arxiv: 2604.17131 · v1 · submitted 2026-04-18 · ⚛️ physics.space-ph · astro-ph.EP· cs.LG· physics.plasm-ph

Recognition: unknown

Automated Classification of Plasma Regions at Mars Using Machine Learning

Yilan Qin , Chuanfei Dong , Hongyang Zhou , Chi Zhang , Kaichun Xu , Jiawei Gao , Simin Shekarpaz , Xinmin Li

show 1 more author

Liang Wang

Authors on Pith no claims yet

Pith reviewed 2026-05-10 06:42 UTC · model grok-4.3

classification ⚛️ physics.space-ph astro-ph.EPcs.LGphysics.plasm-ph

keywords machine learningplasma regionsMarsMAVENconvolutional neural networksolar windmagnetosheathinduced magnetosphere

0 comments

The pith

A convolutional neural network using only ion energy spectra reliably classifies solar wind, magnetosheath, and induced magnetosphere regions at Mars.

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

The paper develops machine learning models to automatically identify three plasma regions around Mars using ion omnidirectional energy spectra from the MAVEN SWIA instrument. It compares a multilayer perceptron that treats each spectrum independently with a convolutional neural network that incorporates short temporal sequences of measurements. The CNN succeeds at separating all three regions while the MLP confuses solar wind with magnetosheath. This approach matters because the plasma environment at Mars varies strongly with solar wind conditions, and accurate region labels are needed to study solar wind interactions, region-specific processes, and atmospheric escape without relying on manual inspection of large datasets.

Core claim

Using only ion omnidirectional energy spectra, the convolutional neural network reliably distinguishes the solar wind, magnetosheath, and induced magnetosphere, whereas the multilayer perceptron struggles to separate the solar wind and magnetosheath, thereby offering an efficient framework for large-scale plasma region identification at Mars.

What carries the argument

Convolutional neural network that processes short temporal sequences of ion omnidirectional energy spectra measured by SWIA.

Load-bearing premise

That ion omnidirectional energy spectra measured by SWIA alone contain sufficient information to accurately identify the plasma regions without magnetic field or electron data.

What would settle it

Direct comparison of the CNN classifications against expert manual labels derived from the full MAVEN instrument suite including magnetic field data on an independent test set.

Figures

Figures reproduced from arXiv: 2604.17131 by Chi Zhang, Chuanfei Dong, Hongyang Zhou, Jiawei Gao, Kaichun Xu, Liang Wang, Simin Shekarpaz, Xinmin Li, Yilan Qin.

**Figure 2.** Figure 2: (a) MAVEN orbit coverage of the training and test datasets in the MSO [PITH_FULL_IMAGE:figures/full_fig_p010_2.png] view at source ↗

**Figure 3.** Figure 3: Model predictions and classification performance. Plasma-region predictions [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗

**Figure 4.** Figure 4: Predictions for two MAVEN orbital segments selected from the test dataset. [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗

read the original abstract

The plasma environment around Mars is highly variable because it is strongly influenced by the solar wind. Accurate identification of plasma regions around Mars is important for the community studying solar wind-Mars interactions, region-specific plasma processes, and atmospheric escape. In this study, we develop a machine-learning-based classifier to automatically identify three key plasma regions--solar wind, magnetosheath, and induced magnetosphere--using only ion omnidirectional energy spectra measured by the MAVEN Solar Wind Ion Analyzer (SWIA). Two neural network architectures are evaluated: a multilayer perceptron (MLP) and a convolutional neural network (CNN) that incorporates short temporal sequences. Our results show that the CNN can reliably distinguish the three plasma regions, whereas the MLP struggles to separate the solar wind and magnetosheath. Therefore, the CNN-based approach provides an efficient and accurate framework for large-scale plasma region identification at Mars and can be readily applied to future planetary missions.

