EESD special issue: AI and data‐driven methods in earthquake engineering – (Part 1)

Abstract

The rapid development of artificial intelligence (AI) technologies has become foundational to worldwide multidisciplinary research, thereby establishing “AI for science” as a new paradigm. Due to the vast potential, an increasing number of researchers are investigating the application of AI and data-driven methods to earthquake engineering. In fact, over the last few years, the area has arguably become one of the research frontiers in earthquake engineering. Some of the sub-topics that have been investigated include ground motion characterization, geotechnical engineering, seismic behavior and the health status of engineering structures, and risk and resilience of distributed systems. Despite the explorations to date, additional substantial investigations are needed to facilitate meaningful integration with fundamental earthquake engineering knowledge. Big data, intelligent algorithms, and the availability of large scale computing systems are pillars of AI development. These technologies have provided new opportunities for earthquake engineering research. For example, in the field of ground motion characterization, deep learning models have aided in the discovery of new physical laws. In the domains of structural design and regional earthquake impact assessment, intelligent algorithms are capable of replacing some time-consuming and labor-intensive conventional steps in the overall workflow. An increase in the accuracy of predictive models with embedded AI technologies can lead to safer and more reliable infrastructure systems and communities. However, previous studies have underscored a variety of critical challenges, such as insufficient high-quality data, the need for physics-informed and domain knowledge-infused models, and the importance of explainable AI. This special issue is dedicated to current research on newer and more refined techniques and approaches that cover novel AI and hybrid models that combine machine learning and physics. The topics covered include ground motion modeling, geotechnical engineering, structural analysis, seismic design, and multi-hazard assessment that considers earthquakes and other hazards. Emphasis is also placed on bridging the gap between existing physics-based approaches with new data-driven techniques. As a reflection of the overall level of interest in the topic, we received a diverse set of interesting articles that were considered for publication. After rigorous review, we are pleased to publish approximately 2-dozen articles in this special issue. To ensure the timely dissemination of those articles accepted earlier in the review process, the issue is being published as a pair, that is, Part 1 and Part 2. Part 1, which is presented here, includes the first 13 articles that focus on ground motion characterization, structural analysis, and damage evaluation. Part 2 will be published in the very near future and will cover topics in the areas of geotechnical engineering, risk and resilience, structural analysis and damage evaluation. The first four articles in the Part 1 issue examine the application of AI and data-driven methods to ground motion characterization, site classification, and selection and scaling of ground motion records. Sreenath, Podili, and Raghukanth develop a hybrid machine learning model for predicting earthquake parameters in shallow crustal regions of Europe. A data set of 9682 records corresponding to 512 events and 1331 stations spanning almost all of Europe was collected. The model inputs include features such as moment magnitude (Mw), Joyner-Boore distance (RJB), shear wave velocity (Vs30), and hypocentral depth, while the outputs are ground motion characteristics, including peak ground acceleration (PGA), significant duration (T90), cumulative absolute velocity (CAV), and arias intensity (Ia). The prediction model is a weighted hybrid machine learning model that employs a shallow neural network, deep neural network (DNN), gated recurrent unit (GRU), support vector (SVM), and random forest regression (RFR), with weights calculated using the likelihood (LH), log-likelihood (LLH), and Euclidean distance (EDR) methods. Furthermore, the study explains the uncertainty of the prediction results based on this hybrid model, and compares the proposed non-parametric hybrid prediction model with other regional analysis models. The outcomes demonstrate that the proposed model exhibits superiority, as evidenced by the coefficient of determination (R2) values ranging from 0.84–0.91 for all output variables, reflecting a good agreement between the predicted and recorded values. Ji, Zhu, Yaghmaei-Sabegh, Lu, Ren, and Wen reformulate the site-classification problem as an image recognition task, which is then solved using deep convolutional neural networks (DCNNs). The site topographic slope values and mean horizontal-to-vertical spectral ratios (HVSRs) of the ground motion recordings are combined into a single image that is, used as the input or set of features. Using the images from 1649 sites in Japan, the DCNN model with four convolutional layers is trained with the goal of predicting the National Hazard Reduction Program (NHRP) site classes. The trained DCNN model achieves average recall rates for site class C and D based on five-fold cross validation in excess of 70%. To improve the recall rate for site class E, an ensemble of classifiers is adopted. This resulted in a greater than 60% average recall rate for site class E while maintaining the site classification accuracy based on the original DCNN. The generality of the DCNN model is demonstrated by performing classifications on European sites where approximately 60% accuracy was achieved. Fayaz, Torres-Rodas, Medalla, and Naeim present an interpretable Gaussian process regression (GPR) method for efficiently evaluating ground motion scale factors (SF) with application to steel frames. To quickly compare the engineering demand parameters (EDPs) of steel moment frame structures under scaled and unscaled ground motions, a data-driven surrogate model is developed using GPR. The surrogate model takes Sa (T1) and SF as input, and peak floor acceleration (PFA) and peak story drift ratio (PSDR) as output. Furthermore, the Shapley additive explanation (SHAP) method is used to explain the GPR model's decision-making process, reducing the machine