Neutrino Fingerprints: Image-Based Encodings of IceCube Events for CNN Direction Reconstruction

arXiv:2606.02788 · astro-ph.IM, cs.LG · Submitted 2026-06-01 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Neutrino Fingerprints: Image-Based Encodings of IceCube Events for CNN Direction Reconstruction".

Jane: Here's the gist of the paper. The gist This work introduces neutrino fingerprints, compact 72 × 72 × 3 images encoding IceCube event data, to enable convolutional neural networks to reconstruct incoming neutrino directions with a mean angular error of 1.10 rad Detector and Data Characteristics The IceCube Neutrino Observatory is equipped with a cubic kilometre of Antarctic ice containing 5 160 photomultiplier detectors deployed on 86 vertical strings The detector layout is sparse and irregularly distributed in three dimensions, which makes it challenging for learning algorithms that assume regular grids Each of the 140 million IceCube events is recorded as a variable–length sequence of pulses produced when Cherenkov photons reach the photomultiplier detectors For the Kaggle data, every pulse is described by four quantities: the time within the event window (ns), the integer sensor id of the detector that registered it, the measured charge (photoelectrons), and a Boolean auxiliary flag indicating reduced–quality readout In our work we restrict to the three features time, sensor id, and charge Pulse multiplicities are highly variable, from only a few to more than 105, which complicates direct use in fixed–size models and motivates our image–based fingerprint encoding Event Representation and Fingerprint Construction The core idea is to cast the irregular, sparse detector response into a fixed–size image representation, enabling the direct use of convolutional networks Each event is distilled into a compact neutrino fingerprint: a 72 × 72 × 3 image where every pixel corresponds to a single detector The mapping from detector index d ∈ (0,..., 5159) to image coordinates is given by a deterministic row–major ordering From the variable–length pulse sequence of each event we summarize every detector d by two scalars: a total charge Qd and an earliest arrival time Td To stabilize intensities across events with very different multiplicities and energies, we normalize each variable over the active detector set of the event D The RGB channels are defined as Rd = Norm(Td), Gd = Norm(Qd), Bd = Rd − Gd The red channel emphasizes gradients of photon arrival times, the green channel spotlights relative brightness, while the blue channel encodes their contrast This blue "opponent" channel Bd highlights boundaries where timing and light yield diverge, which often occur near the edges of the Cherenkov light cone and thus carry strong directional information Model Architecture and Training Setup We use a standard ResNet-18 architecture for image recognition as it has proven to be a competitive architecture in image recognition The total number of parameters for this architecture is 11.188.074 which is comfortable lower than the total number of training examples Since we are working with a regression task (we aim to predict the zenith and azimuth angle) we implement a final layer with only two nodes which output the predicted angles Our activation function was selected through trial-and-error and we found that GeLU

13: (out of ReLU, SiLU

14: and ELU

15: ) provided the best performance for our task We cast direction reconstruction as regression on the unit sphere S 2 We minimize the thee losses shown in Table 1 In all experiments we used Adam with learning rate 5 × 10−5, batch size 500, and dropout 0.1 which were selected by a small manual sweep Since arccos has a derivative which diverges as x→±1, we confine the cosine to

-1 + ε, 1 - ε: via clamp(x) = minmax(x, −1 + ε), 1 − ε with ε = 10−6 The Clamped Angular loss instead uses a straight-through clamp (ε = 10−5): its forward value equals clamp, but within the clamped band it passes a unit gradient (∂clamp/∂x = 1) rather than the zero gradient of an ordinary clamp Results and Conclusion Our single, compact CNN—trained without ensembling or advanced optimisation (no LR schedules, cosine restarts, or augmentations)—achieves an MAE of 1.10 rad on the Kaggle dataset For context, the top entries in the public IceCube–Neutrinos in Deep Ice Kaggle competition, which combined graph neural networks and transformer modules with heavy ensembling, reached ≈ 0.95 rad MAE on the held-out leaderboard

