PINN ME: A Physics-Informed Neural Network Framework for Accurate Milne-Eddington Inversions of Solar Magnetic Fields Facilities: Hinode/SOTSP astropy; numpy; matplotlib; pytorch; sunpy
Robert Jarolim ∗, Momchil E. Molnar ∗, Benoit Tremblay, Rebecca Centeno, Matthias Rempel · 2025
Read PINN ME: A Physics-Informed Neural Network Framework for Accurate Milne-Eddington Inversions of Solar Magnetic Fields Facilities: Hinode/SOTSP astropy; numpy; matplotlib; pytorch; sunpy by Robert Jarolim ∗, Momchil E. Molnar ∗, Benoit Tremblay, Rebecca Centeno, Matthias Rempel (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-13924.txt.
Introduction
Reading the Sun's magnetic field from its light is an inverse problem: from polarized spectra, infer the field that made them. Robert Jarolim and co-authors bring physics-informed neural networks to that task in this 2025 paper, under a standard simplifying model of the solar atmosphere. Their network maps time and place coordinates to the field parameters that best reproduce what telescopes see.
The network builds the instrument's blur into the fit, and its built-in coupling across space and time damps noise. Tests on synthetic spectra from a realistic simulation, and on real observations, show clean, coherent maps that resolve fine features in both strong and weak field areas. The method also stays light on memory for wide fields of view.
Catalog record
| CREATOR | Robert Jarolim ∗, Momchil E. Molnar ∗, Benoit Tremblay, Rebecca Centeno, Matthias Rempel |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 40 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-13924
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
PINN ME: A Physics-Informed Neural Network Framework for Accurate Milne-Eddington Inversions of Solar Magnetic Fields Facilities: Hinode/SOTSP astropy; numpy; matplotlib; pytorch; sunpy Robert Jarolim ∗, Momchil E. Molnar ∗, Benoit Tremblay, Rebecca Centeno, Matthias Rempel Abstract Spectropolarimetric inversions of solar observations are fundamental for the estimation of the magnetic field in the solar atmosphere. However, instrumental noise, computational requirements, and varying levels of physical realism make it challenging to derive reliable solar magnetic field estimates. In this study, we present a novel approach for spectropolarimetric inversions based on Physics Informed Neural Networks (PINNs) to infer the photospheric magnetic field under the Milne-Eddington approximation (PINN ME). Our model acts as a representation of the parameter space, mapping input coordinates ( t , x , y t,x,y ) to the respective spectropolarimetric parameters, which are used to synthesize the corresponding stokes profiles. By iteratively sampling coordinate points, synthesizing profiles, and minimizing the deviation from the observed stokes profiles, our method can find the set of Milne-Eddington parameters that best fit the observations. In addition, we directly include the point-spread-function to account for instrumental effects. We use a predefined parameter space as well as synthetic profiles from a radiative MHD simulation to evaluate the performance of our method and to estimate the impact of instrumental noise. Our results demonstrate that PINN ME achieves an intrinsic spatio-temporal coupling, which can largely mitigate observational noise and provides a memory-efficient inversion even for extended fields-of-view. Finally, we apply our method to observations and show that our method provides a high spatial coherence and can resolve small-scale features both in strong- and weak-field regions. † † ∗ R. Jarolim and M. E. Molnar contributed equally to this study. I Introduction Figure 1: Overview of the PINN ME Milne-Eddington inversion method. The neural network acts as a representation of the parameter space, and maps pixel coordinates ( t,x,y ) of the input image sequence to the respective plasma parameters. For model training, the parameters are used as input to the fully-differentiable ME forward model which yields the corresponding spectral line profiles of the Stokes vector ( I,Q,U,V ). The difference (loss) between the predicted Stokes vectors and the observed data is used to update the network weights. The model is iteratively updated until the difference between the modeled and observed stokes vectors converges to a minimum. From the trained model, the individual pixel coordinates can be queried to obtain the final inversion result. The solar magnetic field is the omnipresent component that shapes the solar atmosphere, the heliosphere, and drives space weather ( Owens & Forsyth 2013 ) . The magnetic field in the lower solar atmosphere is studied through its signatures imprinted on the polarization of light ( Casini et al. 2017 ) . For