This manuscript investigates the effective dimensionality of wind field data for offshore wind turbine surrogate modelling. Using the Lillgrund simulation dataset, it compares low-order spatial descriptors against full grid inputs for fatigue (ResMLP) and power (LightGBM) prediction. The study identifies a task-dependent trade-off, concluding that low-order descriptors suffice for fatigue but not for power. Performance is evaluated using five-fold group-based cross-validation to ensure separation of simulation realizations. The topic is relevant and the experiments include multiple baselines and ablations. However, several issues need revision to improve rigor:

1. The study uses a residual MLP for fatigue prediction but switches to LightGBM for power prediction. This inconsistency introduces a confounding variable. It is unclear if the performance drop in power prediction using descriptors is due to information loss or the inductive bias of the tree-based model compared to the neural network.
2. The baseline comparison for the raw input relies on models (MLP/LightGBM) that treat grid points as independent features, ignoring spatial topology. To rigorously prove that the 31 descriptors are sufficient, the baseline should employ convolutional neural network (CNN) or even more complex models. A CNN would better exploit the spatial structure of the raw grid, potentially raising the performance ceiling and challenging the claim of descriptor sufficiency.
3. The claim that the 31 spatial descriptors are sufficient is currently conflated with the specific inductive biases of the chosen models (ResMLP for fatigue and LightGBM for power). To demonstrate that these descriptors possess intrinsic predictive value rather than being overfitted to a specific architecture, the authors should validate their competitiveness across a diverse set of standard regressors (e.g., LR, SVM, random forest, or a simple MLP). This is essential to confirm they capture the governing physics independent of the model structure.
4. The study proposes 31 a priori physical descriptors but lacks a comparison with standard data-driven dimensionality reduction techniques. Benchmarking against principal component analysis (PCA) or Autoencoders (AE) is necessary to justify the superiority or efficiency of the manual feature engineering approach.
5. The study relies on synthetic data from aeroelastic simulations with controlled inflow conditions. Real-world offshore conditions often involve non-stationary effects, complex veer, and sensor noise that are not present in the HAWC2Farm dataset. How robust are these 31 deterministic descriptors to measurement noise or incomplete data in real applications?
6. There is a numbering error in Section 4, where subsections are labeled "4.0.1" and "4.0.2" immediately following the main section title.

Comments are attached.