From Surface Deformation to Pressure Fields: Contribution of Machine Learning to Cost-Effective CO2 Injection Monitoring.

Abstract

Summary In this paper, we explore the possibility of recovering the pressure distribution in the reservoir from the surface deformation data using a machine learning (ML) approach. We chose to use a convolutional neural network (CNN) with an encoder-decoder architecture as it has proven to be very efficient in the task of pattern recognition. We evaluate our ML approach on a synthetic dataset replicating the measurements from the In Salah CO2 injection site in Algeria. In this task, the challenging part is the creation of the training set. We extract properties from the pressure distribution synthetic map, such as pressure shape and dynamic, that allows us to generate random pressure distributions with a lot of randomness. We then calculate their corresponding surface displacement using a generalized Geertsma’s solution. We demonstrate that a ML learning approach can be an effective tool to recover the pressure distribution in a reservoir from the surface deformation analysis. We also highlight the need for a good knowledge of the geology and a deep understanding of the shape/dynamic of the data.