A deep learning model to extract the interphase’s characteristics in microstructures using macroscopic responses

Abstract

This study addresses the challenge of directly measuring the mechanical and geometrical properties of the interphase region in multiphase microstructures due to its small volume. Despite the limited volume, the interphase’s properties can dramatically affect the macroscopic responses, such as elastic modulus in two directions and Poisson’s ratio, because it connects the main parts together. This work proposes a hybrid fusion deep learning model capable of accurately extracting interphase properties, including elastic modulus, Poisson’s ratio, and thickness, using both the microstructural arrangement image and the macroscopic responses of the microstructure as its inputs. To provide the required dataset, 2500 microstructures are generated using the Random Sequential Expansion (RSE) algorithm. Following microstructure generation, homogenization is applied, deriving the effective longitudinal elastic modulus and major Poisson’s ratio through the Rule of Mixture (ROM) method, complemented by the effective transverse elastic modulus obtained from numerical Finite Element (FE) modeling. The hybrid fusion model is trained using 80 % of the dataset, with the remaining instances used for model performance assessment. The R-squared value of 0.94 for the testing dataset demonstrates the model’s high accuracy in predicting interphase characteristics. The proposed model is prooved to be a solid tool for extracting the interphase properties with much less computational costs and time consumption of optimization algorithms and experiments such as atomic force microscopy and nanoindentation.