Source-linked AI summary

Discovering Dual-Origin Slow Wind from Solar Orbiter with Self-Supervised Contrastive Learning

Henry Han, Jorge Yero Salazar

arXiv:2608.22065v1astro-ph.SRcs.AIcs.LG

TL;DR

The paper addresses whether slow solar wind has one coronal source or two channels that overlap in speed and differ mainly in heavy-ion composition. Solar-CDC learns a latent representation with physically initialized, dynamically updated pseudo-labels and a triplet margin objective. On 30,602 Solar Orbiter observations, it achieves stronger geometric separation than standard baselines while recovering physically aligned populations, although the data cannot test Alfvénicity.

  • Problem

    The unresolved problem is separating slow-wind populations that have nearly identical bulk speeds but differ in joint heavy-ion composition using unlabeled in-situ measurements.

  • Method

    Solar-CDC uses a Transformer encoder, triplet margin learning, and dynamically updated pseudo-labels initialized by physical observables to learn a source-oriented latent space.

  • Results

    0.869 silhouette: Solar-CDC outperforms thirty dimensionality-reduction and clustering combinations peaking at 0.454 and recovers a compositionally aligned dual-origin partition.

  • Takeaways & Limitations

    The recovered intermediate-composition population supports the dual-origin hypothesis, while the result depends on physically aware initialization rather than separation alone.

  • Takeaways & Limitations

    The learned objective can separate an unphysical partition, and the silhouette coefficient cannot statistically distinguish k = 2 from k = 3.

Abstract

from arXiv · show

Whether the slow solar wind originates from one coronal source or two distinct channels remains a central open question in heliophysics. Resolving this requires unsupervised separation of two populations that arrive at nearly the same bulk speed and differ mainly in heavy-ion composition. We present Solar-CDC, a self-supervised contrastive deep clustering (CDC) framework that maps plasma observables to a latent space via a Transformer encoder, optimizes a triplet margin loss, and updates pseudo-labels via $k$-means. Theoretically, we prove that neighborhood-preserving embeddings such as t-SNE and UMAP are fundamentally constrained. Preserving the neighbor graph leaves the cross-cluster cut fraction unchanged, and preserving all but a fraction $\varepsilon$ of its links moves that fraction by at most $\varepsilon$. Neither bound depends on the target dimension. A margin objective rewrites the graph and drives the cut fraction to zero. Empirically, on 30,602 Solar Orbiter observations, thirty combinations of dimensionality reduction and clustering peak at a silhouette of $0.454$, whereas Solar-CDC reaches $0.869$. Escaping the geometric bound alone does not guarantee physical validity: TriMap also optimizes triplets and reaches $0.824$, yet its clusters score below chance against the published composition taxonomy. Solar-CDC instead recovers clusters with mean charge-state ratios of $0.080$, $0.160$, and $0.400$, placing the intermediate population inside the window associated with coronal-hole boundaries. Even when the defining charge-state ratio is withheld from the inputs entirely, the model still recovers the taxonomy defined on it. Solar-CDC thus connects self-supervised representation learning to coronal source diagnostics. Importantly, a learning loss recovers physical populations only when driven by dynamically updated physically-aware clusters rather than distances.

1 Introduction

The paper frames dual-origin slow-wind separation as an unsupervised composition problem because the candidate populations overlap in speed. Solar-CDC combines physically initialized pseudo-labels with contrastive representation learning to recover physically meaningful populations.

  • The slow wind may originate from one coronal source or two distinct channels, a central unresolved question in heliophysics.
  • The task requires distinguishing fast wind, streamer-belt slow wind, and boundary-type slow wind using in-situ plasma measurements alone.
  • The two slow-wind populations overlap in bulk speed, so their boundary lies in joint heavy-ion composition rather than any single threshold variable.
  • Solar-CDC learns a source-oriented latent representation and splits the slow wind into populations differing by 4 km s−1 in speed and a factor of 2.5 in oxygen charge state.
  • Solar-CDC alternates representation updates with group re-formation, using a Transformer encoder, contrastive pulls and pushes, and unlabeled pseudo-labels.
  • Physically measured warm-up groups anchor the final partition, while the theoretical and empirical contributions distinguish objective-driven separation from arbitrary geometric structure.

