Solar Wind Classifications at Mars using Machine Learning Techniques

arXiv:2604.08710 · physics.space-ph · Submitted 2026-04-09 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Physics. The day in general physics: instruments, methods and results across the field.

Kai: We're looking at the paper: "Solar Wind Classifications at Mars using Machine Learning Techniques".

Mira: Here's the gist of the paper. Understanding solar wind variability throughout the heliosphere is essential for fundamental space physics and future exploration of the Moon and Mars.

Kai: 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.

Mira: This study applies an unsupervised machine-learning framework combining Principal Component Analysis (PCA) 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.

Kai: The researchers analyzed a multi-year dataset from the MAVEN spacecraft spanning Solar Cycles 24 and 25.

Paper summary: Kai: So we're looking at this paper, "Solar Wind Classifications at Mars using Machine Learning Techniques." The main thing here is that understanding how the solar wind changes across the whole heliosphere is really important for figuring out stuff about exploring Mars and the Moon.

Mira: Right. So the authors are taking data from the MAVEN spacecraft, which has been measuring solar wind upstream of Mars during Solar Cycles twenty-four and twenty-five and they're using a statistical investigation to look at different solar wind regimes at that distance <ref:2604.08710#pg1,Solar Cycles 24 and 25>.

Lev: And what they claim is that they 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 way.

Kai: So it's not just looking for patterns; it's trying to find physical groupings in the data that make sense. The whole point is that this classification reveals distinct slow, fast, intermediate, and compressed solar wind regimes whose timing and organization are strongly affected by solar activity.

Mira: And they’re using PCA to reduce the fourteen-dimensional feature space down to six dimensions before applying K-Means clustering in that reduced space. That's a pretty heavy computational lift for this kind of analysis, isn't it?

Lev: Yeah, yeah, but the goal is to get those physically interpretable modes out loud. The PCA decomposition reveals specific modes: PC1 captures bulk velocity and temperature stuff, PC2 deals with density and magnetic field strength, PC3 isolates the transverse magnetic field orientation, and PC5 is driven by the magnetic field z-component.

Kai: That’s useful because it connects the math back to what we actually see in the plasma physics. So Cluster three and four are interesting because they show intermediate bulk velocities and enhanced magnetic field magnitudes compared to Cluster zero which suggests transitional solar wind conditions bridging slow and fast wind states.

Mira: And then you have Clusters one and two representing the fast solar wind streams, which display elevated bulk velocities, reduced proton densities, increased Alfvén speeds, and significantly higher proton temperatures <ref:2604.08710#pg1>. Meanwhile, Cluster five stands out because it has markedly enhanced proton density and strong magnetic field magnitude across multiple parameters.

Paper summary: Lev: That Cluster five sounds like what we’d call compressed or disturbed plasma conditions; it suggests episodic extreme events are being captured there by the clustering algorithm. The paper uses this classification to show how different solar activity levels modulate the occurrence of these regimes, which is key because solar minima dominate with long-lived slow wind, while solar maxima show frequent transitions between regimes.

Kai: So the structure of the data—the way these regimes organize themselves over time—is directly linked to the behavior of the Sun itself. That’s a big connection for understanding how space weather affects distant environments like Mars.

Mira: It really is, and it ties into what we know from earlier work, like Feldman, Landi, and Schwadron's two thousand five delineation between fast and slow wind originating in different source regions depending on the solar cycle phase <ref:2604.08710#pg2>.

Lev: But the paper's limitation is that it relies on a synthesized upstream driver dataset provided by Halekas and colleagues for its input variables. So, while the machine learning framework is solid, the accuracy of those classifications depends heavily on how well that initial data represents reality.

Kai: And this opens up a question for future work: could we apply this exact same framework to unprocessed solar wind data collected by MAVEN and see if we get similar classifications at Mars? That’s where the real comparison comes in.

