General comments

The manuscript presents a machine-learning ground-motion prediction model for Italy based on a Mixture-of-Experts (MOE) framework combined with XGBoost regressors. The topic is timely and potentially relevant for the ground-motion and natural hazards communities, particularly in light of the growing interest in data-driven and hybrid approaches for seismic ground-motion prediction.

The manuscript is clearly written, technically detailed, and supported by an extensive set of figures and statistical analyses. The use of large Italian strong-motion datasets and the attempt to systematically compare the proposed framework with both empirical and machine-learning reference models are appreciated.

However, despite these positive aspects, the current version of the manuscript presents substantial methodological and interpretative shortcomings that prevent a reliable assessment of the proposed model’s actual predictive capability and scientific added value. In particular, the validation strategy, the interpretation of residual variability, the definition and documentation of the datasets and predictors, and the level of physical interpretability of the results require major revision.

For these reasons, I recommend reconsideration after major revisions. The issues raised below are structural rather than cosmetic, but in my view they can be addressed with a careful redesign of parts of the analysis and a more cautious interpretation of the results.

Major comments

1. Validation strategy and event-wise independence

The most critical issue concerns the validation strategy adopted throughout the manuscript. The authors rely on random splits of the dataset into training, validation, and test subsets (70/15/15). In ground-motion modeling, this approach is not appropriate, as recordings from the same earthquake are inherently correlated through shared source, path, and rupture characteristics.

With random splitting, records from the same seismic event can appear simultaneously in both training and test sets, leading to event-level information leakage. This typically results in overly optimistic performance metrics that reflect within-event learning rather than true generalization to unseen earthquakes.

Although the manuscript repeatedly refers to “independent testing” and “generalization to independent earthquake events” (including in the abstract), no strict event-wise validation strategy (e.g., leave-one-event-out, leave-multiple-events-out, or grouped cross-validation by event) is explicitly described or demonstrated.

Action required:

The authors must clearly document the validation strategy and, if not already done, redesign the evaluation using an explicit event-wise split. All performance metrics (RMSE, correlation, residual distributions, intra- and inter-event residuals) should be recomputed under this framework. Without event-wise independence, the reported performance improvements cannot be considered reliable.

1. Interpretation of aleatory variability and residual reduction

The manuscript emphasizes a strong reduction in residual dispersion (σ) and improved residual concentration compared to reference models. However, the treatment of aleatory variability remains largely descriptive and is not sufficiently grounded in established GMPE practice.

In flexible machine-learning models, reductions in residual variance may arise from overfitting, implicit smoothing, or exploitation of event-specific patterns rather than from a genuinely improved representation of source, path, and site effects. The manuscript does not assess whether the inferred variability levels are physically plausible, nor does it benchmark them against reference σ, τ, and ϕ values from established Italian or pan-European GMPEs.

Action required:

The authors should (i) clarify how residual components are estimated in the presence of complex ML architectures, (ii) compare the resulting variability levels with published GMPE variability models, and (iii) discuss whether the observed reductions are physically meaningful or may reflect methodological artifacts.

1. Definition and documentation of predictors

The manuscript does not clearly and consistently document the full set of predictors used in the MOE-XGB model. While some variables are mentioned (e.g., magnitude, distance, Vs30, elevation, coordinates), a complete and unambiguous list of input features is missing, as is a discussion of their relative roles.

Action required:

The authors should provide a clear table listing all predictors used, their definitions, data sources, and preprocessing steps. This is essential for reproducibility and for assessing the physical consistency of the model.

1. Lack of interpretability analysis

Given the complexity of the proposed MOE-XGB framework and the strong claims regarding predictive performance, the manuscript lacks a systematic interpretability analysis. No global or local assessment of variable importance is provided, and the role of individual predictors in controlling ground-motion behavior remains unclear.

Action required:

The authors should include an explicit interpretability analysis (e.g., SHAP-based global importance such as beeswarm plots, partial dependence or accumulated local effects plots) to clarify which predictors dominate the predictions across different intensity measures and periods. This would substantially improve the scientific value of the work and help distinguish physically meaningful patterns from purely data-driven behavior.

1. Dataset definition and currency (ITACA)

The manuscript refers to the ITACA dataset but does not clearly specify which version of the ITACAext flatfiles is used. Recent releases (e.g., ITACAext flatfile 2.0; Lanzano et al., 2024) provide updated metadata and intensity measures.

Action required:

The authors should explicitly state the exact dataset version used and justify their choice if a non-latest release is adopted. Clear and precise data citation is essential for transparency and reproducibility.

1. Overstatement of applicability to seismic hazard

The manuscript repeatedly suggests applicability to seismic hazard assessment and PSHA. However, no hazard-oriented application is demonstrated, and key aspects such as spatial correlation, rupture geometry, or integration into PSHA workflows are not addressed.

