Poster Presentations:
PhD Student
Stanford University
Presentation Title: Modeling uncertainty of full-scale specimens that employ spines and force-limiting connections
Co-Author: Barbara Simpson
Abstract: To tackle the challenge of modeling uncertainty in lateral force-resisting systems, particularly those employing less established components, a global sensitivity analysis is deployed. The study employs numerical models of full-scale physical tests conducted on specimens equipped with elastic spines and force-limiting connections at Japan's E-Defense shake table. By varying primary models and generating secondary models within predetermined uncertainty groups, both model class and parameter uncertainties are comprehensively addressed. The focus lies in estimating peak story drift ratio and floor acceleration through numerical means and comparing them with experimental results. Metrics like coefficient of variation, Pearson correlation coefficient, and Sobol index shed light on the impact of different model features on result dispersion. This research elucidates the pronounced sensitivity of peak floor acceleration to modeling uncertainty, especially influenced by assumptions in spine-to-frame connections. The implications extend to shaping future design methodologies for enhanced lateral-force resisting systems.
PhD Student
University of California, Los Angeles
Presentation Title: Global Sensitivity Analysis of a Bridge Column Featuring SMA and ECC: considering variations in material properties
Abstract: The development of advanced materials that are more ductile, damage-tolerant, and have self-centering capabilities than conventional materials has led research efforts to focus on creating seismically resilient structural members. To that end, a variety of novel bridge column concepts that utilize emerging advanced materials have been developed and studied over the past few decades. One of the successful applications of novel bridge columns is where advanced materials such as Shape Memory Alloys (SMAs) and Engineered Cementitious Concrete (ECC) are included in the plastic hinge region of the column. This novel column concept leverages the self-centering capability of SMA and damage-tolerant properties and the high tensile capacity of ECC, allowing for low residual drift response and minimal plastic hinge damage even under large earthquakes. Even though the SMA-reinforced ECC column is a proven seismically resilient bridge column concept, identifying the effect of key material model parameters that define the hysteretic response of SMA and ECC on the global response is necessary for the optimal design of this novel concept or determining the extent of material uncertainty for numerical simulation of structures that include SMA and ECC. This study aims to investigate the effect of variation in mechanical properties of SMA and ECC on the seismic response of a bridge column featuring SMA and ECC materials in its plastic region. QuoFEM, the SimCenter tool, is planned to be used to conduct sensitivity analyses to quantitatively measure the influence of material modeling parameters on the seismic response of the column.
PhD Student
University of California, San Diego
Presentation Title: Surrogate Modeling of Nonlinear Structural Systems with Long Short-Term Memory (LSTM) Networks for Probabilistic Performance-Based Seismic Assessment
Co-Authors: Joel Conte; Zhen Hu
Abstract: Performance-based seismic design of structural systems depends on computationally expensive high-fidelity finite element (FE) models for predicting the structural response to seismic excitation. For risk-based assessment, the FE simulations must run thousands of times at different realizations of the uncertainty sources, such as uncertain earthquake ground motions and uncertain structural model parameters. Consequently, data-driven machine learning (DDML) surrogate models have gained prominence as fast emulators for predicting structural response in probabilistic analyses. This paper leverages Long Short-Term Memory (LSTM) networks, known for capturing dynamic causal inference, as global surrogate models. Two LSTM variants, the sequence-to-sequence (Seq-2-Seq LSTM) model and the Non-linear Autoregressive Model with Exogenous Input (NARX-LSTM), were evaluated as efficient metamodels for accurate structural dynamic response prediction at minimal computational cost. LSTM surrogate models are trained using band-limited white noise (BLWN) ground motion sequences of different magnitudes to emulate the seismic response behavior of a nonlinear inelastic multi-degree-of-freedom (MDOF) system modeled using the Giuffrè-Menegotto-Pinto (GMP) steel material constitutive model. Next, LSTM models are used as emulators for a reinforced concrete (RC) structural wall by training and testing the LSTM models on experimental data from a quasi-static reversed-cyclic test of the RC wall. An OpenSees model of the wall under quasi-static and dynamic loading is also emulated with LSTM models. Using surrogate modeling, three seismic risk assessment scenarios are explored. The first scenario fixes the ground motion and considers uncertainty in structural model parameters only. The second scenario fixes the structural model parameters and accounts for the earthquake ground motion uncertainty. Finally, the third scenario considers uncertainty in both the earthquake ground motion and the structural model parameters.
