. This study develops data-driven machine learning models by integrating controlled laboratory experiments with existing datasets (<i>N</i> = 396) to predict nondimensional WSP (φ = H/h). Four algorithms including KStar, M5P, Reduced Error Pruning Tree (REPT), and Random Forest were implemented using eight hydraulic and geometric parameters such as width ratio (α), relative depth (β), discharge ratio (Q_r), aspect ratio (δ), convergence angle (θ), relative distance (Xr), bed slope (S), and Froude number (Fr).Model performance was evaluated using validation strategies including standard 10-fold cross-validation and leakage-controlled profile-based grouped cross-validation, to assess generalization. Sensitivity analysis revealed that relative depth, discharge ratio, and Froude number exhibited the strongest correlations with water surface profile (WSP).Among the models, Random Forest achieved the highest predictive accuracy (R² = 0.981, RMSE = 0.0477), while REPT performed comparably in error metrics (R² = 0.975, MAE = 0.0286). Wilcoxon signed-rank testing confirmed statistically significant differences between model performances (<i>p</i> &lt; 0.01). Overall, tree-based models outperformed instance-based approaches, offering robust and interpretable predictions for complex non-prismatic channel geometries. The proposed framework enhances the reliability of flood risk assessment and supports improved river management practices.</p>