2 Related Work: Why Standard Tools Fall Short

Standard thresholds and neighborhood-preserving projections cannot separate slow-wind populations tangled in composition space. Solar-CDC addresses this by actively separating populations without labels or physics-damaging augmentations.

  • Traditional thresholds and bulk-parameter clustering separate fast from slow wind but struggle because candidate slow-wind populations overlap almost perfectly in bulk speed.
  • Predefined-label composition schemes motivate removing dependence on human-assigned class labels.
  • t-SNE, UMAP, and PHATE preserve local neighborhoods, keeping mutually neighboring slow-wind populations tangled rather than actively separating them.
  • TriMap escapes the neighborhood-preservation bound through triplet relationships but produces a geometrically tight partition that does not match the composition taxonomy.
  • Standard image augmentations do not transfer to seven-dimensional composition vectors because altering charge-state ratios changes the physical meaning of the measurement.
  • Solar-CDC combines observable-based initialization, an alternating triplet-loss loop, label-free learning, and no augmentation to separate overlapping populations.

3 Solar Orbiter Data and Feature Space

The study uses Solar Orbiter in-situ measurements combining bulk plasma properties and heavy-ion composition. Its seven physical observables provide the feature space for distinguishing solar-wind populations.

  • Solar Orbiter observations from January 2022 through April 2023 combine SWA-PAS bulk plasma measurements with SWA-HIS ion composition.
  • The model input consists of seven physical observables drawn from bulk plasma properties and ion composition.
  • Proton bulk speed separates fast from slow wind but cannot distinguish the two slow-wind sources, while proton density is higher in slow wind.

4 Methods

Solar-CDC learns a physically informed latent representation and partition from unlabeled plasma observations through warm-started, iteratively refreshed pseudo-labels and contrastive training. Its margin objective separates compositionally distinct slow-wind populations without manual thresholds or ground-truth labels.

  • Representation learning: Solar-CDC maps seven plasma observables into a latent space where distances reflect coronal-source similarity rather than bulk-speed similarity.A Transformer encoder produces the latent codes, whose dimension is not required to be smaller than the input dimension.
  • Representation learning: The method jointly learns latent codes and clusters intended to recover fast wind, boundary-type slow wind, and streamer-belt slow wind.The partition is defined over latent codes rather than directly in the original observable space.
  • Contrastive objective: Triplet training pulls each observation toward a same-group real observation and pushes it from a different-group observation by a fixed margin.The margin imposes relative distance constraints that break cross-boundary neighbor links rather than merely drawing points toward cluster centers.
  • Iterative refinement: Every p epochs, k-means reruns on updated latent codes, refreshing the pseudo-labels that define subsequent positive and negative triplets.This feedback loop lets the grouping and representation refine one another during training.
  • Physical separation: The contrastive loop separates two slow-wind streams that bulk speed cannot distinguish, without manual velocity thresholds or ground-truth labels.The learned boundary is placed according to joint charge-state structure.
  • Pseudo-label initialization: Warmup initializes pseudo-labels with k-means on standardized observables, anchoring the first assignments to physical plasma properties instead of random encoder noise.After initialization, physical knowledge is not supplied through external labels; subsequent supervision is self-supplied.

5 A Separability Bound for Neighborhood-Preserving Embeddings

The section shows that neighborhood-preserving embeddings cannot improve cross-cluster separability when populations overlap in the input neighbor graph, regardless of target dimension. Margin-based embeddings can rewrite that graph and remove cross-cluster neighbors, but physical initialization determines which partition is separated.

  • The target dimension does not determine separability; the embedding objective does.A mapping can separate arbitrary clusters in two dimensions, while neighborhood preservation constrains separation independently of dimension.
  • Exact κ-neighborhood preservation leaves the cut fraction unchanged for every partition.The result also applies to graph functionals such as conductance, modularity, and spectral-clustering objectives.
  • ε-approximate neighborhood preservation changes the cut fraction by at most ε.This bound follows from the limited number of altered neighbor links and applies to methods such as t-SNE and UMAP.
  • A margin-satisfying embedding drives the cut fraction to zero when κ is smaller than the smallest cluster size.The margin places every observation at least α farther from other clusters than from any same-cluster member.
  • At κ = 10, the raw cut fraction is β10 = 0.064, while t-SNE and UMAP return 0.071 and 0.084 and do not improve it.PHATE and PCA return 0.148 and 0.174; none of these four methods moves the cut fraction downward.
  • Solar-CDC reduces the cut fraction to β10 = 0.001, but the margin objective alone does not ensure a physically meaningful partition.The learned objective separates the partition it receives, so warmup from physical composition ratios determines which partition is sharpened.

