Convolutional neural networks for the in-situ investigation of blistering on plasma-exposed metal surfaces

Abstract

The exposure of metal surfaces to energetic particles originating for example from a plasma can lead to the formation of blisters. These features are caused by hydrogen-filled gas bubbles near the surface and can influence the diffusion and retention of hydrogen atoms in the bulk material during plasma exposure. No comprehensive theory of the underlying processes governing these effects has been developed so far. A major constraint to the investigation of blisters has been the effort that is required to identify and characterize a statistically relevant ensemble of blisters. Therefore, the analysis of a limited number of blisters in isolated observations before and after plasma exposure has been state of the art so far. In this article, a framework is presented that allows to automatically analyze blisters in image data with the aid of convolutional neural networks (CNNs) that are trained with artificial training data. It thereby resolves the limitations of existing analysis methods that prevented the large-scale and in-situ investigation of blistering up to now. Automated routines are used to identify and localize blisters, estimate important blister parameters such as size, and track individual blisters over time. Apart from discussing the principles underlying this approach and describing the implementation of the framework in detail, this article also presents first results acquired with a molybdenum sample that was exposed to a low-temperature deuterium plasma. The results indicate a blister identification accuracy of the framework of more than 80% within the considered parameter ranges and overall demonstrate the feasibility of the approach in general.