4: While our fingerprint CNN therefore falls short of absolute SOTA by about 0.15 rad, it provides a clean convolutional baseline on the Kaggle dataset This work advances a methodological claim: a deterministic event–to–RGB encoding, where detector timing, brightness, and their discrepancy are mapped into image channel, yields effective direction reconstruction with conventional convolutional networks The representation is intrinsically interpretable as each detector corresponds to a single pixel In conclusion, CNNs are not obsolete for IceCube-like reconstruction tasks: the proposed neutrinofingerprint encoding provides an interpretable and efficient baseline that establishes a solid reference while leaving room for future advances in model and encoding design Acknowledgments Brecht Verbeken acknowledges support from ENRICH (European Network for radiation-detection-based Research and Innovation addressing increasing societal CHallenges) through COST Action CA24131 Vincent Ginis acknowledges support from the Research Foundation – Flanders (FWO) under grants No.

Tom: G032822N and G0K9322N References

1: The IceCube Collaboration.

Jane: Low energy event reconstruction in icecube deepcore

4: Habib Bukhari, Dipam Chakraborty, Philipp Eller, Takuya Ito, Maxim V.

Tom: Shugaev, and Rasmus Ørsøe.

Paper summary: Tom: So we're wrapping up this look at "Neutrino Fingerprints: Image-Based Encodings of IceCube Events for CNN Direction Reconstruction." This paper takes massive, messy IceCube event data and turns it into compact images so a neural network can figure out where the neutrinos are coming from.

Jane: The authors are focused on how they take the timing, the charge, and even their difference—that blue channel—and encode them into seventy-two by seventy-two by three image spaces to handle those sparse detector responses. It’s about fitting those scattered data points into a fixed size for a convolutional network.

Tom: What's interesting is that this deterministic mapping from the detector index to the pixel coordinate, combined with the RGB encoding based on normalized time, charge, and their difference, creates an interpretable representation of what's happening physically in the signal.

Jane: The paper shows that by using these specific color channels—timing gradients in red, brightness in green, and contrast in blue—we get a signal that makes sense to a computer looking for directions.

Lu: What's important is they make the process deterministic; every detector gets a fixed pixel on the image, and the mapping from detector index to coordinate is defined clearly row-major.

Lalam: This work suggests that by using these physical properties to create an event fingerprint, we can open up avenues for designing better ways to process this type of sparse, high-dimensional input in the future. We need to think about how those fingerprints might integrate with other signal processing techniques next.

Tom: The implication here is that you don't always need the most complex AI setup to get a good result; sometimes a clever data encoding based on physical properties is enough to build something functional and deployable.

Lu: They are essentially proving that you can use a standard convolutional network architecture on this structured image data to get an angular error of one point one zero rad just with this setup without needing super complicated optimization or huge models to get started on your project.

Meng: The paper shows that you don't always need some massive graph network; you can use a standard ResNet18 architecture for the image recognition part, which keeps inference cost manageable when you scale things up to real data streams.

Lalam: This work suggests that the idea of creating an event fingerprint is really viable, opening up avenues for designing better ways to process this type of sparse, high-dimensional input in the future. We need to think about how those fingerprints might integrate with other signal processing techniques next.

Conclusion: Tom: So, wrapping up this look at "Neutrino Fingerprints: Image-Based Encodings of IceCube Events for CNN Direction Reconstruction." This paper takes massive, messy IceCube event data and turns it into compact images so a neural network can figure out where the neutrinos are coming from.

Jane: The authors are focused on how they take the pulse timing, the charge, and even their difference—that blue channel—and encode them into seventy-two by seventy-two by three image spaces to handle those sparse detector responses. It’s about fitting those scattered data points into a fixed size for a convolutional network.

Jane: The paper shows that by using those specific color channels—timing gradients in red, brightness in green, and contrast in blue—we get a signal that makes sense to a computer looking for directions.

Lu: What’s important is they make the process deterministic; every detector gets a fixed pixel on the image, and the mapping from detector index to coordinate is defined clearly row-major.

Meng: From what I see on the data handling side, this approach simplifies things because it forces a structure onto the input data right away, which helps when you’re trying to build something that runs reliably in production.

Lalam: This work suggests that by using those physical properties to create an event fingerprint, we can open up avenues for designing better ways to process this type of sparse, high-dimensional input in the future. We need to think about how those fingerprints might integrate with other signal processing techniques next.

