In July 2026, physicists at ETH Zurich and EPFL unveiled a particle detector that does something almost perverse by the standards of the field: instead of subdividing a detector into millions of separate light-collecting channels, each wired to its own sensor, it uses a single solid block of scintillator material, one camera, and an AI model. The device, called PLATON, borrows a camera technology built for photography — the plenoptic, or light-field, camera — and pairs it with ultra-sensitive photon sensors and a transformer neural network to reconstruct exactly where inside that block a particle passed through, in three dimensions, from a sparse scatter of individual photons. It’s a striking substitution of software for hardware. And the researchers themselves didn’t stop at particle physics — they’ve already filed patents applying the same architecture to PET medical scanners. That’s a real, announced medical pivot. It just isn’t the one this piece originally set out to explore.
Scientific Foundation
The core problem PLATON solves is one of geometry and inference. In a conventional particle detector, a scintillator crystal is diced into millions of tiny segments, each coupled to its own optical fiber and photon detector, so that whichever segment lights up tells you roughly where the particle passed. PLATON discards the segmentation entirely. It couples one large, undivided block of scintillator to a plenoptic camera — a micro-lens array sitting in front of a single-photon avalanche diode (SPAD) sensor called SwissSPAD2 — which records not just how much light arrives, but the direction it came from. Each micro-lens acts as a tiny camera capturing a slightly different view of the same scintillation flash; combining those views lets the system reconstruct the full light field and triangulate the flash’s 3D origin.
The AI does the heavy lifting on sparse data. Because only a handful of photons might be detected from a faint event, the team trained a transformer network — the same architectural family behind large language models — to process the entire photon pattern at once via self-attention, learning that a photon landing at one point on the sensor is statistically linked to photons landing elsewhere in a pattern consistent with a specific 3D origin. That global, all-at-once view, rather than a channel-by-channel scan, is what lets the system reconstruct particle tracks from photon-starved data. Simulations suggest the prototype could match or exceed today’s best detectors while being dramatically easier to scale, and the ETH team — led by researchers including Sgalaberna, Dieminger, and Alonso-Monsalve — has already filed three patents adapting the concept for positron emission tomography, covering both the scanner design and the neural-network-based image reconstruction.
Cross-Domain Connection
The tempting move is to slide sideways from PET scanners to another medical imaging problem: restoring sight to blind patients through retinal or cortical visual prostheses, which also depend on turning sparse, indirect signals into usable spatial information via modern AI. And there is a real methodological kinship here, worth naming precisely. Visual prostheses — devices that electrically stimulate surviving retinal or cortical cells to produce points of light called phosphenes — have, in just the past few years, moved toward exactly the kind of end-to-end, learned, physically-grounded modeling that PLATON exemplifies. Researchers have built differentiable phosphene simulators that let a neural network optimize stimulation patterns directly against a biologically realistic model of what a patient will actually perceive, and deep reinforcement learning approaches that search the enormous space of possible electrode stimulation patterns to find ones that render a target image as faithfully as a handful of phosphenes allow. Both PLATON and modern prosthetic vision research, in other words, are abandoning hand-engineered signal-processing pipelines in favor of networks trained against a model of the underlying physical or biological process.
But the honest version of this connection has to acknowledge that the two problems run in opposite directions. PLATON is a decoding problem with a clean answer: given rich photon data, infer where a well-defined physical event actually happened. It’s an inverse problem with ground truth available from simulation, which is exactly what makes it tractable for supervised transformer training. Prosthetic vision is, in a real sense, the reverse: an encoding problem. The device starts with a rich, high-resolution camera image of the world and has to compress it down into a tiny handful of electrode signals — then hand that impoverished signal to a damaged visual system whose exact response can’t be directly observed or fully modeled, only approximated. The genuinely hard, unsolved part of visual prosthetics isn’t reconstructing a signal computationally; it’s predicting how a specific patient’s remaining retinal or cortical neurons will interpret sparse stimulation, a perceptual and neuroscientific problem far softer than “which voxel did this photon pattern come from.”
What Remains Undemonstrated
No published research connects PLATON specifically to visual prosthetics, and there’s no evidence anyone in either field is aware of the other’s work. The real, concretely announced medical translation of this technology is PET imaging, not vision restoration — that distinction matters and shouldn’t be blurred. The broader claim this piece can honestly support is narrower than the original spark suggested: both fields are riding the same general wave in computational imaging, replacing dense physical instrumentation with algorithmic inference trained against a model of the underlying process, but the specific technique, the direction of inference, and the nature of the “ground truth” available are all different enough that a direct transfer between them is speculative at best, and nobody has proposed it in print.
Why It Matters
Naming that shared wave still matters, even without a direct technology transfer. Across wildly different domains, expensive physical resolution — millions of scintillator segments, dense grids of stimulating electrodes — is being partially substituted with algorithmic inference, because modern neural networks are good enough at exploiting statistical structure in sparse data to make that trade worthwhile. Recognizing the pattern helps calibrate expectations in both directions: PLATON’s transformer approach is powerful because its target problem is well-posed, and prosthetic-vision researchers reaching for similarly sophisticated deep learning tools are, correctly, tackling a harder and less well-posed problem that likely needs different tools, not the same ones borrowed wholesale.
Human Dimension
There’s something quietly striking about a research group in a Swiss physics department filing patents for a camera meant to catch neutrinos, and, a continent-spanning field away, another set of researchers training networks to help a blind person recognize a doorway from a dozen points of light. Neither group is thinking about the other. But both are making the same fundamental bet: that in 2026, it’s often smarter to let an algorithm infer the missing picture than to keep building more physical hardware to capture it directly. That bet is paying off differently in each field, for reasons specific to what each is trying to see.
Sources:
1. ETH Zurich, Department of Physics — “Neutrinos caught on camera: Testing the first prototype of a new elementary particle detector” — https://www.phys.ethz.ch/news-and-events/d-phys-news/2026/04/neutrinos-caught-on-camera.html
2. Phys.org — “Neutrinos caught on camera: Testing the first prototype of a new elementary particle detector” — https://phys.org/news/2026-04-neutrinos-caught-camera-prototype-elementary.html
3. ScienceDaily — “Scientists built a camera that can track invisible particles in 3D” — https://www.sciencedaily.com/releases/2026/07/260716023610.htm
4. Tech Times — “ETH Zurich Cuts Neutrino Detector From Millions of Optical Fibers to One Block and an AI Camera” — https://www.techtimes.com/articles/321013/20260720/eth-zurich-cuts-neutrino-detector-millions-optical-fibers-one-block-ai-camera.htm
5. SciTechDaily — “Seeing the Invisible: Scientists Develop New Way To Track Particles in 3D” — https://scitechdaily.com/seeing-the-invisible-scientists-develop-new-way-to-track-particles-in-3d/
6. eLife — “Towards biologically plausible phosphene simulation for the differentiable optimization of visual cortical prostheses” — https://elifesciences.org/articles/85812
7. arXiv — “Deep Learning-Based Perceptual Stimulus Encoder for Bionic Vision” — https://arxiv.org/pdf/2203.05604
8. arXiv — “Learning to See via Epiretinal Implant Stimulation in silico with Model-Based Deep Reinforcement Learning” — https://arxiv.org/html/2606.03118
Idea originated at artificialideas.org. Article researched and written by Claude Sonnet 5. Published at artificialideas.org.
Artificial Ideas 