learning model's black-box nature and providing engineering explanations for the model. Finally, the analysis results using the GPR-based surrogate model show that SFs between 0.5 and 3.0 are allowable for PSDR, and SFs between 0.6 and 2.0 are allowable for PFA. The paper by Hu, Liu, and Xie tackles the issue of reducing the computational expense in generating fragility curves within the performance-based earthquake engineering framework. The adopted strategy involves selecting a subset of ground motion records with the goal of producing fragility functions that are consistent with those generated by a much larger record-set. This goal is achieved through the use of dimensionality reduction techniques that extract those ground motion features with the greatest influence on structural performance. Once the new feature space is established, a hybrid genetic algorithm and simulated annealing fuzzy c-means (FCM) clustering are employed to extract a subset of ground motions from the much larger set until there is fragility curve convergence. Two steel frame buildings (4-story and 8-story) and a large set of single-degree-of-freedom systems are used in a case study to demonstrate the proposed methodology. It is shown that the combination of dimension-reduction and FCM-based clustering enables the fragility analysis to be performed with 40%–60% of the larger suite of site-specific ground motions that would typically be used, with minimal loss of accuracy. The following five articles aim to apply AI and data-driven methods to the advancement of structural analysis, including seismic responses, collapse behavior, and reliability analysis. Kim, Kwon, and Song propose a deep learning-based approach named “Earthquake Response Deep neural network (ERD-net)” for near-real-time predictions of the seismic response of hysteretic systems with degradation and pinching effects. First, a new Bouc-Wen class model, termed a modified Bouc-Wen-Baber-Noori (m-BWBN) model, is proposed to introduce the yield strength as an explicit model parameter. The feasible parameter domains are also specified to promote the practical use of the m-BWBN model. Second, a seismic demand database is constructed using nonlinear response history analyses that adopts the m-BWBN model and a large ground motion set. Third, the ERD-net and detailed training methodologies are proposed to learn the effects of the complex hysteretic characteristics on the peak seismic responses. Numerical examples of reinforced concrete (RC) structures are introduced to test the prediction performance and applicability of the proposed DNN model. The results show that the ERD-net outperforms the existing regression-based prediction methods. For example, the ERD-net has a lower relative error of 3.1% when compared to the coefficient method specified in ASCE 41-17, which has a relative error of 7.9%. The source codes, data, and trained models are available for download at http://ERD2.snu.ac.kr. Zhong, Nararro, Govindjee, and Deierlein employ a new surrogate modeling technique, called probabilistic learning on manifolds (PLoM) as an alternative to the time-consuming nonlinear response history analysis (NLRHA). The authors first illustrate the concept and theory of the PLoM algorithm. Essentially, the PLoM method provides an efficient stochastic model to develop mappings between random variables, which can then be used to efficiently estimate the structural responses for systems with variations in design/modeling parameters or ground motion characteristics. Then, the PLoM algorithm is used in 2 case studies of 12-story buildings for estimating probability distributions of structural responses. Both examples show good agreement between the PLoM model estimations and verification data sets. The authors conclude that the PLoM algorithm is a promising approach for modeling structural seismic response. An illustrative case study demonstrated that the proposed method reduces computational costs by a factor of approximately 5, while maintaining a reasonable level of accuracy with a relative error of 5% in the predicted EDPs. Guo, Enokida, Li, and Ikago propose a physics-data combined modeling method for predicting the nonlinear seismic response of structures by introducing the deep residual network into the classical numerical integration method. The deep residual network replaces the classical time-stepper in the time-step numerical integration method to accelerate numerical calculation. Different networks such as MLP (Multi-Layer Perceptron), CNN (Convolutional Neural Network), and RNN (Recurrent Neural Network) are explored in the deep residual network structure, and the learning performance of the time-stepping schemes for the nonlinear state variables of the system is compared and analyzed. The results indicated that CNN and MLP outperform RNN. The proposed Physics-DNN hybridized integration (PDHI) time-stepping scheme has important advantages compared to the current pure data-driven method, such as being flexible in incorporating known time-invariant physics information, effectively circumventing the demand for structural seismic response data, and efficient training and validation. Finally, the method's performance was verified through a series of numerical and experimental examples. Bijelic, Lignos, and Alahi present an integrated methodology for augmenting training samples and generating data-driven estimates of the collapse performance of structures. Described as an automated collapse data constructor (ACDC), the data-augmentation step leverages the understanding of the ground motion properties and their relationship to the structural collapse performance assessment methodology. An important feature of the ACDC is that it is agnostic to the machine learning or statistical model that is, used to estimate the collapse performance parameters. The data-driven collapse classifier (D2C2) facilitates the transformation of the augmented collapse data from a regression to a classification problem. By doing so, the collapse capacity of the structure under consideration is estimated in a manner that is, consistent with the incremental dynamic analysis (IDA) methodology, which is the primary evaluation engine. The integrated ACDC-D2C2 framework is demonstrated using a case study that generates an extensive data set of collapse responses for four-story and eight-story steel moment resisting frames. The results show that