Wear 255 (2003) 708–713 Communication Artificial neural network predictions on erosive wear of polymers Z. Zhang ∗ , N.-M. Barkoula, J. Karger-Kocsis, K. Friedrich Institute for Composite Materials (IVW GmbH), University of Kaiserslautern, Erwin Schroedinger Str. 58, 67663 Kaiserslautern, Germany Abstract In the present paper, an artificial neural network (ANN) approach was applied to the erosive wear data of three polymers, i.e. polyethylene (PE), polyurethane (PUR), and an epoxy modified by hygrothermally decomposed polyurethane (EP-PUR). Three independent datasets of erosive wear measurements and characteristic properties of these polymers were used to train and test the neural networks. For the first two material examples, the impact angle of solid particle erosion and some characteristic properties were selected as ANN input variables. Whereas the third one, material compositions, i.e. epoxy and HD-PUR weight contents, were also involved as additional ANN input variables. In all these cases, the output parameter was the erosive wear rate. Acceptable ANN predictive qualities were reached, demonstrating that ca. 35–80% of the randomly selected test dataset had a coefficient of determination B ≥ 0.9 for these three cases, respectively. Ranking of the importance of characteristic properties to erosive wear rate could offer some information about which property has a stronger relationship to wear in each polymer case. Even though the ANN approach is only a phenomenological method, a well-trained ANN is believed to be also of help for a mechanistic understanding of the problem considered. © 2003 Elsevier Science B.V. All rights reserved. Keywords: Artificial neural networks (ANN); Erosive wear; Polymer; Prediction 1. Introduction The correlations between wear resistance and character- istic properties of polymers have been discussed in terms of various semi-empirical equations by some pioneers. These include, e.g. the Ratner–Lancaster equation [1,2], i.e. the relationship of the single pass abrasion rate with the recip- rocal of the product of ultimate tensile stress and strain, or an equation used by Friedrich [3] to correlate the erosive wear rate of polymers with the quotient of their hardness to fracture energy. Although these equations are quite helpful to estimate the wear behavior of polymers in some special cases, wear normally is very complicated, and it therefore depends on many more mechanical and other parameters. This means that simple functions cannot always cover all the prevailing mechanisms under wear. For predictive purposes, an artificial neural network (ANN) approach has, therefore, been introduced recently into the field of wear of polymers and composites by Velten et al. [4] and Zhang et al. [5]. An ANN is a computa- tional system that simulates the microstructure (neurons) of biological nervous system. The most basic compo- nents of ANN are modeled after the structure of the brain. ∗ Corresponding author. Tel.: +49-631-2017213; fax: +49-631-2017196. E-mail address: zhang@ivw.uni.kl.de (Z. Zhang). Inspired by these biological neurons, ANN is composed of simple elements operating in parallel. ANN is the simple clustering of the primitive artificial neurons. This clustering occurs by creating layers, which are then connected to one another. How these layers connect may also vary. Basically, all ANN have a similar structure of topology. Some of the neurons interface the real world to receive its input, and other neurons provide the real world with the network’s output. All the rest of the neurons are hidden from view. As in nature, the network function is determined largely by the interconnections between neurons, which are not simple connections, but some non-linear functions. Each input to a neuron has a weight factor of the function that determines the strength of the interconnection and thus the contribution of that interconnection to the following neurons. ANN can be trained to perform a particular function by adjusting the values of these weight factors between the neurons, either from the information of outside the network or by the neu- rons themselves in response to the input. This is the key to the ability of ANN to achieve learning and memory. The multi-layered neural network is the most widely ap- plied neural network, which has been utilized in the most of the research works for materials science, reviewed by Zhang and Friedrich [6]. Backpropagation algorithm can be used to train these multi-layer feed-forward networks with differentiable transfer functions to perform function approx- imation, pattern association, and pattern classification. The 0043-1648/03/$ – see front matter © 2003 Elsevier Science B.V. All rights reserved. doi:10.1016/S0043-1648(03)00149-2