Senior Research Lecturer
Teesside University and University College London (UCL)
Presentation Title: Bayesian Neural Networks based Structural Demand Estimation Surrogate Models
Co-Author: Adam Zsarnóczay
Abstract: This study aims to develop efficient and accurate surrogate models for estimating seismic demand of steel moment frame (SMF) buildings using Bayesian neural networks (BNNs). Seismic demand estimation is crucial for designing and analyzing structures in all three-temporal phases of an earthquake event, including pre-event (for seismic demand hazard analysis and preliminary designs), during-event (for early warning and rapid response), and post-event (for risk management and retrofitting analysis). However, traditional methods for estimating seismic demand, such as using finite element models or mechanics-based models, can be time-consuming and computationally expensive. Additionally, for probabilistic decision-making, it is often necessary to have a fast and reliable estimate of the structural demands rather than the precise values derived from physics-based models. To overcome these challenges, the study uses surrogate models, which provide a more efficient and cost-effective solution. Non-linear time-history analyses (NLTHA) are conducted on ~10,000 SMFs, which are randomized by sampling different design parameters, loading conditions, and geometrical and material properties. Hence the analyzed SMFs encompass different design and capacity levels. These SMFs are analyzed under a statistically balanced dataset of ground motions on TACC-Stampede high performance cluster using NHERI Simcenter’s Design-Safe framework. For each NLTHA, the peak interstory drift ratio, peak roof drift ratio, peak floor acceleration spectrum, and peak base shear force are recorded as the target demand parameters. The resulting big data is used to train hierarchical BNNs with different groups of hyper-priors. The hierarchical approach takes into account building-to-building, record-to-record, and event-to-event variabilities to capture different sources of uncertainties and provide more accurate predictions of seismic demands. The resulting surrogate models can be used to develop city-wide structural demand-based shake maps after an earthquake event to locate vulnerable areas of the city under specific earthquake conditions. To improve engineering interpretability, the models are explained using explainable artificial intelligence (XAI) based Shapley additive explanations (SHAP).
Research Professor
Johns Hopkins University
Presentation Title: Seismic risk assessment of structures using manifold learning-based surrogate modeling
Co-Authors: Alexandros Taflanidis; Michael Shields
Abstract: Performance-Based Earthquake Engineering (PBEE) is a framework for the seismic design and evaluation of structures that focuses on predicting the performance of a structure during an earthquake. Modern PBEE, considers uncertainties, both aleatory (due to natural ground motion record variability) and epistemic (due to modeling assumptions and idealizations). Propagating the various sources of uncertainty requires repeated model evaluations which allow for the calculation of the probability distribution of the engineering demand parameters (EDPs). However, even with the aggregation of computational power into supercomputers and the development of sophisticated solvers, calculating the statistics of the EDPs can be computationally expensive, especially for problems that entail the detailed characterization of performance/risk using nonlinear dynamic analyses. We propose a surrogate modeling framework based on nonlinear manifold learning to accelerate the quantification of seismic risk of structural systems in PBEE The method utilizes Grassmannian diffusion maps, a nonlinear dimension reduction technique and geometric harmonics to learn a functional between the input parameter space to the high-dimensional response space. A nine-story steel moment-resisting frame with stochastic structural properties is used as a testbed. We select the seismic fragility surface as a measure of the structure's seismic performance since it provides an estimate of the probability of entering specified damage states for given levels of ground shaking. To assess the seismic risk of the model we are employing Incremental Dynamic Analysis (IDA), a computationally demanding tool for evaluating the variability in the seismic demand and capacity of non-deterministic structural models.
Associate Professor
University of Utah
Presentation Title: Identification of Main Predictors of Collapse Capacity on Steel Buildings using Several Sensitivity Analysis Techniques.
Co-Author: Prakash Gaire
Abstract: Machine learning (ML) models were used to predict global collapse of steel moment resisting frame (MRF) buildings subjected to seismic loading, and the results used to assess the pros and cons of different sensitivity analyses techniques. Five steel MRF buildings of varying heights were evaluated, considering nonlinear behavior and uncertainties. Incremental dynamic analyses were used to generate a database of more than 25,000 nonlinear analyses, which was used to train and test ML algorithms, such as logistic regression, random forest, AdaBoost, K-Nearest Neighbors (KNN), and XGBoost, among others. Then, confusion matrices identified the methods providing more accuracy, precision recall, and F1-score values. Thereafter, several sensitivity analysis techniques were applied on the results from the most promising ML methods to identify the features with the largest contributions to collapse, including information gain, permutation importance, and global sensitivity analysis (GSA). The permutation feature analysis calculated the model's performance reduction when the values of a single feature were randomly shuffled, resulting in outcomes similar to those of information gain, where ground motion features controlled collapse. Conversely, the GSA modified all the features at the same time, and the average of 100 realizations was considered. First, the Saltelli sampling method was used, assuming uniform distribution between default lower and upper bounds. Then, bounds at mean ± one standard deviation were considered, as well as the original distributions. For both bounds, the GSA sensitivity indices from a Sobol analysis were compared to those obtained from information gain and permutation feature techniques.
