For the training set preparation, five different extraction rates and five different inflow rates (in order to account for the variability in groundwater replenishment) were selected such that they include values from low to high ranges. Additionally, an adaptive pumping scheme was applied during the higher extraction rates by deactivating the sea-side pumps in order to reproduce farmers' adaptation on saline water intrusion. If five extraction rates were to be distributed over the 60-year period considering yearly time step, the total number of unique patterns would be 5⁶⁰. The same number of patterns could be generated for five inflow rates as well. If we consider both inflow and extraction patterns together, the number of total patterns would become 25⁶⁰. Clearly, there was an obvious need to reduce the number of patterns. For that purpose, inflow and extraction rates were treated as fixed pairs, meaning, the first extraction pattern would always concur with the first inflow pattern, the second extraction pattern with the second inflow pattern, and so on. This reduced the total number of patterns from 25⁶⁰ to 5⁶⁰. These patterns were then applied to the first 12 years of the 60-year period. In order to further reduce the size of the data, we considered a 3-year time step for the management problem, which essentially implies that the inflow and the extraction rates are constant for 3 successive years. This was a valid assumption in our case mainly for three reasons: (1) we were interested in long-term planning and management, (2) constant pumping rates can arise from water quota systems, (3) since inflow rates depend on the longer time scales, the 3-year duration follows the dynamics of the groundwater recharge more closely. These 12-year patterns with 3-year time step were then repeated four more times to extend the entire management period of 60 years. This technique reduced the number of patterns drastically to 5⁴ = 625. Data from OGS were extracted at a fixed interval of 3 years, thereby creating 20 points for each 60-year pattern. Thus, the final size of the training data became 625 × 20 = 12,500. Similar methodology was applied for generating the other three datasets as well. For testing and validation sets, four different inflow and four different extraction rates were chosen, and the size of each one of the datasets was 4⁴ × 20 = 5,120. Furthermore, the inflow and the extraction rates for testing and validation were chosen such that they lie within their range used in training. The extrapolation set was created with two values of inflow and two values of extraction rates such that one value is lower than the lowest and the other value is higher than the highest used in training. Thereby, the size of the extrapolation set was 2⁴ × 20 = 320. Different inflow and extraction rates used for the data preparation in this study are summarized in Table 3.