Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper presents the prediction of flow rates through the RTHG by utilizing hybrid models based on gradient boosting algorithms. Some major issues need to be addressed:

1. While the manuscript presents a comprehensive benchmark of gradient-boosting variants, the entire training/test corpus originates from a single laboratory flume. Please discuss the extent to which the optimized SSA-GBMs can be extrapolated to field-scale canals where approach-channel aspect ratios, turbulence spectra, sediment loads and three-dimensional effects differ markedly. Providing an additional paragraph on expected scaling limitations and recommending minimum Froude or Reynolds thresholds for reliable deployment would strengthen the engineering relevance.
2. The SHAP analysis correctly identifies h/B and S as dominant drivers, yet the manuscript does not verify whether the learned response surfaces behave plausibly beyond the observed parameter ranges. A short section containingphysics-based sanity plots (e.g., predicted Q vs h³/² for fixed geometry and variable slope) would reassure readers that the black-box models do not violate basic hydraulic limits such as critical-flow or orifice-type asymptotes when extrapolated.  
3. Although SSA is convincingly shown to outperform CMA-ES, PSO and GA for LightGBoost, the search budgets (population size, iteration limit, random seeds) are not reported. Please add a concise table in the Supplementary Material specifying these settings so that the community can reproduce the tuning exercise. Moreover, include convergence curves for the other GBMs (CatBoost, NGBoost, etc.) to justify the claim that SSA remains superior across all five algorithms.
4. The bootstrap-derived confidence intervals and R-factor are useful, yet they only reflect aleatoric uncertainty stemming from data scatter. A brief discussion of epistemic uncertainty, originating from model structure (choice of GBM) and from the limited parameter space explored, should be added. Reporting 95 % prediction intervals generated by NGBoost’s native probabilistic output and comparing them with the bootstrap CIs would demonstrate how much of the total uncertainty is captured, thereby clarifying the robustness of the claimed “lowest uncertainty”for NGBoost-SSA.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

This manuscript presents innovative and scientifically valuable study on predicting and quantifying the uncertainty of flow rates through Rectangular Top-Hinged Gates using hybrid gradient boosting models. However, some aspects of the manuscript should be improved to enhance its quality and practical applicability. Specific recommendations include:

1. Expand the background by adding a brief description of the governing hydraulic processes, such as vena contracta formation, flow separation, and interactions between gate geometry and energy dissipation. Giving readers a mechanistic understanding will better justify why machine learning modeling is needed for these nonlinear relationships.
2. Incorporate more recent literature on CatBoost, LightGBM, NGBoost, and hybrid optimization models applied in civil or hydraulic engineering. Highlight gaps such as the absence of boosting-metaheuristic frameworks in hydraulic gate systems. This will update and modernize the literature review.
3. Include full specifications of all sensors (flow meters, pressure transducers, angle measurement tools), their accuracy levels, calibration protocols, and how measurement uncertainties were propagated. Provide material specifications of the gate model and tank dimensions.
4. Clearly state whether any data cleansing or preprocessing was applied. Describe outlier detection criteria, normalization/scaling methods, and splitting strategy (train–test percentages and whether stratification was used).
5. Include a table listing all tuned hyperparameters for each model, their search ranges, justification for the range boundaries, and metaheuristic algorithm configuration (population size, iterations, mutation/crossover rates).
6. Include details on the number of experimental repetitions for each condition, how variability was handled, and whether averaged Q values were used. Explain whether control experiments were performed to validate equipment consistency.
7. Provide details on the computing environment, including Python or MATLAB versions, ML libraries (scikit-learn, CatBoost, LightGBM, NGBoost versions), CPU/GPU specifications, and runtime considerations. This is standard for ML-focused papers.
8. Interpret why certain dimensionless parameters appear more influential (such as h/B dominating due to its direct control on upstream energy head). Discuss how the ML-based ranking compares with expected hydraulic theory. This strengthens scientific validity.
9. Analyze potential hydraulic reasons such as increased turbulence intensity, secondary flow effects, or measurement errors at high velocities. Discuss how these nonlinearities challenge the ML models and whether data augmentation or physics-informed adjustments could mitigate the issue.
10. Add comparisons with existing empirical discharge equations. Even if crude, such benchmarks contextualize ML improvements and highlight reliability gains over classical formulations.
11. Provide learning curves or validation curves demonstrating absence of overfitting. Include k-fold CV metrics and discuss consistency across folds. This gives confidence in generalization.
12. Add a comparative reflection showing how your results align or differ from past studies on weirs, sluice gates, and other flow measurement structures. Highlight what new scientific contributions your hybrid approach provides relative to previous methods.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have well replied my comments. I think the current version can be accepted.