Prediction of grain size distribution in microstructure of polycrystalline materials using one dimensional convolutional neural network (1D-CNN)

Abstract

This paper presents a machine learning (ML) framework for online prediction of grain size distribution in microstructures of polycrystalline materials. The testing and training databases for the ML model are comprised of guided wave (GW) responses generated through localized excitation. The GW model is first developed numerically and later validated using experiments. In numerical simulations, Voronoi tessellation is used to generate a synthetic microstructure with different grain size distributions using a concept of regularity parameter. Finite element (FE) simulations are performed thereafter using a commercial FE package ANSYS-APDL. The generated wave responses reveal that wave characteristics are highly sensitive to grain size distributions. It helps the ML model to extract wave-based features influenced by grain size distribution. To classify the grain size distribution into three separate classes, a multi-headed 1-D CNN-based ML model is developed. The ML model is first trained using a set of simulated GW response of a polycrystalline material of Inconel-600. Next, it is tested with experimentally obtained GW responses of similar specimen. The grain size distribution of the Inconel-600 specimen is obtained using an electron backscatter diffraction (EBSD) scanning electron microscope. The developed multi-headed 1-D CNN model predicts the grain size distribution of the experimental specimen with high accuracy.