Characterization of influential parameters on friction in the nanometric domain using experimental and machine learning methods

Friction is a ubiquitous phenomenon of great research interest in engineering practice. Fundamental frictional features of two solids in contact and in relative motion are governed by microscopic single asperity contacts at their interface. A structured multidisciplinary approach to the experimental determination of friction in the nanometric domain is presented in this work. The dependence of nanoscale friction on process parameters comprising the materials in relative motion, normal forces, sliding velocities and the temperature conditions is studied experimentally by employing scanning probe microscopy. The data hence attained from multidimensional experimental measurements on thin-film samples is used for the development of machine learning-based models. In fact, due to the stochastic nature of the considered phenomena, conventional regression methods yield poor predictive performances, prompting thus the usage of the machine learning numerical paradigm. Such an approach enables obtaining an insight into the concurrent influence of the process parameters on nanoscale friction. A comparative study allows thus showing that, while the best typical regression models result in coefficients of determination (R) of the order of 0.3, the predictive performances of the used machine learning models, depending on the considered sample, yield R in the range from 0.54 to 0.9. The proposed method, aimed at accomplishing an in-depth insight into the physical phenomena influencing nanoscale frictional interactions, will be complemented next with advanced studies based on genetic programming-based artificial intelligence methods. These could, in fact, allow obtaining a functional description of the dependence of nanoscale friction on the studied variable parameters, thus enabling not only true nanoscale friction prediction but also an important tool for control purposes.