6 Results

Solar-CDC achieves sharply separated clusters while showing that geometric tightness alone does not establish physical validity. Its recovered populations align with composition-based solar-wind structure, including when the defining oxygen ratio is withheld.

  • 0.869 is Solar-CDC’s optimal silhouette, compared with a 0.454 peak across thirty neighborhood-preserving baseline combinations.The optimal configuration uses one attention head and a 2-D latent code; the grid mean is 0.795.
  • Solar-CDC’s latent dimension significantly affects performance, whereas attention-head count does not.The Friedman test gives χ2 = 22.3, p = 1.8 × 10−4 for latent dimension and χ2 = 7.4, p = 0.19 for attention heads.
  • TriMap reaches a 0.824 silhouette, but its geometrically tight clusters are less physically meaningful than lower-silhouette neighborhood-preserving embeddings.Neighborhood-preserving methods achieve NMI 0.33–0.42, while Solar-CDC is the only compared method reported to achieve both geometric separation and physical agreement.
  • 0.321 NMI and 0.653 accuracy remain achievable when O7+/O6+ is withheld, showing that the recovered boundaries are not memorized from that single ratio.On the same six observables, k-means reaches NMI 0.311, statistically indistinguishable from Solar-CDC, while Solar-CDC retains a matched-accuracy advantage.
  • The three recovered clusters have mean O7+/O6+ ratios of 0.080, 0.160, and 0.400, matching coronal-hole, boundary, and streamer-belt windows.The two slow populations differ by only 4 km s−1 in bulk speed but by a factor of 2.5 in oxygen charge state.

7 Discussion

The recovered intermediate-composition population supports the dual-origin hypothesis, but its Alfvénic identity cannot be tested without magnetic and velocity-fluctuation data. Cross-method silhouette comparisons are indicative rather than absolute, and model selection still relies on external reference labels.

  • The recovered intermediate-composition slow-wind population supports the dual-origin hypothesis despite nearly overlapping bulk speeds.The evidence is extracted without labels and separates the population from streamer-belt wind by composition.
  • Alfvénicity remains untested because the seven observables omit magnetic and velocity fluctuations.Cluster 1 is therefore designated boundary type, while its identification as Alfvénic slow wind remains plausible rather than established.
  • Silhouette comparisons between Solar-CDC and dimensionality-reduction baselines are indicative rather than absolute because each method is scored in its own representation space.The internal Solar-CDC ablations use identical protocols and are directly comparable.
  • Configuration choice currently relies on external reference labels, leaving a fully unsupervised criterion for physical correctness as an open challenge.The discussion also identifies incorporating magnetic fluctuation statistics as the most valuable extension.

8 Conclusion

Solar-CDC recovers composition-ordered solar-wind clusters while surpassing conventional dimensionality-reduction and clustering combinations. The paper argues that margin-based objectives and physically anchored initialization overcome geometric constraints, while identifying broader validation needs.

  • 0.869 silhouette on 30,602 Solar Orbiter observations surpassed thirty dimensionality-reduction and clustering combinations, which peaked at 0.454 with zero distribution overlap.
  • Recovered clusters were monotonically ordered across every composition variable, mirroring the established charge-sequence of coronal source regions.
  • Neighborhood-preserving embeddings cannot reduce a partition’s cross-cluster cut fraction below its input-metric baseline, regardless of target dimensionality.
  • A margin objective can rewrite the graph and drive the cut fraction to zero, but success depends on initialization anchored in physical plasma measurements.
  • Future work proposes testing broader archives, integrating magnetic-field and fluctuation data, and applying the domain-agnostic bound to other near-neighbor populations.

A Supplementary Results

Table 7 compares silhouette across the full configuration grid using seven observables versus those observables plus two additional input columns.

  • Table 7 evaluates 30,602 observations with k = 3, 200 epochs, and p = 4 across grids with seven or nine observables.
Loading 2608.22065v1…