Irregular microstructure-property linkage for cast alloys by a novel deep learning approach: Application on cast austenitic stainless steel

Abstract

The conventional approach of extracting microstructure parameters is inaccessible to accurately describe the irregular microstructure in cast alloys. In the present paper, we have proposed a new model based on convolutional neural network (CNN) to directly link the irregular microstructure with micrographs and property relationship for such alloys. By taking 10 × 10 units through the adaptive average pooling layer after the best channel selection on the convolutional layer, the model has succeeded to establish the microstructure-toughness relationship for the cast austenitic steel solutioning treated at different temperatures. It shows a prediction accuracy with R2 of 0.84, and the predicated median error to actual testing impact toughness is as small as ± 2 J. In addition, intermediate representations in the model are visualized, showing that microstructure features such as volume fraction, morphology and distribution of irregular ferrite are captured in the 10 × 10 units.