Seismic Analysis of Concrete Faced Rockfill Dams Incorporating the Material Deterioration and Construction Sequence

Although concrete-faced rockfill dams serve the community and the ecosystem in many ways, they have the potential to generate catastrophic results in case of a dam breach due to the large water body that is stored in the reservoir. Therefore, it is required to conduct an investigation of possible failures through accurate finite element analysis for the prevention and mitigation of hazards. Although the requirement of utilizing an accurate finite element model was identified in previous studies, the incorporation of the construction sequence has not been highlighted explicitly. Similarly, even though the importance of identifying the material parameters at a dam site has been mentioned in previous studies, the variation of these parameters over the time period has not been investigated. Clearly, the deformations over time indicate significant variations. Due to the impracticality of conducting laboratory tests for existing structures, displacement-based back analysis has been used by previous research studies to obtain the material parameters. However, the accuracy of these algorithms and the machine-learning approaches have not been explicitly discussed. Moreover, the parameter selection and the data set generation for the back analysis have not been justified. Many studies have been conducted to investigate the failure of the concrete face under seismic loading conditions, but minor attention has been given to properly defining the failure of the rockfill layers since laboratory experiments have proved that a failure in the rockfill layers could create massive failures in the concrete face leading to catastrophic failures of the structure. Therefore, a study has been conducted with the focus of identifying the failure of rockfill layers using accurate finite element models with initial configurations that align with the actual deformations of the structure while incorporating the time-dependent variation of the material parameters derived through justified back-analysis methods. Initially, a comparison of deformations resulting with and without construction stage analysis using comprehensive finite element models was conducted. A comparison of the 2D and 3D model results to assess the accuracy of the finite element model predictions was conducted to check the validity of adopting a 2D model for static behavior prediction. The sensitivity analysis was conducted to identify the effect of the material parameter variations on the dam deformations using a data set covering the 9-dimensional hyperspace of the Duncan and Chang parameters with statistical methods and the SHAP methods. A machine learning model comparison was conducted with the same data set to obtain the optimum machine learning model to be included in the SHAP method and the back analysis. The deformation-based back analysis was conducted using a hybrid algorithm including the particle swarm optimization and the identified machine learning model to derive the values of the selected parameters using the sensitivity analysis that produced deformations aligning with the actual deformations. With the observed variation of the deformations during construction and until the current period, the back analysis was conducted for different time periods to derive the time-dependent material parameter variations. Seismic analysis was conducted using the 3D model updated with the optimized material parameters for different magnitudes of the earthquake that was selected based on the design response spectrum of the region. The failure surfaces of the rockfill material were identified using the derived principal stress through nonlinear time history analysis and were compared with the deformation contours for validation. Results showed that the deformations obtained through construction stage inclusion and the nonlinear material model in the FE analysis agree well with the in-situ measurements, warranting the need for such advancements in Finite Element Analysis. The deformations obtained under construction stage analysis of 2D, and 3D models had insignificant variations therefore the 2D model was used for the data set generation. In the machine learning model comparison, the XGBoost algorithm proved to be highly accurate, and the accuracy of the Artificial neural network’s algorithm and support vector regression algorithm predictions were minimal. The sensitivity analysis showed parameters bulk modulus number (Kb), friction angle (∅), density (γ), exponent (n), and modulus number (K) to be sensitive to dam deformations. Furthermore, the SHAP method provided a detailed explanation regarding the impact on the deformations of each material parameter through trained machine learning methods. Therefore, the method could be adopted to conduct sensitivity analysis. From the back analysis, the time-dependent variation of the material parameters was identified where the friction angle (∅) and density (γ) showed minor variations while the bulk modulus number (Kb), exponent (n), and modulus number (K) parameters depicted significant variation over time. Finally, in the seismic analysis, the failure surface generated considering the developed stresses and the safety factors aligned well with the failures defined using the threshold deformations. This study provided a clear framework for the development of finite element models of CFR dams highlighting the significance of incorporating the construction stage analysis in the initial configuration and the accuracy of the model prediction through nonlinear material parameter behavior. Furthermore, through the machine learning model comparison, the XGBoost algorithm was identified to be an optimum model to be utilized in the back analysis. The possibility of conducting a detailed sensitivity analysis using the SHAP method was highlighted in the study through comparison with traditional methods. The derived time-dependent variation of the material parameters was normalized and presented using standardized equations to be incorporated in CFR dam analysis in the future. Finally, the proposed methodology for the failure surface definition under dam breach conditions provides a new approach that can be incorporated towards real-time cascade effects and in the seismic analysis of CFR dams.