Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This paper introduced an early warning method of structural damage from induced seismic events. The Long Short-Term Memory neural networks were employed to predict full S-wave accelerograms from initial P-wave, the predicted S-wave motion was then used to

The predicted as input to a suite of fragility curves 19 in real-time to estimate the probability of structural damage for masonry buildings. This research is very novel and has good engineering practicability. However, some major issues still need to be improved:

（1）Line 103-104, it introduced that this paper has the following contributions in the research area of ML methods in earthquake engineering. However, how the actionable damage alerts are conducted and what is the effective early classification of structural damage are not seen in the paper.

（2）In Figure 3, what is the evidence and specific values of the event’s trigger on and off.

  (3) In Line 128, 20 earthquake event records were selected, but only nine induced earthquakes were useful. The Fig.2 still showed 20 records, why? It is not explained how the training set and test set of LSTM model are allocated, and the 9 data are obviously insufficient.

  (4) The description of LSTM network architecture is very simple. The number of layers, number of neurons, hyper-parameters, and trigger on and trigger off are not clear. And the performance comparison between the LTSM model and traditional P-wave feature models (such as SVR, CNN) is missing. It is recommended to increase contrast test to prove the superiority of LSTM.

  (5) It is emphasized many times that the proposed method that couples a deep learning approach with structural fragility assessment is real-time. The whole process including P-wave acquisition, LSTM model input, S-wave prediction, fragility curves and the probability analysis of structural damage is how to achieve real-time warning.

  (6) The original data come from USA induced earthquakes, but the fragility curves derived for TFM and URM real buildings in Alsace, France (Fig.6) were directly referenced. How to explain the matching problem of information such as region and structure type, and are the fragility curves obtained according to the predicted S-wave accelerograms?

  (7) How to convert the predicted PGA into real-time probability of structural damage and early warning standard is not clear.

Author Response

The authors would like to thank the reviewer for his/her valuable time and attention to review the submitted manuscript. Please find attached below our detailed response to the reviewer concerns. 

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

This work integrates LSTM-based prediction of S-wave accelerograms with fragility-curve-based structural damage alerts for induced seismicity. The work is novel in its application focus and has the potential to make an impact in real-time earthquake early warning systems.

That said, several areas require improvement before the manuscript can be considered ready for publication:

• The dataset used (20 records from 9 earthquakes) is too limited to train a deep model with 500 neurons. This raises concerns about overfitting and generalization. Consider expanding the dataset through synthetic ground motion generation, physics-based simulations, or international data sources.
• Only LSTMs are tested. To strengthen the contribution, please compare your results with simpler methods and modern architectures (CNNs, CNN-LSTM hybrids, Transformers, or TCNs). Without baselines, it is hard to judge the true advantage of your approach.
• Hyperparameters (epochs, neuron count, learning rate) appear arbitrary and lack justification. Provide a rationale for their selection and consider hyperparameter tuning.
• The data preprocessing and split into training/validation/testing sets should be described in greater detail to ensure reproducibility.
• Current metrics (MSE/MAE) are not sufficient for time-series similarity. Consider reporting SMAPE.
• Quantify the error margins in relation to fragility-based risk thresholds, since this is the main practical application of your work.

Overall, the study is a goood step toward integrating AI into earthquake early warning, but it needs stronger methodological grounding, comparative evaluation, and a larger dataset to reach its full potential.

Comments on the Quality of English Language

The English is understandable but could be improved for clarity and flow. Some sentences are long and overly technical, which makes comprehension harder for a multidisciplinary audience. 

Author Response

The authors would like to thank the reviewer for his/her valuable time and attention to review the submitted manuscript. Please find attached below our detailed response to the reviewer's concerns. 

Author Response File: Author Response.docx

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors The suggestions of the first review should be taken seriously.

Author Response

Please find below attached our paper improvements. 

Author Response File: Author Response.docx

Reviewer 2 Report

Comments and Suggestions for Authors

The manuscript addresses a crucial and timely issue at the intersection of earthquake engineering and artificial intelligence. The integration of LSTM-based prediction of S-wave accelerograms with fragility-curve-based structural alerts is a promising direction with clear potential to advance real-time early warning systems. The revisions have improved clarity in several areas, particularly in the description of the preprocessing pipeline, the train/validation/test split, and the rationale for hyperparameter choices. Figures and tables are clear, well-structured, and add value to the presentation.

However, one substantive concern remains that must be addressed before the paper can reach its full potential:

Comparative baselines – The study currently evaluates only LSTM models. Without benchmarks against simpler regressions, CNNs, CNN-LSTM hybrids, Transformers, or TCNs, it is not possible to determine whether LSTM is truly the most effective approach. Including at least a brief comparative experiment, or at minimum a direct comparison with results from the literature applied to the same dataset, would significantly strengthen the validity and impact of the contribution.

Author Response

Please find below attached our paper improvements. 

Author Response File: Author Response.docx

Round 3

Reviewer 2 Report

Comments and Suggestions for Authors

The mere application of a specific model or technique, without rigorous comparison, validation, or novel methodological advancement, does not in itself constitute a scientific contribution.

Although the authors have added a comparison between LSTM and standard RNNs, this evaluation remains confined to the same family of recurrent architectures. The original concern regarding the lack of comparative baselines with fundamentally different architectures—such as CNNs, CNN-LSTM hybrids, TCNs, Transformers, or even simpler regression models—remains valid. Modern state-of-the-art earthquake early warning (EEW) systems increasingly employ CNN-based and Transformer-based models due to their strengths in local feature extraction and parallel processing capabilities (e.g., Hsu & Huang, 2021; Munchmeyer et al., 2021).

Suggestion: While the current focus on RNNs is justified, it would enhance the study’s impact to include, as part of future work, a comparison with non-RNN architectures (e.g., CNNs or Transformers), or at minimum, a discussion of how existing results from such models perform on comparable datasets. This would help more clearly position the proposed LSTM-based framework within the broader landscape of contemporary EEW approaches.

Author Response

Please find below our response to the reviewer's concerns. 

Author Response File: Author Response.docx