Physics-Informed Machine Learning for Polymer Drag Reduction: Bridging Laboratory Turbulent Pipe Flow and Field-Scale Crude Oil Pipeline Operation

Polymeric drag-reducing agents (DRAs) can significantly suppress turbulent friction in crude oil pipelines. However, the design of field dosing strategies still relies on empirical rules because laboratory-scale drag-reduction (DR) measurements fail to transfer quantitatively to industrial diameters, fluids, and shear histories. In this study, we present an integrated laboratory–field–machine-learning framework that closes this gap. Systematic recirculating-loop experiments (Reynolds number of 2 × 104–1.2 × 105, four pipe diameters, 3–200 ppmw of a polyalphaolefin DRA in a model oil) established concentration–Reynolds response surfaces, first-order mechanical degradation kinetics with a wall-shear-scaling rate constant kd ∝ τw1.25, and a normalized scale-up correlation validated over an eight-fold range of controlled laboratory diameters (12.7–102.0 mm) and, after a two-parameter transfer calibration, applicable to a 711 mm trunk line. A 90-day campaign on a 720 km, 711 mm crude oil trunk line (129,600 SCADA records) yielded a mean field DR of 22.9% and a mean pumping-energy saving of 22.9%. A gradient-boosting model trained jointly on laboratory and field data predicts DR with an R2 = 0.984 and an RMSE = 0.89 percentage points, while SHAP analysis ranks Reynolds number, concentration, and pipe diameter as the variables with the largest marginal contribution. A physics-informed hybrid network constrained by the Virk maximum-drag-reduction asymptote reduces extrapolation error by 34% relative to a purely data-driven network, and a sequence-to-one LSTM forecasts field DR 24 h ahead with an RMSE ≈ 1.4 points, outperforming persistence by more than a factor of two. Bayesian multi-objective optimization of the trained surrogate identifies dosing schedules that add two to three percentage points of energy savings per DRA-treated segment, corresponding to a modeled net benefit of 1.85 M US$ yr−1 and 17.7 kt CO2 yr−1. The framework demonstrates that physics-constrained machine learning can convert routine laboratory rheometry and SCADA streams into quantitatively predictive, economically deployable drag-reduction engineering.