engleski

A comparison of conventional and machine-learning based methods for estimation of materials' parameters of cyclic strain hardening behavior

Introduction/Purpose For calculations and simulations of cyclic behavior and fatigue life of components and structures, materials’ cyclic stress-strain curves and cyclic parameters must be known. Cyclic strain hardening behavior of most metallic materials is successfully described with Ramberg-Osgood equation. Corresponding parameters, besides experimentally, can be estimated based on monotonic properties which is faster, cheaper and sufficiently accurate in early stages of product development. Methods Nowadays, two main methodologies exist for estimation of cyclic strain hardening parameters based on monotonic properties of materials: widely used empirical estimation methods and machine-learning based methods, mainly artificial neural networks (ANNs). ANNs enable easier manipulation with larger datasets, larger number of input variables and facilitate capturing complex relationships among input and target variables. Results ANNs were developed for estimation of cyclic strain hardening parameters and behavior of unalloyed, low-alloy and high-alloy steels and evaluated on an independent set of data. Comparisons to experimental values and values obtained by empirical methods have shown that ANNs are more successful than empirical methods both for estimation of cyclic strain hardening parameters and behavior of all considered steel subgroups. Developed ANNs were further applied to estimation of cyclic parameters and behavior of a single material, a widely used low-alloy steel 42CrMo4, and for a set of austenitic stainless steels - a subgroup of high-alloy steels group for which conventional methods perform rather poorly. In both of these particular cases, better results were obtained using developed ANNs. Conclusions Evaluations showed that when developed correctly, ANNs are more successful than empirical methods for estimation of strain hardening parameters and behavior of particular materials. By including even more data in developing ANNs for the given purpose, it is possible to obtain a quick, robust and efficient solution that can be successfully used for the estimation of cyclic strain hardening parameters and behavior from monotonic properties.

estimation methods ; monotonic properties ; cyclic strain hardening ; artificial neural networks

nije evidentirano