oa Real-Time Monitoring of Plume Migration with Distributed Pressure Sensors Using Machine Learning

This work introduces a general-purpose inversion model. A model capable of addressing various reservoir types is developed, captured by a very broad training set. A machine learning model is trained with the U-FNO architecture developed in Wen et al (2022) to predict a single deterministic full saturation map over time using input well log data and multilevel pressure measurements. This direct inversion approach is then used to generate a stochastic ensemble of full saturation maps over time representing a posterior. The direct inversion approach is shown to be effective, achieving close statistical agreement with the ground truth. The plume error over the 2D radial 500 sample test set was 1.72%. The benefits of this approach over traditional history matching are computational efficiency, the ability to generalize out of sample, and not being dependent on priors. The produced ensemble of saturation maps can learn the height and footprint of the plume and reasonably reconcile observed pressure data with predicted pressure data over time while incorporating uncertainty quantification. This approach enables real-time plume monitoring by providing accurate envelopes of saturation maps.