Authors: D. Lloveras1,2, L. Balmaceda1,2, F. Iglesias3,4, Y. Machuca3,4, M. Sánchez Toledo3, H. Cremades3,4, C. Sasso5, P. Sahani5, Y. de Leo6,7 & T. Nieves-Chinchilla1 1 Heliophysics Science Division, NASA Goddard Space Flight Center, MD, USA 2 Physics and Astronomy Department, George Mason University, VA, USA 3 Grupo de Estúdios en Heliofísica de Mendoza, Universidad de Mendoza, Arg. 4 Consejo Nacional de Investigaciones Científicas y Técnicas, Arg. 5 Istituto Nazionale di Astrofisica, Italy. 6 INAF-Astrophysical Observatory of Catania, Catania, Italy 7 Institute of Physics, University of Graz, Graz, Austria

Coronal Mass Ejections (CMEs) are a primary driver of severe space weather. For this reason, their accurate and timely identification in coronagraph imagery is essential for improving our current forecasting capabilities of their impact. Recently, we addressed the scarcity of annotated training data for deep learning applications by developing a robust automated detection tool: a Mask R-CNN architecture fine-tuned on a novel synthetic CME dataset. This dataset was constructed by combining real quiet-Corona observations with synthetic CMEs generated via the Graduated Cylindrical Shell (GCS) forward model. Having successfully demonstrated this architecture’s capability to identify and segment CME outer envelopes in running difference images, our current work focuses on cross-instrument evaluation. In this presentation, we analyze the performance and generalizability of our synthetically-trained Mask R-CNN when applied directly to observational data from the Metis coronagraph onboard the Solar Orbiter (SolO) mission. We evaluate the tool’s detection accuracy on real-world Metis CMEs, apply a CME tracking system, and calculate CME parameters of interest such as central position angle (CPA), angular width (AW), and apex.