SVD-ANN-Based Processing of Melting PCM Bridge Experiments in Microgravity: Sensitivity to Image Defects and Data Repair Algorithms

Abstract

We analyze the sensitivity to image defects of the processing algorithm proposed by Salgado Sánchez et al. (Microgravity Sci. Technol. 37, 12, 2025) to evaluate melting bridge experiments in the context of the MarPCM microgravity project (Porter et al., Acta Astronaut. 210, 212–223, 2023). The algorithm uses the projection of input images onto the first m singular vectors (modes), obtained via Singular Value Decomposition (SVD), of the original (non-defective) image database. The resulting set of m amplitudes is then used as input for an Artificial Neural Network (ANN) that is trained to give the corresponding liquid fraction as an output. For the analysis presented here, the images are modified to generate a new database that includes rotated images, which represent optical misalignment, overexposed and underexposed images, which represent incorrect exposure time and/or aperture settings in the camera, noisy images and gappy images, which model the presence of dead pixels, bubbles and large reflections that compromise certain regions of the image. The results suggest that only relatively large defects are a concern for processing the experiment and that the most critical case is that of gappy images. Data repair algorithms based on SVD can be used to correct the defective images and reconstruct the missing information, which then allows for accurate processing.