Introduction and development of machine learning models for LIBS for thermonuclear fusion

Due to its potential in contactless and non-interfering diagnostics, laser-induced breakdown spectroscopy (LIBS) has become an important element of research in the PWI (plasma wall interactions) area, which has been under the intensive investigation of numerous teams from Europe and China. The advantages of this effort demonstrated successful analysis of the surface materials in EAST and WEST tokamaks and deployment of the remotely controlled LIBS head to conduct LIBS measurements at FTU. In 2024 an experiment with remotely controlled LIBS head at JET finally proved that LIBS constitutes a highly relevant technique for ITER. Despite these achievements, applying LIBS to the next-step fusion device remains challenged by uncertainties. These uncertainties stem from the undetermined morphology of co-deposits in ITER and potential issues with controlling laser beam and plasma parameters. These control issues can lead to inaccuracies in estimating the chemical composition of plasma-facing components. Additionally, the vast amount of measurement data expected in ITER could further complicate the diagnostic performance. On a positive note, recent years have seen the emergence of new tools that capitalize on large datasets. These tools are artificial intelligence methods, particularly artificial neural networks and convolutional neural networks. These methods enable deep learning, a technique highly adept at identifying patterns in data. Models built using these methods have already demonstrated superior performance in various scientific, industrial, and information technology tasks. Nevertheless, these methods have not yet been thoroughly adapted in the PWI community, and improperly recognized terminology leads to misunderstanding and confusion. The main goal of this contribution is to thoroughly present the machine learning (ML) models to an audience of PWI scientists so that they can gain an understanding of the available tools and methods of ML. This knowledge is supported by exemplary analysis of the chemical composition of brass-alloys, which shows the capabilities of deep learning for more complex mixtures which were analysed in previous investigations.