Support vector regression model for the prediction of buildings’ maximum seismic response based on real monitoring data Authors: Dongwang Tao, Shizhe Fang, Haixu Liu, Jianqi Lu, Jiang Wang, Qiang Ma Published: 2024 ABSTRACT AbstractMaximum drift ratio (MDR), one of the engineering demand parameters (EDPs), provides fundamental physical value for predicting building damage. Existing machine learning based prediction models mainly rely on numerical simulation data or structural experiments and are not appropriate for prediction of seismic response of real structures. The New Earthquake Data (NDE1.0) is the most comprehensive publicly available dataset of actual structural seismic response observations. Currently the prediction models using NDE1.0 are mainly based on linear or log-linear regression. In this study, based on the NDE1.0 flatfile, we develop a full-feature support vector regression (SVR) based MDR prediction model (SVR-MDR), treating all the available 41 characteristic parameters including structural information as input feature. To improve the model’s efficiency and practical applicability, we also establish a reduced-feature SVR model (RSVR-MDR) by selecting 10 fundamental parameters based on SHapley Additive exPlanations (SHAP) values and the accessibility of features. Our results demonstrate that SVR-MDR model outperform other machine learning models such as kernel ridge regression and decision tree models, and SVR-MDR and RSVR-MDR models outperform conventional loglinear regression and multinomial models, because SVR can map the complex nonlinear function of multiple variables and consider the available information of buildings especially the fundamental frequency. The proposed RSVR-MDR model have promising potential application for post-event seismic damage assessment and post-event emergency response in near real time.Keywords: Structural response, Engineering demand parameters, Maximum drift ratio, Support vector regression, Intensity measureSubject terms: Natural hazards, Civil engineering Introduction Earthquake is one of the most destructive natural hazards faced by modern society and human beings. The effects from earthquake include strong ground shaking, surface rupturing and ground failure, among which strong ground shaking is most destructive. The civil engineering structures under strong ground shaking are exerted on strong horizontal inertial force and may be damaged or destroyed. Meanwhile the destruction and collapse of building structures under strong shaking will directly lead to casualties, property damage, and overall functional loss. The direct average economic losses are up to 14.7 billion USD per year resulted from ground-shaking-related damage to building structures in the United States1. Rapid post-earthquake assessment of buildings’ seismic damage is paramount to post-event emergency response and economic loss assessment2,3. The structural seismic responses and engineering demand parameters (EDPs) such as building’s maximum drift ratio (MDR) which is defined as the maximum relative displacement between the top and bottom floors normalized by building height, peak floor accelerations, maximum inter-story drift ratio, residual story drift are fundamental physical values for estimating building damage. For a big city, there are hundreds of thousands of buildings and most of the buildings have incomplete or missing design information. The key step for post-event damage assessment for urban building clusters is to rapidly predict the buildings’ maximum seismic responses using the ground motions and incomplete building information. Currently the approaches for estimating structural seismic responses and EDPs are mainly divided into three categories: mechanics-based, purely data-driven and hybrid approaches4. Mechanics-based approach is the most reliable solution, requiring both the mathematical model such as finite-element model (FEM) or shear and flexural beam model or multiple-degree-of-freedom (MDOF) model for modeling building structure and the site-specific ground motion history recorded at the basement of the analyzed building4–7. Miranda and Reyes combined the Timoshenko beam theory and the spectral displacement of the fundamental period of the structure to estimate the amplification factor and the MDR5. Lu et al. proposed SimCenter Workflow for seismic damage simulation and loss prediction of urban buildings, which combined actual building inventories, MDOF models and nonlinear time history analysis6. Ghahari et al. compared Federal Emergency Management Agency (FEMA) P-588, American Society of Civil Engineers (ASCE) 7-169 simplified equations, and response spectrum method to predict peak floor acceleration and found that response spectrum method and gave accurate estimation result7. However the