"Sweet spots" in the unconventional reservoirs such as organic-rich mudrocks are zones with high productivity. However, identifying such regions in unconventional reservoirs depends non-only on their petrophysical and but also on their geomechanical properties. Supervised learning methods can help in integrating numerical simulation and legacy field data in sweet-spot identification workflows and enhance their analysis in complex reservoirs. The objectives of this paper are to: (i) demonstrate the use of supervised learning in parameter selection and evaluation for fracture design and (ii) provide non-linear models for sweet-spot analysis in complex reservoirs.

We used fracture simulator that combines with fracture deformation with fluid-flow in discrete fracture networks. We started by selecting different geomechanical rock properties related to its fracability. We then used quasi-random design approach to obtain wide variation in aforementioned properties and performed 200 fracture simulations using the hydraulic fracturing simulator. We used the short-term Stimulated Reservoir Volumes (SRV) obtained at the end of numerical simulations, to quantify the performance of hydraulic fracturing operations. We used supervised learning techniques like support vector machines, decision trees, and random forests to perform parameter ranking and create non-linear regression models that can correlate the SRV to formation geomechanical properties.

The inputs for the analysis are: initial aperture, toughness, dilation angle, closure stress, and friction coefficient of initial fractures, stress anisotropy, shear modulus and a ratio of the reservoir rock. We analyzed the results using β-linear and multinomial regression, support vector machines, decision trees, and random forests. The linear models and non-linear models can explain up to 89.1% of output variance. The classification accuracy of support vector machines was at most 35% higher than other algorithms like random forests. Parameter rating using non-linear models showed that stress anisotropy and dilation angle demonstrated the highest effect on SRVs. Shear modulus and fracture toughness show minimal effect on the SRV but these parameters might still be useful they could be correlated to other formation parameters.

The outcomes of this paper demonstrated that parameters pertaining to unpropped fracture conductivity play a significant role in determining the success of hydraulic fracturing treatments. We have also compared the performances of supervised machine learning algorithms in assessing the impact of rock properties on fracturing treatments. Such supervised machine learning algorithms can help integrate field legacy data and numerical simulation outputs to develop proxy models that improve sweet-spot analysis and production estimates in unconventional reservoirs.

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