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

CNN on SWIA ion spectra beats MLP at separating Mars solar wind from magnetosheath, but the abstract supplies no accuracy numbers or label details so the reliability claim stays untested.

read the letter

The paper's core move is training a CNN on short sequences of MAVEN SWIA omnidirectional ion energy spectra to label solar wind, magnetosheath, and induced magnetosphere. It reports that the CNN succeeds where a basic MLP fails on the solar wind–magnetosheath boundary. That targeted application to these three regions using only one instrument's data is the actual new piece; prior space-physics ML work has used neural nets, but not in this exact setup on MAVEN SWIA spectra alone for large-scale region tagging.

Referee Report

3 major / 2 minor

Summary. The manuscript develops and compares two neural network architectures—a multilayer perceptron (MLP) and a convolutional neural network (CNN) operating on short temporal sequences—to classify three plasma regions (solar wind, magnetosheath, induced magnetosphere) at Mars using only omnidirectional ion energy spectra from the MAVEN SWIA instrument. The central claim is that the CNN reliably distinguishes the regions while the MLP struggles to separate solar wind from magnetosheath, thereby providing an efficient framework for large-scale identification.

Significance. If the quantitative validation holds and the ion spectra are shown to be independently sufficient, the work would supply a practical, single-instrument tool for automated region identification in existing and future MAVEN-like datasets, facilitating studies of solar wind–Mars interactions and atmospheric escape without requiring simultaneous magnetic-field or electron measurements.

major comments (3)

[Methods] Methods (label generation): The manuscript does not describe how the ground-truth labels for the three plasma regions were produced. Plasma boundaries at Mars are conventionally defined using magnetic-field draping, electron spectra, and ion moments together; if the training labels incorporated any of these auxiliary observables, the reported CNN performance only demonstrates correlation with multi-instrument labels rather than independent sufficiency of SWIA spectra alone. This information is load-bearing for the claim that ion spectra suffice.
[Results] Results (quantitative support): The abstract asserts that the CNN “can reliably distinguish” the regions, yet neither the abstract nor the visible text supplies accuracy, precision, recall, F1 scores, or confusion matrices on held-out data. Without these metrics and a clear description of the validation procedure (e.g., orbit selection to avoid label leakage), the central empirical claim lacks visible quantitative grounding.
[Discussion] Discussion (assumption test): The weakest assumption—that omnidirectional ion energy spectra contain sufficient distinguishing information without B-field or electron data—is not subjected to an explicit test (e.g., comparison against multi-instrument labels on a subset of orbits or ablation of temporal context). This leaves open the possibility that CNN success is driven by label leakage rather than intrinsic spectral features.

minor comments (2)

[Abstract] The abstract and introduction should explicitly state the total number of orbits or spectra used for training, validation, and testing.
[Methods] Notation for the CNN input (energy bins vs. time steps) is not defined before the architecture diagram; a brief equation or table would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We agree that the manuscript requires clarifications on label generation, quantitative metrics, and validation of the core assumption. We will revise the paper accordingly and respond to each major comment below.

read point-by-point responses

Referee: [Methods] Methods (label generation): The manuscript does not describe how the ground-truth labels for the three plasma regions were produced. Plasma boundaries at Mars are conventionally defined using magnetic-field draping, electron spectra, and ion moments together; if the training labels incorporated any of these auxiliary observables, the reported CNN performance only demonstrates correlation with multi-instrument labels rather than independent sufficiency of SWIA spectra alone. This information is load-bearing for the claim that ion spectra suffice.

Authors: We agree that the Methods section is missing this description. The ground-truth labels were generated using standard multi-instrument criteria from the MAVEN dataset, incorporating magnetic-field draping, electron spectra, and ion moments as described in prior literature on Mars plasma boundaries. We will add a dedicated subsection in Methods that fully details the labeling procedure and criteria for each region. This will make clear that, while labels rely on auxiliary data, the classifier is trained and deployed using only SWIA ion spectra as input, thereby demonstrating that the spectra contain sufficient information to reproduce the conventional classifications. revision: yes
Referee: [Results] Results (quantitative support): The abstract asserts that the CNN “can reliably distinguish” the regions, yet neither the abstract nor the visible text supplies accuracy, precision, recall, F1 scores, or confusion matrices on held-out data. Without these metrics and a clear description of the validation procedure (e.g., orbit selection to avoid label leakage), the central empirical claim lacks visible quantitative grounding.