response spectrum method requires knowledge of the seismic loading and building modal properties. To balance computation efficiency and prediction accuracy, Scaini et al. used single-degree-of-freedom oscillator as a first order approximation of building typology and estimated the MDR rapidly for near-real-time damage estimation10. Shan et al. proposed a seismic performance prediction method based the equivalent SDOF hysteresis model of the structure, which is obtained through FEM11. However, it is noted that mechanics-based approach is time-consuming and requires a large amount of computational resources, which hinders the usage for rapid post-earthquake assessment. Purely data-driven approach uses linear regression or machine learning (ML) models to establish relationships between ground motion intensity measure (IM) or time history and the EDPs of building structure. Unlike mechanics-based approaches that do not require pre-existing data for training, ML approaches need amounts of high-quality data to train effectively. Shen et al. used deep neural network to establish relationships between pseudo-acceleration spectra and structural safety, and the training and test accuracy were above 93%12. However, this method needs over tens of thousands of samples to train the deep learning model. The other main problem of purely data-driven approach is the lack of interpretability. The hybrid approach is a combination of mechanics-based and purely data-driven approaches, combining the advantage of interpretability of mechanics-based solution and applicability of the purely data-driven solution4, and the approach has attracted the attention of researchers and has strong application prospect in rapid post-event emergency response. Guan et al. proposed ML based linear elastic MDOF with known yield strength model to predict peak story drift based on building information, modal parameters, spectral intensity parameters and nonlinear pushover analysis parameters, and the results showed that the proposed hybrid model had better performance than the purely data driven model4. Wen et al. used convolutional neural network to predict the MDR and peak floor acceleration by taking the ground motion time history and five easily available building parameters including story height, story number, fortification intensity, longitudinal column distance and transverse span as inputs13. Demertzis et al. conducted extensive comparative evaluation of 15 machine learning algorithms including light gradient boosting machine, gradient boosting regressor, random forest regression to predict 3D R/C buildings’ seismic response and found that light gradient boosting machine was a good choice for EDPs prediction according to generalizability and computation efficiency14. However, the data used in the former hybrid approach research were obtained by numerical simulation of a series of FEM models rather than real structure response under ground shaking. Real structural seismic response data in buildings have made fundamental contributions to the seismic design of buildings and provide basic data for establishing relationships between ground motions and EDPs. Nowadays there are two main structural seismic response databases: the US Centre for Engineering Strong Motion Data (CESMD) and Japanese Building Research Institute (BRI) strong motion network. Perrault and Guéguen used California earthquake records in CESMD to improve EDPs prediction models by using multiple logarithmic linear function and ground motion parameters including peak acceleration, cumulative absolute velocity, velocity response spectrum corresponding to the fundamental period and displacement response spectrum15. Abdelmalek‑Lee and Henry constructed a model combining kriging and XGBoost based on a database of responses of 207 buildings from 35 earthquake events using CESMD earthquake data16. Astorga et al. constructed the New Earthquake Data (NDE1.0) of instrumented buildings, which contained 118 buildings and 8520 recordings from 2737 events17. The NDE1.0 database provides excellent earthquake data for the prediction of structural seismic response and EDPs. We download the original xlsx file provided by Astorga et al. (2020) which included 147 buildings and 12,293 recordings from 3081 events. We use the original xlsx file dataset to establish MDR prediction models. In this study, a full-feature support vector regression (SVR) model is used to predict MDR of building structures under ground motion based on the earthquake data from NDE1.0 flatfile. In the full-feature SVR-MDR model, 41 characteristic parameters including 6 structure information, 9 spectrum information, 16 duration information, and 10 other earthquake information parameters are chosen as input features. Considering the importance and availability