• capillary channels
• flow boiling
• flow pattern maps
• image processing
• machine learning
• pulsating heat pipes
• semi-supervised learning
• two-phase flows

This work proposes the implementation of machine learning techniques dedicated to the recognition of flow regimes for confined capillary two-phase (liquid-vapor) flows. The goal is to generate flow maps that are as objective as possible, seeking to overcome the limitations imposed by a classification of regimes affected by the author's subjective judgment. The adopted approach involves the construction and comparison of six semi-supervised architectures, obtained by combining autoencoders, Principal Component Analysis (PCA) techniques, and clustering capable of exploiting visual and numerical information. The reference database comes from flow boiling tests in microgravity conditions, obtained from a parabolic flight campaign conducted by the ENEA research center in Rome. The data were reprocessed defining dimensionless groups formulated from the physical phenomena involved to obtain generalizable results. This required the development of specific algorithms designed for the recognition and tracking of vapor in the images from the experiments, implementing innovative solutions capable of handling a wide spectrum of different flow regimes. At the end of the work, the generated flow maps and the errors resulting from the analysis procedures involved were critically evaluated.