Cosmic rays hitting the upper atmosphere produce showers of particles. When a single muon from such a shower reaches an underground detector, its rate peaks in summer. Warm air expands, so the parent pions and kaons more often decay into muons before they collide with air. That picture is well measured. When several muons arrive together, though, the rate often peaks in winter. MINOS saw closely spaced muon pairs peak in winter while widely spaced ones peaked in summer. NOvA saw the winter peak at the surface, and it grows stronger at higher multiplicity. No atmospheric model explains this. We propose to stop guessing which hand-picked event property drives the effect. Instead, we let the atmosphere supply the training labels. A random forest (an ensemble of decision trees that vote) is trained on real data to predict the atmospheric state from each event's features. Because summer and winter events are mixtures of the same underlying event types in different proportions, the forest is pushed toward the features that separate the summer-loving population from the winter-loving one. It needs no Monte Carlo simulation and no particle labels.
The second half of the idea turns this discovery tool into a particle-identification test. One live but untested explanation of the winter excess is mundane: high-multiplicity events may contain non-muon tracks (punch-through hadrons, electromagnetic debris near shower cores) that are counted as muons, and the amount of such debris may itself depend on air density. If the anomaly survives a strict, data-calibrated muon purification, it is shower physics. If it shrinks, a share of it is misidentification. Either answer narrows the space of models, and either answer gives a systematic correction for experiments that treat the seasonal muon flux as a background.
Core Claim
A classifier trained to predict short-timescale atmospheric fluctuations from event-level and track-level features can isolate a cleaner winter-peaking and summer-peaking split of multiple-muon events than the one-dimensional cuts on track separation, multiplicity or zenith angle used so far. Combined with track-level muon purification, the same machinery decides whether the winter inversion belongs to true muon bundles or to misidentified tracks.
The key methodological choice is to train against day-scale weather anomalies rather than against "summer versus winter." Detector drift is slow, so it cannot imitate a cold front.
Background and Positioning
The anomaly itself is what motivates this work. MINOS multiple-muon seasonal variation fit sinusoids to rates in bins of track separation and found that at the Far Detector (2100 meters water equivalent, or mwe) muons more than 8 m apart peak at $t_0 = 184.8 \pm 6.5$ days, while muons 0.6–4.5 m apart peak at $t_0 = 27.6 \pm 8.9$ days. At the Near Detector all separations peak in winter. The only simulation used was an isothermal atmosphere in CORSIKA, the standard air-shower simulation code. NOvA Near Detector multiple muons found anti-correlation with temperature that strengthens with multiplicity. NOvA on the surface measured $V = 5.86\% \pm 0.05\%$, $\alpha_T = -1.14 \pm 0.02$ and a barometric coefficient of $-0.08 \pm 0.01\%$/mbar for 15 or more muons. It also found the effect about 30% larger at the highest multiplicities and about 20% larger near the horizon. Every one of these analyses splits data along one or two hand-chosen axes. None asks a multivariate question, and none checks muon purity as a function of season.
The single-muon literature supplies the atmospheric inputs and a warning. MINOS single-muon modulation shows that the answer depends on how the atmosphere is summarized. Weighting the effective temperature by muon arrival angle moves $\alpha_T$ from 0.428 to 0.352, and in both cases the systematic error (about 0.05) is more than ten times the statistical error (0.003). Double Chooz supplies the satellite-profile approach. We use their effective-temperature constructions as candidate targets, not as conclusions.
Another line of work constrains what simulation can deliver. LHAASO's muon attenuation measurement finds that QGSJET-II-04 and SIBYLL 2.3d predict attenuation lengths significantly longer than measured. Any explanation that rests on one hadronic model carries a model systematic of unknown size. This is why the core method is trained on data rather than on simulation.
The tree-ensemble method comes from three sources. Random Forest for MAGIC established the forest score for particle separation, at about 2 ms per event. The same paper names Monte Carlo mismatch, including atmospheric mismatch, as its main weakness. Interpretable BDTs for the Majorana Demonstrator show that Shapley attributions (a game-theory method for assigning a prediction to individual input features) can recover physical mechanisms from a tree model. Their boosted trees rejected 8.6% more single-escape-peak background than the hand-tuned cut. The deep-learning alternative, CNN-GNN gamma/hadron separation, has been tested only on simulation at one zenith angle.
