Comparative study of spherical indentation test results using neural network models

Abstract

Mapping of indentation test results to material properties is not straightforward due to highly nonlinearity and complexity of contact problem involved in indentation tests. A viable approach for the above reverse analysis through neural network models requires a collection of database relating the test results to material properties. Extensive finite element simulations have been carried out to establish the comprehensive database of spherical indentation curves for a wide range of elasto-plastic materials obeying power law strain-hardening. Characteristics of indentation curves depend on properties of indented material (E*, Y, n) and the ratio of indentation depth and radius of indenter tip (h/R). For a fixed ratio of h/R, indentation curves are normalized to ensure the same response of indentation curve for the same set of material properties. Indentation parameters relating elasto-plastic material properties and the normalized indentation parameters, namely, the normalized maximum load (Pmax/Yh2 max) and the ratio of the remaining and total work done (WR/WT) have been identified. In the present study, two neural network models, Artificial Neural Networks (ANNs) and Support Vector Machines (SVMs) are constructed, trained and verified to map the indentation parameters to elasto-plastic material properties through the statistical learning theory. The comparative study on the two neural network models based on dual indenter technique for several combinations of databases established from simulated indentation tests is highlighted. Identified material properties based on the proposed neural network models are presented. Comparative study on efficiency and accuracy of both neural network models is included in the presentation. The neural network models can be extended for other applications in material characterization and other nonlinear reverse analyses.