Tom: Looking at the title "Neutrino Fingerprints: Image-Based Encodings of IceCube Events for CNN Direction Reconstruction," it’s clear they are tackling a fundamental problem of turning chaotic sensor readings into something a computer can actually map to the sky.

Jane: The authors are using this image encoding idea to solve that mapping problem directly, showing how the physical characteristics of each event can be represented by an RGB image instead of just a list of numbers.

Lu: What’s important is that they are making it deterministic; every detector gets a fixed pixel on the image, and the mapping from detector index to coordinate is defined clearly row-major.

Lalam: It means we can start thinking about how these fingerprints could be used for classification too, not just direction reconstruction, which is a bigger application area.

Tom: So the implication here is that you don't always need the most complex AI setup to get a good result; sometimes a clever data encoding based on physical properties is enough to build something functional and deployable.

Jane: It gives us a solid, understandable reference point for how we should structure these kinds of representations when dealing with irregular detector layouts in massive sensors like IceCube.

Lu: This work lays out exactly how the RGB channels emphasize different physical aspects, like timing gradients versus brightness contrast in a way that makes the signal more obvious to the CNN.

Meng: It shows that you don't always need some massive graph network; you can use a standard ResNet18 architecture for the image recognition part, which keeps inference cost manageable when you scale things up to real data streams.

Lalam: The idea of creating an event fingerprint is really viable, opening up avenues for designing better ways to process this type of sparse, high-dimensional input in the future. We need to think about how those fingerprints might integrate with other signal processing techniques next.

Brecht Verbeken, Floriano Tori, Vincent Ginis

Vrije Universiteit Brussel · School of Engineering and Applied Sciences, Harvard University

astro-ph.IM, cs.LG

Submitted: 2026-06-01

Updated: 2026-06-01

Comments: 6 pages, 1 figure

Code: https://github.com/Integrated-Intelligence-Lab/neutrino-detection-experiment

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

The gist: The gist This work introduces neutrino fingerprints, compact 72 × 72 × 3 images encoding IceCube event data, to enable convolutional neural networks to reconstruct incoming neutrino directions with

Key concepts

Neutrino Fingerprint: A compact 72x72x3 image representation of an IceCube event. This image is created by mapping the sparse detector response onto a fixed grid, where each pixel represents a single detector. The three color channels encode different aspects of the data: timing gradients (red), relative brightness (green), and contrast between timing and light yield (blue).
Event Encoding: The process of summarizing variable-length pulse sequences from an event into fixed-size scalar values for each detector. For every detector, two key scalars are calculated: the total charge ($Q_d$) and the earliest arrival time ($T_d$). These scalars are then normalized across all active detectors to ensure intensities remain comparable regardless of how many pulses an event produced.
RGB Channel Definition: The three color channels in the fingerprint represent distinct physical information. The red channel, $R_d$, is based on the normalized earliest arrival time ($T_d$), emphasizing timing gradients. The green channel, $G_d$, represents the normalized total charge ($Q_d$), highlighting relative brightness. The blue channel, $B_d = |R_d - G_d|$, captures the contrast between these two features, which is particularly useful for identifying boundaries in the light cone.
Clamped Angular Loss: A specialized loss function used during training to handle the angular regression task. Since the cosine function's derivative becomes zero at its limits ($ ext{x} = ext{±1}$), this loss uses a straight-through clamp. This means that while the output is clamped to a specific range, it passes a unit gradient within that band during backpropagation, ensuring effective learning near the boundaries of the unit sphere.

Terminology

Transformer — 4× in this paper · explained in 2 episodes
Loss Function — 3× in this paper · explained in 2 episodes
Activation Function — 2× in this paper · explained in 1 episode
Adam — 2× in this paper · explained in 2 episodes
Benchmark — 2× in this paper · explained in 2 episodes
Core — 2× in this paper · explained in 2 episodes
Reinforcement Learning — 2× in this paper · explained in 4 episodes
Transformers — 2× in this paper · explained in 2 episodes
Convolutional Neural Networks (CNNs) — 1× in this paper · explained in 3 episodes
Hilbert Space — 1× in this paper · explained in 2 episodes
IEEE International Conference on Computer Vision Workshops — 1× in this paper
Inverse Problem — 1× in this paper · explained in 2 episodes

Summary

The gist

This work introduces neutrino fingerprints, compact 72 × 72 × 3 images encoding IceCube event data, to enable convolutional neural networks to reconstruct incoming neutrino directions with a mean angular error of 1.10 rad

Detector and Data Characteristics

The IceCube Neutrino Observatory is equipped with a cubic kilometre of Antarctic ice containing 5 160 photomultiplier detectors deployed on 86 vertical strings The detector layout is sparse and irregularly distributed in three dimensions, which makes it challenging for learning algorithms that assume regular grids Each of the 140 million IceCube events is recorded as a variable–length sequence of pulses produced when Cherenkov photons reach the photomultiplier detectors For the Kaggle data, every pulse is described by four quantities: the time within the event window (ns), the integer sensor id of the detector that registered it, the measured charge (photoelectrons), and a Boolean auxiliary flag indicating reduced–quality readout In our work we restrict to the three features time, sensor id, and charge Pulse multiplicities are highly variable, from only a few to more than 105, which complicates direct use in fixed–size models and motivates our image–based fingerprint encoding

Event Representation and Fingerprint Construction

The core idea is to cast the irregular, sparse detector response into a fixed–size image representation, enabling the direct use of convolutional networks Each event is distilled into a compact neutrino fingerprint: a 72 × 72 × 3 image where every pixel corresponds to a single detector The mapping from detector index d ∈ (0,..., 5159) to image coordinates is given by a deterministic row–major ordering From the variable–length pulse sequence of each event we summarize every detector d by two scalars: a total charge Qd and an earliest arrival time Td To stabilize intensities across events with very different multiplicities and energies, we normalize each variable over the active detector set of the event D The RGB channels are defined as Rd = Norm(Td), Gd = Norm(Qd), Bd = Rd − Gd The red channel emphasizes gradients of photon arrival times, the green channel spotlights relative brightness, while the blue channel encodes their contrast This blue opponent channel Bd highlights boundaries where timing and light yield diverge, which often occur near the edges of the Cherenkov light cone and thus carry strong directional information

Model Architecture and Training Setup

We use a standard ResNet-18 architecture for image recognition as it has proven to be a competitive architecture in image recognition The total number of parameters for this architecture is 11.188.074 which is comfortable lower than the total number of training examples Since we are working with a regression task (we aim to predict the zenith and azimuth angle) we implement a final layer with only two nodes which output the predicted angles Our activation function was selected through trial-and-error and we found that GeLU [13] (out of ReLU, SiLU [14] and ELU [15]) provided the best performance for our task We cast direction reconstruction as regression on the unit sphere S 2 We minimize the thee losses shown in Table 1 In all experiments we used Adam with learning rate 5 × 10−5, batch size 500, and dropout 0.1 which were selected by a small manual sweep Since arccos has a derivative which diverges as x→±1, we confine the cosine to [-1 + ε, 1 - ε] via clamp(x) = minmax(x, −1 + ε), 1 − ε with ε = 10−6 The Clamped Angular loss instead uses a straight-through clamp (ε = 10−5): its forward value equals clamp, but within the clamped band it passes a unit gradient (∂clamp/∂x = 1) rather than the zero gradient of an ordinary clamp

Results and Conclusion

Our single, compact CNN—trained without ensembling or advanced optimisation (no LR schedules, cosine restarts, or augmentations)—achieves an MAE of 1.10 rad on the Kaggle dataset For context, the top entries in the public IceCube–Neutrinos in Deep Ice Kaggle competition, which combined graph neural networks and transformer modules with heavy ensembling, reached ≈ 0.95 rad MAE on the held-out leaderboard [4] While our fingerprint CNN therefore falls short of absolute SOTA by about 0.15 rad, it provides a clean convolutional baseline on the Kaggle dataset This work advances a methodological claim: a deterministic event–to–RGB encoding, where detector timing, brightness, and their discrepancy are mapped into image channel, yields effective direction reconstruction with conventional convolutional networks The representation is intrinsically interpretable as each detector corresponds to a single pixel In conclusion, CNNs are not obsolete for IceCube-like reconstruction tasks: the proposed neutrinofingerprint encoding provides an interpretable and efficient baseline that establishes a solid reference while leaving room for future advances in model and encoding design

