The origin of the slow solar wind is a central question in heliophysics. The fast wind is confidently linked to coronal holes, but the slow wind may originate both in the streamer belt and at the boundaries of coronal holes. The problem is that these two channels produce plasma with nearly identical speeds, so they can only be distinguished by the composition of heavy ions, which 'freezes in' near the Sun. Existing approaches either rely on expert thresholds for one or two variables or use dimensionality reduction followed by clustering. The former miss the multidimensional boundary; the latter preserve point neighborhoods and thus cannot separate entangled populations. This work stands apart: nearby in the archive are studies on predicting solar wind with deep networks and on modeling slow wind sources, but no one has tackled the unsupervised separation of two slow populations in this way.
Solar-CDC combines contrastive learning with deep clustering. First, k-means is applied to seven standardized physical observables (speed, density, oxygen and carbon charge state ratios, iron-to-oxygen ratio, and mean oxygen charge) to obtain initial pseudo-labels. Then a transformer encoder maps observations into a latent space where a triplet loss function pulls points from the same pseudo-cluster together and pushes points from different clusters apart. Every four epochs, pseudo-labels are updated by re-clustering the latent codes, forming a self-correcting loop. A key detail: a positive pair is not an augmentation of a single observation (as in computer vision) but another real observation assigned by the current grouping to the same source. Theoretically, methods that preserve the nearest-neighbor graph cannot change the fraction of 'cut' edges, whereas a margin-based objective can reduce it to zero.
On 30,602 Solar Orbiter observations, Solar-CDC achieved a silhouette score of 0.869, while the best of 30 combinations of dimensionality reduction and clustering yielded only 0.454. Thirty hyperparameter grid configurations converged to three non-empty clusters, with the optimal model having one attention head and a two-dimensional latent space. The three found populations have mean O7+/O6+ ratios of 0.080, 0.160, and 0.400—exactly matching the thresholds for coronal hole, its boundary, and streamer belt. The two slow populations differ in speed by only 4 km/s, but in oxygen charge state by a factor of 2.5. When the defining O7+/O6+ ratio is excluded from the input, the model still recovers the taxonomy (NMI 0.321 vs. 0.000 for random partitioning). TriMap, which also optimizes triplets, achieves a silhouette of 0.824, but its clusters are physically meaningless (adjusted Rand index −0.021).
This work shows for the first time that separating the two slow solar wind populations is possible without any labels, relying solely on plasma composition. The theoretical boundary explains why popular visualization methods like t-SNE and UMAP cannot amplify weak boundaries in data: they preserve neighborhoods rather than rewriting them. This challenges the common practice of clustering in reduced dimensions for problems where populations are entangled in the original space. Unlike neighboring work on solar wind prediction, the focus here is not on prediction but on isolating physically distinct sources. The method is independent of specific observables and can be applied to any data where the signal is distributed across multiple variables.
A natural next step is to include magnetic field data to directly test the Alfvénicity of the intermediate population. Historical archives from ACE, Wind, and Ulysses missions would allow testing the robustness of the separation across different solar cycle phases. Since the theoretical boundary is not tied to heliophysics, the method could be transferred to spectral archives, geochemical analyses, and single-cell sequencing data, where populations are also nearest neighbors in the measured space.
The results directly affect models of the coronal magnetic field and our understanding of how the solar wind fills the heliosphere. For space weather forecasting, it is important to know which type of slow wind will reach Earth, as different sources carry different disturbances.
The immediate next step is to add magnetic fluctuation statistics to the model and check whether the found boundary population matches the Alfvénic slow wind. Then, the model should be trained on data from multiple missions simultaneously to assess transferability.
For heliophysics, the exact topology of coronal sources of the slow wind remains an open question: which structures contribute and how this changes with the activity cycle. This work does not solve that problem but provides a tool for investigating it—a way to objectively identify populations by composition without a priori assumptions. The connection to the broader problem of understanding coronal heating and wind acceleration is still indirect, but the obtained clusters can serve as a benchmark for testing models.
This section is addressed to the author of the paper and their colleagues. Neighbouring works were found by meaning-based search across our archive, not by word overlap. We do not review: we show the field around the work.
What the work does
The authors developed Solar-CDC, a self-learning contrastive deep clustering algorithm that, without labels, separated 30,602 Solar Orbiter observations into three physically meaningful groups. They proved theoretically that standard dimensionality reduction methods (t-SNE, UMAP) fundamentally cannot enhance the separation inherent in the data. The resulting clusters correspond precisely to the taxonomy based on oxygen charge states, including an intermediate population associated with coronal hole boundaries.
Where its strength is
The strength of the work lies in combining contrastive learning with self-correcting clustering, which allows separating populations with nearly identical speeds based on the multidimensional composition of heavy ions. The theoretical proof of the limitations of neighborhood-preserving methods provides a clear justification for the architecture choice.
Where it could go next
Worth trying
Why this may matter
If the method proves capable of separating populations without key variables, it will pave the way for automatic classification of large datasets from future missions such as Parker Solar Probe and Solar Orbiter, and will allow linking observed wind types to physical conditions in the corona.
Nearby works in our archive
The archive is sparse around this work: the closest review is only 0.54, against a usual 0.64. It looks like few have drilled nearby.
The speed difference between the two slow populations is only 4 km/s, about the speed of a walking person, while the wind itself travels at about 400 km/s.
The idea of hidden populations distinguishable only by subtle chemical traces echoes the motif from James S. A. Corey's 'The Expanse,' where the protomolecule leaves elusive signatures that require nontrivial analysis.