PIST-Net: A Multimodal Deep Learning Framework for Red Giant Stellar Parameter Estimation

Abstract Red giant stars are pivotal for probing the formation and evolution of the Galaxy. Conventional methods for determining their parameters typically rely on high-precision, resource-demanding observations over long durations. In this work, we present a novel deep learning approach, the Photometric Image Spectral Transformer Net, which utilizes low-resolution spectra from LAMOST Data Release (DR) 10 and photometric data from SDSS DR17 for multimodal stellar parameter estimation. The proposed approach employs two specialized architectures: SP-Net for spectral data and ResNet for photometric images, which are subsequently fused using a transformer encoder for cross-modal integration. Our results demonstrate that this approach achieves promising performance in estimating five key physical parameters, including effective temperature ( T eff ), surface gravity ( <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mi>log</mml:mi> <mml:mi>g</mml:mi> </mml:math> ), metallicity ([M/H]), mass and radius. Notably, the model achieves particularly strong results for T eff and [M/H], with mean absolute errors of 0.025 and 0.021, respectively. This significantly outperforms models relying on a single modality. This study highlights the advantages of multimodal learning techniques in improving the accuracy of stellar parameter estimation, while reducing reliance on time-consuming observational methods. All code and data used in this study are available at GitHub ( https://github.com/yuexiuya1/PIST-Net ).