of input features, the reduced-feature SVR for MDR prediction model (RSVR-MDR) model is also developed. Finally, the performance of SVR-MDR, RSVR-MDR are analyzed and compared to linear regression models. Dataset and methodology Seismic response database of building structures (NDE1.0) The NDE1.0 is a database of building response records and corresponding earthquake information, which provides a shared and accessible data source for studies on seismic hazard, damage prediction and other fields. The information of NDE 1.0 is from the US CESMD and Japanese BRI, and provides post-processed ground motion IMs, EDPs, building details and earthquake characteristics17. In terms of building typology, the database includes 77 steel frame, 45 reinforced concrete (RC), 11 steel-reinforced concrete (SRC), 10 masonry and 4 wooden structures. The number of records and events for each type of building are shown in Table 1. SRC buildings record the most earthquake events and earthquake responses, and the maximum and minimum peak top accelerations (PTAs) and MDRs of the SRC buildings are shown in Table 2. Table 1The number of events and records for each type of building.TypologyThe number of eventsThe number of recordsSteel frame6581358Reinforced concrete15054558Steel-reinforced concrete20506219Masonry52134Wood1324 The number of events and records for each type of building. Table 2PTA and MDR range of SRC buildings.Building IDPTA-min (cm/s2)MDR-minPTA-max (cm/s2)MDR-maxAKB12.323.29E−05249.711.09E−03ANX0.324.77E−07596.845.30E−03CHB4.937.62E−06374.946.86E−03EDG5.582.28E−05298.821.08E−03IWK1.652.31E−06578.455.43E−03MST1.733.25E−06219.196.18E−03NGN5.531.72E−05194.318.42E−04NGY0.956.12E−0745.793.32E−04SMZ1.123.26E−06539.503.37E−03TDS0.997.53E−06531.195.37E−03THU3.194.28E−06908.241.06E−02 PTA and MDR range of SRC buildings. The parameters used to predict EDPs such as PTAs and MDRs in the NDE1.0 flat-file can be classified into 4 classes, namely (1) structure information, (2) spectrum information, (3) duration information and (4) other earthquake information. The parameter and the description of the data are displayed in Table 3. Structure information includes 6 parameters B_lat, B_long, H, N, f1 and f2; spectrum information includes 9 parameters Sa1, Sv1, Sd1, Sa2, Sv2, Sd2, Avg_Sa, Avg _Sv and Avg_Sd; duration information includes 16 duration parameters such as DZX, DE, DU, DSa, DSv, DSd and DB; other earthquake information includes 10 parameters E_lat, E_long, M, Dep, PGA, PGV, PGD, Ia, CAV and DP. Table 3Parameters and their descirptions in NDE1.0.ParametersParametersLatitude of the buildingB_latLongitude of the buildingB_longHeight of the buildingHTotal number of storiesNFundamental frequencyf1co-seismic minimum frequencyf2Pseudo-spectral acceleration of f1Sa1Pseudo-spectral velocity of f1Sv1Spectral displacement of f1Sd1Pseudo-spectral acceleration of f2Sa2Pseudo-spectral velocity of f2Sv2Spectral displacement of f2Sd2Mean between Sa1 and Sa2Avg_SaMean between Sv1 and Sv2Avg_SvMean between Sd1 and Sd2Avg_SdLatitude of the epicenterE_latLongitude of the epicenterE_longMagnitude of earthquakeMEpicentral distanceDepPeak ground accelerationPGAPeak ground velocityPGVPeak ground displacementPGDArias intensityIaCumulative absolute velocityCAVDestructive potentialDPZhou and Xie (1984) durationDZXTime interval between 0.01 m/s and Ia = 0.125 m/s using the Husid diagramDETotal time between the first and the last exceedance of the acceleration threshold, i.e. 0.05 g, 0.10 g, 0.15 g and 0.20 gDB1, DB2, DB3, DB4Sum of the time intervals during which acceleration exceeds the acceleration threshold, i.e. 0.05 g, 0.10 g, 0.15 g and 0.20 gDU1, DU2, DU3, DU4Duration corresponding to (5–75)% and (5–95)% of total energy associated with ground motion acceleration.DSa1, DSa2Duration corresponding to (5–75)% and (5–95)% of total energy associated with ground motion velocity.DSv1, DSv2Duration corresponding to (5–75)% and (5–95)% of total energy associated with ground motion displacement.DSd1, DSd2 Parameters and their descirptions in NDE1.0. Support vector regression Support vector machine (SVM) is a supervised learning method that can be used for both classification and regression tasks. SVR is a powerful nonlinear regression method, and introduces nonlinear kernel to model nonlinear function of multiple variables and uses insensitive loss function to accept small deviation from objective function18. SVR identifies support vectors and enables the model to capture the underlying regression function effectively. SVR has been proven to be effective and powerful in real value function estimation19. For a given training data set , where refers to the vector of characteristic parameters and refers to corresponding true target value, a linear regression function is used to fit the