The Varamin aquifer is crucial in central Iran, with groundwater level decreased by over 35 m over the last three decades. The sustainability of water supply in Tehran and Varamin regions relies on the proper management of this aquifer. This study aims to simulate and predict the groundwater levels in the aquifer based on a combination of machine learning and metaheuristic algorithms. Surface discharge, hydraulic conductivity, and groundwater levels were initially employed to cluster the aquifer, thereby identifying homogeneous aquifer areas and selecting a base observation well within each cluster. An artificial neural network (ANN) model was then used to simulate groundwater levels in each cluster. Input variables such as GWL with one- and two-month lags, monthly precipitation, monthly temperature, monthly evaporation, and monthly aquifer withdrawal were considered. An arithmetic optimization algorithm (AOA) was developed to optimize the simulation performance of ANN. Results showed that the aquifer can be divided into four clusters that each cluster exhibited significant annual decline in groundwater level ranging from 0.5 to 1.0 m. The ANN-AOA hybrid model demonstrated reasonable performance in estimating groundwater levels in the clusters, with MAE and RMSE values of 1.46 m and 1.69, respectively, for the cluster1, 0.57 m and 0.37 m, respectively, for the cluster2, 0.21 m and 0.29 m, respectively, for the cluster3, and 0.34 m and 0.47 m, respectively, for the cluster4. The hybrid model improved the performance of the ANN in clusters remarkably, reducing the RMSE and MAE values by 20-30 cm and 30-35 cm, respectively. This study demonstrates the effectiveness of machine learning models optimized by metaheuristic algorithms in managing groundwater resources, which can be applied to similar areas.