Comparative analysis of regression techniques for dual-fidelity surrogates in concrete gravity dams

The probabilistic seismic assessment of a structure considers inherent uncertainties in seismic load and material variables, which are challenging for traditional deterministic approaches to address. However, hydraulic structures under extreme loads can fall into the nonlinear domain, with responses influenced by foundations and reservoirs. Therefore, accurately representing structural systems requires highly complex and time-consuming models, limiting the ability to conduct numerous analyses required by probabilistic approaches. This limitation often leads to choosing between a smaller set of high-fidelity analyses or adopting a lower-fidelity model to generate the necessary number of observations. When compared to high-fidelity models, lower-fidelity models are less precise and increase the epistemic uncertainty associated with probabilistic studies. Nevertheless, lower-fidelity models are easier to implement while maintaining important information about the structure response. For this reason, these models are frequently adopted in research and industry. This study identifies the best of 181 machine learning regression algorithms to convert accessible lower-fidelity observations into high-fidelity equivalent predictions. Geometric variations are considered alongside material and seismic uncertainties to enhance robustness and ensure that a single algorithm can be used for the evaluation of different dams without repeating the procedure for different geometries. To assess the applicability of the algorithms and identify potential limitations, fragility curves were generated via the proposed methodology for studying dams located in Eastern and Western North America. These predicted fragility curves were compared with those generated via only high-fidelity analysis. The comparison demonstrates that the proposed methodology successfully captures variations caused by the considered uncertainties.