Spectral Handling and Estimation of AGN Parameters (SHEAP), The first AGN fitting GPU-based code

Abstract

In the coming years, the number of discovered active galactic nuclei (AGN) is expected to increase significantly due to upcoming spectroscopic surveys. This growth will challenge current analysis and modeling techniques, requiring scalable methods for large, heterogeneous datasets with diverse signal-to-noise ratios, spectral resolutions, and host-galaxy contamination. We present SHEAP (Spectral Handling and Estimation of AGN Parameters), a spectral-fitting framework designed to analyze large AGN samples efficiently while preserving physical interpretability, reproducibility, and robust uncertainty estimation. SHEAP uses JAX, a Python GPU-powered framework, to implement a flexible model with modular components, including continuum, host galaxy, FeII pseudo-continuum, and multi-component emission lines, together with parameter tying and physically motivated constraints. By combining gradient-based optimization with automatic differentiation, vectorization, and just-in-time compilation, SHEAP achieves stable convergence in blended regions, such as H$β$, while substantially reducing runtime. We compare SHEAP measurements with literature results and public fitting pipelines across four samples covering the CIV, MgII, H$β$, and H$α$ regions. We find good agreement for the main AGN spectral parameters, with $\sim85$--$100%$ of objects lying within the $\pm0.3$ dex band and reduced chi-square distributions close to unity. Relative to the runtime reported by \citet{2026Bernal} using \texttt{pPXF}, the fitting stage requires only $\sim1.7%$ of the computational time, corresponding to an improvement of approximately $100$ times. These results show that \texttt{SHEAP} delivers reliable AGN spectral decompositions at substantially lower computational cost, making it suitable for massive spectroscopic datasets.