Transformer-Based Generative Model for Automated Field Development Planning and Optimization in Hydrocarbon Reservoirs

As global demand for energy is increasing, exploration and new discoveries of oil and gas reservoirs continue, with optimal field development becoming more critical. Field development planning requires a comprehensive understanding of the reservoir, which usually involves developing computer-based simulation models and using time-intensive optimization techniques. Recent progress in generative artificial intelligence (AI) indicates potential to support field development optimization (FDO). With this research, we investigate the application of generative AI in FDO by developing a transformer-based generative model for generating optimal future wells that maximize the net present value (NPV). A new transformer architecture is developed, and the corresponding model is trained and applied to generate two outputs: (i) a probability map, providing spatial scores that guide the optimization of well types and locations, and (ii) a well control map that specifies the optimal operating settings at each location. The architecture includes an encoder module, which captures complex patterns and dependencies within the numerical simulation data to support well placement decisions. It also includes an embedding layer that converts raw data into dense vectors and encoding layers that add spatial and temporal context to improve model understanding of reservoir dynamics. The proposed approach was evaluated using the PUNQ-S3 benchmark reservoir simulation model using various input scenarios with different numbers and configurations of existing wells. The results demonstrate promising performance, providing instant predictions that closely align with the outcomes of the differential evolution (DE) optimization algorithm coupled with high-fidelity numerical simulation outputs, which demonstrate a strong understanding of the reservoir system behavior.