Physics‐Guided Machine Learning for Friction Coefficient Prediction in Statically Loaded Radial Journal Bearings

This study combines an analytic Stribeck baseline with machine‐learned corrections (physics‐guided machine learning) and examines what each part contributes. A full‐factorial campaign on a bearing with D = 25 mm, B = 30 mm and ISO VG 10 mineral oil provided 675 friction‐coefficient measurements over 27 shaft speeds (0–200 rpm), five loads (143–553 N) and five oil temperatures (25°C–65°C). For all 25 load–temperature conditions the friction minimum lies at 30–32 rpm, while the Stribeck number at the minimum varies 10.5‐fold: the Hersey number does not collapse the regime transition, and a baseline in the Stribeck number alone explains only 59.5% of the test variance (RMSE = 0.0301). Learning the residual of this baseline reduced the test RMSE to 0.0034 (least‐squares‐boosted trees) and 0.0037 (RBF support‐vector regression), with R 2 ≥ 0.994. A difference‐based estimate of the random measurement component is 1.0 × 10 −3 in µ at 0–5 rpm and below 1 × 10 −4 above 20 rpm, 0.9%–1.5% of µ in every speed band; the model errors remain 5–17 times larger, so the prediction accuracy is limited by the model and not by the torque measurement. The monotonicity projection reduces Stribeck‐trend violations by exactly its blend weight, while the test RMSE changes by less than 0.9% over the full weight range.