Abstract
Recent advancements in spectropolarimetric instrumentation, such as the new facilities at the GREGOR and DKIST telescopes, have generated vast amounts of data with each observation. This increase in data volume results in longer processing times, heightened demands on computational resources, and an expanded carbon footprint, complicating scientific development timelines. The numerical inversion codes used for data analysis, based on radiative transfer models, are inherently complex. Modern projects focused on the solar atmosphere and its magnetic field require additional assumptions, significantly increasing processing times for each pixel. To address this challenge, new methods are being developed, leveraging modern data processing algorithms from statistics and machine learning. We are testing a 1D convolutional neural network model inspired by the 1D parallel atmosphere model of radiative transfer to enhance spectropolarimetric inversions and achieve significant reductions in processing times, as demonstrated in previous studies. Our approach aims to integrate physical constraints into the learning process, allowing the model to not only replicate inversions but also gain insights into the underlying physics. The data for our project was synthesized using state-of-the-art codes for magnetohydrodynamics (MURaM) and radiative transfer under non-local thermodynamic equilibrium (NICOLE). Preliminary results without physical constraints show loss rates approaching 10-3 order of magnitude and Pearson correlations of the order of 0.8 on average along different optical depths inverted in the process for thermodynamic quantities.
Co-authors:
Nicolás Morales, Santiago Vargas-Domínguez, Sergiy Shelyag