The manuscript provides a new valuable method for recognizing geochemical anomalies from high-dimensional geochemical survey datasets. It can be considered for publication after moderate revisions. All the suggestions have been marked on the attached pdf file. The suggestions include the following aspects:

1. Please further polish the English of the manuscript.

2. The training patches are only 134. It is too few for training the deep learning models used in geochemical anomaly recognition. This needs to furtehr explanations.

3. The limitations of the new method should be discussed in the Results and Discussion Section.

This contribution proposes a hybrid approach that integrates geological knowledge (specifically ore-controlling faults) into Deformable Convolutional Networks (GK_DCN) to recognize anisotropic geochemical anomalies. The authors carry out a case study in the Southern Tianshan Au-Cu polymetallic ore district to investigate the model's ability to capture complex anomaly patterns and enhance interpretability through visualization techniques. This study offers a solid workflow for coupling data-driven and knowledge-driven approaches, providing potential references for mineral prospectivity mapping in structurally complex terrains. The methodology and argumentation of this paper are comparatively reliable, however, the robustness of the experimental design requires further strengthening. I think it will be of general interest to readers of EGUsphere after a major revision.

The main issues are as follows:

1. Considering the training set is notably small (only 134 samples), it is likely insufficient to support the generalization capability of a deep learning model with high-dimensional features (39 channels). Although the authors introduced geological constraints for regularization and performed multi-metric evaluations, the risk of overfitting has not been entirely mitigated. I strongly recommend implementing k-fold cross-validation to strengthen the robustness of the experimental results. Furthermore, regarding the negative sampling strategy, the authors state that samples were randomly selected from barren regions. However, the absence of discovered deposits does not equate to the absence of mineralization. Random selection risks inadvertently labeling undiscovered concealed deposits as negative, thereby introducing label noise. I suggest establishing a buffer zone around known deposits to exclude potential mineralization areas when selecting negative samples.

1. The authors incorporate a “distance to faults” rule directly into the loss function to constrain model training. The significant performance improvement observed confirms that fault structures provide critical information. However, this raises an important question: the study currently only addresses the influence of known (mapped) faults. How do concealed or blind faults affect the model's judgment? Since the loss function relies on the distance to known faults, there is a risk that the model might suppress valid anomalies controlled by unmapped blind faults. In other words, if fault information is incomplete, what is the extent of the potential bias in the results? I believe that discussing these limitations and the potential influence of hidden structures would significantly enhance the generalizability and practical value of the proposed method.

Specific comments:

Line 8: Delete the duplicate “that”.

Line 13: replace “fault” by “faults”.

Line 27: replace “movement” by “migration”.

Line 29: replace “ore materials” by “economic minerals”.

Line 242: replace “points” by “deposits”.

Line 257: The specific value of λ in Equation 9 is not provided. Since this parameter balances the data loss and the geological constraint, it is critical for reproducibility. How do different values of λ affect the model's performance?

Lines 311-316: IDW is an isotropic smoothing interpolation method. This approach may artificially weaken the anisotropic features of the data before it is even input into the model. Given that the core of this study is to utilize DCN to capture irregular and anisotropic patterns, why was the isotropic IDW method chosen for preprocessing instead of Kriging or other methods capable of preserving directionality?

Lines 317-322: The positive samples are extracted as 9*9 grid patches centered on known deposits. If the distance between two deposits is less than the patch dimensions, their corresponding samples will spatially overlap. If one of these overlapping samples is assigned to the training set and the other to the validation set, the model may achieve artificially inflated accuracy by simply memorizing local background features rather than learning generalizable metallogenic rules. Did the authors implement any de-duplication steps or use a spatially blocked split (e.g., splitting by geographic region) to prevent this data leakage? Please clarify the splitting strategy.

Line 344: replace “perspective” by “prospectivity”.

Line 372: There appears to be a contradiction between the text and the figures regarding the PC2 scores. The text states that low PC2 values correspond to the Au-Cu mineralization assemblage. However, the legend in Figure 11 explicitly states that yellow (high values) represents high concentration, and Figure 9b clearly shows that most Au-Cu deposits are located in areas with high PC2 scores.

Line 469: replace “anomalous” by “anomalies”.