Two Watts is All You Need: Enabling In-Detector Real-Time Machine Learning for Neutrino Telescopes Via Edge Computing

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, some spanning cubic kilometers of ocean water or polar ice, sit in places where power is precious, bandwidth is limited, and the physics is extraordinary.

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 most violent events in the cosmos: exploding stars, merging black holes, and the blazing cores of distant galaxies. Catching them in real time matters — but the detectors generate data at 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, called GPUs — accurate, but hungry for electricity. Simpler processor-based methods, running 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 now 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 — a combination that wasn’t previously possible in resource-constrained detector environments.

How It Works

The heart of the approach is a carefully engineered neural network matched to unusual hardware. The Google Edge TPU is a microchip designed for machine learning inference at the network’s edge (think smart cameras or mobile devices), consuming only 2 watts for the chip itself and 3 watts for the full development board. Squeezing a physics reconstruction algorithm onto it requires rethinking the model from the ground up.

The team’s pipeline has three major stages:

1. 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.

2. Recursive network with residual convolutional embedding — The core is a recursive neural network, which applies the same transformation repeatedly to structured input — well-suited for 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 the network can work with. Residual connections preserve information across that compression step.

3. 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 numbers tell a clear story. The Edge TPU matches GPU-based reconstruction accuracy while consuming the same power as the simple CPU regression detectors 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 extend 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.

With power budgets measured in watts rather than kilowatts, in-detector intelligence has historically meant crude threshold cuts, not neural networks. This work opens a door to genuine real-time physics discrimination — separating signal neutrinos from background muons, flagging rare high-energy events for rapid follow-up alerts — without a power cable to the outside world.

The team frames this as a proof of concept, acknowledging that the current architecture is shaped by hardware constraints rather than optimized for peak reconstruction performance. Future iterations could explore 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. There’s also a natural path toward deployment on actual IceCube hardware 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. Pushing intelligence to the edge isn’t just an engineering convenience — for some of the most extreme physics experiments ever built, it may be the only option.

Bottom Line: Two watts is genuinely enough. This proof-of-concept shows that edge AI hardware can match GPU-level accuracy in neutrino reconstruction at a fraction of the power cost, unlocking real-time in-detector machine learning for experiments that were previously too power-constrained to benefit from it.