Regression of Stellar Atmospheric Parameters from Spectra Using Machine Learning

Stellar atmospheric parameters, such as effective temperature (Teff), metallicity ([Fe/H]), and surface gravity (log g), can be estimated from spectroscopic data through supervised learning methods. In this work, we investigate the regression of these three parameters from stellar spectra of the SDSS DR12 survey using a multi-task residual Multi-Layer Perceptron. The spectra were resampled onto a common 4000-point wavelength grid via cubic spline interpolation, shifted to the stellar rest frame, and standardized, while the target variables were scaled using RobustScaler. The dataset was divided into training, validation, and test sets with object-level separation in order to avoid data leakage. Gaussian noise augmentation was applied only to the training set. The proposed architecture combines a shared residual backbone with compact task-specific branches, allowing the joint learning of common and parameter-dependent spectral representations. Model selection was carried out through Bayesian optimization. The obtained results show mean absolute errors of 59.1 K for Teff, 0.100 dex for [Fe/H], and 0.131 dex for log g on the held-out test set. Compared with neural baselines trained under the same data split and preprocessing pipeline, the proposed approach achieved competitive or lower errors while using only 0.54 million parameters, indicating that residual multi-task MLPs provide an efficient alternative for stellar parameter estimation from large spectroscopic surveys.