Prediction of Friction Coefficient in Elastohydrodynamic Lubricated Contacts through a Multi-Source, Few-Shot Transfer Learning Artificial Neural Network Approach

This study examines the potential for use of experimentally measured traction coefficient under elastohydrodynamic lubrication (EHL) conditions under different conditions for training an artificial neural network that could potentially be used as a generic initial assessment tool. In this study a physics-informed Artificial Neural Network is trained on 4,519 experimentally EHL friction measurements from ten different sources with a few-shot-ridge regression adapter calibrated on approximately 30 points per target dataset. Three input representations are compared: a full dimensional set (8 parameters), a dimensionless set based on Hamrock and Dowson’s groups and a two parameter reduced set comprising the modified Hersey number and slide-to-roll ratio. The network incorporates a custom Dual Stribeck Activation output layer that enforces the physical shape of the Stribeck curve with non-negative friction and a well-defined EHL plateau throughout training instead of a post-hoc constraint. Against three independent holdout targets, the adapted model reaches R^2 = 0.984 based on the data from Zapletal 2018 with RMSE = 9.0% of mean CoF and R^2 = 0.934 based on the data provided in Lu, et al 2006 from those 30 measurements alone. Out-of-distribution diagnostics potentially can isolate the physical root cause of each base-model failure, and SHAP attribution shows the network recovers physically correct feature weightings including piezo-viscosity dominance for Lu, et al 2006, and load dominance for Simpson et al, 2024 without any mechanistic encoding of these dependencies.