KAN-Payne: A Controlled Evaluation of Kolmogorov–Arnold Networks for Stellar Spectral Emulation and Label Recovery

Abstract

Kolmogorov–Arnold networks (KANs) replace the fixed activations of a multilayer perceptron (MLP) with learnable univariate edge functions. We evaluate KANs as Payne-style label-to-flux emulators using the public 1000-spectrum Kurucz grid released with The Payne. Two capacity-matched KAN–MLP pairs (approximately 2.3 and 18–19 million parameters) were tuned separately, trained for the same 120,000-update budget across three seeds, and evaluated on a fixed 200-spectrum holdout. At the smaller capacity, the KAN did not outperform the tuned MLP. At the larger capacity, KAN-L achieved the lowest flux errors among the models trained from scratch: its mean absolute error was 35% below the best small model, all three large-MLP layouts remained less accurate, and its residual tails were lighter. In injection–recovery tests, KAN-L improved the recovery of effective temperature, surface gravity, and metallicity at all the tested signal-to-noise ratios relative to the small matched MLP; the best large-MLP control was not included in this test. The learned KAN response shapes were reproducible across seeds. These gains required substantially greater training and inference costs. In an APOGEE Data Release 17 deployment, surface-gravity offsets decreased by approximately 0.09 dex for all the tested emulators when calibrated APOGEE Stellar Parameter and Chemical Abundances Pipeline (ASPCAP) values were replaced by the raw spectroscopic scale, indicating a common reference-scale contribution. KANs are therefore not efficient drop-in replacements for MLP emulators, but the large KAN tested here improved spectral emulation accuracy and residual-tail robustness.