A surrogate modeling framework is developed in this study to circumvent the computational burden of high-fidelity numerical groundwater models for arid coastal aquifers. Two different surrogate models, namely, artificial neural network (ANN) and Gaussian process model (GPM) are trained to replace the computationally expensive numerical flow and transport model OpenGeoSys. A novel time-dependent training scheme is introduced which helps the surrogates in tracking the discrete-time state-space trajectories of the high-fidelity model, thereby making them suitable for variable-time simulations. The surrogates are also tested in the extrapolation range corresponding to some extreme boundary conditions such as a very high rate of extraction. Both the surrogates show comparable accuracy in efficiently approximating the numerical model response; however, ANN is found to be much faster than GPM for the size of the data used. The trained surrogates are then used in developing a long-term planning and management framework for analyzing feasible management scenarios in the coastal aquifer of Oman.