Mira: Exactly. The paper shows the value of unsupervised classification techniques because they allow us to characterize these solar wind variations in a way that is both data-driven and physically grounded simultaneously. It provides a foundation for comparative studies across different heliocentric distances and planetary environments, which is exactly what we need for missions like ESCAPADE going to Mars in two thousand twenty-seven <ref:2604.08710#pg2>.

Lev: So the implication for hardware development is that if you’re building something on real hardware, like error correction systems, you have to account for the overhead of processing these complex classifications before you can trust the input data. The sorting and classification step itself adds complexity.

Kai: It's about using this kind of modeling to characterize solar wind variability in a way that's physically grounded, which is something we need when we move from simulations to actual measurements on spacecraft.

Paper summary: Mira: And it moves us toward understanding the hybrid magnetosphere environment Mars experiences, as mentioned in the introduction, because characterizing these flow regimes at MAVEN’s distance helps map out how that interaction works.

Lev: So the title itself, "Solar Wind Classifications at Mars using Machine Learning Techniques," points to a very specific application of advanced statistics to solve a real exploration problem. It's less about discovering new physics and more about applying robust tools to interpret existing observational data for future mission planning.

Kai: I think that’s fair. The authors are using this framework as a tool, not necessarily inventing a totally new physical theory of the solar wind itself, but rather providing a better way to categorize the observed behavior across different solar cycles.

Mira: And they do manage to show how the relative occurrence and temporal organization of these distinct regimes are strongly modulated by solar activity. That modulation is what makes the whole analysis relevant for understanding long-term space weather trends.

Lev: So, in short, this paper provides a way to take a massive amount of multi-dimensional solar wind data and sort it into six physically meaningful buckets—slow, fast, intermediate, compressed—and show how those buckets shift depending on whether the Sun is at its minimum or maximum phase.

Kai: And the next step is taking that classification scheme and applying it directly to real MAVEN data to see if those same regimes show up when we look at the solar wind environment near Mars. That’s a really tangible path forward.

Mira: So, to wrap up, this paper gives us a data-driven way to categorize the solar wind based on measurable plasma and magnetic field parameters, revealing how those categories change with the solar cycle phase. It frames solar wind variability in a way that directly informs future missions like those targeting Mars.

Lev: And for anyone building models or error correction systems, the point is that you need to be ready for these transitions; the data isn't static. The classification scheme highlights where those transitions happen most often, which means where you might expect the hardest problems to occur in real hardware.

Kai: That’s what we’re looking at here: using machine learning not just to predict things, but to provide a physically interpretable lens through which we can view the complex dynamics of solar wind across our entire solar system.

Conclusion: Kai: So we're wrapping up on this piece about "Solar Wind Classifications at Mars using Machine Learning Techniques."

Mira: Yeah, we've been looking at how they used machine learning to sort the solar wind data from MAVEN into these distinct regimes.

Lev: It’s interesting because they didn't just look for a single pattern; they built a system that groups the plasma and magnetic field measurements together based on similarity.

Kai: The authors are using Principal Component Analysis and K-Means clustering to find six clusters that correspond to slow, fast, intermediate, and compressed solar wind conditions.

Mira: What's important here is how they tie those mathematical groupings back to real physical states like bulk velocity and magnetic field orientation.

Lev: The numbers show that the relative occurrence of these regimes shifts dramatically depending on whether we are in a solar minimum or a solar maximum phase.

Kai: So, the title itself points to using this classification method not just for Earth, but specifically to figure out what the environment is like when you get close to Mars.

Mira: It suggests that understanding those solar wind patterns is key because it helps us predict how the plasma environment at Mars might behave during different parts of a solar cycle.

Lev: If we want to build anything for a mission going to Mars, we need that kind of data-driven way to categorize the input so we don't get blindsided by unexpected flow conditions.

Kai: So, it’s about using these tools to map out the solar wind variability in a way that is physically grounded while still being data-driven.

Mira: And the next thing we need to figure out is if those same classifications hold up when we feed them actual unprocessed data collected by MAVEN and try to apply them directly at Mars.

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

West Virginia University Department of Physics and Astronomy · Centre for mathematical Plasma Astrophysics Department of Mathematics KU Leuven Centre for mathematical Plasma Astrophysics Laboratory for Atmospheric and Space Physics Smead Aerospace Engineering Sciences Department University of Colorado Boulder Laboratory for Atmospheric and Space Physics Institute for the Study of Earth Oceans, and Space University of New Hampshire Department of Astronomy Boston University Department of Electrical and Computer Engineering University of Illinois at Urbana-Champaign California Institute of Technology Department Of Earth Planetary and Space Science · University Iowa

physics.space-ph

Submitted: 2026-04-09

Updated: 2026-11-06

Comments: Submitted to Solar Physics

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 88/100

The gist: Understanding solar wind variability throughout the heliosphere is essential for fundamental space physics and future exploration of the Moon and Mars.

Key concepts

Principal Component Analysis (PCA)
PCA is a mathematical tool used here to simplify complex data. It takes many variables (like velocity and temperature) and reduces them into a smaller set of independent 'modes' that capture the most important variations in the solar wind data, making it easier to interpret physically.
K-Means Clustering
This machine learning technique groups similar data points together based on their characteristics. In this study, it was used to automatically sort the complex solar wind measurements into six distinct clusters, each representing a specific physical regime or type of solar wind condition.
Solar Wind Regimes
These are different types of solar wind flow observed in space, categorized here as slow (Cluster 0), fast (Clusters 1 and 2), intermediate (Clusters 3 and 4), and compressed/disturbed (Cluster 5). These classifications help scientists understand how the solar wind changes over time.
Solar Cycle Modulation
This refers to how the overall pattern of solar activity, which repeats roughly every 11 years, influences the observed solar wind. The study showed that during solar minima, slow winds dominate, while during maxima, more frequent transitions between different regimes occur.

Terminology

Summary

Understanding solar wind variability throughout the heliosphere is essential for fundamental space physics and future exploration of the Moon and Mars. p. 0

Where the paper says it

Gross5 · Farzad Kamalabadi6 · Marco Velli7 · Jasper S. Halekas8 © The author(s) •••• Abstract Understanding solar wind variability throughout the heliosphere is essential for fundamental space physics and future exploration of the Moon and Mars.

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. p. 1
Where the paper says it

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.

Methodology Overview

This study applies an unsupervised machine-learning framework combining Principal Component Analysis (PCA) 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 researchers analyzed a multi-year dataset from the MAVEN spacecraft spanning Solar Cycles 24 and 25. p. 1

Where the paper says it

This study utilizes these methods to investigate how the solar wind can be classified at Mars using over 10 years of data from the MAVEN mission. Through the use of an unsupervised machine-learning technique which combined Principal Component Analysis (PCA) and K-Means clustering, we analyzed a multi-year dataset from the MAVEN spacecraft spanning Solar Cycles 24 and 25.

They used the publicly available synthesized upstream driver dataset provided by Halekas and colleagues. p. 4
Where the paper says it

The spacecraft samples the solar wind a substantial fraction of the time, producing a large dataset of upstream solar wind measurements. For this study, we use the publicly available synthesized upstream driver dataset provided by Halekas and colleagues, which combines Level 2 measurements from the Solar Wind Ion Analyzer (SWIA) and Magnetometer (MAG) instruments (Connerney et al.

The input variables included plasma and magnetic field parameters such as proton number density (np), solar wind bulk velocity (vp), proton temperature (Tp), and magnetic field components (Bx, By, Bz) All input variables were normalized using z-score normalization before the application of Principal Component Analysis to ensure numerical stability. p. 4
Where the paper says it

MAVEN Dataset and Data Normalization The MAVEN dataset analyzed in this study comprises plasma and magnetic field parameters characterized by heterogeneous physical units and markedly different numerical ranges, including proton and alpha particle number densities (np, nα), solar wind bulk and component velocities (vp, vx, vy, vz), proton temperature (Tp), and magnetic field components (Bx, By, Bz), as well as several derived quantities. These include the magnetic field magnitude, |B| = q B2 x + B2 y + B2 z , the proton entropy ratio, Sp = Tpn −2/3 p , the Alfv´en speed, VA = |B|/ √µ0mpnp, where µ0 = 4π × 10−7 , H, m−1 is the magnetic permeability of free space and mp is the proton mass, and a temperature–velocity ratio defined as Tratio = (vp/258)3.

PCA was employed as a linear, physically interpretable dimensionality-reduction technique to reduce the fourteen-dimensional feature space to six dimensions. p. 6
Where the paper says it

The resulting high dimensionality complicates both visualization and statistical analysis, and may increase the influence of noise in subsequent clustering procedures. To address these issues, PCA was employed as a linear, physically interpretable dimensionality-reduction technique.

K-Means clustering was then applied in the reduced six-dimensional PCA space to identify groupings based on similarity. p. 7
Where the paper says it

K-Means Clustering and Sensitivity Analysis Once in the reduced six-dimensional PCA space, K-Means clustering was applied to identify potentially interesting groupings of the dataset within the feature space. K-Means is an unsupervised learning algorithm that partitions a dataset into a predefined number of clusters based on similarity.

The number of clusters was determined through a systematic sensitivity study using inertia and silhouette scores, which suggested k=6 as the optimal compromise. p. 8
Where the paper says it

In summary, inertia is the sum of the squared distances between each sample and its assigned cluster centroid; lower inertia values indicate tighter, more compact clusters. To assess the robustness of the K-Means clustering results, we performed a systematic sensitivity study by varying the number of clusters k in the K-Means algorithm.

PCA Structure and Physical Interpretation

The PCA decomposition revealed physically interpretable modes that correspond to different aspects of solar wind dynamics. PC1 is mostly composed of the bulk velocity terms (vp, vx), proton temperature (Tp), and proton entropy ratio (Sp), capturing variations in large-scale flow speed and thermodynamic state. p. 9

Where the paper says it

The PCA separates the feature set into physically interpretable modes. PC1 is mostly composed of the bulk velocity terms (vp, vx), proton temperature (Tp), and proton entropy ratio (Sp), indicating that this component primarily captures correlated variations in large-scale flow speed and thermodynamic state.

PC2 is governed by proton and alpha particle number densities (np, nα), magnetic field magnitude (B), and Alfv´en speed (VA), reflecting changes in plasma density and overall magnetic field strength. p. 9
Where the paper says it

PC1 is mostly composed of the bulk velocity terms (vp, vx), proton temperature (Tp), and proton entropy ratio (Sp), indicating that this component primarily captures correlated variations in large-scale flow speed and thermodynamic state. PC2 is governed by proton and alpha particle number densities (np, nα), magnetic field magnitude (|B|), and Alfv´en speed (VA), reflecting changes in plasma density and overall magnetic field strength.

PC3 is strongly controlled by the transverse magnetic field components (Bx, By), isolating magnetic field orientation and fluctuation structure. p. 9
Where the paper says it

PC2 is governed by proton and alpha particle number densities (np, nα), magnetic field magnitude (|B|), and Alfv´en speed (VA), reflecting changes in plasma density and overall magnetic field strength. PC3 is strongly controlled by the transverse magnetic field components (Bx, By), isolating magnetic field orientation and fluctuation structure.

PC5 is overwhelmingly driven by the magnetic field z-component (Bz), representing dominant variability associated with magnetic field polarity. p. 9
Where the paper says it

PC4 is dominated by the y-component of the velocity (vy) and the temperature ratio (Tratio), suggesting sensitivity to directional flow structure coupled with thermodynamic contrasts. PC5 is overwhelmingly driven by the magnetic field z-component (Bz), representing dominant variability associated with magnetic field polarity.

These modes correspond to bulk flow dynamics, density against magnetic pressure balance, magnetic field geometry, and thermodynamic structure. p. 9
Where the paper says it

PC1 is mostly composed of the bulk velocity terms (vp, vx), proton temperature (Tp), and proton entropy ratio (Sp), indicating that this component primarily captures correlated variations in large-scale flow speed and thermodynamic state. PC2 is governed by proton and alpha particle number densities (np, nα), magnetic field magnitude (|B|), and Alfv´en speed (VA), reflecting changes in plasma density and overall magnetic field strength.

Table 2 reports the mean and standard deviation of selected plasma and magnetic field features for each of the six clusters identified by the K-Means algorithm. p. 5
Where the paper says it

To mitigate this effect, all input variables were normalized using z-score normalization. For each feature x, the normalized value xnorm was computed as xnorm = x − µ σ , (1) where µ and σ denote the mean and standard deviation of the feature calculated over the full dataset.

Cluster 3 and 4 exhibit intermediate bulk velocities and enhanced magnetic field magnitudes relative to Cluster 0, suggesting transitional solar wind regimes that bridge slow and fast wind conditions. p. 1
Where the paper says it

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.

Fast solar wind streams are primarily represented by Clusters 1 and 2, which display elevated bulk velocities, reduced proton densities, increased Alfv´en speeds, and significantly higher proton temperatures and entropy proxy values. p. 4
Where the paper says it

MAVEN Dataset and Data Normalization The MAVEN dataset analyzed in this study comprises plasma and magnetic field parameters characterized by heterogeneous physical units and markedly different numerical ranges, including proton and alpha particle number densities (np, nα), solar wind bulk and component velocities (vp, vx, vy, vz), proton temperature (Tp), and magnetic field components (Bx, By, Bz), as well as several derived quantities. These include the magnetic field magnitude, |B| = q B2 x + B2 y + B2 z , the proton entropy ratio, Sp = Tpn −2/3 p , the Alfv´en speed, VA = |B|/ √µ0mpnp, where µ0 = 4π × 10−7 , H, m−1 is the magnetic permeability of free space and mp is the proton mass, and a temperature–velocity ratio defined as Tratio = (vp/258)3.

Cluster 5 stands out due to its markedly enhanced proton density, strong magnetic field magnitude, and large variability across multiple parameters, indicative of compressed or disturbed plasma conditions. p. 4
Where the paper says it

MAVEN Dataset and Data Normalization The MAVEN dataset analyzed in this study comprises plasma and magnetic field parameters characterized by heterogeneous physical units and markedly different numerical ranges, including proton and alpha particle number densities (np, nα), solar wind bulk and component velocities (vp, vx, vy, vz), proton temperature (Tp), and magnetic field components (Bx, By, Bz), as well as several derived quantities. These include the magnetic field magnitude, |B| = q B2 x + B2 y + B2 z , the proton entropy ratio, Sp = Tpn −2/3 p , the Alfv´en speed, VA = |B|/ √µ0mpnp, where µ0 = 4π × 10−7 , H, m−1 is the magnetic permeability of free space and mp is the proton mass, and a temperature–velocity ratio defined as Tratio = (vp/258)3.

Solar Cycle Modulation

The temporal analysis segmented the dataset by Solar Cycles 24 and 25 to investigate the influence of large-scale solar activity. During the minimum of Solar Cycle 24, Cluster 0 dominates, reflecting extended intervals of relatively low and stable magnetic field strength, bulk velocity, proton density, temperature, and Alfv´en speed. p. 9

Where the paper says it

PC1 is mostly composed of the bulk velocity terms (vp, vx), proton temperature (Tp), and proton entropy ratio (Sp), indicating that this component primarily captures correlated variations in large-scale flow speed and thermodynamic state. PC2 is governed by proton and alpha particle number densities (np, nα), magnetic field magnitude (|B|), and Alfv´en speed (VA), reflecting changes in plasma density and overall magnetic field strength.