sample data, assuming that all sample data can be fitted with permissible error, given by18 where is the weighted combination of individual parameters, is a fixed feature-space transformation of the input , is the intercept, and is the permissible error. The optimization goal is to minimize the following regularized function: where is a penalty parameter, is the error between predicted and true value with the tolerance , given by As the fitting error is allowed within a certain range, the upper and lower slack variables , are introduced so that the optimization goal is to minimize the loss function written as follows: with the condition where is a specified constant, also known as the penalty parameter, used to represent the penalty degree of the sample when it exceeds the permissible error. Slack variables , represent the training errors on the upper and lower sides of the regression function respectively. Since Lagrange multiplier method can be used to solve extreme value problems under constrained conditions, the above objective function can be converted into a Lagrangian: where , , and are Lagrange multipliers. Setting the derivative of the above Lagrangian with respect to , , and to zero gives: And the corresponding vector is the support vector when either or does not equal to 0. Assuming , it can be deduced that: where is the kernel function. The prediction of can be derived by Prediction results of SVR Training and test results of SVR-MDR model In this section, earthquake response data of SRC buildings in NDE1.0 is used to train SVR-MDR model. The MDR is selected as target output of the model, and the 41 parameters including 6 structure information parameters, 9 spectrum information parameters, 16 duration information parameters, and 10 other earthquake information parameters are selected as input feature. The training and test datasets are split in a 9:1 ratio using the StandardScaler function of the scikit-learn, a free and open-source machine learning library for Python. The normalized probability distribution of MDRs in the training and test datasets is shown in Fig. 1. It demonstrates that the test dataset has almost the same distribution as the training dataset. Normalized probability distribution of MDR in the training and test dataset. Four criteria are introduced to evaluate the predictive performance of SVR model, including the standard deviation of the error (the median absolute relative deviation (MARD) and determination coefficient () and the fraction of the data set whose relative difference is no more than 10% (). The expressions for MARD, , are given by where is the predicted value, and is corresponding true value; where is the mean of . MARD indicates the central tendency of errors. It measures how well the predicted value fits the true value. The value of is used to determine a model’s goodness of fit, and the best possible score of is 1.0 and it can be negative when the model is arbitrarily worse. High value of indicates good fitting effect of the model. The above expression of represents the normalized error between the predicted value and the true value of a sample point in the test set. is the percentage of sample points for which are in the range of [− 10%, 10%]. To obtain the best SVR model, the hyper parameters (C, γ) for Gaussian kernel is determined using grid search and 10-fold cross validation. The n-fold cross validation divides the training dataset into n mutually exclusive and complementary subsets and each time uses n-1 subsets as training set and the remaining one subset as testing set, and the optimal hyper parameters are determined by selecting the model giving the minimal mean squared error (MSE) for the all n subsets. The predictive performance for the training and test datasets are shown in Fig. 2a,b respectively. The solid line of 1:1 represents that the predicted value is equal to the true value, and the dashed red lines represent one standard deviation. The prediction model generally approaches a 1:1 linear proportional relationship. The standard deviation of errors for training and test dataset are 0.181 and 0.202 respectively. The calculated for the train and test datasets are 97.1% and 91.2% respectively, indicating that the SVR-MDR model has high accuracy in predicting MDR. Distribution of predicted and true MDR using SVR-MDR model with all 41 input features (a) Training dataset; (b) Test dataset. The performance of the SVR-MDR model is compared with two other machine learning models: kernel ridge regression (KRR) and decision tree (DT). Radial basis function kernel is used in the KRR and the hyperparameters are chosen as α = 1, γ = 0.01, and the maximum depth of the decision tree is 10. The results for KRR model and DT model for the training and test sets are presented in Figs. 3 and 4 