Postdoctoral Scholar
University of California, Berkeley
Presentation Title: High-dimensional forward uncertainty quantification using surrogate model extracted from dimensionality reduction
Co-Authors: Sang-ri Yi; Ziqi Wang
Abstract: Uncertainty quantification of physical systems with high-dimensional uncertainties remains challenging, especially when the underlying computational models involve high-fidelity simulations. Surrogate models combined with dimensionality reduction are often used to handle this problem, but the applications can be limited when the input uncertainties are genuinely high-dimensional. To address this challenge, this study presents a framework to construct a stochastic surrogate model from the results of dimensionality reduction in forward uncertainty quantification. The overarching hypothesis is that the high-dimensional input space augmented by the output of a computational model admits a low-dimensional representation. This assumption can be met by numerous uncertainty quantification applications with physics-based computational models. The proposed framework extracts a surrogate model from the results of dimensionality reduction in the augmented input space, which differs from a sequential application of dimensionality reduction followed by surrogate modeling. This feature becomes desirable when the original, non-augmented input space is genuinely high-dimensional. The final product of the proposed framework is a stochastic surrogate model that approximates the original computational model. The proposed approach is demonstrated through uncertainty quantification problems with high-dimensional input uncertainties and will be compared to surrogate models in the Quantified Uncertainty with Optimization for the Finite Element Method (quoFEM) tool provided by the SimCenter.
Assistant Professor
Rensselaer Polytechnic Institute
Presentation Title: Surrogate-based Seismic Risk Assessment of Large-scale Transportation Networks Considering Component Damage Correlation
Co-Author: Qiling Zou
Abstract: Seismic risk assessment of transportation networks plays a crucial role in developing effective risk mitigation strategies. Simulation-based approaches such as Monte Carlo method are often used to propagate the uncertainties from intensity measure and damage state associated with each component within the network. However, this presents significant computational challenges as the network performance needs to be evaluated many times, particularly when dealing with large-scale transportation networks. In addition, due to common engineering practice usually adopted for components in close proximity, damages experienced by different components, depending on their intensity levels, can be correlated. Nevertheless, such damage correlation was either disregarded or merely addressed in oversimplified ways in many existing seismic risk studies. To tackle the above challenges, this study proposes an efficient surrogate-based approach to facilitate efficient seismic risk assessment of large-scale transportation networks considering damage correlation between components. More specifically, A graph neural network is trained as a surrogate model of the original network model and then used to speed up network performance evaluation under probabilistic component conditions and seismic intensities. Copula functions are used to characterize the correlation between damage states of components with known marginal distributions. A real-world transportation network in Southern California region is used to demonstrate the efficiency and effectiveness of the proposed approach for the seismic risk assessment.
PhD Student
University of California, San Diego
Presentation Title: Effect of Uncertainty in RC Walls on Seismic Responses of Buildings with Force-limiting Connections
Co-Author: Georgios Tsampras
Abstract: The uncertainty in the reinforced concrete (RC) wall model parameters is propagated to the seismic responses of a 12-story building with force-limiting connections between the floor system and the structural RC walls. Preliminary analysis showed that the seismic response of the longitudinal reinforcing steel in flexure-dominant base-mechanism RC walls significantly influences the seismic response of the walls. Therefore, the reinforcing steel bar behavior is explicitly modeled through a fiber-based modeling approach, in which the response of the steel fibers is simulated using the Giuffrè-Menegotto-Pinto (GMP) uniaxial steel constitutive law in OpenSees. A joint probability density function of the parameters of the GMP material model for the ASTM A615 Grade 60 reinforcing steel available in the literature is used to perform an uncertainty propagation analysis through Monte Carlo simulation. Dozens of thousands of nonlinear numerical earthquake simulations are performed using parallel computation in Python. The uncertainty propagation analysis is also performed for a building model with conventional high-stiffness unbounded-strength connections to evaluate the effects of the inclusion of force-limiting connections on the distribution of the seismic responses. The histograms of the mean peak base shear, base moment, roof drift, roof acceleration, and strains at the base responses are presented. The use of force-limiting connections reduces the mean value of the distribution of the seismic responses and the dispersion of the distribution of the base shear and acceleration responses. This results in higher confidence in the prediction of the seismic responses of the building model with force-limiting connections compared to conventional connections.
