A Deep Learning-Based Integration Method for Hybrid Seismic Analysis of Building Structures: Numerical Validation
Round 1
Reviewer 1 Report
Utilizing machine learning methods for use in seismic analysis is an emerging field that shows a lot of promise. This paper capitalizes on that idea; however, it fails to show that the developed method offers any improvement over current analysis techniques. The abstract proposes that the method used in this paper overcomes typical machine learning shortcomings in seismic response predictions. How this is accomplished has not been made clear. Furthermore, little information was provided on the specific application of analyzing a building with a base isolation system. It is the understanding of the reviewer that this machine learning model, which took a considerable amount of time to develop, would have to be re-developed for a base isolation system with different design properties than the one used to develop it. Importantly, isolation systems are designed for a specific building and it is highly unlikely that the exact same components will be used on another building. The following are comments and questions for use in improving this paper:
- Could this model be expanded to apply to more than one base isolation system?
- How is this model able to perform better than standard analysis techniques? Elaborate on how this paper aims to increase further the accuracy of conventional NTHA.
- The introduction proposes that ‘laws of physics’ can be encoded in the network architecture. Physics informed neural networks would seem to be a great application here. It is not clear how that has been incorporated in this model. Training with a traditional seismic analysis technique negates the reason to have a machine learning model for prediction.
- Although this model has been shown to be accurate for predicting synthesized data while having been trained on synthesized data, there is no evidence that it would be able to provide this level of accuracy when trained on or predicting real data with wider variation. In Ln 227, it is stated that experimental data would be more challenging and this needs to be shown to work in this case.
- The title addresses the methodology as an analysis tool. The base isolation system parameters change several times throughout the analysis process to achieve the desire behavior. The methodology proposed locks the base isolation parameters, therefore becoming a verification tool. There are property modification factors that need to be considered in design in order to capture the potential range of isolation system behavior due to aging, temperature, etc.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This is a very interesting study! This article would be very helpful in the field of earthquake engineering.
I have a few comments and I wish the authors to address them in their revised version.
1-Nomenclature section should be added before the introduction to include all the acronyms and symbols used in this manuscript.
2- Abstract in line 14, are must be were.
3- Abstract in line 16, do you mean its computation time or its time computation.
4- Abstract in line 22 and 23 can be rephrased to be as a concluding remarks.
5- Figure 3, h needs to be predefined in the text.
6- Figure 7 the axis labels are not clear to be read.
7- Figure 8, is there any other method to differentiate the lines of each training run? It is not easy to follow.
Figure 12, there must be a method to represent the hybrid analysis line from the conventional analysis.
Figure 13. Same issue with Figures 12 and 8. You may report R2 for each case between the hybrid and the conventional or to select one part of each case and maximize it in the same Figure.
8- Conclusions: point c (lines 514-515) you must mention the value of the acceptable limit from the earthquake engineering practice.
9- references: in lines 587, requires formatting.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors present a hybrid seismic analysis in application to computing the full non-linear response of building structures according to Japanese engineering standards. The study combines the advantages of non-linear time history analysis (NTHA) and machine learning models (MLM). The paper presents a framework for generating artificial data, optimizing network architecture, training MLM, and finally testing it. For MLMs, all trainings were performed using the TensorFlow machine learning library under Python 3.6. The term hybrid is related to the time integration algorithm of the NTHA which is applied in method based on the operator splitting technique. Instead of adopting simplified analytical models, the authors applied recurrent neural networks (RNNs), which were trained using available experimental data. This approach made predictions possible on new data at each time step of the hybrid seismic analysis.
In the case study, three base-isolated buildings of 5, 10, and 15 stories were considered which were modelled using the software Structural Earthquake Response Analysis 3D. The superstructure was simulated using analytical models whereas the isolation level by MLM. A lumped mass model (LMM) was adopted for the superstructure. The isolation layer was formed by natural rubber bearing (NRB), lead rubber bearing (LRB), and oil damper (Oil). ). The combination of both NRB and LRB devices was assumed to perform a bilinear hysteresis behaviour. The force developed in the oil damper device was assumed to depend only on the relative velocity of its edges.
Two ground motions (GMs) were used to generate the training data used to develop the MLMs, and four ground motions for analysis of structural responses of the three building models using the proposed hybrid seismic approach and comparison with conventional analysis. All GMs were intentionally derived from famous Japan Earthquake records with different amplitudes, frequency contents, and durations (Kobe 1995 NS, Taft 1952 EW, Tohoku 1978 NS, El Centro 1940 NS).
The developed MLMs simulate the isolation layer (NRB + LRB + oil damper) of the studied buildings by mapping their displacement time history to the corresponding shear force time history. They were tested, then implemented to perform a total of twelve hybrid seismic analyses. The proposed hybrid seismic analysis computes the full dynamic response in terms of acceleration, velocity, and displacement at each degree of freedom. In addition, this type of analysis does not have limitations on a specific building structure, as the MLM can be saved and used again in any new building model. It also deserves a distinction for its very short computation time, which proves its efficiency for engineering practice.
Minimization of network architecture was a crucial criterion for efficient implementation of MLM in hybrid seismic analyses. Undoubtedly a challenge for future analyzes is the development of reliable and accurate MLMs.
An objective study of this paper reveals that the authors presented very precisely the assumptions of the hybrid seismic analysis method as well as the advantages and possibilities of its use in civil engineering.
On assessing the paper very positively, I have a few minor comments.
- The paper briefly deals with current world research of base isolation structure;
- The article briefly deals with the issue of FEA of a building.
- References could be more extensive.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
Authors present a paper about an application of artificial neural networks in a seismic analysis computing the full nonlinear response of building structures. This study is the first step to develop machine learning model simulating structural component or group of components whose only available data are those obtained at well-designed experimental tests. The obtained results reveal that the hybrid analysis, taking advantages of both mechanics-based and data-driven methods, is an efficient tool for building-response simulation.
In general, the article presents an interesting study of an important both scientific and real problems. The obtained results are very promising to next studies.
The title of the article corresponds to the Abstract and Results.
The paper is sufficiently referenced (consists of 36 articles relevant to the article) and well written and can be published as is.
Round 2
Reviewer 1 Report
The concerns raised in the previous revision were not really addressed, these are limitations of the study that would require significant additional work
