The question
In deviated and horizontal wells, drilled cuttings tend to settle and build up, which slows drilling, can stick the pipe and raises pressures downhole. How well the hole is being cleaned depends on the drilling fluid, the well's geometry and how the well is being drilled, but the concentration of cuttings downhole cannot be measured directly while drilling.
What I built
Using experimental flow-loop data published by researchers at the University of Tulsa, I trained machine-learning models to predict cuttings concentration from quantities known at the surface: fluid density, yield point, plastic viscosity, flow rate, temperature, inclination, hole eccentricity, pipe rotation and rate of penetration. Exploratory analysis pointed to fluid density, flow rate and pipe rotation as the variables that mattered most.
I compared random forest, gradient boosting and AdaBoost models, tuned them with randomized and grid searches, and combined the best of them in a stacked model. A final notebook uses the model to look for drilling settings that improve hole cleaning, the kind of quick check a drilling engineer could run when planning a mud program.
Where it went
The work took third place in the 2022 SPE Nigeria student paper contest, won the undergraduate division of the 2022 SPE Africa regional student paper contest, and was published in the proceedings of the 2022 SPE Annual Technical Conference and Exhibition. The notebooks are on GitHub.