Action required:

The authors should substantially tone down claims related to seismic hazard applications or explicitly demonstrate how the proposed model could be integrated into PSHA frameworks. At present, such claims appear overstated relative to the analyses performed.

Minor comments and technical issues

• The Results section contains considerable redundancy, with multiple figures conveying similar information without adding new insight. A reduction or consolidation of figures is recommended.
• Performance metrics are sometimes averaged across periods and intensity measures without clear justification of their engineering relevance.
• Terminology such as “MORI dataset” is potentially misleading and should be replaced with a clearer description of the data sources and compilation criteria.
• Minor typographical and grammatical issues are present throughout the manuscript and should be corrected in a revised version.

Recommendation

In its current form, the manuscript does not meet the methodological standards required for publication. However, I believe that, with substantial revision, the study could become a meaningful contribution. I therefore recommend reconsideration after major revisions, addressing in particular the validation strategy, interpretability, dataset definition, and the physical interpretation of residual variability.

This manuscript attempts to construct a more accurate predictive model for spectral acceleration (SA) by combining the Mixture-of-Experts (MoE) framework with XGBoost, and evaluates its performance using earthquake ground motion data from Italy. In the experimental design, the authors reuse the training framework of a previous GPR study, including the data partition strategy and validation dataset design, allowing the results to be compared directly under a consistent experimental setting. Subsequently, a more comprehensive dataset is introduced to further extend the comparison. This research strategy—first ensuring strict comparability and then validating the model with a larger dataset to test generalization—appears logically clear and methodologically reasonable. In addition, the additional test conducted on independent earthquake events further strengthens the credibility of the experimental conclusions. Nevertheless, the current version of the manuscript still requires several clarifications and additions to improve methodological completeness and presentation clarity.

1. In Lines 391–395, the authors state that the earthquake events used to evaluate the model's generalization capability are not included in the training, validation, or testing datasets, thereby effectively avoiding potential data leakage. This design is convincing, and the results shown in Fig. 13 are generally consistent with the previously reported error metrics and scatter plot distributions, providing mutual support for the conclusions. However, the training phase currently adopts a record-based data partition strategy. Considering that records from the same earthquake event may exhibit inherent correlations, it would be beneficial to include an additional experiment where the training and validation processes are conducted using an event-based partition strategy (similar to that used in the final testing stage). Such a comparison would help more comprehensively demonstrate the model’s generalization capability at the event level and strengthen the overall evaluation of the MoE model performance.
2. The authors are encouraged to include an analysis of input feature importance. Without feature importance analysis or sensitivity evaluation, the proposed model may appear as a black-box ensemble structure rather than a predictive tool supported by physical and statistical interpretation. In the context of earthquake engineering, where interpretability is increasingly emphasized, reporting only error metrics may be insufficient. At minimum, the authors could provide feature importance rankings, partial dependence relationships, or response trends under input perturbations to demonstrate that the model behavior is consistent with established ground motion attenuation patterns.
3. In Section 3.2, which discusses Bayesian optimization, the description of the hyperparameter optimization process should be clarified. The manuscript currently states that Bayesian optimization was used for parameter tuning, but it does not clearly specify which hyperparameters were included in the optimization procedure. The authors are encouraged to provide a list of the optimized hyperparameters, such as learning rate, tree depth, and regularization parameters, along with a brief description of the search ranges or optimization settings. This additional information would allow readers to better understand the model training process.
4. In the section concerning “SA in earthquake ground motion,” the manuscript does not clearly explain the basis for the selection of SA period ranges. Since the segmentation of SA periods is closely related to the model output structure and its engineering interpretation, it would be helpful if the authors clarified how these period intervals were determined—for example, whether they are based on engineering standards, statistical characteristics of the dataset, or configurations used in existing Italian ground motion models. If different period ranges correspond to different submodels or shared parameter structures, further explanation would strengthen the completeness of this section.
5. The manuscript also shows some inconsistencies in notation and formatting. For example, the use of hyphens and dashes, italicization of variables, and the definition and usage of abbreviations are not entirely consistent. In addition, the font sizes in some figures and tables appear relatively small, which may affect readability. The authors are encouraged to standardize the notation and adjust the figure and table font sizes in the final version to improve overall clarity.

In summary, the proposed MoE structure demonstrates potential innovation and application value for earthquake ground motion prediction. The experimental framework is logically structured, and the inclusion of independent event testing is a positive aspect of the study. However, several issues remain regarding methodological clarity, interpretability analysis, and completeness of experimental descriptions. If the authors address the points above—particularly by including event-based training comparisons, providing feature importance analysis, clarifying the details of Bayesian hyperparameter optimization, explaining the basis for SA period segmentation, and improving the consistency of notation and formatting—the manuscript could be considered after minor revision.