Link to Latest Report: Coming Soon
Background:
Current approaches for post-event structural damage assessment rely on the following methods: on-site inspections performed by qualified personnel, which can take many weeks or even months and create tremendous long-term adverse impacts on the community; monitoring systems, based on accelerometer or on remote sensing, which are presently very sparse and cannot provide reliable real-time measures of structural damage; and computation-based methods. The latter include: (i) pre-computed empirical fragility functions based on scalar intensity measures (IMs), which have the potential of providing real-time information on structural performance, but are very coarse and involve significant uncertainties (e.g. “Prompt Assessment of Global Earthquakes for Response” (PAGER) system) (Wald et al., 2010); and (ii) detailed nonlinear time-history analyses, which can provide detailed understanding of the structural behavior, but require computational capabilities that are prohibitive when interested in real-time regional-scale damage assessments. Having efficient computational tools that can provide key information on structural performance in real time or near real-time after a major earthquake would drastically change the current post-event management paradigm and substantially help enhance community resilience to major disruptive events.
Objective:
The objective is to eventually provide ML-based model that can be effectively utilized for a broad range of structures and across areas of different seismicity. Specifically, multi-span bridges with single and multi-column bents adopting reinforced concrete (RC) columns and prestressed concrete (PC) girders will be utilized as case study highway bridges. The definition of the final set of structures will be informed by a literature survey providing data on the bridge typologies (e.g., material, geometry, etc.) most utilized in medium and high seismic areas across the U.S. As detailed later in this proposal, USGS seismic hazard maps will be utilized to inform the selection of the areas and corresponding search criteria for key ground-motion parameters with the goal of developing ML-based models that can be used in both moderate and high seismicity zones.
Scope:
The proposed project includes several tasks
- Task 1
A literature review will be conducted to select a set of (at least) five highway bridges to utilize as case studies for this project. The bridges will have: (i) reinforced concrete columns and feature the design for moderate and high-seismicity zones (e.g., longitudinal reinforcement ratio, volumetric transverse reinforcement ratio, etc.); (ii) single and multi-column bents; (iii) different types of superstructure and column-to-deck connection types; and (iv) both conventional cast-in-place and ABC column-to-bent-cap connections. Reduced-order numerical models will be developed in OpenSees (McKenna, 2011) to conduct nonlinear dynamic analyses. - Task 2
The objective of this task is to establish a format-consistent database of real and simulated earthquake ground motions. The open-access databases currently available to the research community will be used, including the PEER NGA-West2 database (Ancheta et al., 2014) and the Kyoshin Network (K-NET)(NIED, 2019). Effort will be devoted to integrating data from recent large magnitude events not yet cataloged, including the M7.1 Ridgecrest, CA (2019); M7.2 North Maluku, Indonesia 2019; M7 Izmir, Turkey 2020; and the 2023 Turkey–Syria earthquakes. The database of real records will be augmented with a large population of simulated ground motions generated from fully deterministic physics-based wave-propagation models. The use of simulated motions is motivated by the scarcity of near-field real records from large magnitude events. As such, the current real records databases cannot constitute a solid and statistically significant basis for constraining data-driven ML models. The PI has extensive experience in the generation, analysis, and validation of simulated earthquake motions for use in engineering applications. - Task 3
The activities of this task will mainly relate to carrying out nonlinear time-history analyses of the selected bridge models that will be developed and validated in Task 1 with the ground-motion datasets collected in Task 2. Key engineering demand parameters (EDPs), in addition to the drift in the two horizontal directions (transverse and longitudinal to the bridge), will be monitored during the analyses. Such parameters include accelerations, column axial forces, unseating, decks’ midspan bending moment at the decks’ midspan and the deck-to-bent-cap connection, etc. The objective is to identify IMs or combinations of them that show the strongest correlation with the response of the bridges, particularly when subject to near-field strong motions. The output of this task will be a set of EDPs and their location. - Task 4
The objective of this task is to perform feature selection and develop ML models for the prediction of the bridge EDP(s).Feature selection will allow decreasing the modeling computational cost through the reduction of the number of input variables and, most importantly, identify the IM(s) most relevant to a broad range of bridge responses. A key outcome of this activity will be the possibility to contrast the efficiency and sufficiency of several IMs for limit state prediction in bridges subject to near field ground motions. With the approach proposed in this research, this analysis will be conducted with the consideration of highly nonlinear relationships that would not be possible to include with classical regression-based approaches. The use of logistic regression will be investigated, with two standard regularization methods, the L1 and L2 regularization. For the investigation of vector valued IMs, the scaling of records to the primary IM parameter will be performed and then using regression analysis, the effect of additional IMs will be investigated. For the ML-based feature selection, several supervised ML models will be utilized, including Decision Trees and Perceptrons (Géron, 2019). These models will be trained on the data collected in Task 1 and Task 2 and use K-fold cross-validation to perform feature selection from the final models. State-of-the art Deep Neural Networks (DNNs), as illustrated in Figure 3, will be utilized for this task. The model will take as input the ground-motion IM(s) from a single event and will produce the EDP and location. A quick re-training (fine-tuning) structure will be also implemented so that when new bridges and/or ground motions are added to the database, the models can be easily updated. An essential part of this task is to ensure the real-time feasibility of the proposed ML model. The final trained model must be capable of making predictions in real-time or near-real time if the expectation is for these systems to be used in disaster scenarios. As a part of the system evaluation of the models, run-time on low-power systems will be considered as one of the success metrics. - Task 5
The validation will be conducted with the testing subset, counting 20% of the instances, which are independent of the training dataset, (i.e., ground motions that have not been used to train the ML models). This subset will be used to assess the performance of the developed ML models with unknown data. The 20% of the datasets obtained from Tasks 1 and 2 will be selected with a cluster random sampling technique, to ensure a homogeneous sampling over the populations of simulated and real records. These motions (in the form of IMs) will be used as an input for each of the developed structure-specific ML models. The output vectors containing the EDPs and their respective location will be compared with the output vectors obtained from Task 1. Upon completion, this task will allow to assess the capability of the developed ML models to predict the response of 3-D bridges subject to near-field motions as compared to the response obtained from traditional detailed finite element models. In addition to the above evaluation, it will be ensured that the proposed models are generalizable using Cross Validation testing. The training data will be divided into 5 folds, where at any given time 4 will be used for training and 1 will be used for validating the model. The training and validation process will be repeated 5 times, leaving out a different subset each time. This 5-fold Cross Validation will help in ensuring the generalizability of the final model towards a more practical and feasible future implementation.
Research Team:
Principal Investigator: Floriana Petrone
Co-Principal Investigator: Mohamed Moustafa