PLATON Particle Detector Swaps Millions of Fibers for AI

Forget the headline everyone else is running today: “one camera replaces millions of detector fibers.” That framing is technically true and almost completely misses the point of the PLATON particle detector, the ETH Zurich/EPFL prototype that just reconstructed a full neutrino interaction at 200-micrometer resolution using a single, unsegmented block of scintillating plastic. The real story is what’s doing the reconstructing: a Transformer neural network — the same model family that powers large language models — repurposed to untangle scattered light instead of scattered words. And the even bigger surprise: this device, built to hunt neutrinos and dark matter, looks likely to reach a hospital PET scanner before it ever sees a collider.

I’ve spent enough time around scintillator arrays at CERN to know how absurd this sounds. Traditional neutrino and dark-matter detectors work by physically dicing a scintillating volume into thousands or millions of individual fibers or pixels, each wired to its own photosensor, because that segmentation is how you normally infer where a particle track passed. PLATON throws the fibers away entirely. It replaces physical segmentation with computational segmentation — a plenoptic (light-field) camera and an AI model do algorithmically what wires and connectors used to do by brute force.

What Is the PLATON Particle Detector, Really?

Picture a solid block of plastic scintillator with no internal wiring at all — just one continuous slab that glows wherever a charged particle deposits energy. Every conventional design, from dark-matter TPCs to massive neutrino observatories like JUNO, gets its 3D position information by segmenting that active volume into thousands of individually read-out cells. PLATON keeps the block whole and instead points a specialized plenoptic camera at it — a light-field imaging system built around a micro-lens array paired with SPAD sensors. Unlike an ordinary camera pixel, each SPAD (single-photon avalanche diode) can register the arrival angle of a single scintillation photon, not just its landing spot, which is the raw directional data a conventional detector camera simply throws away.

Inside the PLATON Particle Detector: A Transformer for Photons

Here’s the part every “AI replaces fibers” headline skips: the actual neural network. PLATON’s reconstruction algorithm is built on a Transformer, the same architecture family underneath GPT-style large language models. In an LLM, a Transformer’s attention mechanism learns which words in a sentence are statistically related to which other words, regardless of how far apart they sit. PLATON’s Transformer does the same trick on light: it learns which scattered scintillation photons, captured across the plenoptic camera’s field of view, are correlated with the same original particle track, then reconstructs the 3D path computationally instead of relying on physical fiber-by-fiber tagging.

That’s a genuinely counterintuitive re-use of the architecture — sequence modeling repurposed for spatial photon correlation — and it’s why I think Transformer-based reconstruction is going to quietly show up across detector physics over the next few years. The SPAD sensors underneath are their own story worth flagging: silicon photon counters have been creeping into roles PMTs used to own for a decade (see our SiPM vs PMT breakdown for why silicon hasn’t fully won yet), and PLATON’s angle-sensitive SPAD array is arguably the most ambitious use case yet. Squeezing directional photon information — not just hit counts — out of a scintillator block also means the usual thin-detector statistics headaches, the kind we’ve covered in our piece on Landau-distribution quirks in thin silicon, get reshaped rather than eliminated.

The published numbers back up the ambition. In the case-study event reconstruction described in the team’s Nature Communications paper, PLATON achieved 200-micrometer spatial resolution — a fifth of a millimeter — reconstructing a full neutrino interaction from nothing but photon-correlation data. Scale simulations up to a practical 1-cubic-meter detector volume, and the projected resolution settles to a few millimetres, which the authors note is already on par with state-of-the-art segmented plastic scintillator detectors. In other words, you can drop the millions of fibers and still match the performance of the detectors that need them.

Why a Hospital Will Get the PLATON Particle Detector Before CERN Does

Here’s the twist nobody leads with: PLATON was conceived to catch neutrinos and dark-matter candidates, yet its most plausible next stop is an oncology imaging suite, not a collider hall. Positron emission tomography (PET) scanners already work by detecting scintillation light from annihilation photons, which makes PET a smaller, faster, commercially tractable proving ground for exactly the same single-block-plus-AI approach. A PET-sized detector volume is a rounding error next to a deep-underground neutrino experiment, and a hospital procurement cycle moves in months where a flagship physics experiment moves in decades.

That ordering says something about how expensive frontier physics hardware gets funded now: medical imaging isn’t just a spinoff market, it’s becoming the fastest path to de-risking a novel PLATON-style detector concept before physics collaborations bet a decade of beam time on it. I’d bet plenoptic scintillator blocks show up in commercial PET systems well before any collaboration commits one to a tonne-scale dark-matter search.

Watch three things over the next two years: whether PLATON-derived designs actually reach a commercial PET prototype, whether a neutrino or dark-matter collaboration commits real beam time to a full-scale block, and whether other labs start bolting Transformers onto their own scintillator arrays now that the trick is public. None of that requires the fiber count to matter anymore — it requires the network to keep getting better at finding correlations in scattered light, and if the last few years of Transformer progress are any guide, that part isn’t the bottleneck.