Real-time updating of geological properties in reservoirs through monitoring data is critical for optimizing hydraulic fracturing. In this work, the ML-Physics method, which inherits the advantages of machine learning (ML) and physical modeling, has been developed to determine the geological parameters with limited computational time, including far-field geostress, geostress in the vicinity of faults and rock material properties. The ML-Physics method consists of two stages: the preparation stage and the operation stage. During the preparation stage, which occurs before the in situ operation and is not constrained by computational time limitation, a surrogate artificial neural network (ANN) model is established. This ANN model captures the implicit relationship between geological parameters and monitoring data. During the operational stage, which is constrained by real-time requirements for the convenience of engineers in field work, inverse analysis with physical modeling and the surrogate ANN model is employed to determine geological parameters. In the context of ML-Physics method, the initial values of inverse iteration are chosen based on surrogate ANN results. The obtained geological parameters can be used for hydraulic fracturing analysis. This ML-Physics method demonstrates superior performance in terms of both accuracy and computational efficiency.
Read full abstract