Convolutional neural network for automated quantitative analysis of non-destructively acquired three-dimensional corrosion pit morphology data

Abstract

We have trained an autoencoder-decoder type convolutional neural network for automated quantitative analysis of three-dimensional corrosion-pit morphology data. We have used pit morphology data, acquired non-destructively using scanning white light interferometry from AISI 304 stainless steel specimens subjected to different surface modification treatments, followed by potentiodynamic polarization corrosion testing. The autoencoder was designed to compress the high-dimensional input pit morphology data into a latent space of two variables, from which the decoder produces reconstructions of the input data. Analysis of the compressed data in the latent space led to the identification of a unique vector which encompassed most of the data points. Application of the model to artificially constructed semi-ellipsoidal pits with independently quantifiable morphological characteristics showed that the vector was analogous to ‘effective sharpness’ of pits, helping define a ‘sharpness/bluntness’ metric for arbitrarily shaped corrosion pits, that can be linked to their propensity to transition into stress corrosion cracks.