Predicting the contact stress required for metal gaskets to achieve ultra-low leak rates is challenging due to complex factors like surface structure. In this study, a machine learning (ML) framework is proposed to accurately predict this target stress. Using an experimental dataset with various materials, surface topographies (Ra, RSm), and geometries, linear models (e.g., OLS, LASSO) were compared with nonlinear models (e.g., GBDT, NGPR). In these results, nonlinear models, particularly Nonlinear Gaussian Process Regression (NGPR), provided superior predictive accuracy (R² > 0.8). Feature importance analysis revealed that gasket stiffness is the most dominant factor, and the mean spacing of profile elements (RSm) is more influential than arithmetic mean roughness (Ra). This ML-based approach enables quantitative prediction of contact stress, facilitating the design of highly reliable sealing systems.
Satoshi Nakazato (Wed,) studied this question.