What all of these lack is a way to train without trusting simulated labels. That gap is what we fill. In Cherenkov telescopes the atmosphere is a nuisance to be corrected for. Here it is the label.
Research Program
The statistical basis is the known result for classification on mixed samples. Suppose events come from two populations with feature distributions $p_A(x)$ and $p_B(x)$, and suppose the atmospheric state $s$ changes only the mixing fraction $f(s)$. Then the event density is
$$p(x \mid s) = f(s)\,p_A(x) + (1-f(s))\,p_B(x).$$
Any classifier that separates high-$s$ from low-$s$ data is then a monotone function of $p_A(x)/p_B(x)$. Here $x$ is the event's feature vector and $A$ and $B$ are the summer- and winter-loving populations. Monotonicity holds because the density ratio between two values of $s$ depends on $x$ only through $p_A/p_B$. The forest therefore learns the same ordering of events that a fully supervised classifier would, even though no event carries a population label.
The weakness is signal strength. Modulations are about 5%, so the forest's score will be only slightly better than chance per event, and the method lives on large samples. NOvA has $7.64\times10^6$ surface showers and MINOS has $1.53\times10^9$ single muons.
The program has three parts.
The first part builds the targets. From ECMWF reanalysis profiles (37 pressure levels, as used by MINOS and NOvA) we compute hourly effective temperature, angle-weighted effective temperature, and surface pressure. We then apply a high-pass filter that keeps only fluctuations on timescales of about 1–10 days. Detector gain drift, efficiency loss and hall temperature move over months, so they cannot correlate with high-passed weather. Training on these targets removes the main confounder by construction. The seasonal sinusoid is used only at evaluation time, never as a label.
The second part builds features and trains the forest. Event-level features are multiplicity, zenith, the distribution of pairwise track separations, track-length and energy-deposition statistics, and the density of hits near the shower core. Track-level features are energy loss per unit length along the track, straightness, whether the track ends inside the detector, and whether a delayed decay electron follows (a "Michel" electron, which tags a stopped muon). A random forest regressor predicts the high-passed atmospheric anomaly. Training uses alternating month-long blocks, so that no single weather system appears in both training and test data. Shapley attributions then show which features carry the atmospheric information.
The third part is the purification test. A track-level muon classifier is trained on data-defined control samples. Stopping muons tagged by a Michel electron are the muon sample. Tracks attached to dense shower cores that fail range and straightness criteria are the debris sample. The forest score is then cross-checked on a simulated tag-and-probe sample, purely as a sanity check. We recompute multiplicity using only tracks above a strict muon-purity threshold and refit $V$, $t_0$ and $\alpha_T$. We also train a second forest only on events with the debris-like tracks removed, to see whether the atmospheric signal the forest finds survives.
Baselines
The naive reference is the published analysis itself: a sinusoid or a linear effective-temperature regression on rates binned by one variable. For NOvA surface data the reported values are $V = 5.86\% \pm 0.05\%$ and $\alpha_T = -1.14 \pm 0.02$ for 15 or more muons NOvA surface. For MINOS the separation split gives winter-peaking pairs ($t_0 = 27.6 \pm 8.9$ days) against summer-peaking ones ($t_0 = 184.8 \pm 6.5$ days) MINOS multiple-muon. We will reproduce these numbers on the same event selection before any comparison.
The strongest competitor is not a published method but an optimized version of the same idea. We will search over all one- and two-dimensional rectangular splits of the same feature set and pick the split that maximizes the difference in $\alpha_T$ between its two halves. This search runs on the same training months and has the same number of candidate evaluations as the forest's hyperparameter search. It represents what a careful physicist would do by hand, and it is the real threat.
To beat it, the forest-defined top and bottom score quantiles must show a larger $\alpha_T$ difference, measured in units of its statistical error, on held-out months. With no reported result found on Lacuna for a multivariate seasonal split, we fix the bar in advance. The forest wins if its contrast exceeds the best rectangular split's by more than the spread of contrasts across bootstrap resamples.
