Physics-Informed Data Assimilation of Experimental Measurements for Hydrodynamic Journal-Bearing Pressure Prediction

Accurate prediction of hydrodynamic pressure distribution is essential for journal-bearing performance analysis. Although conventional two-dimensional Reynolds-equation-based models generally provide reasonable predictions, agreement with experimental measurements varies with bearing geometry and operating conditions. Good agreement is observed for shorter bearings, while discrepancies increase at larger bearing length-to-diameter (L/D) ratios and lower Ocvirk numbers. Experimental characterisation of the complete pressure field is also limited by the number of circumferential sensor locations. This study investigates a physics-informed neural network (PINN) framework for assimilating sparse experimental pressure measurements into Reynolds-equation-based pressure prediction. The Reynolds equation and associated boundary conditions are imposed as physical constraints, while seven circumferential pressure measurements are incorporated to refine the pressure field and infer bearing-state parameters. Of 27 operating cases, 22 are used for training and experimental data assimilation, while five complete cases are withheld for validation. The results show that the benefit of assimilation depends on the initial theory–experiment discrepancy. For L/D=0.25, where conventional predictions agree well with measurements over the investigated Ocvirk-number range of 0.217–1.46, assimilation produces only minor changes. In contrast, at L/D=0.5, where discrepancies become more pronounced towards the lower end of the investigated Ocvirk-number range, assimilation substantially improves agreement. For a data-trained low-Ocvirk-number condition, the experimental pressure residual (based on mean square error) decreases from 1.537 to 0.180, an approximately 88% reduction, while the PDE residual decreases from 0.0302 to 0.0250. For the unseen validation cases with substantial experiment-theory deviation, the experimental pressure residual decreases from 0.0903 to 0.0621, corresponding to an approximately 31% reduction, demonstrating improved agreement with experimental measurements at an operating condition not included during training. These results demonstrate that physics-informed data assimilation can preserve Reynolds-equation behaviour, where the conventional formulation is adequate, while improving pressure prediction where its simplifying assumptions become increasingly limiting.