respectively. The predictive performance metrics of SVR-MDR, KRR and decision tree are shown in Table 4. The KRR model exhibits greater variability, particularly when predicting larger MDRs, indicating lower prediction accuracy than our SVR-MDR model. Though the DT model shows much smaller dispersion, the significant difference between the standard deviations of prediction error of the training and test sets suggests potential overfitting. In comparison, the SVR-MDR model not only demonstrates superior predictive accuracy but also maintains greater stability across both training and test sets, highlighting better performance in predicting MDR than KRR and DT models. Distribution of predicted and true MDR using KRR model with all 41 input features (a) Training dataset; (b) Test dataset. Distribution of predicted and true MDR using DT model with all 41 input features (a) Training dataset; (b) Test dataset. Table 4Predictive performance metrics of SVR-MDR, KRR and DT.MetricsSVR-MDRKRRDTTrain setTest setTrain setTest setTrain setTest set 0.9210.9040.7260.7350.9620.976D10%97.1%96.4%95.9%94.0%99.8%97.5%MARD0.0230.0250.0390.0440.0130.028 Predictive performance metrics of SVR-MDR, KRR and DT. Training and test results of RSVR-MDR model There are 41 input features in the initial SVR-MDR model, and each feature exerts different degrees of influence on the SVR-MDR performance. To quantify the importance of each feature, the SHAP method proposed by Lundberg and Lee is employed20. SHAP is an explainability tool for ML aiming to evaluate the impact of each feature. It employs Shapley value as the evaluation index, which reflects the average change in the prediction outcome when a specific feature is added to all possible combinations of other features20. Due to the relatively high computational cost of applying SHAP for SVR-MDR, 604 sets of data, about 10% of total SRC earthquake response data, are randomly selected. The absolute SHAP values of each feature on these 604 samples are ranked as shown in Fig. 5. SHAP values of 41 input features: (a) structural information; (b) other earthquake information; (c) spectrum information; (d) duration information. The fundamental frequency f1 and co-seismic minimum frequency f2 contribute most to the prediction in the 6 structural information parameters. Magnitude, with the highest SHAP value of 85.71, contribute most to the prediction, indicating magnitude’s dominant influence on MDR prediction. In the spectral information parameters, the average of pseudo-spectral acceleration (Avg_Sa) between the fundamental f1 and co-seismic minimum frequency f2 contributes most to prediction. DSa2, the duration corresponding to (5–95)% of total energy associated with ground motion acceleration, contributes most to prediction in the 16 duration information parameters. It is noted that the selection of input feature should not only be based on the SHAP value rankings, but also on the practical availability of input feature. For example, although the fundamental frequency f1 has lower SHAP value than the co-seismic minimum frequency f2, f1 is more accessible than f2. Since the former can be easily obtained by empirical formula or modal analysis; on the other hand, f2 necessitates building response records under specific ground motion21. Consequently, co-seismic minimum frequency and dependent parameters, such as SA2, SV2, SD2, Avg_Sa, Avg_Sv, Avg_Sd, are not easily obtained after earthquake and are not considered in the reduced SVR model. A total of 10 input features are finally selected based on their significance and ease of acquisition, including 3 building information parameters structure height H, number of stories N, and fundamental frequency f1, 4 other earthquake information parameters PGV, PGD, epicentral distance Dep and magnitude M, and 3 spectrum information parameters SA1, SV1, SD1. The RSVR-MDR model is then trained with these 10 input features. The predictive performance of the RSVR-MDR model is shown in Fig. 6. The standard deviation of errors for training and test dataset are 0.192 and 0.213 respectively, almost the same as those in the SVR-MDR model. The calculated for the train and test datasets are 97.0% and 95.7% respectively, also almost the same as those in the SVR-MDR model. In this regards, the RSVR-MDR with 10 features (H, N, f1, PGV, PGD, Dep, M, SA1, SV1, SD1) has almost the same predictive performance as the SVR-MDR model with all the 41 features. It is noted that although duration information parameters DSa2, and DSv2 have higher SHAP value than building information parameters such as structural height H, number of stories N, the corresponding predictive performance with 10 features (DSa2, DSv2, f1, PGV, PGD, Dep, M, SA1, SV1, SD1) is worse than that using 10 features (H, N, f1, PGV, PGD, Dep, M, SA1, SV1, SD1), as