the number of ground motions used in the IDA-based assessment can be reduced by more than 50%. Additionally, feature-importance analysis is used to benchmark the data-driven model against engineering knowledge and produce a new period-dependent measure of ground motion duration. Das and Tesfamariam present an efficient reliability analysis method using the probability density evolution method (PDEM) and stochastic spectral embedding (SSE) based surrogate model. The PDEM is used to estimate the structural response's probability density function (PDF), which is derived based on the principle of probability conservation where generalized density evolution equations (GDEEs) are decoupled from the physical system. To reduce the computation burden via approximating the original response surface, the SSE is trained by a few observations and enables output prediction as spectral representation. Three numerical results show that the proposed SSE-based PDEM can estimate failure probability using a very small number of representative points without compromising accuracy compared with Monte Carlo simulation (MCS). For example, 106 samples of random variables are required for MCS. In contrast, the probability of failure using the proposed method matches that of MCS with only 100 to 150 samples. The last four articles apply AI and data-driven methods to damage detection and evaluation problems using images and response records from monitoring. The paper by Miao, Ji, Wu, and Gao focuses on the issue of strength and stiffness degradation of seismically damaged RC flexure-controlled columns. Using images of the damaged columns as the input features, a predictive model is constructed to estimate the degraded strength and stiffness of the damaged columns in a post-earthquake context. A database is constructed using images from physical experiments and the accompanying hysteretic curves. The latter is used to extract strength and stiffness reduction factors. A multi-output DCNN model is constructed that integrates adaptive feature fusion of the strength and stiffness degradation parameters. The performance of the DCNN model in predicting the shear and strength degradation is benchmarked against that of the Japanese Guideline for Post-earthquake Damage Evaluation and Rehabilitation (JBDPA) and another model developed in a previous study. In general, the predictive accuracy of the DCNN model is significantly higher (orders of magnitude reduction in root mean square error) than the other two approaches, demonstrating the overall viability of the image-based strength and stiffness estimation. Wang, Hornauer, Yu, McKenna, and Law address the issue of accurate characterization of building inventories for seismic risk assessment. The focus of their paper is on using artificial intelligence technologies to detect the presence of a soft-story vulnerability. The authors developed an automated pipeline that segments street view images to identify soft-story buildings. The Mask Region-Based Convolutional Neural Network (Mask R-CNN) is used to create building and soft-story detector models. The workflow also includes an algorithm that performs semi-automatic annotation of each image. A data set comprised of 566 images of soft-story buildings and 736 images with non-soft-story buildings is used to train the classification model. The FRONTERA computational platform at the Texas Advanced Computing Center was leveraged for this purpose. Compared to previously implemented methods based on full image classification, the proposed instance segmentation approach produces superior predictive performance across multiple metrics including accuracy, precision, recall, and F1 scores. Asjodi, Dolatshahi, and Burton propose a three-dimensional fragility surface for the rapid assessment of earthquake damage to RC shear wall structures based on visual damage features. The compressed cracking index and crushing variables are the variables of the three-dimensional fragility surface. The core work is mapping image-based damage to these two variables. Cracking and crushing patterns in damage images are separated using convolution kernel-based filters, and shear and flexural cracking are distinguished using Gaussian mixture modeling. Shear and flexural cracking measurements for cracking damage are compressed into a cracking index using principal component analysis (PCA). The crushing measurement is used directly as the crushing variable for crushing damage. The Box-counting method is then applied to generate a three-dimensional fragility surface based on the cracking and crushing features, achieving more than 95% fitting accuracy. This study successfully establishes a correlation between visible structural damage and structural fragility. The paper by Chou, Liu, and Chang uses multi-target deep learning to estimate the story drift ratios and damage levels in building structures using responses measured by acceleration-based sensors. The level of damage is quantified in terms of the ratio between the pre-event and post-event stiffnesses (or “stiffness ratio”). The benefit of the multi-target model is that full profile story drift ratios can be estimated, which captures the spatial correlation of the demands across the building height. The story drift prediction model also utilizes two long short-term memory (LSTM) networks with physics-guided loss functions. The performance of the two models is demonstrated using a three-story shear building. MCS is used to generate realizations of the baseline structure with variations in the structural properties (e.g., stiffness and damping ratio). Response history analyses are performed on the structural model realizations to generate a robust data set of responses that are used as the training samples. The trained model is evaluated using responses generated by an eight-story structure tested using a shake table. Both the story drift and stiffness ratio models produced high accuracy in their predictions, demonstrating the effectiveness of the proposed approach. As noted earlier, the papers presented here represent the first part (Part 1) of this special issue. The second part (Part 2) will be published in the very near future and will feature the application of AI and data-driven methods to geotechnical engineering and risk and resilience problems, in addition to more articles on structural analysis and damage evaluation. The authors have nothing to report.