Two Watts is All You Need: Enabling In-Detector Real-Time Machine Learning for Neutrino Telescopes Via Edge Computing
Authors
Miaochen Jin, Yushi Hu, Carlos A. Argüelles
Abstract
The use of machine learning techniques has significantly increased the physics discovery potential of neutrino telescopes. In the upcoming years, we are expecting upgrade of currently existing detectors and new telescopes with novel experimental hardware, yielding more statistics as well as more complicated data signals. This calls out for an upgrade on the software side needed to handle this more complex data in a more efficient way. Specifically, we seek low power and fast software methods to achieve real-time signal processing, where current machine learning methods are too expensive to be deployed in the resource-constrained regions where these experiments are located. We present the first attempt at and a proof-of-concept for enabling machine learning methods to be deployed in-detector for water/ice neutrino telescopes via quantization and deployment on Google Edge Tensor Processing Units (TPUs). We design a recursive neural network with a residual convolutional embedding, and adapt a quantization process to deploy the algorithm on a Google Edge TPU. This algorithm can achieve similar reconstruction accuracy compared with traditional GPU-based machine learning solutions while requiring the same amount of power compared with CPU-based regression solutions, combining the high accuracy and low power advantages and enabling real-time in-detector machine learning in even the most power-restricted environments.
Concepts
The Big Picture
Imagine trying to run a hospital’s MRI machine on a car battery, buried under a mile of Antarctic ice. That’s roughly the engineering challenge facing scientists who want to use machine learning at neutrino telescopes. These enormous instruments span cubic kilometers of ocean water or polar ice, planted in places where power is scarce and bandwidth is thin.
Neutrino telescopes like IceCube in Antarctica, and the upcoming KM3NeT in the Mediterranean, hunt for ghostly particles called neutrinos. These particles stream in from some of the universe’s most violent events: exploding stars, merging black holes, blazing galactic cores. Catching them in real time matters, but the detectors generate roughly 3,000 events per second, and next-generation experiments will push that rate eight to thirty times higher.
Current machine learning approaches run on powerful graphics cards (GPUs). They’re accurate but hungry for electricity. Simpler processor-based methods on standard CPUs use far less power but sacrifice precision. Scientists have been stuck choosing between fast-and-dumb and smart-but-power-starved.
A team from Harvard and the University of Washington has found a third path: deploy machine learning directly inside the detector, on a chip that sips just two watts of power.
Key Insight: By combining a custom neural network architecture with aggressive quantization and deployment on a Google Edge TPU, the researchers achieved GPU-level reconstruction accuracy at CPU-level power consumption. That combination wasn’t previously possible in the remote, power-limited environments where these detectors operate.
How It Works
At the center of the approach is a neural network carefully engineered to match unusual hardware. The Google Edge TPU is a microchip built for machine learning inference at the network’s edge (think smart cameras or mobile devices). It consumes only 2 watts for the chip itself, 3 watts for the full development board. Squeezing a physics reconstruction algorithm onto it means rethinking the model from the ground up.
The team’s pipeline has three major stages:
-
Data preprocessing — Raw detector hits (photons arriving at optical modules) are transformed into a compact representation. The network receives time-series data from each optical module, encoding hit patterns in a way the Edge TPU’s constrained memory can handle.
-
Recursive network with residual convolutional embedding — The core is a recursive neural network, which applies the same transformation repeatedly to structured input. This structure suits the hierarchical geometry of thousands of optical modules. The inputs first pass through a residual convolutional embedding: a set of pattern-detecting filters that compress the full detector geometry into a compact summary. Residual connections preserve information across that compression step.
-
Quantization — Standard neural networks use 32-bit floating point numbers; the Edge TPU requires 8-bit integer quantization, shrinking numerical precision to a coarser grid (the only format the chip handles natively). Done naively, this destroys accuracy. The team adapted a fine-tuning procedure that retrains the network after quantization, recovering most of the lost precision.
The two test detectors, WaterHex and IceHex, mimic realistic next-generation telescope geometries: 114 strings of 60 optical modules each (6,840 total), inspired by the KM3NeT layout but testing water and ice media separately.
Why It Matters
The Edge TPU matches GPU-based reconstruction accuracy while consuming the same power as the simple CPU regression methods currently used for real-time triggers. That’s not an incremental improvement. It collapses a tradeoff that has constrained the field for years.
The implications reach well beyond IceCube. Experiments like TAMBO in Peru and GRAND, a proposed radio-based detector, envision solar-powered remote stations where every watt is spoken for. TRIDENT in China will be thirty times larger than IceCube.
When power budgets are measured in watts rather than kilowatts, in-detector intelligence has historically meant crude threshold cuts, not neural networks. This work makes genuine real-time physics discrimination possible: separating signal neutrinos from background muons, flagging rare high-energy events for rapid follow-up alerts, all without a power cable to the outside world.
The team frames this as a proof of concept. The current architecture is shaped by hardware constraints rather than optimized for peak reconstruction performance. Future iterations could try newer edge AI chips, different quantization schemes, or hybrid designs where an edge processor handles first-level selection and passes only the most interesting events to more powerful off-site systems. Deployment on actual IceCube hardware is a plausible next step as part of the IceCube-Gen2 upgrade.
Neutrino astronomy is entering a data-rich era, and the algorithms that process those data need to live where the data are born. For some of the most extreme physics experiments ever built, pushing intelligence to the edge may be the only option.
Bottom Line: Two watts is genuinely enough. Edge AI hardware can match GPU-level accuracy in neutrino reconstruction at a fraction of the power cost, making real-time in-detector machine learning viable for experiments that were previously too power-constrained to use it.
IAIFI Research Highlights
This work connects deep learning model design with experimental particle physics constraints, reshaping GPU-era neural network methods into an architecture built around the power and memory limits of remote detector hardware.
The full pipeline (custom recursive network with residual convolutional embedding, post-training quantization, and fine-tuning recovery) deploys neural networks on 8-bit integer edge hardware without significant accuracy loss. The technique applies broadly beyond physics.
Real-time in-detector ML reconstruction lets neutrino telescopes issue rapid alerts for rare astrophysical events and separate signal from background at the source, improving sensitivity to cosmic neutrino sources and new physics signatures.
Future directions include extending this framework to next-generation telescopes like IceCube-Gen2 and TRIDENT and exploring newer edge AI chips with higher throughput; the full paper is available at [arXiv:2311.04983](https://arxiv.org/abs/2311.04983).