Machine-Learning-Assisted Modeling of Diverting Agent Performance for Multiple Fracture Propagation

This study discusses the development of a diverting agent (DA) performance prediction model for simultaneous hydraulic fracture propagation based on numerical and machine learning approaches. The filtrate capacity of a particulate DA plug was quantified by the filtrate coefficient obtained from the analyses of past filtrate experimental data, with a parametric study performed to evaluate the influence of the filtrate coefficient on multiple fracture propagations. We developed a wellbore-fracture coupled model that considered filtration by a DA for multiple hydraulic fractures. The proposed model solved flux redistribution and simultaneous fracture propagation after diversion. The filtrate performance of the DA can be adjusted and controlled by the filtrate coefficient. Furthermore, we developed a prediction model using a machine learning approach to evaluate the performance of the DA. The model was constructed using four different algorithms: MLP, SVC, RF, and AdaBoost, with each model evaluated and compared using five evaluation indicators. Furthermore, a feature importance analysis was conducted to assess the contribution of the experimental parameters employed in this study. The numerical modeling results were validated against the analytical solutions for a plane-strain Khristianovic- Geertsma de Klerk (KGD) model. A series of numerical simulations were conducted to investigate the multi-fracture growth patterns under different filtrate coefficients during fracturing treatments. The parametric study showed that a DA based on Butane-diol vinyl alcohol co-polymer had sufficient filtration capacity, equivalent to that of ball sealers when the filtration coefficient was less than 10 [mL/min^0.5], with almost no filtrate capacity when the filtration coefficient was over 1×10^6[mL/min^0.5]. As for the machine learning models, all models showed over 80% model scores; however, the RF and AdaBoost models, which are ensemble learning algorithms, provided better performance in terms of the five evaluation indicators compared to the other two models. Through the feature importance analysis, we calculated the contribution of each experimental parameter to the filtrate performance of DA. The results of this study clearly demonstrate the influence of the filtrate coefficient on the diversion process. To the author's knowledge, this study is the first published paper to link the filtrate coefficient and actual multifracture propagation. The DA process developed in this study helps evaluate the diversion performance of particulate diverting agents. Furthermore, the machine learning model clarified the ambiguous performance evaluation of DA and enabled the prediction of the DA filtrate performance from complicated physical and chemical processes.