Sliding Friction and Wear of Polyphenylene Sulfide Matrix Composites: Experimental and Artificial Neural Network Approach

In recent years the consumption of polymer based composites in many engineering fields where friction and wear are critical issues has increased enormously. Satisfying the growing industrial needs can be successful only if the costly, labor-intensive and time-consuming cycle of manufacturing, followed by testing, and additionally followed by further trial-and-error compounding is reduced or even avoided. Therefore, the objective is to get in advance as much fundamental understanding as possible of the interaction between various composite components and that of the composite against its counterface. Sliding wear of polymers and polymer composites involves very complex and highly nonlinear processes. Consequently, to develop analytical models for the simulation of the sliding wear behavior of these materials is extremely difficult or even impossible. It necessitates simplifying hypotheses and thus compromising accuracy. An alternative way, discussed in this work, is an artificial neural network based modeling. The principal benefit of artificial neural networks (ANNs) is their ability to learn patterns through a training experience from experimentally generated data using self-organizing capabilities. Initially, the potential of using ANNs for the prediction of friction and wear properties of polymers and polymer composites was explored using already published friction and wear data of 101 independent fretting wear tests of polyamide 46 (PA 46) composites. For comparison, ANNs were also applied to model the mechanical properties of polymer composites using a commercial data bank of 93 pairs of independent Izod impact, tension and bending tests of polyamide 66 (PA 66) composites. Different stages in the development of ANN models such as selection of optimum network configuration, multi-dimensional modeling, training and testing of the network were addressed at length. The results of neural network predictions appeared viable and very promising for their application in the field of tribology. A case example was subsequently presented to model the sliding friction and wear properties of polymer composites by using newly measured datasets of polyphenylene sulfide (PPS) matrix composites. The composites were prepared by twinscrew extrusion and injection molding. The dataset investigated was generated from pin-on-disc testing in dry sliding conditions under various contact pressures and sliding speeds. Initially the focus was placed on exploring the possible synergistic effects between traditional reinforcements and particulate fillers, with special emphasis on sub-micro TiO2 particles (300 nm average diameter) and short carbon fibers (SCFs). Subsequently, the lubricating contributions of graphite (Gr) and polytetrafluoroethylene (PTFE) in these multiphase materials were also studied. ANNs were trained using a conjugate gradient with Powell/Beale restarts (CGB) algorithm as well as a variable learning rate backpropagation (GDX) algorithm in order to learn compositionproperty relationships between the inputs and outputs of the system. Likewise, the influence of the operating parameters (contact pressure (p) and sliding speed (v)) was also examined. The incorporation of short carbon fibers and sub-micro TiO2 particles resulted in both a lower friction and a great improvement in the wear resistance of the PPS composites within the low and medium pv-range. The mechanical characterization and surface analysis after wear testing revealed that this beneficial tribological performance could be explained by the following phenomena: (i) enhanced mechanical properties through the inclusion of short carbon fibers, (ii) favorable protection of the short carbon fibers by the sub-micro particles diminishing fiber breakage and removal, (iii) self-repairing effects with the sub-micro particles, (iv) formation of quasi-spherical transfer particles free to roll at the tribological contact. Still, in the high pv-range stick-slip sliding motion was observed with these hybrid materials. The adverse stick-slip behavior could be effectively eliminated through the additional inclusion of solid lubricant reservoirs (Gr and PTFE), analogous to the lubricants used in real ball bearings. Likewise, solid lubricants improved the wear resistance of the multiphase system PPS/SCF/TiO2 in the high pv-range (≥ 9 MPa·m/s). Yet, their positive effect, especially that of graphite, was limited up to certain volume fraction and loading conditions. The optimum results were obtained by blending comparatively low amounts of Gr and PTFE (≈ 5 vol.% from each additive). An introduction of softer sub-micro particles did not bring the desired ball bearing effect and fiber protection. The ANN prediction profiles for PPS tribo-compounds exhibited very good or even perfect agreement with the measured results demonstrating that the target of achieving a well trained network was reached. The results of employing a validation test dataset indicated that the trained neural network acquired enough generalization capability to extend what it has learned about the training patterns to data that it has not seen before from the same knowledge domain. Optimal brain surgeon (OBS) algorithm was employed to perform pruning of the network topology by eliminating non-useful weights and bias in order to determine if the performance of the pruned network was better than the fully-connected network. Pruning resulted in accuracy gains over the fully-connected network, but induced higher computational cost in coding the data in the required format. Within an importance analysis, the sensitivity of the network response variable (frictional coefficient or specific wear rate) to characteristic mechanical and thermo-mechanical input variables was examined. The goal was to study the relationships between the diverse input variables and the characteristic tribological parameters for a better understanding of the sliding wear process with these materials. Finally, it was demonstrated that the well-trained networks might be applied for visualization what will happen if a certain filler is introduced into a composite, or what the impacts of the testing conditions on the frictional coefficient and specific wear rate are. In this way, they might be a helpful tool for design engineers and materials experts to explore materials and to make reasoned selection and substitution decisions early in the design phase, when they incur least cost