Condition-Aware Simulation-to-Real Translation for Horizon Cherenkov Neutrino Searches

Telescopes that hunt for ultra-high-energy tau neutrinos look sideways at the horizon. A neutrino skims through the Earth's crust or the ocean, produces a tau lepton, and the tau decays in the air, starting an upward-going particle shower whose flash of Cherenkov light a camera can catch. Every classifier that separates these hypothetical showers from the far more numerous ordinary cosmic-ray showers is trained only on simulation, because no such neutrino has ever been observed. Near the horizon, simulation is at its weakest. Light travels through roughly a hundred kilometres of dense, hazy, variable ground-level air, and the haze, the sky glow and the viewing angle all change continuously over a night. We propose to learn, without paired examples, a translation from simulated camera images to realistic ones. The translation is steered by measured continuous conditions: atmospheric transmission, night-sky brightness and pointing angle. It is trained and validated on the one real population that horizon telescopes record in abundance, which is background cosmic-ray showers.

The central design choice is a split between what can be copied from data and what must be learned. Stochastic, additive effects such as night-sky photon noise and stray starlight are taken directly from real empty camera frames and overlaid onto simulated showers, so nothing has to be learned for them. Only the deterministic, condition-dependent part of the instrument and atmosphere response is learned: dimming, blurring, timing smear and gain nonlinearity as functions of the measured conditions. This split targets a known failure of cycle-consistent translation. When one domain carries much more random noise than the other, a two-way generator learns to hide that noise in invisible pixel patterns so that it can reconstruct the input. The split also lets us impose per-event physical constraints, such as total light decreasing monotonically as transmission drops, which distribution-level methods cannot guarantee.

Core Claim

A conditional, unpaired translator restricted to the deterministic part of the response can match real horizon background images across a continuous range of atmospheric and pointing conditions better than three alternatives: raw simulation, noise overlay alone, and unconditional translation. The real noise is overlaid rather than learned, and the translator is constrained to be monotone in transmission. A neutrino classifier trained on translated simulation should then predict the real background pass rate correctly and keep a signal efficiency we can verify in a controlled setting. What is new is not "CycleGAN for physics". The new parts are three. The translation is conditioned on continuous, externally measured physical quantities. The noisy part of the domain gap is removed from the learning problem altogether. And the result is judged by downstream physics quantities, background rate and effective sensitivity, rather than by image realism.

Positioning

Prior work falls into three groups.

Work that supplies the method. CycleGAN-driven transfer learning for electronics response emulation in high-purity germanium detectors introduced CPU-Net, an unpaired cycle-consistent translator from simulated to real germanium detector pulses. It raised drift-time agreement, measured as overlap (intersection over union) between distributions, from 39.5% to 62.3% on single-escape events and from 5.4% to 22.5% on double-escape events. This is our closest precedent and our existence proof that unpaired translation captures response features that first-principles simulation misses. CPU-Net is one-dimensional and unconditional, though: one fixed simulated distribution is mapped to one fixed real distribution. The same paper also reports that runs converge to different local optima, a warning we take seriously.

Work that supplies continuous conditioning. - AttriCtrl maps a scalar through a sinusoidal embedding and a small network into tokens that drive a diffusion model, giving smooth intensity control. We reuse its value encoder. - CamEdit controls physical camera parameters continuously, but it is trained on synthetic pairs, which do not exist for us. - Property Controllable Variational Autoencoder via Invertible Mutual Dependence enforces monotone, invertible links between latent axes and labelled properties. We borrow the monotonicity idea and apply it to a measured condition, transmission, rather than a learned latent. - Path-independent Flow Matching for Multi-parameter Generative Dynamics builds unpaired multi-parameter transport whose result does not depend on the order in which parameters are changed. It does so only at the level of distributions and explicitly not per sample. We treat it as an alternative backbone and as a source of a consistency test. Its per-sample gap is why we add per-event constraints.

Work that motivates the gap. - Discrimination of Cosmic Rays and Neutrinos at Large Zenith Angles Using Machine Learning trains stacked ensembles on idealized CORSIKA showers in discrete zenith and energy bins. The authors state that real detector effects are not captured. - Convolution and Graph-based Deep Learning Approaches for Gamma/Hadron Separation in Imaging Atmospheric Cherenkov Telescopes names the simulation-reality gap as its main limitation. - The MAGIC horizon searches (Limits on the flux of tau neutrinos from 1 PeV to 3 EeV with the MAGIC telescopes; Search for tau neutrinos with the MAGIC telescopes: improving selection criteria*) document ground-layer extinction and variable aerosols. They also chose a conservative "zero-background" cut, which our method aims to make less conservative without losing safety.

Research Program

Data. The real domain is MAGIC horizon data: about 30 hours pointed at the ocean (seaON) and 9.2 hours pointed just above the horizon (seaOFF). Since no candidate survived in seaON, essentially all recorded showers in both sets are cosmic-ray background. Each event carries measured conditions $c$:

Pedestal frames are binned by these conditions and form the noise library. The simulated domain is the existing ANIS, PYTHIA and CORSIKA chain for both protons and tau-induced showers, rendered without night-sky noise. Access to MAGIC data requires collaboration agreement. If that is unavailable, the same program applies to Trinity demonstrator data once it exists; the file reports no operational numbers for Trinity yet.

Model. Write a simulated image as $x$, a real pedestal frame drawn at matching conditions as $n_c$, and the translator as $T_\theta$. The realistic image is formed as

$$ \hat{y} = T_\theta(x, c) + n_c , $$

where $T_\theta$ is a conditional U-Net with AttriCtrl-style value embeddings of $c$. A discriminator compares $\hat{y}$ against real background images at the same conditions. The reverse map $T^{-1}_\phi$ is applied only after the real noise has been subtracted from the real image's statistics. In practice it is applied to the denoised (image-cleaned) real images that IACT analyses already compute, so the round-trip never has to carry the random noise. Three further losses complete the objective. The first is a monotonicity penalty requiring each event's total charge in $T_\theta(x,c)$ to be non-increasing in the transmission component of $c$. The second is a small-displacement penalty that keeps the image centroid and major axis close to the simulated values, so geometry is preserved. The third, borrowed as a test from PiFM, is a commutation check: changing transmission then zenith should give the same result as the reverse order.

A flow-matching backbone in place of the adversarial one is a planned variant. It is aimed at the run-to-run instability CPU-Net reports.

Downstream use. A single classifier, conditioned continuously on $c$, is trained on translated proton and tau-shower simulation. It replaces both the binned ensembles and the fixed cut on the Hillas-based discriminant (Hillas parameters are the standard moments of shower image shape).

Controlled pseudo-real benchmark. Real neutrino images do not exist, so we also build a sim-to-sim benchmark where truth is known. A second simulation with withheld physics plays the role of "real": a different aerosol profile, a broadened optical blur, and nonlinear gain. The translator never sees its parameters. This is the only place where signal efficiency after translation can be measured exactly. Its weakness is that hidden effects absent from both simulations stay invisible, and we say so.

Baselines

The naive reference point is a classifier trained on raw simulation and applied directly to real data, as in the published MAGIC chain. Its strongest reported operating point is the cut $\log_{10}(Y') > 2.35$, a "zero-background" region with no surviving candidates in 30 hours and a 90% confidence upper limit of $2.0 \times 10^{-4}$ GeV cm$^{-2}$ s$^{-1}$ (Limits on the flux of tau neutrinos from 1 PeV to 3 EeV with the MAGIC telescopes). Beating it means a higher signal efficiency at a background rate validated on held-out real data to be equally consistent with zero, which translates to a lower projected limit for the same exposure.

The strongest existing machine-learning discriminator is the stacked ensemble of Discrimination of Cosmic Rays and Neutrinos at Large Zenith Angles Using Machine Learning, with reported simulation-only AUC "generally exceeded 0.94". We retrain it on our simulation with the same features, the same events and a matched hyperparameter search budget, then apply it to real background. The test is whether its real-data background pass rate matches what simulation predicts. A simulation-only AUC is not evidence of real-data performance, and on Lacuna we found no reported real-data result for any horizon Cherenkov classifier.

The strongest adaptation competitor is real pedestal overlay alone: our model with $T_\theta$ set to the identity. It is standard practice and handles additive noise exactly.

Two further baselines are an unconditional CPU-Net-style translator plus overlay, and domain randomization over transmission, blur and gain.

Three ablations isolate the key ingredients: removing the conditioning on $c$, replacing overlay with full two-way cycle consistency on noisy images, and dropping the monotonicity penalty.

All methods share the same simulated events, the same real training nights (with whole nights held out), five seeds each, and equal GPU budgets. Beating a baseline means lower distributional mismatch on held-out nights across seeds, measured as Wasserstein distance and CPU-Net-style overlap on Hillas Size, Length, Mean Time and the discriminant. It also means a predicted background pass rate within statistical error of the observed rate where the baseline fails this test.

Why not just…?

Why not just overlay real pedestal noise on simulation? This is the most serious threat. If additive noise is the whole gap, overlay will match real background, and our learned component will add nothing. We will report this honestly. The experiment that settles it is a comparison of overlay-only against the full model on held-out nights, stratified by transmission. If the mismatch remaining after overlay grows with haze, the gap is a multiplicative or blurring effect that overlay cannot represent. The MAGIC papers' emphasis on ground-layer extinction suggests this, but they do not measure it.

Why not just randomize the simulation parameters? Randomization makes a classifier tolerant of variation, but it does not tell it which variation is real on a given night. It therefore widens background distributions exactly where the faint signal sits. It also cannot produce effects that are not parameterized in the generator. Our conditioned model uses the measured $c$ instead of sampling it. The pseudo-real benchmark, where the hidden nonlinear gain is outside the randomized family, directly tests whether randomization recovers signal efficiency.

Why not just align classifier features between simulation and real data adversarially, skipping image generation? Feature alignment on background alone can distort how the signal class is represented, and there is no real signal to catch the distortion. We include domain-adversarial training as a baseline. Its signal-efficiency error on the pseudo-real benchmark, where truth is known, is the decisive comparison. Translated images have one further advantage: physicists can inspect them and run them through the established Hillas analysis.

Why not just bin and train separate classifiers? Binning in zenith, energy and transmission multiplies the number of models and leaves the best behaviour at intermediate conditions undefined. We test this by training on a sparse grid of transmission values and evaluating on nights between the training values.

Proposed Experiments and Decisive Checks

The first check is on the pseudo-real benchmark. The method must recover tau-shower efficiency at fixed background within a few percent of truth, and at least one baseline, overlay or randomization, must miss it clearly. Otherwise the learned component is unjustified.

On held-out real MAGIC nights, the translated proton simulation must match real background in Hillas-parameter overlap and Wasserstein distance better than overlay alone, with the advantage concentrated at low transmission.

The classifier's predicted background pass rate on held-out seaON nights must agree with the observed count. This is the safety check that permits a looser cut than $\log_{10}(Y') > 2.35$.

Per event, the monotonicity in transmission and the commutation check should hold, and the variance across seeds of the downstream efficiency should be smaller than for the unconditional adversarial translator.

The final number is the projected point-source sensitivity under the same exposure, compared against the published limit. If the gain over overlay alone is within seed variance, we conclude that noise overlay is sufficient for horizon searches, which is itself a useful result for Trinity's analysis design.