Pitting corrosion prediction from cathodic data: application of machine learning

What is it about?

In a previous work, we demonstrated the effectiveness of machine learning in predicting the localized corrosion resistance of a material by considering as input data the physicochemical properties of its environment [‎1]. With the present work, we improve the results by considering as input data, cathodic currents. The reason of such an approach is to have input data that integrate both the surface state of the material as well as the physicochemical properties of its environment. In order to constitute input data, we carried out electrochemical experiments. We performed voltammetric scans in a very cathodic potential region. We constituted an experimental table where for each experiment we note the current values recorded at a low polarization range and the pitting potential observed in the anodic region. The study concerns a carbon steel used in a nuclear installation. The properties of the chemical solutions are close to that of the cooling fluid used in the circuit. The experimental table was submitted to two neural networks: a recurrent network and a convolution network. The convolution network gives better pitting potential predictions. Results also prove that the prediction by observing cathodic currents is better than that obtained by considering the physicochemical properties of the solution. The originality of the study lies in the use of cathodic currents as input data. These data contain implicit information on both the chemical environment of the material as well as its surface condition. This approach appears to be more efficient than considering the chemical composition of the solution as input data. The objective of this work remains, at the same time, to seek the optimal neuronal architectures and the best input data.

Featured Image