Hydrodynamic fluid bearings have become increasingly sophisticated, and the resulting design requirements have gradually exceeded the capabilities of conventional exhaustive analytical approaches. This development has increased interest in alternative methods, including data-driven modeling. Although black-box surrogate modeling requires evaluation of the physical validity of its predicted solutions, assessing problems that lack theoretical solutions is extremely difficult. In this study, a surrogate deep learning model was developed for Rayleigh step bearing, for which the analytical solution is known, and its predictive accuracy was quantitatively evaluated. The results presented that, when the dataset was sufficiently large, the correlation coefficient between the prediction error rate on the validation data, which constituted part of the dataset, and the prediction error rate across the entire theoretical solution distribution was high. The model also achieved a practically acceptable approximation accuracy when Gaussian noise was intentionally added to the training data. These findings demonstrate the effectiveness and robustness of the proposed surrogate models and further present that the Rayleigh step bearing can function as a benchmark tool for verifying the methodological validity of various evaluation approaches.
Science Journal 