Probabilistic methods, such as fragility analysis, have been developed as a promising alternative for the seismic assessment of dam-type structures. However, given the costly reevaluation of the numerical model simulations, the effect of the model parameters likely to affect the seismic fragility of the system is frequently overlooked. Acknowledging the lack of the thorough exploration of different machine learning techniques to develop surrogates or metamodels that efficiently approximate the seismic response of dams, this study provides insight on viable metamodels for the seismic assessment of gravity dams for use in fragility analysis. The proposed methodology to generate multivariate fragility functions offers efficiency while accounting for the most critical model parameter variation influencing the dam seismic fragility. From the analysis of these models, practical design recommendations can be formulated. The procedure presented herein is applied to a case study dam in northeastern Canada, where the polynomial response surface of order 4 (PRS O4) came up as the most viable metamodel among those considered. Its fragility is assessed through comparison with the current safety guidelines to establish a range of usable model parameter values in terms of the concrete-rock angle of friction, drain efficiency, and concrete-rock cohesion.