| Presentation Title: | Physics Informed Neural Networks for Spectropolarimetric Inversions |
| Author(s): | R. Jarolim, M. Molnar, B. Tremblay, R. Centeno |
Abstract:
Spectropolarimetric inversions of solar observations are fundamental for the estimation of the magnetic field in the solar atmosphere. To expedite solar inversions, a reasonable assumption for photospheric spectral lines is that they form under the Milne-Eddington (ME) approximation, where the source function is posed to vary linearly with optical depth, describing the atmosphere with a few (nine) free parameters. Despite the simplifications in current inversions, observational data typically exhibits noise and instrumental effects, which limits the success of estimating the solar plasma parameters.
We present a novel method for spectropolarimetric inversions of solar data under the ME approximation with a Physics Informed Neural Network (PINN) to estimate the solar magnetic field. Building on synthetic spectral line profiles, we demonstrate that our approach can reliably solve complex configurations in a computationally-efficient way. Our method intrinsically enables spatial and temporal coupled inversions that can overcome limitations of noise, directly provides uncertainty estimates, and is straight-forward to extend to different instrument configurations. Our method also accounts for the instrumental point-spread-function, further informing the spatial coupling.
Finally, we apply our method to observations from Hinode SOT/SP and compare our results to other state-of-the-art ME inversion codes. Our method operates with minimal memory requirements, makes efficient use of spatio-temporal relations, and is adjustable to observations from new instruments. This study provides a first step to utilize PINNs for spectropolarimetric inversions, and demonstrates the potential to address long standing challenges.