A machine learning approach for determination of coefficient of friction from ring compression test

Abstract

A framework is defined for determining the coefficient of friction (COF) using numerical simulations of ring compression tests and machine learning techniques, utilizing von Mises plasticity with the three-parameter Swift law material model. A big dataset, containing 18750 rows and 20 columns, is created from 3750 numerical simulations of ring compression test in ABAQUS using MATLAB scripts. A dense feedforward neural network is trained to predict the values of COF from geometrical changes in the cross-section of the ring and the applied load. When the machine learning model is tested and the desired accuracy is obtained, we collect the measurements of the cross-section of the deformed ring from the experiments, and put the measurements, along with the applied load, into the trained machine learning model based on the numerical simulations and obtain the value of predicted COF. The average error of the predictions is around 4%. By this method, the COF is determined from only one ring compression test by stopping the test at an arbitrary compression ratio, regardless of the geometry of the ring, without utilizing the calibration curves, and with or without prior knowledge about the material properties of the ring.