Distributed Acoustic Sensing (DAS) is a fiber optics method that is revolutionizing the unconventional reservoir monitoring technology with substantial spatial coverage, high frequency data acquisition, and broad cable deployment options including hazardous/harsh environments compared to traditional geophysical methods such as point sensors (i.e., geophones). However, a single well equipped with fiber cannot acquire the far-field strain response since the sensitivity of this technique is restricted to a region near the monitor well. In this paper, we develop an Artificial Intelligence (AI) algorithm to estimate the magnitude of the far-field DAS response for any spatio-temporal input. Moreover, we identify a discontinuity in displacement results following fracture hit, which is interpreted as an effect of rock plastic deformation, and for the first time we demonstrate that it may be related to fracture width. Therefore, the output of our algorithm is used to estimate such geometric property along time in multiple locations.

We generate the tangent displacement component (uy) (parallel to monitor well) using an in-house code based on Displacement Discontinuity Method (DDM). Several monitor wells are incorporated in the simulation of physical scenarios characterized by single and multiple hydraulic fractures. For each specific scenario we train and test an Artificial Neural Network (ANN) with position and time as input variables, and axial displacement as output. The Machine Learning (ML) model is designed with 7 hidden layers, 100 the maximum number of neurons per layer and hyperbolic tangent as activation function. Finally, predicted uy is used to: (1) obtain Distributed Acoustic Sensing (DAS) data deriving it sequentially in space and time; and (2) estimate fracture width based on discontinuity magnitude.

Training stage is performed avoiding overfitting and minimizing ANN loss function. In the testing phase, error between true and predicted variables is negligible in the entire waterfall plot region, except at initial time steps where fracture treatment starts at operation well and magnitude of axial displacement collected at monitor well is very small on the order of 10-6 or even lower. In this case, we suspect that these tiny supervisor values may have minimal impact on the loss function, and consequently weights and biases of regression model are barely updated to consider the effect of such outputs. Regarding fracture width estimation, error reduces consistently along time at all locations reaching values near 0%.

To the best of our knowledge this is the first work that creates a ML algorithm able to estimate strain fields generated during hydraulic fracturing treatments merely based on position and time inputs. The model developed with synthetic data is an incentive for the deployment of multiple monitor wells in the field to enhance beyond the near wellbore region geometric characterization of created fracture systems, and possibly identify critical patterns associated with fracture propagation that ultimately can lead to production optimization.

You can access this article if you purchase or spend a download.