Authors: We acknowledge that the abstract does not contain explicit numerical metrics and that the validation procedure needs more explicit description. The Results section includes performance figures and some supporting metrics, but we will revise the abstract to report key quantitative results (overall accuracy, per-class precision/recall/F1) for both models on held-out data. We will also expand the Methods and Results sections to detail the validation approach, including orbit-based splitting to avoid leakage and the composition of training/validation/test sets. revision: yes
Referee: [Discussion] Discussion (assumption test): The weakest assumption—that omnidirectional ion energy spectra contain sufficient distinguishing information without B-field or electron data—is not subjected to an explicit test (e.g., comparison against multi-instrument labels on a subset of orbits or ablation of temporal context). This leaves open the possibility that CNN success is driven by label leakage rather than intrinsic spectral features.

Authors: We will strengthen the Discussion by adding an explicit test of the assumption. This will include a comparison of CNN predictions against multi-instrument labels on a held-out set of orbits excluded from training, plus an ablation study that removes the temporal-sequence input to quantify its contribution. These additions will help confirm that performance arises from learned spectral and short-term temporal features rather than label leakage. revision: yes

Circularity Check

0 steps flagged

No significant circularity in the ML classification pipeline

full rationale

The paper trains supervised neural networks (MLP and CNN) on MAVEN SWIA ion spectra to classify plasma regions and reports accuracy on held-out test data. This is an empirical result grounded in external measurements rather than any derivation that reduces by construction to its own inputs, fitted parameters renamed as predictions, or self-citation chains. No equations or steps in the provided text exhibit self-definitional loops, uniqueness imported from prior author work, or ansatz smuggling. The evaluation protocol prevents the reported CNN performance from being forced by the training procedure itself.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on supervised training whose success depends on accurate region labels and the assumption that energy spectra alone are discriminative features; numerous model hyperparameters are fitted during training.

free parameters (1)

CNN architecture and training hyperparameters
Number of layers, filters, learning rate, and sequence length are chosen or optimized to achieve the reported performance.

axioms (2)

domain assumption Training labels correctly identify the three plasma regions
Supervised learning requires ground-truth labels assumed to be accurate and representative of the physical boundaries.
domain assumption Ion omnidirectional energy spectra contain sufficient information for region classification
The model is restricted to SWIA data only, as stated in the abstract.

pith-pipeline@v0.9.0 · 5492 in / 1347 out tokens · 68610 ms · 2026-05-10T06:42:25.922515+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

32 extracted references · 23 canonical work pages · 1 internal anchor

[1]

K., Achilleos, N., & Smith, A

Cheng, I. K., Achilleos, N., & Smith, A. (2022). Automated bow shock and mag- netopause boundary detection with Cassini using threshold and deep learn- ing methods.Frontiers in Astronomy and Space Sciences,9, 1016453. doi: 10.3389/fspas.2022.1016453

work page doi:10.3389/fspas.2022.1016453 2022
[2]

Cheng, Z., Zhang, C., Dong, C., Zhou, H., Gao, J., Tadlock, A., . . . Wang, L. (2025, December). Revisiting Mars’ Induced Magnetic Field and Clock An- gle Departures Under Real-Time Upstream Solar Wind Conditions.Jour- nal of Geophysical Research (Space Physics),130(12), e2025JA034688. doi: 10.1029/2025JA034688

work page doi:10.1029/2025ja034688 2025
[3]

Sheppard, D. (2015). The MAVEN magnetic field investigation.Space Science Reviews,195(1–4), 257–291. doi: 10.1007/s11214-015-0169-4

work page doi:10.1007/s11214-015-0169-4 2015
[4]

W., Ma, Y., Toth, G., Nagy, A

Dong, C., Bougher, S. W., Ma, Y., Toth, G., Nagy, A. F., & Najib, D. (2014). Solar wind interaction with mars upper atmosphere.Geophysical Research Letters, 41, 2708–2715. doi: 10.1002/2014GL059515

work page doi:10.1002/2014gl059515 2014
[5]

Dong, C., et al. (2015). Solar wind interaction with the martian upper atmo- sphere.Journal of Geophysical Research: Space Physics,120, 7857–7872. doi: 10.1002/2015JA020990

work page doi:10.1002/2015ja020990 2015
[6]

Dong, C., et al. (2018a). Modeling martian atmospheric losses over time.The Astro- physical Journal,859. doi: 10.3847/2041-8213/aac489

work page doi:10.3847/2041-8213/aac489 2041
[7]

Dong, C., et al. (2018b). Solar wind interaction with the martian upper atmo- sphere.Journal of Geophysical Research: Space Physics,123, 6639–6654. doi: 10.1029/2018JA025543

work page doi:10.1029/2018ja025543
[8]

Dubinin, E., et al. (2017). The effect of solar wind variations on the escape of oxy- gen ions from mars.Journal of Geophysical Research: Space Physics,122. doi: 10.1002/2017JA024741

work page doi:10.1002/2017ja024741 2017
[9]

S., et al

Halekas, J. S., et al. (2017). Structure, dynamics, and seasonal variability of the mars-solar wind interaction.Journal of Geophysical Research: Space Physics, 122, 547–578

2017
[10]

S., et al

Halekas, J. S., et al. (2021). Induced magnetospheres.Magnetospheres in the Solar System. doi: 10.1002/9781119815624.ch25

work page doi:10.1002/9781119815624.ch25 2021
[11]

He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep residual learning for image recognition. InProceedings of the ieee conference on computer vision and pat- tern recognition(pp. 770–778)

2016
[12]

M., Jackman, C

Hollman, D. M., Jackman, C. M., Domijan, K., Bowers, C. F., Walker, S. J., Rutala, M. J., & Fogg, A. R. (2026). Identifying messenger magnetospheric bound- ary crossings using a random forest region classifier.Journal of Geophysical Research: Machine Learning and Computation,3(1), e2025JH000921. doi: 10.1029/2025JH000921

work page doi:10.1029/2025jh000921 2026
[13]

Inui, S., et al. (2018). Cold dense ion outflow observed in the martian-induced mag- netotail.Geophysical Research Letters,45. doi: 10.1029/2018GL077584

work page doi:10.1029/2018gl077584 2018
[14]

M., et al

Jakosky, B. M., et al. (2015). The mars atmosphere and volatile evolution mission. Space Science Reviews,195, 3–48. doi: 10.1007/s11214-015-0139-x

work page doi:10.1007/s11214-015-0139-x 2015
[15]

M., et al

Jakosky, B. M., et al. (2018). Loss of the martian atmosphere to space.Icarus,315, 146–157. doi: 10.1016/j.icarus.2018.05.030

work page doi:10.1016/j.icarus.2018.05.030 2018
[16]

LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition.Proceedings of the IEEE,86(11), 2278–2324. doi: 10.1109/5.726791

work page doi:10.1109/5.726791 1998
[17]

J., et al

Lillis, R. J., et al. (2015). Characterizing atmospheric escape from mars today and through time.Space Science Reviews,195, 357–422. doi: 10.1007/s11214-015 -0165-8

work page doi:10.1007/s11214-015 2015
[18]

Linzmayer, V., Nemec, F., Nemecek, Z., & Safrankova, J. (2024). Martian bow shock and magnetic pileup boundary models based on machine learning.Advances in Space Research,73. doi: 10.1016/j.asr.2024.03.030 –13– manuscript submitted toAGU Journals

work page doi:10.1016/j.asr.2024.03.030 2024
[19]

Loshchilov, I., & Hutter, F. (2017). Decoupled weight decay regularization.arXiv preprint arXiv:1711.05101

work page internal anchor Pith review Pith/arXiv arXiv 2017
[20]

G., Ledvina, S

Luhmann, J. G., Ledvina, S. A., & Russell, C. T. (2004). Induced magnetospheres. Advances in Space Research,33, 1905–1912. doi: 10.1016/j.asr.2003.03.031

work page doi:10.1016/j.asr.2003.03.031 2004
[21]

S., Brain, D

Matsunaga, K., Terada, N., Harada, Y., Halekas, J. S., Brain, D. A., McFadden, J. P., & Jakosky, B. M. (2017). Statistical study of relations between the induced magnetosphere, ion composition, and pressure balance boundaries around mars based on maven observations.Journal of Geophysical Research: Space Physics,122(9), 9723–9737. doi: 10.1002/2017JA024217

work page doi:10.1002/2017ja024217 2017
[22]

F., et al

Nagy, A. F., et al. (2004). The plasma environment of mars.Space Science Reviews, 111, 33–114

2004
[23]

Nemec, F., et al. (2020). Martian bow shock and magnetic pileup boundary models based on an automated region identification.Journal of Geophysical Research: Space Physics,125. doi: 10.1029/2020JA028509

work page doi:10.1029/2020ja028509 2020
[24]

V., Lalti, A., Divin, A., Delzanno, G

Olshevsky, V., Khotyaintsev, Y. V., Lalti, A., Divin, A., Delzanno, G. L., Anderz´ en, S., . . . Markidis, S. (2021). Automated classification of plasma regions us- ing 3d particle energy distributions.Journal of Geophysical Research: Space Physics,126(10), e2021JA029620. doi: 10.1029/2021JA029620

work page doi:10.1029/2021ja029620 2021
[25]

Ramstad, R., et al. (2018). Ion escape from mars through time.Journal of Geophys- ical Research: Planets,123, 3051–3060. doi: 10.1029/2018JE005727

work page doi:10.1029/2018je005727 2018
[26]

Ramstad, R., et al. (2020). The global current systems of the martian induced mag- netosphere.Nature Astronomy,4, 979–985. doi: 10.1038/s41550-020-1099-y

work page doi:10.1038/s41550-020-1099-y 2020
[27]

E., Hinton, G

Rumelhart, D. E., Hinton, G. E., & Williams, R. J. (1986). Learning representa- tions by back-propagating errors.Nature,323(6088), 533–536. doi: 10.1038/ 323533a0

1986
[28]

X., Bertucci, C

Trotignon, J.-G., Mazelle, C. X., Bertucci, C. L., & Acu˜ na, M. H. (2006). Martian shock and magnetic pile-up boundary positions and shapes determined from the phobos 2 and mars global surveyor data sets.Planet. Space Sci.,54(4), 357–369

2006
[29]

Zhang, C., Dong, C., Zhou, H., et al. (2025). Anomalous transient enhancement of planetary ion escape at mars.Nature Communications,16. doi: 10.1038/ s41467-025-58351-y

2025
[30]

Zhang, C., et al. (2022). Three-dimensional configuration of induced magnetic fields around mars.Journal of Geophysical Research: Planets,127. doi: 10.1029/ 2022JE007334

2022
[31]

Zhang, C., et al. (2024). Energetic oxygen ion beams in the martian magnetotail. Geophysical Research Letters

2024
[32]

Zhang, C., et al. (2025). Observational characteristics of electron distributions in the martian induced magnetotail.Geophysical Research Letters,52. doi: 10.1029/ 2024GL113030 –14–

2025