PhD Student
University of California, Davis
Presentation Title: Towards the Quantitative Validation and Uncertainty Quantification of Liquefiable Geosystems
Co-Author: Katerina Ziotopoulou
Abstract: The seismic performance assessment of liquefiable geosystems often relies on non-linear deformation analyses (NDAs). The effectiveness of an NDA evaluation hinges on numerous technical and non-technical factors, such as the quality and extent of site investigation, the choice of constitutive model, and the calibration procedure, to only name a few. When experimental data or well-instrumented case histories are available, the process of validation can gauge how well an NDA captures the mechanisms of interest before upscaling or extending to the problem at hand. This comparison can focus on response metrics like accelerations, displacements, and excess pore water pressures. Deviations between simulation results and expected soil or geosystem behaviors have been predominantly assessed qualitatively through the visual inspection of data, thereby lacking rigorous quantification. Furthermore, quantifying the uncertainty in both numerical and experimental responses is essential for understanding the likelihood of expected outcomes. In this study, the experimental measurements of the LEAP-2017 and LEAP-2020 projects, coupled with their numerical simulations, are investigated to quantify uncertainties in both domains for multiple response metrics. The validation process employs two signal processing-based validation metrics to measure the discrepancy between simulations and experiments. Furthermore, the sensitivity of simulation responses to input parameters is evaluated. For the numerical simulations, the PM4Sand constitutive model is adopted. The uncertainties identified in this study could inform further research with uncertainty quantification tools like QUO-FEM and EE-UQ. This information serves to provide a more comprehensive representation of the uncertainties inherent in earthquake soil-structure interactions.
PhD Student
University of California, Berkeley
Presentation Title: Enhancing EDP Generation: Direct utilization of residual drift analysis results
Co-Authors: Adam Zsarnóczay; Dimitrios Konstantinidis
Abstract: In FEMA P-58 performance evaluations, generating statistically simulated Engineering Demand Parameters (EDPs) is the starting point, influencing all subsequent calculations. One set of particularly impactful EDPs is the Residual Interstory Drift ratios (RIDs) since they inform the distinction between cases of building replacement versus restoration. The marginal distribution of residual drifts typically deviates substantially from lognormality, especially at low shaking intensities where the results are zero-inflated. Hence, it is common to utilize the FEMA P-58 residual drift estimation equations to generate statistically simulated RID realizations. This approach, however, discards existing RID analysis results, which better represent the RID distribution when sufficiently detailed analysis models are used. We propose an improved approach for generating simulated EDPs capable of accommodating the zero-inflated RID marginals, enabling better agreement with the structural analysis results. This approach accommodates the zero-inflated nature of the RID marginals while preserving the correlation structure between peak transient and residual drifts. It leads to an accurate representation of the RID marginal variance and does not require an estimate of story yield drift. We demonstrate the efficacy of the proposed method using 5760 analysis results of 18 three-dimensional archetype steel buildings subjected to various shaking intensities. The results suggest that using our method enables the generation of EDP realizations that better represent the distributions of the structural analysis results. Better representation of the RID distribution leads to a more accurate estimation of the probability of building replacement, reducing bias in the loss estimation results.
PhD Student
University of California, Berkeley
Presentation Title: Incorporating Expert Knowledge for Bayesian Model Averaging in Structural Engineering: A Sammon's Mapping Approach
Co-Author: Tracy Becker
Abstract: In numerical analysis, engineers face multiple choices on which numerical elements or modeling assumptions to use. However, the uncertainties coming from these decisions are not well understood. One method is to apply weighting, representing the probabilities, to various model options, and Bayesian Model Averaging (BMA) using experimental data are planned to be developed and used in QuoFEM to determine these weights. Specifying prior model probabilities is a challenging aspect of BMA, especially when data is lacking or incomplete, which is common in structural engineering. Fortunately, models often have a physical or semi-physical basis, which provides a certain level of "expert knowledge" regarding their applicability and predictive capabilities. In the context of applying BMA through the QuoFEM application, this study proposes the use of Sammon's mapping, a high-dimensional information visualization technique, to incorporate and quantify the prior knowledge into the prior model probability using the numerical models under consideration. This technique is employed to project a collection of simulation models onto a two-dimensional map. Subsequently, the relative distances between points on this map are used as proportions to estimate the hyperparameters of the Dirichlet distribution, a commonly used prior distribution in BMA. To demonstrate the effectiveness of the approach, three simulation models related to lead-rubber bearings are considered. The posterior model probabilities obtained using the proposed prior, conditioned on a subset of force-history response data, are compared with those conditioned on a more comprehensive and informative dataset but only employing a uniform prior. When the two align, it implies that adjustments to the prior can, to some extent, compensate for deficiencies in the collective data.