SDOFMv2: A Multi-Instrument Foundation Model for the Solar Dynamics Observatory with Transferable Downstream Applications

Foundation models have emerged as a promising paradigm in heliophysics by enabling transferable representations across diverse solar analysis tasks. Recent efforts such as SDOFM, Solaris, and Surya have demonstrated the potential of self-supervised learning for Solar Dynamics Observatory (SDO) data. In this work, we introduce SDOFMv2 , a family of foundation models trained on more than 500,000 observations from the Atmospheric Imaging Assembly (AIA) and Helioseismic and Magnetic Imager (HMI) instruments spanning most of Solar Cycle 24 (2010 – 2018). The framework includes modality-specific and joint AIA+HMI models, a region-aware masking and reconstruction objective that explicitly distinguishes solar-disk, prominence, and background regions, and a denoising refinement stage designed to improve representation quality for challenging magnetic-field observations. Qualitative and quantitative evaluations demonstrate substantial improvements in reconstruction fidelity relative to SDOFMv1. Extensive ablation studies and downstream evaluations against SDOFMv1, Solaris, and Surya reveal that modality-specific and joint foundation models exhibit complementary strengths across reconstruction, solar activity regression, and solar wind classification tasks, providing one of the most comprehensive benchmarking studies of heliophysics foundation models to date. We further evaluate the learned representations on multiple downstream tasks, including missing-channel reconstruction, solar radio flux ( F10.7 F 10.7 ) regression, and solar wind classification, while comparing against several state-of-the-art heliophysics foundation models. Across these evaluations, SDOFMv2 achieves competitive or superior performance, demonstrating the effectiveness of multimodal masked-autoencoding representations for solar observations. Besides, this study provides a benchmark evaluation framework that facilitates systematic evaluation of representation-learning approaches for solar observations. To facilitate reproducibility and future benchmarking efforts, model weights and source code are publicly available at github.com/Joaggi/sdofmv2 .