Review Reports

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The BIM-ML framework proposed in this paper effectively combines Building Information Modeling (BIM) with Machine Learning (ML) to address the issue of energy consumption prediction for office buildings in the early design stage, with HVAC system type, WWR, and operation schedule as key influencing factors. By generating datasets through a reduced-factorial Design of Experiments (DOE), it significantly reduces the computational time of traditional simulations while maintaining high prediction accuracy (R² of RF model > 0.97). This framework shortens the simulation time through automated data generation and ML prediction, providing reliable support for sustainable design decisions. The core contribution of the research lies in generating structured datasets through reduced-factorial DOE and comparing the performance of various ML algorithms, demonstrating that this workflow can significantly improve analysis efficiency while maintaining high accuracy. Specific suggestions are as follows:

1. In the exposition part of 1. Introduction, the review of related research lacks depth and criticality. For example, it fails to point out the specific challenges of data interoperability in existing BIM-ML methods and does not analyze the applicable conditions of different ML algorithms in building energy consumption prediction.
2. The innovative positioning is ambiguous, and the boundaries with existing research are unclear. Although the research gap is pointed out in the introduction, in 1.4. BIM-based data generation and integration with Machine Learning, the proposed "BIM-ML integration framework" does not significantly differ from existing works such as [17, 18, 19] at the methodological level. Lines 120-122 "(1) generation of parametric datasets for model training and (2) embedding of predictive analytics directly within design workflows - key goals addressed in this study." are common goals of such research rather than the unique contributions of this paper. It is suggested to think and supplement "What is the most unique and generalizable methodological contribution of this paper compared to existing BIM-ML research?" It is recommended to directly compare the methods of this study with the existing works in the cited literature to highlight the unique contributions of this work.
3. In Section 2.3 Reduced-factorial framework, line 227 "Level balance and approximate orthogonality were verified through iterative frequency checks", but no specific verification standards and processes are provided. Standard methods in the field of building performance optimization (such as D-optimal design or Bayesian optimization) can obtain higher-quality experimental designs with fewer samples. The authors did not explain why this simplified DOE method was chosen and did not compare it with standard methods. It is suggested that the authors supplement the reasons for choosing this simplified DOE method and add a quantitative assessment of the quality of the experimental design.
4. In terms of the clarity of the research methods in 2. Materials and Methods, although the DOE framework is well-structured, more details on the specific implementation of the coupling rules between HVAC systems and orientations will improve the repeatability of the experiment. 1. It is suggested to add a composition structure diagram of the supplementary method framework to better display the method process of the system.
5. In 2.9.1. Model training and 2.9.3. Model validation, lines 380-381 mention "with RF achieving the highest accuracy and lowest error rates across both EUI and OE predictions." After comparing four algorithms, the paper only points out that RF performs best in EUI and OE predictions, but does not mention any hyperparameter optimization process. It is suggested that the authors must explain whether hyperparameter optimization was conducted. If so, please describe in detail the method used (such as grid search, Bayesian optimization) and the search space. If not, it must be clearly stated in the limitations and the conclusion should be revised to "Under the default parameter settings of the selected algorithms, RF performs best."
6. Regarding the presentation of charts, the charts are generally clear and informative. However, in 3.2. Feature importance analysis, it is suggested to add an explanation of the figure caption for Figure 3 at lines 456-459, specifically stating what the "168-model" and "210-model" datasets refer to, to facilitate readers' understanding. Additionally, it is recommended to improve the color scheme of Figure 3 by using colors with greater chromatic contrast to highlight the comparison results; and the font sizes of the titles of Figure 3 (a) and (b) are different, which needs to be corrected.
7. In terms of experimental design and limitations, all simulations in the paper are based on the warm and humid climate of Orlando (ASHRAE 2A zone). It is suggested to clearly state in 5. Conclusions or the limitations section that the research conclusions may not be directly applicable to other climate types (such as cold climates or extremely dry climates), provide supplementary explanations, and propose regional adaptation suggestions.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Author,

The submitted article seems to be an interesting and trend work. But it needs further improvements to be evaluated as an article. Please try to consider the below comments and answer the following questions in your article:

• Could you provide more details about the structure of your BIM-ML framework and how the integration between Revit, Insight, and Weka was implemented?

• What specific validation steps were taken beyond 10-fold cross-validation to ensure the robustness of your predictive models?

• How was the reduced-factorial Design of Experiments (DOE) designed, and why was this approach chosen over other sampling techniques?

• Can you elaborate on the dataset size and diversity ? do the 210 parametric simulations adequately represent real-world variability?

• What measures were taken to prevent overfitting, especially given the high R² values reported for all models?

• How do you plan to address the lack of external validation or comparison with other machine learning frameworks to confirm the generalizability of your approach?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Comments and Suggestions for Authors

1. Although the literature review section is comprehensive, it is not closely connected with the "research gap" and fails to fully highlight the necessity of this study.
2. The conclusion section focuses more on summarization and does not fully emphasize the practical application value of this framework in engineering practice.
3. The paper does not comprehensively discuss its limitations, such as "only for office buildings in Florida" and "not considering real-time weather changes", etc.
4. The formats of some references are not uniform. Please standardize them in accordance with the publication requirements of MDPI.
   
   

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

Comments and Suggestions for Authors

Dear Authors,

Many thanks for your response. The revision work is improved but still needs to answer the following questions:

• How does the proposed framework generalize beyond office buildings in Orlando, Florida, given that the training data and DOE are region and typology specific?

•  Is a dataset of 210 simulations with 14 input parameters sufficient to justify the reported high R² values, and how is overfitting especially for Random Forest explicitly addressed?

• How are early-design uncertainties (e.g., HVAC system type and operational schedules) handled, and how sensitive are the predictions to assumptions that may not be fixed at this stage?

• Why were default hyperparameters used for all ML models, and would systematic tuning alter the comparative performance or conclusions?

• Beyond 10-fold cross-validation, was any external or unseen-case validation performed to demonstrate real-world predictive robustness of the framework?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf