arxiv: 2604.08710 · v1 · submitted 2026-04-09 · ⚛️ physics.space-ph

Recognition: unknown

Solar Wind Classifications at Mars using Machine Learning Techniques

Catherine E. Regan , Silvia Ferro , Austin M. Smith , Alvin J. G. Angeles , Nicholas A. Gross , Farzad Kamalabadi , Marco Velli , Jasper S. Halekas

Authors on Pith no claims yet

Pith reviewed 2026-05-10 16:56 UTC · model grok-4.3

classification ⚛️ physics.space-ph

keywords solar windMarsMAVENK-means clusteringsolar activityregime classificationunsupervised learningspace physics

0 comments

The pith

Unsupervised machine learning on MAVEN data classifies solar wind at Mars into four regimes modulated by solar activity.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper applies Principal Component Analysis and K-Means clustering to a normalized multi-parameter dataset of solar wind measurements collected by MAVEN at Mars across Solar Cycles 24 and 25. This data-driven process identifies four recurrent regimes labeled slow, fast, intermediate, and compressed. Their relative frequencies and the way they appear in sequence over time vary markedly with the overall level of solar activity. A reader would care because these regimes control how the solar wind interacts with the Martian environment and because the same approach can be repeated at other heliocentric distances.

Core claim

An unsupervised framework that first reduces a normalized multi-dimensional solar wind dataset with Principal Component Analysis and then applies K-Means clustering recovers four physically interpretable regimes—slow, fast, intermediate, and compressed—whose occurrence rates and temporal sequencing are strongly organized by solar activity.

What carries the argument

The combination of Principal Component Analysis followed by K-Means clustering performed on a normalized set of solar wind parameters measured by MAVEN.

If this is right

Four regimes (slow, fast, intermediate, compressed) are recovered directly from the data without manual boundary definitions.
The fraction of time each regime occupies changes systematically with solar activity level.
The sequence in which regimes appear is organized according to the phase of the solar cycle.
The same classification can be applied to future MAVEN or other spacecraft intervals to track regime evolution.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The regime labels could be used to bin MAVEN atmospheric measurements and test whether each regime produces a statistically distinct response in the Martian ionosphere or exosphere.
Repeating the pipeline at 1 AU with ACE or Wind data would show whether the same four-regime structure persists or whether new intermediate states appear closer to the Sun.
If the compressed regime is driven by stream interaction regions, its occurrence should correlate with the number of high-speed stream encounters observed at Mars.

Load-bearing premise

The four K-means clusters truly represent distinct physical solar wind regimes rather than being produced by the particular normalization, feature choices, or the selected number of clusters.

What would settle it

Repeating the identical analysis after swapping to a different normalization scheme or forcing K to three or five yields regime boundaries and properties that no longer match the published four-regime description.

Figures

Figures reproduced from arXiv: 2604.08710 by Alvin J. G. Angeles, Austin M. Smith, Catherine E. Regan, Farzad Kamalabadi, Jasper S. Halekas, Marco Velli, Nicholas A. Gross, Silvia Ferro.

**Figure 1.** Figure 1: Representative multi-panel time series of solar wind parameters during the minimum of Solar Cycle 24. Panels show (from top to bottom): magnetic field magnitude |B|, proton velocity vp, proton density np, proton temperature Tp, and Alfv´en velocity VA. Data correspond to the interval 1 October 2018 – 15 January 2019; they illustrate typical behavior and do not show the full dataset. 2019), highlighting ty… view at source ↗

**Figure 2.** Figure 2: Explained variance ratio of the principal components computed from the normalised MAVEN dataset. The first six components account for the majority of the total variance and are therefore used as the reduced feature space for K-Means clustering. has been applied on over 50 years of data and is found to be extremely accurate. The presence of these parameters may introduce correlated variance into the PCA spa… view at source ↗

**Figure 3.** Figure 3: Sensitivity of K-Means clustering results using six retained principal components (PCA = 6). The left panel shows the silhouette score as a function of the number of clusters (k), while the right panel displays the corresponding K-Means inertia. The dotted line segments in the inertia panel indicate the change in slope around k = 6, highlighting the elbow that motivates the adopted number of clusters [PIT… view at source ↗

**Figure 4.** Figure 4: Heatmap showing the percentage contribution of the original features to the first six principal components. The values in the heatmap indicate the percentage contribution of each feature to the respective principal component. This highlights that a majority of each PC is dominated by 1 to 3 original features. strength effects, and plasma thermodynamics. Together, the PCA separates the dataset into modes co… view at source ↗

**Figure 5.** Figure 5: Multi-panel time series of solar wind parameters during the minimum of Solar Cycle 24. Panels show (from top to bottom): magnetic field magnitude |B|, proton velocity vp, proton density np, proton temperature Tp, and Alfv´en velocity VA. Data correspond to the interval 1 October 2018 – 15 January 2019. Points are colored by cluster assignment. coinciding with enhancements in magnetic field strength and pro… view at source ↗

