A Deep Learning Framework for Predicting Stellar Spectral Energy Distributions from Multiband Photometric Data

Abstract In this work, we present a carefully designed deep learning framework for reconstructing stellar spectral energy distributions (SEDs) from multiband photometric data spanning the near-ultraviolet to near-infrared. The framework comprises an autoencoder for latent space learning and a deep residual convolutional neural network (DRCNN) that performs nonlinear mapping in latent space. When trained to reconstruct BOSZ theoretical stellar spectra based on ATLAS9 model atmospheres from inputs at a resolution of 2000 to a resolution of 10,000, the framework achieves a typical reconstruction precision of approximately 0.001–0.007 over the wavelength range 3000–8000 Å. Even after accounting for observational uncertainties, the framework remains robust and delivers stable reconstruction performance. We further explore several extreme test cases in which the input information is limited to BOSZ fluxes at only 200, 43, or as few as 4 carefully selected narrowband photometric filters and find that the reconstructed BOSZ spectra remain consistent with the theoretical models at the level of approximately 0.002–0.03, depending on wavelength. When applied to Javalambre Physics of the Accelerating Universe Astrophysical Survey (J-PAS) Early Data Release photometry, the trained DRCNN reconstructs Gaia DR3 XP spectra with a typical consistency of approximately 0.01–0.12 across the wavelength range, with the reconstruction uncertainty dominated by XP spectral observational uncertainties at blue wavelengths and by J-PAS photometric errors at longer wavelengths, thereby confirming the stability and robustness of the framework under realistic observational conditions. Overall, the framework provides a statistically robust approach for constructing stellar SEDs from high-precision multiband photometry and demonstrates the possibility of improving spectral resolution in a statistical application.