Optical Processing of Melting PCM Bridges in Microgravity Using SVD and ANNs

Abstract Both Singular Value Decomposition (SVD) and Artificial Neural Networks (ANNs) can be powerful tools for image processing. Here they are applied in the context of the “Effect of Marangoni Convection on Heat Transfer in Phase Change Materials” (MarPCM) microgravity experiment [Porter et al. (Acta Astronautica 210, 212–223, 2023)], which investigates the use of thermocapillary (Marangoni) convection to expedite melting of organic Phase Change Materials (PCMs) in cuboidal and cylindrical domains. The processing of the cylindrical “melting bridge” experimental images is particularly challenging due to the converging lens effect caused by the curved interface and the refractive index of the liquid PCM. A combination of SVD and ANNs is used to propose an algorithm to process these images. The network is trained on a set of synthetic images of the melting bridge, generated via ray-tracing [Martinez et al. (Advances in Space Research 72, 1915–1928, 2023)] then projected onto the eigenmodes associated with the largest singular values of the image database, which includes snapshots of the melting process in all representative cases. Two optimal algorithm architectures are described, characterized by the number of SVD modes considered in the projection and the hyperparameters of the ANN. The performance of the algorithm is analyzed in terms of its ability to associate images with the correct liquid fraction. The processing strategy is tested by applying it to images obtained from ground experiments using the scientific prototype of the MarPCM cuboidal cell.