Well placement optimization is probably one of the most challenging problems the oil & gas industry has to face particularly for green fields development. The complexity of the task comes from different factors: huge number of possible solutions, increasing size of reservoir models, difficulties to correctly modeling and handling uncertainty, complex well trajectories and operational constraints, … The optimization problem for multiple wells is also extremely hard to solve due to the exponential number of possibilities to test. Although several optimization methods have been proposed in scientific and technical articles, the most used approach in industry remains the manual optimization performed by well trained reservoir engineers. The workflow proposed here aims at assisting reservoir engineers to better perform their choices on main factors affecting the optimal well placement and to be able to test much more well configurations. To this end, a machine learning algorithm is trained on simulated data to be able to evaluate the performance of new possible well locations and configurations. Differently to previous experimental design approaches, complex geological features based on expert knowledge are extracted from the geological model in order to find the main factors driving the production mechanisms. Both properties along well trajectories and properties corresponding to connected reservoir zones are used as features in the machine learning method. Clustering techniques are used to select the training data to simulate used by the machine learning algorithm. The method also handle geological uncertainty that is a critical factor in well placement optimization particularly for green fields. A large number of realistic horizontal well trajectories honoring the different operational constraints is generated, then once the machine learning method is validated a very large number of these trajectories combinations can be evaluated in seconds.
The result of the approach are illustrated in a realistic synthetic green field where the objective is to optimize the position of three wells two producers and one injector under uncertainty (corresponding to several geological realizations). A test on real field case has also been performed, but will be presented in more details in a future work. Data analytics methods allow us to evaluate millions of possible well placement configurations on many geological realizations by performing only a few thousands full field reservoir simulations. We demonstrated this could help engineers placing wells in very profitable positions minimizing risk while maximizing long term production.