Hybrid machine learning framework for multi-well trajectory optimization in an unconventional field

The well placement problem in the petroleum industry is usually solved by integrating an optimizer with a reservoir simulation model. This is a time-consuming approach requiring thousands of simulation runs depending on the complexity of the reservoir system. This paper introduces two well placement optimization models: a Fast-Marching Model (FMM) based on fluid flow physics and a Hybrid Model integrating reservoir simulations and FMM results with a gradient boosting algorithm. These models prioritize speed and accuracy when integrated with genetic algorithm optimization, demonstrated using a synthetic unconventional field. From 290 randomly selected locations across the synthetic unconventional field, simulation runs were performed to determine cumulative gas production and used as ground truth. Using reservoir properties at these locations as inputs, Relative Opportunity Ranking (ROR) maps were generated using the FMM and Hybrid Model approaches. The ROR maps were then linked with a genetic algorithm to determine optimal well locations, iteratively updating ROR maps with penalty maps until all wells were appropriately placed. Results indicated that the standalone FMM generated ROR maps were highly correlated with the simulation results. Time complexity analysis revealed that both models were significantly faster than traditional methods, with the FMM independent of simulation and the Hybrid Model requiring only about 290 simulation runs. Ultimately, these models have shown tremendous potential for integration into current reservoir engineering workflows, reducing decision-making times in both greenfield and brownfield scenarios.