shown in Fig. 7. In this regards, in the combination of multiple features, the building information parameters are more important than duration information parameters. Distribution of predicted and true MDR using RSVR-MDR with 10 selected input features (H, N, f1, PGV, PGD, Dep, M, SA1, SV1, SD1) (a) Training dataset; (b) Test dataset. Distribution of predicted and true MDR using RSVR-MDR with 10 selected input features (DSa2, DSv2, f1, PGV, PGD, Dep, M, SA1, SV1, SD1) (a) Training dataset; (b) Test dataset. Comparison with the conventional methods Conventional empirical prediction methods for building response To show the advantage of the proposed RSVR-MDR model, we compare the proposed model with single parameter and multi parameters linear regression models. As pointed out by Perrault and Guéguen15, the EDPs follow a lognormal distribution for a uniform distribution of ground motion parameters, and the single parameter linear regression model for predicting MDR is given by where and are coefficients obtained by linear regression, is the residual variability, and IM is the intensity measure. The predictive models using PGV, PGD, SD1 are denoted as PGV-MDR, PGD-MDR, SD1-MDR. Perrault and Guéguen15 studied the relationship between IMs and EDP using data obtained from CESMD and proposed an empirical prediction method employing not only PGA, CAV but also spectrum information parameters, and the multi parameter linear regression model for predicting MDR is given by where to are coefficients, and the model is denoted as PG2015 hereinafter. It is essential to recognize that the models developed by Perrault and Guéguen15 used different datasets from ours, and this variation could introduce systematic biases when applied to our specific dataset. To mitigate this issue, we establish the corresponding single parameter and multi parameters linear regression models based on the NDE1.0 dataset. Comparison results and discussion The prediction results of MDR using PGV-MDR, PGD-MDR, SD1-MDR and PG2015 models are shown in Fig. 8. The standard deviation of errors for PGV-MDR, PGD-MDR, SD1-MDR and PG2015 are 0.276, 0.285, 0.265 and 0.243 respectively, while the standard deviation of error for RSVR-MDR is 0.213 on the test dataset. The other three metric , and for assessment prediction performance is shown in Table 5. It can be found that RSVR-MDR has better prediction performance than single parameter models PGV-MDR, PGD-MDR, SD1-MDR and multi parameters linear regression models PG2015. Although the PG2015 model has the closest predictive performance to our proposed model, considering the difficulty in obtaining parameters Avg_Sv that depends on f2, the proposed RSVR-MDR demonstrates much higher potential real application for predicting buildings’ seismic response after earthquake. Predictive performance results of (a) PGV-MDR; (b) PGD-MDR; (c) SD1-MDR; (d) PG2015 model. Table 5Predictive performance metrics of RSVR-MDR and other regression models.MetricsRSVR-MDRPGV-MDRPGD-MDRSD1-MDRPG2015Train setTest set 0.9120.8810.8170.8040.8300.858D10%97.0%95.7%88.8%88.0%91.0%91.5%MARD0.0250.0260.0400.0370.0320.028 Predictive performance metrics of RSVR-MDR and other regression models. Prediction performance of SVR models for other building types The SVR-MDR and RSVR-MDR models are verified on other building types to assess their predicting performance. For each type of building, the training and test datasets are also split in a 9:1 ratio. Two ways are employed to process building type during training. One way is to train a model using the entire dataset, treating the typology as a feature; the other way is to train separate model using data from each typology individually. The results are shown in Figs. 9 and 10. The predictive performance metrics , MARD, , for the SVR-MDR and RSVR-MDR model treating building type differently are displayed in Table 6. The results demonstrate that training separate model for each type achieves higher accuracy than training one model with building type as an input feature. Predicted and true MDR of other type of buildings using SVR-MDR. (a) Trained entirely; (b) Trained separately. Predicted and true MDR of other type of buildings using RSVR-MDR (a) Trained entirely; (b) Trained separately. Table 6Predictive performance metrics of SVR-MDR and RSVR-MDR trained in two ways.ModelsBuilding typesTrain entirelyTrain separately D10%MARD D10%MARDSVR-MDRSteelTrain set0.2890.79588.4%0.0350.1100.97299.4%0.021Test set0.2900.81190.1%0.0390.1980.87794.9%0.029RCTrain set0.3380.61485.9%0.0440.2030.86295.9%0.023Test set0.3470.58884.4%0.0450.2440.77593.2%0.030MasonryTrain set0.4450.09868.5%0.0690.1460.93298.2%0.025Test set0.5730.33943.5%0.1060.320-0.54769.2%0.057WoodTrain set0.2100.78895.0%0.0330.0930.957100.0%0.025Test