A Scientific Machine Learning Model for Real-Time Transient Prediction of Cuttings Bed Height in Long Horizontal Laterals

Abstract

Poor hole cleaning remains the root cause of stuck pipe, elevated torque and drag, and non-productive time while drilling long laterals. Yet most cuttings transport prediction models are either too computationally expensive for real-time decisions or too limited in generalizability when based purely on field data. This study develops a scientific machine learning framework to predict the transient cuttings transport phenomena under varying operating conditions and enable real-time hole cleaning monitoring. A physics-informed neural network (PINN) model was developed using a modified multilayer perceptron with random Fourier feature embedding and tanh activation, which enables the PINN to resolve high-frequency input variations. The model takes input operational, fluid, and cuttings parameters. Its composite loss combines measurement-based data to fit with a physics residual derived from the Navier-Stokes equations that captures gravitational settling, buoyancy, and annular drag. The PINN is trained and validated on hydraulic numerical models published in the literature. On two wells held out of training, the PINN reproduces the bed blockage, suspended cuttings concentration, annular pressure and phase velocities of the reference simulator. Bed blockage is recovered with a coefficient of determination between 0.84 and 0.94 and a mean absolute percentage error below 5%, and the remaining channels lie between 0.95 and 0.97. The model returns a full clean-up transient in milliseconds, which neither empirical correlations nor numerical simulators can do at rig-floor speed, and its physics residual and solids mass balance can be checked without access to a simulator. It was observed that higher annular velocity and pipe rotation reduce bed blockage, with diminishing returns beyond roughly 120 rev/min, while a higher drilling rate and coarser cuttings raise the suspended loading and lengthen the clean-up cycle. This framework is aimed at drilling engineers who need to optimize hole cleaning efficiency both during well planning and during real-time operations. In the planning phase it can be used to select the pump rate, pipe rotation and mud rheology that keep the lateral clean within the fracture gradient at the shoe. On the rig, because a full clean-up transient is returned in milliseconds, the engineer can quantify the circulating bottoms-up before the string is pulled, rank the intervals where pack-off risk is building, and test a change in rate or rotation before deciding. In day-to-day use the engineer enters the current rate, rotation, drilling rate and mud properties and get the bed height along the wellbore, how long a bottoms-up will take, and how much bed is left when it finishes. It therefore equips drilling engineers with an interpretable, real-time prediction tool that has direct potential to reduce non-productive time caused by poor hole cleaning.