In contrast to solar maximum conditions in Cycle 24 (Figure 6), Cluster 5 occurs repeatedly across multiple parameters, reflecting active and dynamically evolving solar wind configuration. p. 1
Where the paper says it

Conversely, at solar maximum, we observed rapid transitions between plasma regimes, with fast and compressed regimes becoming more prevalent, highlighting the increased complexity of the heliospheric plasma environment. While each principal component incorporates multiple physical parameters, the combined analysis of PCA loadings, cluster-averaged plasma characteristics, and temporal variability indicates that the resultant groups constitute a physically coherent classification of the solar wind at Mars.

The maximum of Solar Cycle 25 is characterized by a solar wind environment that remains organized into distinct regimes while showing strong temporal variability. p. 9
Where the paper says it

Taken together, these diagnostics suggest that k = 6 provides a reasonable compromise between maximizing cluster separation at low k and avoiding over-fragmentation at higher k, while retaining sufficient granularity to capture distinct solar wind regimes across the full dataset. Equivalent sensitivity analyses performed for individual solar-cycle phases (Solar Cycle 24 minimum and maximum, and Solar Cycle 25 maximum) exhibit similar trends in both inertia and silhouette score, supporting the robustness of the adopted number of clusters with respect to the temporal subset of the data.

Overall, the relative occurrence and organization of these regimes are strongly modulated by solar activity. p. 1
Where the paper says it

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.

Solar minima are dominated by long-lived slow solar wind conditions with reduced variability, whereas solar maxima are characterized by frequent transitions between regimes and increased large-scale structuring. p. 1
Where the paper says it

Conditions at solar minimum were characterized by extended slow solar wind intervals with reduced variability and infrequent compressed states. Conversely, at solar maximum, we observed rapid transitions between plasma regimes, with fast and compressed regimes becoming more prevalent, highlighting the increased complexity of the heliospheric plasma environment.

Regime Characterization Summary

The six identified clusters correspond to physically meaningful solar wind regimes that persist across different phases of the solar cycle. Cluster 0 corresponds to classical slow solar wind, characterized by low bulk velocity, weak magnetic field strength, and cool plasma. p. 4

Where the paper says it

MAVEN Dataset and Data Normalization The MAVEN dataset analyzed in this study comprises plasma and magnetic field parameters characterized by heterogeneous physical units and markedly different numerical ranges, including proton and alpha particle number densities (np, nα), solar wind bulk and component velocities (vp, vx, vy, vz), proton temperature (Tp), and magnetic field components (Bx, By, Bz), as well as several derived quantities. These include the magnetic field magnitude, |B| = q B2 x + B2 y + B2 z , the proton entropy ratio, Sp = Tpn −2/3 p , the Alfv´en speed, VA = |B|/ √µ0mpnp, where µ0 = 4π × 10−7 , H, m−1 is the magnetic permeability of free space and mp is the proton mass, and a temperature–velocity ratio defined as Tratio = (vp/258)3.

Fast solar wind conditions are primarily captured by Clusters 1 and 2, exhibiting elevated flow speeds and high proton temperatures. p. 9
Where the paper says it

The PCA separates the feature set into physically interpretable modes. PC1 is mostly composed of the bulk velocity terms (vp, vx), proton temperature (Tp), and proton entropy ratio (Sp), indicating that this component primarily captures correlated variations in large-scale flow speed and thermodynamic state.

Clusters 3 and 4 represent intermediate or transitional regimes, marked by moderate enhancements in magnetic field strength and plasma parameters. p. 7
Where the paper says it

Consequently, each principal component corresponds to a linear combination of the original physical variables and reflects correlated variations among multiple plasma parameters rather than the behavior of individual quantities in isolation. In particular, the leading components capture coupled plasma properties such as relationships between particle densities and flow velocities, as well as correlations between magnetic field strength and Alfv´en speed.

