The magnetosheath is the region of shocked, turbulent and highly variable plasma that forms downstream of Earth’s bow shock, where the supersonic solar wind first collides with the planet’s magnetic shield. Its character depends on the orientation of the interplanetary magnetic field relative to the shock normal. When the field points mostly perpendicular to the shock, the resulting quasi-perpendicular magnetosheath tends to be comparatively orderly. When the field is mostly parallel to the shock normal, the quasi-parallel magnetosheath becomes markedly more turbulent, hosting different instabilities, transient structures and wave activity. Distinguishing these two regimes is essential for studying magnetic reconnection, turbulence and energy transfer throughout the magnetosphere system.

Classifying these regions at scale has proven stubbornly difficult. While earlier studies identified quasi-perpendicular and quasi-parallel intervals using bow shock models, upstream solar wind monitors such as ACE and Wind at the L1 Lagrange point, or in-situ plasma measurements, no robust and easy-to-use algorithm existed to systematically classify the entire magnetosheath environment from local data. Upstream monitoring is unreliable because the solar wind varies both spatially and temporally between L1 and Earth, and the propagation time from L1 to the bow shock is ambiguous. Previous in-situ classification efforts showed promise but were restricted to narrow time intervals, specific data products, or particular transient phenomena such as high-speed jets.

To build the new dataset, the team started from an existing catalogue of magnetosheath intervals generated by an unsupervised machine learning algorithm applied to MMS1 data, and incorporated magnetosheath jet intervals identified in earlier work to include disturbed plasma periods. From this pool they drew a sample deliberately representative of the multidimensional range of solar wind and observational conditions, including variations in solar wind density, speed, mean local time and year. Jet intervals made up roughly 15 to 20 percent of the sample, another 15 to 20 percent captured the surrounding magnetosheath, and the remainder was randomly selected from the longer intervals. An expert then manually reviewed every interval at MMS’s higher native temporal resolution, rejecting about 10 percent of the initial sample that was contaminated by solar wind or magnetospheric plasma.

The citizen scientists themselves examined three-panel figures showing five consecutive one-minute intervals. Each panel displayed ion spectra from the Fast Plasma Investigation, parallel and perpendicular ion temperatures, and a magnetic fluctuation parameter computed from Fluxgate Magnetometer data. Rather than applying fixed thresholds, volunteers were trained in a tutorial to recognize qualitative patterns: quasi-parallel magnetosheath tends to show enhanced high-energy ion flux, coinciding parallel and perpendicular temperatures, and elevated magnetic variability, while quasi-perpendicular intervals show weaker suprathermal flux, greater temperature anisotropy and calmer fields. The interface explicitly avoided implying that all three signatures must be satisfied simultaneously, encouraging volunteers to weigh the overall pattern. In a second workflow, they flagged minutes containing a temporal transition between the two regimes.

Each one-minute interval received 13 independent classifications per workflow, a number chosen through beta testing on 300 subjects where simple intervals reached full agreement with only five classifications but complex environments required the fuller wisdom of the crowd. The researchers computed a normalized consensus score for each interval, ranging from unanimous quasi-perpendicular to unanimous quasi-parallel. In one example from 11 November 2017, volunteers unanimously classified all five minutes as quasi-perpendicular and unmixed, correctly handling an ion spectrum shifted to unusually high energy bands that could have confused simpler methods.

The results are striking. Of all classifications, 23,366 minutes, or 65.8 percent, were high-likelihood quasi-perpendicular, while 7,697 minutes, or 21.7 percent, were high-likelihood quasi-parallel, with a further 2,315 moderate-likelihood quasi-parallel intervals. The distribution aligns with statistical expectations from the Parker spiral orientation of the interplanetary magnetic field at Earth’s orbit: quasi-parallel classifications skew toward the dawn side of the magnetosphere, where the field connects more nearly along the shock normal, and quasi-perpendicular classifications toward the dusk side. The scatter around this pattern is very large, however, meaning any single interval is unlikely to sit at its statistically preferred location.

Transitions between the two regimes were also captured, using both the mixed workflow, which flagged 2,190 minutes with a 0.5 consensus threshold, and a combined method that identified jumps larger than 0.4 between neighboring classifications, yielding 1,248 minutes, for a total of 2,677 unique transition intervals, or 7.5 percent of the sample. To test robustness, the team embedded 2,140 duplicate minutes classified by multiple independent sets of volunteers. Only 81 duplicate minutes exceeded a score difference of 0.25, the equivalent of jumping a full classification category, indicating that 13 classifications per sample are sufficient to smooth out human subjectivity.

Notably, the volunteers appear to have made more nuanced judgments than any fixed threshold could. While high-confidence classifications aligned well with established thresholds from prior work, lower-confidence intervals fell outside the hard thresholds, suggesting that volunteers incorporated contextual information, such as recognizing when suprathermal ion flux shifted toward the upper end of MMS’s energy range during fast solar wind, and used the surrounding magnetosheath structure to judge local variability.

The dataset, publicly available on Zenodo, opens several scientific avenues. It can support statistical comparisons of the two magnetosheath populations, including tests of whether quasi-parallel regions host larger populations of ultra-low frequency waves transmitted through the shock from the foreshock, waves known to influence magnetospheric dynamics. It can also help establish robust connections between magnetosheath regions and transient phenomena such as jets, magnetic holes and hot flow anomalies, whose occurrence rates near regime boundaries remain open questions. Looking ahead, the team plans to train machine learning models on these human labels, extending classifications beyond 2024, and potentially applying transfer learning to other missions such as THEMIS, Cluster and MAVEN, extending the approach as far as Mars.

Subject of Research: Citizen science classification of quasi-parallel and quasi-perpendicular magnetosheath intervals from NASA's MMS mission

Article Title: Shock Detectives: classifying MMS observations of Earth’s magnetosheath via citizen science

Article References: Toy-Edens, V., Mo, W., Raptis, S., Bell, N. C., Dąbrowska, M., Eschweiler, T., Ford, R. A., Gonano, V., Gorokhov, A. C. R. B., Leal Quesada, J. M., Luperchio, A. M., & Pasqua, A. (2026). Shock Detectives: classifying MMS observations of Earth’s magnetosheath via citizen science. Frontiers in Astronomy and Space Sciences, 13, Article 1986629. https://doi.org/10.3389/fspas.2026.1986629

Image Credits: AI Generated

DOI: 10.3389/fspas.2026.1986629

Keywords: citizen science, Zooniverse, MMS, magnetosheath, bow shock, quasi-parallel, quasi-perpendicular, solar wind, machine learning, space physics, Shock Detectives, plasma classification