New AI Foundation Model Learns the Sun Across Instruments and Solar Cycle

A team of researchers has unveiled SDOFMv2, a new family of artificial intelligence foundation models trained on more than 500,000 observations from NASA’s Solar Dynamics Observatory (SDO), designed to learn general-purpose representations of the Sun that can be transferred to a wide range of scientific tasks. The work, published in the journal Solar Physics, was led by Jinsu Hong of Georgia State University, together with Daniela Martin of the University of Delaware and Joseph Gallego of Pennsylvania State University. The models were trained on data from two of SDO’s flagship instruments, the Atmospheric Imaging Assembly (AIA) and the Helioseismic and Magnetic Imager (HMI), spanning most of Solar Cycle 24 between 2010 and 2018, and the pretrained weights and source code have been released publicly to support reproducible research.

The Solar Dynamics Observatory has transformed solar science since its launch, delivering continuous, multi-wavelength monitoring of the Sun and generating roughly 1.5 terabytes of data every day. AIA captures the full solar disk in multiple extreme ultraviolet, ultraviolet, and visible channels, each sensitive to plasma at different temperatures and heights in the solar atmosphere, producing 57,600 images daily. HMI complements this view by measuring the photospheric magnetic field through spectropolarimetric observations, providing magnetograms, Dopplergrams, and continuum images. Together these instruments underpin real-time monitoring of solar flares and coronal mass ejections that can threaten power grids, GPS, and satellite communications. Yet despite this data abundance, analysis methods in solar physics have remained largely task-specific, relying on handcrafted features, curated event catalogs, or models trained independently for each objective, a fragmentation that motivated the search for shared, reusable representations.

Foundation models offer a way out of this fragmentation. Analogous to large language models or vision models trained on massive datasets, a foundation model learns broadly useful structure from unlabeled data and can then be adapted to many downstream tasks with lightweight task-specific heads or fine-tuning. SDO’s archive is uniquely suited to this paradigm because of its long-term continuity, consistent calibration, and multi-channel coverage, which allow self-supervised training without extensive manual labeling. SDOFMv2 builds on earlier efforts including SDOFMv1, the first masked-autoencoder-based foundation model for SDO imagery, as well as Solaris, a forecasting-centered transformer model, and Surya, a large-scale spatiotemporal model trained on native 4096 by 4096 resolution observations with roughly 366 million parameters.

At the heart of SDOFMv2 is a masked autoencoder built on a Vision Transformer backbone. During pretraining, a high fraction of image patches is hidden and the model must reconstruct the missing regions, forcing it to learn global spatial structure rather than local pixel interpolation. The team introduced several solar-specific innovations. A region-aware masking and reconstruction strategy explicitly distinguishes the inner solar disk, informative off-limb structures such as prominences, and nearly empty background regions, which can occupy roughly 20 percent of a square full-disk image. Background areas are strongly down-weighted in the reconstruction loss so the model concentrates its capacity on physically meaningful solar morphology rather than trivial black corners.

The loss function itself was redesigned around the physics of solar imagery. A weighted, region-aware objective balances reconstruction of the disk, prominences, and background, while a Huber loss provides robustness to the extreme statistical imbalance between the quiet Sun and intense, sparse active regions. Because magnetograms contain sparse, high-frequency magnetic structures with weaker spatial correlations than smooth EUV emission, the researchers added a denoising fine-tuning stage in which Gaussian noise is injected into inputs and the model learns to recover the clean signal. This stage, combined with a reduced masking ratio of 10 percent for magnetic data, substantially improved HMI reconstruction quality. A signum-logarithmic transformation was also applied to compress the enormous dynamic range of solar intensities and magnetic fields while preserving sign information.

The framework includes three configurations: an AIA-only model with approximately 114 million parameters, an HMI-only model with about 111 million, and a joint AIA+HMI model with 114 million. Notably, the models were trained on workstation-scale hardware, four compute nodes with a total of eight NVIDIA RTX 5090 GPUs, requiring roughly three weeks and an estimated 670 GPU-hours per model. The team found that data engineering mattered as much as raw compute: converting the dataset to the cloud-optimized Zarr format reduced image loading times from 20 to 40 seconds to under 100 milliseconds, and a RAID 0 array of NVMe SSDs achieving about 21 gigabytes per second kept the GPUs continuously fed. An interleaved monthly splitting strategy ensured that training, validation, and test sets each sampled the full range of solar activity across the cycle.

Qualitative and quantitative evaluations showed substantial improvements in reconstruction fidelity over SDOFMv1, with sharper intensity gradients, better preserved coronal loops, and faithful recovery of off-limb prominence emission that the earlier model largely omitted. Analyses of the learned latent space using principal component analysis revealed that the dominant modes of variation consistently highlight active regions, magnetic concentrations, and prominences, while attention maps from the transformer encoder showed that different attention heads specialize in complementary solar structures and adapt their focus as solar conditions change. Ablation studies indicated that modality-specific models remain strongest for pure reconstruction, and that among joint models the ViT-Base configuration offered the best balance of capacity and performance, with the larger ViT-L variant not surpassing it under equal training budgets.

The true test of a foundation model is transfer, and here SDOFMv2 was evaluated on three downstream tasks against SDOFMv1, Solaris, and Surya in one of the most comprehensive benchmarking studies of heliophysics foundation models to date. In regression of the F10.7 solar radio flux, a widely used ground-based proxy for solar EUV emission, Surya achieved the best overall performance with a coefficient of determination of 0.918, consistent with its high-resolution, forecasting-oriented pretraining, while SDOFMv2’s AIA-based variants remained competitive and outperformed Solaris on mean squared error and R-squared. In classification of solar wind phenomena into streamer belt, sector reversal, coronal hole, and ejecta categories using Parker Solar Probe in situ labels, fine-tuned SDOFMv2-AIA delivered the highest accuracy, precision, and F1-score, and all SDOFMv2 variants outperformed the corresponding SDOFMv1-based results. In missing-channel reconstruction, where an entire AIA wavelength must be inferred from the others, the AIA-only model achieved a mean R-squared of 0.96, improving on SDOFMv1’s 0.94 and exceeding both Solaris and Surya on agreement with ground truth.

The results carry a broader message about how machine learning is reshaping heliophysics. Modality-specific and joint foundation models exhibit complementary strengths, and accurate pixel-level reconstruction turns out not to be a strict prerequisite for useful downstream representations, as the HMI model achieved competitive solar wind classification despite lower reconstruction fidelity. The authors acknowledge limitations, including the reduced spatial resolution of the SDOML dataset, where each pixel corresponds to roughly 1,353 kilometers on the solar surface, and the absence of Dopplergrams and continuum intensity products, and they point toward temporal forecasting objectives and cross-modal alignment as future directions. By releasing model weights, training code, and a benchmark evaluation framework openly, the team has established strong, reproducible baselines that could accelerate space-weather research and lower the barrier to entry for smaller groups hoping to build on solar foundation models.

Subject of Research: A multi-instrument AI foundation model for the Solar Dynamics Observatory with transferable applications in solar physics and space weather

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

Article References: Hong, J., Martin, D., & Gallego, J. (2026). SDOFMv2: A Multi-Instrument Foundation Model for the Solar Dynamics Observatory with Transferable Downstream Applications. Solar Physics, 301(9), Article 137. https://doi.org/10.1007/s11207-026-02740-z

Image Credits: AI Generated

DOI: 10.1007/s11207-026-02740-z

Keywords: Solar Dynamics Observatory, foundation models, machine learning, solar physics, masked autoencoder, space weather, heliophysics, solar magnetic fields, solar wind, self-supervised learning, AIA, HMI