CORONA-Fields: Leveraging Foundation Models for Classification of Solar Wind Phenomena

TL;DR

This paper addresses the challenge of linking solar imagery to in situ solar wind structures by adapting a solar foundation model (SDO-FM) based on MAE embeddings and enriching them with Fourier-encoded spacecraft position and magnetic connectivity. A neural-field classification head (linear or NeRF-inspired skip connections) maps the embeddings to a four-class wind-structure taxonomy, using a two-stage training strategy (freeze-then-fine-tune) and focal loss to mitigate class imbalance. Despite modest overall accuracy (~0.30–0.35), the approach demonstrates the feasibility of transferring image-based representations to heliospheric tasks, with evidence that positional encoding helps organize the embedding space and that end-to-end fine-tuning yields gains. The work lays groundwork for integrated solar-physics pipelines, offering a reproducible framework and pointing to future improvements like incorporating magnetograms and improved labeling to advance reliable space weather forecasting.

Abstract

Space weather at Earth, driven by the solar activity, poses growing risks to satellites around our planet as well as to critical ground-based technological infrastructure. Major space weather contributors are the solar wind and coronal mass ejections whose variable density, speed, temperature, and magnetic field make the automated classification of those structures challenging. In this work, we adapt a foundation model for solar physics, originally trained on Solar Dynamics Observatory imagery, to create embeddings suitable for solar wind structure analysis. These embeddings are concatenated with the spacecraft position and solar magnetic connectivity encoded using Fourier features which generates a neural field-based model. The full deep learning architecture is fine-tuned bridging the gap between remote sensing and in situ observations. Labels are derived from Parker Solar Probe measurements, forming a downstream classification task that maps plasma properties to solar wind structures. Although overall classification performance is modest, likely due to coarse labeling, class imbalance, and limited transferability of the pretrained model, this study demonstrates the feasibility of leveraging foundation model embeddings for in situ solar wind tasks. As a first proof-of-concept, it lays the groundwork for future improvements toward more reliable space weather predictions. The code and configuration files used in this study are publicly available to support reproducibility.

Figures (6)

• Figure 1: Architecture of the proposed model. The framework combines two input streams: (i) solar images processed by a pretrained MAE backbone (SDO-FM) to obtain an embedding layer, and (ii) positional information from the Parker Solar Probe and its magnetic footpoints, encoded through Fourier features. The embeddings and encoded positions are concatenated and passed through the classification head (either a simple linear head or a skip-connection variant), producing logits over the available solar wind classes. During training, predictions are compared with ground-truth labels using the focal loss function.
• Figure 2: Distribution of each solar wind class across each dataset.
• Figure 3: Training loss curves for a linear head with skip-connection. Three model configurations are compared: (1) head only, with a frozen, pretrained backbone, (2) backbone & head with a pretrained backbone fine-tuned during training, and (3) backbone & head with random initialization.
• Figure 4: t-SNE on a randomly selected balanced set of embeddings produced by SDO-FM with PSP's position included (top row) and without PSP's position included (bottom row) at five different perplexities. With positional information included, the embeddings are projected into a lower-dimensional space, as evidenced by the strip-like structure in the t-SNE visualization.
• Figure 5: Examples of correct predictions on the test set for three solar wind classes: coronal hole, sector reversal, and streamer belt. Each column shows an SDOML/AIA image, with crosses indicating the predicted magnetic footpoints from PSP.