**Figure 6.** Figure 6: Multi-panel time series of solar wind parameters during the maximum of Solar Cycle 24. Panels show (from top to bottom): magnetic field magnitude |B|, proton velocity vp, proton density np, proton temperature Tp, and Alfv´en velocity VA. Data correspond to the interval 1 December 2014 – 15 March 2015. Points are colored by cluster assignment. extreme compression events. Together, these patterns reflect an … view at source ↗

**Figure 7.** Figure 7: Multi-panel time series of solar wind parameters during the maximum of Solar Cycle 25. Panels show (from top to bottom): magnetic field magnitude |B|, proton velocity vp, proton density np, proton temperature Tp, and Alfv´en velocity VA. Data correspond to the interval 1 January 2025 – 15 April 2025. Points are colored by cluster assignment. MAVEN upstream solar wind dataset spanning Solar Cycles 24 and 25… view at source ↗

**Figure 8.** Figure 8: Percentage of observations within each cluster over time, during the maxima and minima of solar cycle 24, and maxima of solar cycle 25 [PITH_FULL_IMAGE:figures/full_fig_p019_8.png] view at source ↗

read the original abstract

Understanding solar wind variability throughout the heliosphere is essential for fundamental space physics and future exploration of the Moon and Mars. The Mars Atmosphere and Volatile EvolutioN (MAVEN) spacecraft has provided upstream solar wind measurements at Mars spanning Solar Cycles 24 and 25, enabling a statistical investigation of solar wind regimes at this heliocentric distance. In this work, we apply an unsupervised machine-learning framework combining Principal Component Analysis and K-Means clustering to a normalized, multi-dimensional solar wind dataset to identify recurrent solar wind regimes in a physically interpretable, data-driven manner. The resulting classification reveals distinct slow, fast, intermediate, and compressed solar wind regimes whose relative occurrence and temporal organization are strongly modulated by solar activity. This manuscript is part of the Heliophysics Summer School Machine Learning Special Collection.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

This applies standard PCA plus K-means to MAVEN solar wind data and labels four regimes, but the physical distinctness rests on untested interpretation.

read the letter

The paper takes normalized MAVEN upstream measurements at Mars, reduces them with PCA, and runs K-means to produce four clusters that the authors map to slow, fast, intermediate, and compressed solar wind. The central new element is the extension of this unsupervised approach to 1.5 AU with data spanning Solar Cycles 24 and 25, plus the observation that the clusters' occurrence rates shift with solar activity levels. That temporal link is the most concrete output and could feed into atmospheric-loss calculations for Mars missions. The methods section is transparent about the normalization and the pipeline, which makes the work easy to follow and potentially reusable as a template at other planets. The dataset itself is a real strength; MAVEN provides a long, continuous record that earlier studies lacked at this distance. The soft spots are in the validation. No silhouette scores, no elbow-method justification for K=4, no stability tests under different normalizations or feature subsets, and no side-by-side comparison of cluster centroids against literature thresholds for speed, density, or magnetic field. Without those checks the mapping to named regimes stays interpretive rather than demonstrated, and the stress-test point about possible artifacts from the chosen preprocessing is fair. The claim of strong solar-activity modulation therefore depends on how robust the partitions prove to be. This is the kind of paper that space-physics groups working on planetary solar-wind interactions or on ML classification of in-situ data would read for the occurrence statistics and the Mars-specific example. It is not a methodological advance, but the data product could be cited in modeling papers. It deserves a serious referee because the measurements are high-quality and the application is timely; the current draft would come back with requests for the missing robustness tests, which is standard for this style of work. I would send it to review.

Referee Report

4 major / 2 minor

Summary. The paper applies PCA followed by K-means clustering (K=4) to normalized multi-dimensional MAVEN solar wind observations at Mars to derive a data-driven classification into slow, fast, intermediate, and compressed regimes, asserting that these clusters are physically interpretable and that their occurrence rates and temporal patterns are strongly modulated by solar activity across Solar Cycles 24 and 25.

Significance. If the clusters can be shown to map robustly onto established physical regimes rather than algorithmic artifacts, the work would supply a reproducible, unsupervised framework for solar-wind regime identification at 1.5 AU that could be applied to other heliospheric datasets and would strengthen statistical studies of solar-wind-driven atmospheric escape at Mars.

major comments (4)

[§3] §3 (Methods): the choice of K=4 is presented without any cluster-validation metric (silhouette score, elbow criterion, or gap statistic) or stability test across random initializations; because the central claim rests on the existence of four distinct, physically meaningful regimes, this omission leaves the number of clusters unanchored.
[§4] §4 (Results): the mapping of the four clusters to 'slow', 'fast', 'intermediate', and 'compressed' solar wind is performed by post-hoc inspection of centroid values; the manuscript does not report quantitative comparison of those centroids (proton speed, density, |B|) against literature thresholds or against an independent physical labeling of the same intervals.
[§4] §4 (Results): no sensitivity analysis is shown for the normalization scheme or for alternative feature subsets; K-means partitions are known to change under different scalings, so the reported regime boundaries and solar-activity modulation could be artifacts of the chosen preprocessing.
[§5] §5 (Discussion): the assertion that regime occurrence and temporal organization are 'strongly modulated by solar activity' is supported only by qualitative time-series plots; a statistical test (e.g., Spearman rank correlation with sunspot number or a null model that shuffles cluster labels while preserving occurrence rates) is required to substantiate the modulation claim.

minor comments (2)

[Figures] Figure captions should explicitly state the time interval and number of data points used for each panel.
[Abstract] The abstract states that the method produces 'physically interpretable regimes' but supplies no quantitative support; this phrasing should be softened or moved to the conclusions pending the requested validation.