The supervised baseline is a forest trained on simulated labels. CORSIKA showers generated with QGSJET-II-04 and with EPOS-LHC are given truth labels of muon or non-muon and early-developing or late-developing shower, and trained with the same features and budget. Comparing the two hadronic models shows how much the "physics" ordering depends on the generator.
The key ablation replaces the high-passed weather target with the raw seasonal label. If that version yields a much larger contrast that disappears under the pressure-only and hall-temperature checks below, the extra contrast was detector drift. This isolates the main ingredient of the proposal.
Why not just…?
Why not just keep slicing by track separation, multiplicity and zenith? Those slices already show the effect, and they may be all there is. The rectangular-split baseline above is exactly this approach, run with the same search budget. If it matches the forest, we report that the anomaly is captured by known variables and drop the multivariate claim. NOvA's finding that the effect is larger both at high multiplicity and near the horizon, where multiplicities are lower, suggests to us that no single axis is the right one. That is a hypothesis this comparison tests, not an established result.
Why not just run CORSIKA with realistic month-by-month atmospheres? This is the most serious alternative. MINOS flagged the isothermal atmosphere as the limitation, and if realistic profiles reproduce the winter peak, the physics question may be largely answered. We will run it, with the same profiles under two hadronic models. The data-driven forest still matters for two reasons. LHAASO shows that hadronic models disagree with measured muon attenuation, so a simulated match under one model is not decisive. And the forest supplies a model-independent ordering of events against which simulation can be checked feature by feature. If simulation and the forest agree on which features carry the signal, that agreement is the strongest result this project could produce.
Why not regress rates on several atmospheric variables at once (temperature, pressure, angle-weighted temperature)? This improves the description of the atmosphere, not of the events. It can say which combination of air properties a fixed sample responds to. It cannot say which events within the sample respond. We include it as the atmospheric half of the naive baseline and keep its best target variable for the forest.
Why not use a graph or convolutional network on raw hits? These networks have more capacity, but the per-event signal here is tiny and the decisive outputs are attributions that physicists must check. Tree ensembles with Shapley values have already recovered physical mechanisms in the Majorana analysis. The only image-based IACT network in the file has never been tested on real data. A boosted-tree variant is run as a capacity check. A network is warranted only if the forest saturates.
Decisive Checks
We will use MINOS Near and Far Detector and NOvA surface and Near Detector cosmic data, which requires collaboration access. As a fallback, the full pipeline will be developed on CORSIKA plus detector simulation with injected seasonal profiles, where the true populations are known. That simulation-first stage also measures how many events the weak-label forest needs to recover a known split.
Four outcomes decide the project. The first is whether the forest's score quantiles reproduce MINOS's phase flip on held-out months, with the flip ordered as a function of score rather than of separation alone. The second is whether a forest trained only on pressure anomalies orders events the same way as one trained on temperature anomalies. The barometric and temperature responses have different physical origins, so agreement would point to a single event population responding to air density. The third is whether, after strict muon purification, the surface $\alpha_T$ for 15 or more muons stays near $-1.14$ or moves toward zero. A shift larger than its uncertainty would show that misidentification inflates the anomaly. The fourth is whether the score's modulation tracks atmospheric effective temperature and not detector-hall temperature. Failing that check would mean the forest has learned instrumentation, and the claim would be withdrawn.
References
- Observation of seasonal variation of atmospheric multiple-muon events in the MINOS Near and Far Detectors
- Observation of seasonal variation of atmospheric multiple-muon events in the NOvA Near Detector
- Seasonal variation of multiple-muon cosmic ray air showers observed in the NOvA detector on the surface
- Observation of muon intensity variations by season with the MINOS near detector
- Cosmic-muon characterization and annual modulation measurement with Double Chooz detectors
- Measurement of attenuation length of the muon content in extensive air showers from 0.3 to 30 PeV with LHAASO
- Implementation of the Random Forest method for the Imaging Atmospheric Cherenkov Telescope MAGIC
- Interpretable boosted-decision-tree analysis for the Majorana Demonstrator
- Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation in Imaging Atmospheric Cherenkov Telescopes