Design of novel interpretable deep learning framework for microstructure–property relationships in nickel and cobalt based superalloys

Abstract

Featurization of microstructures is one of the most fundamental challenges in establishing microstructure–property relationships. Conventional machine learning and statistical methods require explicit featurization methods such as image processing, which are difficult to implement for complex and diverse sets of microstructures. To this end, deep learning methods such as convolution neural networks (CNNs) have been used to automate the featurization based on target properties. However, these CNNs do not include composition information limiting them to a single set of compositions. Moreover, these networks are complex and difficult to interpret. To overcome these challenges, a deep learning mixed input network consisting of a convolutional neural network (CNN) for microstructure input and an artificial neural network (ANN) for composition input is developed to predict the Vickers hardness of nickel and cobalt-based superalloys. A unique three-step optimization procedure is employed to reduce the complexity of the network. The network architecture is designed based on hardening models which allows the analysis of contributions of precipitation hardening and solid solution strengthening to the Vickers hardness. The network has been analyzed using synthetically generated controlled microstructures to understand the effect of microstructural features on the hardness. Furthermore, SHAPley additive explanations (SHAP) analysis has been used to understand the effect of composition and assess the interdependence between microstructure and composition in determining hardness.