Cluster 5 isolates strongly compressed and highly variable plasma conditions associated with episodic extreme events. p. 1
Where the paper says it

Introduction Understanding the conditions of the solar wind is of increasing importance with the drive for human exploration of the Moon and Mars (NASA 2025). The change in the solar wind on a temporal and spatial scale is variable, changing over the solar cycle in addition to individual space weather events such as solar flares or co-rotating interaction regions (CIRs).

Future Applications

The framework developed here provides a foundation for future comparative studies across heliocentric distances and planetary environments. The clustering approach offers a natural context for forthcoming multi-point observations by the dual-spacecraft ESCAPADE mission expected at Mars in 2027. p. 1

Where the paper says it

This study utilizes these methods to investigate how the solar wind can be classified at Mars using over 10 years of data from the MAVEN mission. Through the use of an unsupervised machine-learning technique which combined Principal Component Analysis (PCA) and K-Means clustering, we analyzed a multi-year dataset from the MAVEN spacecraft spanning Solar Cycles 24 and 25.

This technique could be applied to unprocessed solar wind data collected by MAVEN to compare whether this is also the case at Mars. p. 1
Where the paper says it

This study utilizes these methods to investigate how the solar wind can be classified at Mars using over 10 years of data from the MAVEN mission. Through the use of an unsupervised machine-learning technique which combined Principal Component Analysis (PCA) and K-Means clustering, we analyzed a multi-year dataset from the MAVEN spacecraft spanning Solar Cycles 24 and 25.

The analysis demonstrates the value of unsupervised classification techniques for characterizing solar wind variability in a data-driven yet physically grounded manner. p. 1
Where the paper says it

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 study used similar tools and techniques used in Amaya et al. p. 4
Where the paper says it

The paper contains the following; section 2 outlines the MAVEN dataset used and the machine learning techniques, section 3 contains our results and interpretation and our analysis is concluded in section 4. Methods This study utilizes data from the Mars Atmosphere and Volatile EvolutioN (MAVEN) mission that has been orbiting Mars since 2014 (Jakosky (2017)), covering more than a full solar cycle (solar cycle 24 to 25).

(2020) to classify solar wind conditions but applied them to data at Mars. p. 4
Where the paper says it

The dataset analyzed in this work therefore provides a robust, physically grounded basis for characterizing the upstream solar wind conditions at Mars, which we further process as described below. MAVEN Dataset and Data Normalization The MAVEN dataset analyzed in this study comprises plasma and magnetic field parameters characterized by heterogeneous physical units and markedly different numerical ranges, including proton and alpha particle number densities (np, nα), solar wind bulk and component velocities (vp, vx, vy, vz), proton temperature (Tp), and magnetic field components (Bx, By, Bz), as well as several derived quantities.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Improved AI system can perform unsupervised classification of solar wind at Mars using a framework combining Principal Component Analysis (PCA) and K-Means clustering to identify recurrent solar wind regimes in a physically interpretable, data-driven manner.

  2. The improved system can classify solar wind regimes into six distinct clusters: slow, fast, intermediate, and compressed states based on the analysis of over 10 years of MAVEN data spanning Solar Cycles 24 and 25.

  3. The system can characterize the relative occurrence and temporal organization of these solar wind regimes as being strongly modulated by solar activity, specifically identifying that Conditions at solar minimum were characterized by extended slow solar wind intervals with reduced variability and infrequent compressed states.

  4. The system can distinguish between plasma characteristics based on cluster analysis, such as Cluster 0 being characterized by relatively low bulk velocity, weak magnetic field strength, low Alfv´en speed, and comparatively cool plasma, which is Dominant during solar minimum.

  5. The improved system can identify fast solar wind streams using Clusters 1 and 2, noting that these display elevated bulk velocities, reduced proton densities, increased Alfv´en speeds, and significantly higher proton temperatures and entropy proxy values.

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.

Related papers