Simulated Author's Rebuttal

4 responses · 0 unresolved

We thank the referee for the constructive comments. We address each major comment point-by-point below, indicating where revisions will be made.

read point-by-point responses

Referee: §3 (Methods): the choice of K=4 is presented without any cluster-validation metric (silhouette score, elbow criterion, or gap statistic) or stability test across random initializations; because the central claim rests on the existence of four distinct, physically meaningful regimes, this omission leaves the number of clusters unanchored.

Authors: The value of K=4 was selected to correspond to the established physical solar wind regimes (slow, fast, intermediate, compressed) documented in the heliophysics literature. To strengthen the justification and address the concern directly, we will add cluster validation metrics including the silhouette score, elbow criterion, and stability tests across multiple random initializations in the revised Methods section. revision: yes
Referee: §4 (Results): the mapping of the four clusters to 'slow', 'fast', 'intermediate', and 'compressed' solar wind is performed by post-hoc inspection of centroid values; the manuscript does not report quantitative comparison of those centroids (proton speed, density, |B|) against literature thresholds or against an independent physical labeling of the same intervals.

Authors: Cluster labels were assigned by comparing centroid values of proton speed, density, and |B| to typical ranges reported in prior Mars and 1 AU studies. We will revise the Results section to include a quantitative comparison table against specific literature thresholds and discuss feasibility of cross-validation with independent labeling methods. revision: yes
Referee: §4 (Results): no sensitivity analysis is shown for the normalization scheme or for alternative feature subsets; K-means partitions are known to change under different scalings, so the reported regime boundaries and solar-activity modulation could be artifacts of the chosen preprocessing.

Authors: Z-score normalization was applied to ensure equal weighting of features, consistent with standard multivariate clustering practice. We agree that sensitivity should be demonstrated; we will add analyses using alternative normalizations (e.g., min-max) and feature subsets in the revised manuscript to confirm robustness of the reported regimes and modulation patterns. revision: yes
Referee: §5 (Discussion): the assertion that regime occurrence and temporal organization are 'strongly modulated by solar activity' is supported only by qualitative time-series plots; a statistical test (e.g., Spearman rank correlation with sunspot number or a null model that shuffles cluster labels while preserving occurrence rates) is required to substantiate the modulation claim.

Authors: The occurrence time series exhibit clear alignment with solar cycle phases. To provide quantitative substantiation, we will incorporate statistical tests such as Spearman rank correlation of regime occurrence rates with sunspot number and a shuffled-label null model in the revised Discussion section. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the derivation chain

full rationale

The paper applies standard unsupervised PCA followed by K-Means clustering directly to a normalized multi-dimensional MAVEN solar wind dataset. No equations, fitted parameters, or self-citations reduce the output clusters, their labeling as slow/fast/intermediate/compressed regimes, or the reported solar-activity modulation to the input data by construction. The regimes emerge from Euclidean partitioning in the chosen feature space and are interpreted post-hoc against known physical thresholds; this interpretation step is independent of any author-defined tautology or prior self-citation load-bearing premise. The central claims therefore rest on the clustering results themselves rather than on any self-referential reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on the assumption that Euclidean clustering in PCA-reduced space yields physically meaningful solar wind regimes; no additional free parameters or invented entities are introduced beyond the standard choice of four clusters.

free parameters (1)

number of clusters K
Set to four to produce the reported regimes; the abstract does not state how this number was selected or validated.

axioms (1)

domain assumption Normalized multi-dimensional solar wind observations can be partitioned into a small number of physically interpretable clusters using K-means in PCA space
Implicit in the choice of unsupervised clustering without additional physical constraints or validation against independent labels.

pith-pipeline@v0.9.0 · 5455 in / 1271 out tokens · 33378 ms · 2026-05-10T16:56:35.633885+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

2 extracted references · 1 canonical work pages

[1]

https://iopscience.iop.org/article/10.1088/0004-637X/745/2/162

DOI. https://iopscience.iop.org/article/10.1088/0004-637X/745/2/162. Liu, D., Rong, Z., Gao, J., He, J., Klinger, L., Dunlop, M.W., Yan, L., Fan, K., Wei, Y.: 2021, Statistical Properties of Solar Wind Upstream of Mars: MAVEN Observations.The Astrophysical Journal113. DOI. Lloyd, S.: 1982, Least Squares Quantization in PCM.IEEE Transactions on Information...

work page doi:10.1088/0004-637x/745/2/162 2021
[2]

SOLA: text.tex; 13 April 2026; 0:08; p

DOI. SOLA: text.tex; 13 April 2026; 0:08; p. 23

2026