Greenfield development planning is a challenging problem due to the complex objective function and the large number of variables. While stochastic algorithms and data-driven offline surrogate models have been used to locate and control wells, these methods often require a large number of runs or fail to reach an accurate optimal solution. For this reason, the current study aimed to develop a dynamic surrogate model updated after each evaluation, referred to as an active learning model. In this study, deepEnsemble and gaussian process (GP) algorithms were applied as an active alternative model. This approach was compared with optimization algorithms directly coupled with real simulations, where algorithms such as genetic algorithms (GA), particle swarm optimization (PSO), complex matrix adaptation evolution strategy (CMA-ES), and differential evolution (DE) were used for comparison. The deepEnsemble active model outperformed the other approaches, achieving a net present value (NPV) of more than 1 E+10 and 17% recovery factor (RF) for a field in southern Iraq for a production scenario. The algorithm suggested shutting in two existing wells and drilling 23 new wells. The algorithm was also tested using a water injection scenario; a stable pressure, NPV of approximately 1.28 E+10, and 25% RF was achieved by suggesting drilling 39 new wells, 17 of which were injection wells. The approach has proven effective in dealing with complex field development problems with a minimum number of runs.