Application of Feed-Forward Neural Networks for Modeling Friction Phenomena with Presence of Vegetable Oil-Based Bio-Lubricants

Reducing the use of petroleum-based oil lubricants in metal forming processes is a major task of modern sustainable industry. In order to meet the expectations of industry, this article presents the friction results of DC04 sheets in a friction pair with 145Cr6 tool steel tested with corn, sunflower, rapeseed, cottonseed and soybean vegetable oils. The friction process variables were also contact pressure and surface roughness of countersamples. The assessment of the lubricant efficiency was based on the friction coefficient value. Complex phenomena in the contact zone make the interpretation of results difficult. Therefore, feed-forward neural networks were used to analyse the relationship between input parameters and coefficient of friction. Different training algorithms and transfer functions were tested to find the optimal architecture of the neural network. The friction coefficient value, depended on the friction conditions and considering all tested oils, ranged between 0.155 and 0.181. The 3-10-1 neural network trained using the Levenberg-Marquardt algorithm and consisted of neurons with a radial basis transfer function provided the lowest mean squared error (MSE = 2.2233×10⁻⁶) and the root mean squared error (RMSE = 0.001491). The prediction quality of this network defined by the coefficient of determination was R² = 0.9697.