High-temperature rheology testing is critical for all fracturing applications to design the well-specific breaker and additive schedule. The rheology depends on the source water quality, testing temperatures, shear profiles, and additives-stabilizers-breaker combinations used for the test. The process for each treatment requires extensive staff to fine-tune the optimal fluid formulations requiring proportional laboratory resources and time. Data analytics intelligent system design can be implemented beyond analytical mathematical correlations to reduce the time and resource requirements.
A total of 820 rheology tests were conducted using the high-pressure/high-temperature (HP/HT) rheometer Chandler 5550 and ISO 13503-1 guideline. Temperatures ranged from 200 to 336°F and fluid systems consisted of borate and metallic crosslinkers. A structured database with 40 input-output features was prepared to digitize each rheology curve by incorporating the source water parameters, laboratory setup details, additive concentrations, and rheology (consistency and behavior indices) results. ML algorithms and techniques were then applied to the database to predict the rheology for given testing parameters.
The algorithm inputs were prepared as the source water quality (i.e., monovalent/divalent ions, minerals, salinity, hardness etc.) and the test temperature. The outputs predicted were set to be the detailed fluid formulation for specified viscosity and fluid stability requirements. Data cleaning and ingestion were done thoroughly to remove nonphysical outliers such as bob-climbing during testing. A detailed parametric correlation study followed and revealed the impact of different parameters, especially divalent ions such as Ca+2 and Mg+2, and total dissolved solids on the rheology. The training set to holdout set ratio was fixed at 90:10 for different trials. Further, 5-fold cross validation was used to choose the hyperparameters for the final model. To predict fluid formulation/target rheology in terms of additive concentrations, which is a continuous quantity, regression-based models were attempted. Ridge regression and ensemble methods such as random forest and boosting type models were trained. Boosting-based models gave an average 88% goodness of fit (R2) for the holdout datasets. For field implementation, the model results were used to create a digital laboratory request for the laboratory technician instead of having the fracturing design engineer manually handle this task. The physics-based data-driven ML model reduced an average HP/HT runs/well from 20 to 5 yielding a 400% laboratory resource savings.
This ML-based workflow is unique and does not exist in the literature. It can enable resource optimization for all large-scale fracturing projects and reduce manual laborious input for generating laboratory requests followed by trial-and-error optimizations with a potential of saving thousands of hours and reduce all the laboratory equipment maintenance costs. The technique can easily be extended to designing cementing fluids, drilling muds, and corrosion properties.
