Spectropolarimetric inversions are usually employed to obtain the physical parameters (e.g., magnetic field, velocity, temperature) of the solar atmosphere that contribute to the observed Stokes spectra. However, this process involves optimization in a multi-dimensional parameter space, which not only demands substantial computational time, but also frequently encounters local minima during the search. Our aim is to develop a spectropolarimetric inversion code leveraging the physics-informed neural networks (PINNs) and a novel gradient-bias annealing (GBA) algorithm with global minimum search capabilities to enable computationally efficient convergence toward near-optimal solutions. Inversions were trained on synthetic data generated by a forward model. The training process incorporated both data regression and cyclic physics-informed constraints to ensure physical consistency. Additionally, we developed a GBA algorithm to enable preferential annealing during optimization. The entire framework was implemented with CUDA for parallel acceleration. We evaluated the effectiveness of the GBA algorithm, PINN, and their combined approach using the Helioseismic and Magnetic Imager (HMI) Stokes spectrum observations. When employed independently, the PINN achieved inversion results comparable to HMI's quick products, via Very Fast Inversion of the Stokes Vector (VFISV), within significantly shorter timescales. The GBA algorithm alone demonstrated superior spectral recovery, despite requiring several hours to process millions of pixels in a single field of view. By initializing the annealing process with the PINN results, we obtained quality-enhanced results on minute-level timescales.
Chen et al. (Mon,) studied this question.