set0.384-2.23075.0%0.0490.279-1.63266.7%0.020RSVR-MDRSteelTrain set0.3190.75085.7%0.0400.2070.89996.1%0.024Test set0.2940.80588.1%0.0370.2020.87198.0%0.026RCTrain set0.3410.60686.8%0.0400.2510.79092.8%0.028Test set0.3430.59784.4%0.0440.2730.71990.6%0.033MasonryTrain set0.3550.42574.8%0.0530.3500.60877.9%0.049Test set0.6830.06347.8%0.1280.286-0.24276.9%0.046WoodTrain set0.1960.81690.0%0.0320.1040.946100.0%0.025Test set0.596-6.77575.0%0.0740.323-2.51766.7%0.018 Predictive performance metrics of SVR-MDR and RSVR-MDR trained in two ways. Conclusion In this paper, first a full-feature SVR-MDR model is trained using data of SRC structures from the NDE1.0 database to predict the MDR of buildings. Then a reduced-feature RSVR-MDR model with much less input features that are selected based on the contribution and accessibility. The performance of SVR-MDR and RSVR-MDR is compared against conventional loglinear methods. Additionally, performance of models with data of other types of building is analyzed and discussed. We can find that (1) the SVR-MDR and RSVR-MDR models demonstrate high accuracy in predicting the MDR of buildings, significantly outperforming traditional loglinear regression methods; (2) the reduced-feature RSVR-MDR model offers nearly the same predictive performance as the full-feature model, while requiring fewer input parameters that are easily accessible; (3) models trained separately on specific building types provide better predictive accuracy than those trained using data from all building types combined. We believe the proposed RSVR-MDR model can give accurate and rapid MDR predictions and provide important data for near real-time seismic damage assessment and post-earthquake decision-making. Although the proposed SVR models have better predictive performance than KRR and DT models, for a more comprehensive comparison, the SVR models need to be compared with other ML methods, especially ensemble learning methods such as gradient boosting machine, extreme gradient boosting. The authors believe that there will be more research using ML or deep learning methods to fully utilize NDE1.0 to predict structural responses in the future. Another limitation of SVR models is that we do not consider sufficient structural information, such as high-order frequencies of the structure, presence or absence of infill walls, building construction age, and seismic fortification intensity of the building. If the above structural information is considered, the SVR models may further reduce the prediction variance. However, due to the limited availability of these structural information, we can only use the basic information such as B_lat, B_long, H, N, f1 and f2 provided by NDE1.0 to establish the optimal model. For further research, we will use the original earthquake records and structural information from CESMD and BRI, conduct modal analysis to obtain high-order frequencies and high-order response spectra, and establish ML models that consider more high-order structural information as features. Acknowledgements This research was partially supported by Scientific Research Fund of Institute of Engineering Mechanics, China Earthquake Administration (Grant No. 2021B08, 2022C07), and the Natural Science Foundation of Heilongjiang Province of China (Grant No. LH2022E119). The authors would like to acknowledge URBASIS program led by ISTerre/Université Grenoble Alpes for providing the NDE1.0 flatfile. Author contributions D.T., Q.M., and J.L. made substantial contributions to the conceptualization and methodology. D.T., S.F. and J.W. made substantial contributions to the formal analysis and visualization. D.T., H.L. and S.F. made substantial contributions to the writing-original draft, review and editing. All authors read and approved the final manuscript. Data availability Data supporting the findings of this study are available from https://www.isterre.fr/annuaire/pages-web-du-personnel/philippe-gueguen/new-earthquake-data-recorded-in-buildings-nde1-0/article/acces-to-the-flatfile.html provided by URBASIS program led by ISTerre/Université Grenoble Alpes. The free and open-source machine learning library for Python scikit-learn is used to establish support vector regression models, available from https://scikit-learn.org/stable/index.html. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement Data supporting the findings of this study are available from https://www.isterre.fr/annuaire/pages-web-du-personnel/philippe-gueguen/new-earthquake-data-recorded-in-buildings-nde1-0/article/acces-to-the-flatfile.html provided by URBASIS program led by ISTerre/Université Grenoble Alpes. The free and open-source machine learning library for Python scikit-learn is used to establish support vector regression models, available from https://scikit-learn.org/stable/index.html.