this, spectropolarimetric inversions are used to determine the plasma conditions which most closely reproduce the observed polarization signal ( del Toro Iniesta & Ruiz Cobo 2016 ) . Methods for magnetic field inversions use an iterative approach to fit the observed spectral line properties while employing different levels of sophistication. Recent methods can account for the varying conditions from the photosphere, which is in local thermodynamic equilibrium (LTE), to the chromosphere which is out of LTE conditions ( de la Cruz Rodríguez & van Noort 2017 ) . However, computational requirements and numerical instabilities are strong limitations for the application of the most advanced inversions methods to the vast amounts of data current generation of solar telescopes produce ( Reardon et al. 2023 ) . A frequently applied approximation for interpreting lines formed in the solar photosphere is the Milne-Eddington (ME) model atmosphere. The ME approximation poses that the source function changes linearly over the formation height of the spectral line under consideration, while the other atmospheric parameters are constant over that region. ME inversions are established in the pipelines of many current generation magnetographs, such as the Helioseismic and Magntic Imager ( Schou et al. 2012 , HMI;) onboard the Solar Dynamics Observatory ( Pesnell et al. 2012 , SDO;) , the Solar Optical Telescope Spectro-Polarimeter ( Tsuneta et al. 2008 ; Lites et al. 2013 , SOTSP;) onboard Hinode, and the Synoptic Optical Long-term Investigations of the Sun ( Keller et al. 2003 , SOLIS;) operated by the National Solar Observatory (NSO), as they provide the real-time speed required for operational needs. Furthermore, their performance has been benchmarked against the different implementations used in the community on realistic radiative magnetohydrodynamic (rMHD) simulations to show high degree of accuracy ( Borrero et al. 2014 ) . To enable ME inversions to utilize the correlations between the neighboring pixels, inversions with spatio-temporal coupling based on sparse-matrix algorithmic approaches ( van Noort 2012 ) have been used to mitigate observational noise and improve the precision of the obtained magnetic fields ( Morosin et al. 2020 ; de la Cruz Rodríguez & Leenaarts 2024 ) . However, these methods typically have large computational memory requirements which are prohibitively challenging for larger field-of-views and longer temporal series (e.g., full-disk observations). Machine learning methods were applied to solar magnetic field inversions and related applications (see Asensio Ramos et al. 2023 , for a review) . A frequent approach is the use of simulated data or existing inversion results to perform supervised training ( Milić & Gafeira 2020 ; Socas-Navarro 2005 ; Asensio Ramos & Díaz Baso 2019 ; Higgins et al. 2021 ; Higgins et al. 2022 ; Gafeira et al. 2021 ) . While this approach can achieve significant accelerations of inversion codes, the quality of the resulting inversion strictly depends on the provided data set. In contrast, in this study, we develop a novel inversion method that provides independent and self-consistent solutions, solely based on the observed Stokes vector. This approach sets the foundation to exceed the quality and to overcome challenges of existing inversion methods (e.g., computational demand of non-LTE inversions). Physics-Informed Neural Networks ( Raissi et al. 2019 , PINNs; ) provide a novel approach to smoothly integrate physics-based models and noisy observational data ( Karniadakis et al. 2021 ) . The method uses a neural network to encode the simulation volume, where input coordinates (e.g., x,y,z ) are mapped to the corresponding physical quantities (e.g., the magnetic field \vec{B}(x,y,z) ). The model training is performed by fitting a given boundary condition and by minimizing the residuals of the physical model (e.g., partial differential equations. Encoding the simulation volume into weights of a neural network can enable smoother solutions, with an intrinsic coupling of adjacent points. In addition, this encoding is highly memory efficient, making it well suited for high-dimensional problems, and can simultaneously encode small- and large-scale features. In solar physics, the concept of a neural representation was used for magnetic field extrapolations ( Jarolim et al. 2023 ; Jarolim et al. 2024a ; Jarolim et al. 2024b ) , tomographic reconstructions ( Bintsi et al. 2022 ; Jarolim et al. 2024c ; Asensio Ramos 2023 ) , and solar flux transport ( Athalathil et al. 2024 ) . Díaz Baso et al. 2024 used a PINN for spectropolarimetric inversions under the Weak-Field Approximation (WFA) of chromospheric Ca II 854.2 nm solar observations. In this study, we present a Physics Informed