Development of Surrogate Fracture Cleanup Model

The paper is devoted to the development of a surrogate model describing Hydraulic Fracture (HF) flowback, which is a short time period after the start of an oil or gas well, during which it is cleaned out of fracturing fluid and set for a long-term production. The model is based on the combination of the in-house mechanistic model and Machine Learning (ML) algorithms. The mechanistic flowback model describes the dynamics of fracture conductivity during the cleanup, considering several hydro- and geomechanical effects. We conducted a series of computations to assemble the synthetic dataset consisting of 10 000 runs covering the range of input parameters typical for Western Siberia’s conventional terrigenous oil reservoirs. We obtained that the constructed surrogate model is sufficiently accurate at solving the forward problem in terms of the Mean Absolute Percentage Error (MAPE) metric showing the difference between the results acquired by the metamodel and the mechanistic simulator. In particular, we obtained MAPE of 5.12% and 6.80% for Newtonian HF fluid in predicting the average fracture aperture and permeability at the end of flowback, respectively, which we assume to be an acceptable match regarding the variables’ scales. The metamodel allows decreasing the CPU time of computations from dozens of minutes for the mechanistic model down to about 0.1 s for the surrogate one.