Acknowledgments

Brecht Verbeken acknowledges support from ENRICH (European Network for radiation-detection-based Research and Innovation addressing increasing societal CHallenges) through COST Action CA24131 Vincent Ginis acknowledges support from the Research Foundation – Flanders (FWO) under grants No. G032822N and G0K9322N

References

[1] The IceCube Collaboration. Low energy event reconstruction in icecube deepcore

[4] Habib Bukhari, Dipam Chakraborty, Philipp Eller, Takuya Ito, Maxim V. Shugaev, and Rasmus Ørsøe. Icecube–neutrinos in deep ice: The top 3 solutions from the public kaggle competition

[5] Nicholas Choma, Federico Monti, Lisa Gerhardt, Tomasz Palczewski, Zahra Ronaghi, Prabhat Prabhat, Wahid Bhimji, Michael M. Bronstein, Spencer R. Klein, and Joan Bruna. Graph neural networks for icecube signal classification

[6] Maximilian Kronmueller and Theo Glauch. Application of deep neural networks to event type classification in icecube

[7] IceCube Collaboration. Fast low energy reconstruction using convolutional neural networks, 2025

[8] The IceCube Collaboration. The icecube data acquisition system: Signal capture, digitization, and timestamping

[9] Peter B. Denton and Irene Tamborra. Invisible neutrino decay could resolve icecube’s track and cascade tension

[10] Ivet Rafegas and Maria Vanrell. Color representation in cnns: Parallelisms with biological vision

[11] Karen Simonyan and Andrew Zisserman. Two-stream convolutional networks for action recognition in videos

[12] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition

[13] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus)

[14] Stefan Elfwing, Eiji Uchibe, and Kenji Doya. Sigmoid-weighted linear units for neural network function approximation in reinforcement learning

[15] Djork-Arné Clevert, Thomas Unterthiner, and Sepp Hochreiter. Fast and accurate deep network learning by exponential linear units (elus)

[16] Bert Verbruggen and Vincent Ginis.

Improvements for AI systems

textbfEnsemble Training for Robustness Recovery of SOTA Performance Stability: The current baseline achieves an MAE of 1.10 rad, falling short of SOTA by about 0.15 rad; implementing ensembling or advanced optimization techniques could bridge this gap to near-SOTA performance by leveraging the heavy ensembling seen in top competition entries [4].
textbfLocality-Aware Encoding Refinement: The current encoding uses a naive 3D-to-2D locality proxy via row-major ordering, which separates physical neighbors; improving this mapping using methods like space-filling curves [16], string×depth rasters, additional color channels would enhance the representation of spatiotemporal structure.
textbfDirectional Uncertainty Calibration: Incorporating alternative loss formulations based on the von Mises–Fisher distribution would extend the cosine objective by learning a concentration parameter and thereby provide calibrated directional uncertainties, allowing the model to output not just a direction but also its confidence level.
textbfArchitectural Modernization: Upgrading the network architecture beyond ResNet-18, perhaps using depthwise-separable/modern CNN blocks, could improve computational efficiency while maintaining or exceeding performance, addressing the potential for modest architectural upgrades mentioned in the discussion.

Abstract

Reconstructing the direction of incoming neutrinos in the IceCube Neutrino Observatory is an important problem in astrophysics. The public IceCube--Neutrinos in Deep Ice Kaggle competition provided 140 million simulated events to benchmark reconstruction techniques. To address this challenge from a novel perspective we introduce neutrino fingerprints compact 72 times 72 times 3 images in which each pixel represents a single detector, with pulse timing and charge statistics encoded as color channels. This representation transforms sparse, irregular pulse data into dense images suitable for convolutional processing. Our ResNet18 model achieves a mean angular error of 1.10 rad, indicating that convolutional networks trained on fingerprints rival more complex architectures while offering an effective, interpretable baseline for IceCube event reconstruction.