Physical AI at the Sensing Edge: A Neuromorphic Camera That Sorts Microparticles Without Inference

A Stream of Light, Not a Stack of Frames
Inside a glass-bottomed microfluidic slide no thicker than a coin, a swarm of colored plastic fragments tumbles through a channel at speeds reaching eighty millimeters per second. Some are barely wider than a human hair. They spin, they wobble, and they ride invisible eddies stirred up by stray air bubbles. A conventional high-speed camera staring at this scene would smear most of them into faint streaks, then bury the operator under gigabytes of mostly empty pixels. Yet the sensor watching this particular channel stays almost silent. It fires only when something changes, registering each fleeting edge of each passing fragment as a tiny burst of asynchronous events. By the time a red fragment has crossed the field of view, the system already knows its size, its speed, and its color. No neural network ran. No frame was reconstructed. The answer was computed, in effect, by light, geometry, and a handful of photographic filters.
That is the central result reported by Joana Teixeira and colleagues at INESC TEC and the University of Porto. Rather than throwing a larger model at a harder sensing problem, the researchers moved the intelligence out of software and into the optics. The outcome is a microfluidic platform that classifies and tracks heterogeneous microparticles in real time while generating more than two hundred forty times less data than a comparable high-speed camera. For systems architects who have spent years cramming deep learning onto edge hardware, this is a notably different bet, and it is worth understanding exactly how it works.
Why Watching Fast, Small Things Is So Hard
To appreciate the result, consider first what conventional instruments are up against. Microfluidics underpins drug discovery, single-cell analysis, environmental screening, and much else, and nearly all of it depends on watching particles move through tiny channels. Standard CMOS and CCD cameras capture global snapshots at a fixed rate, so faster dynamics force a painful set of tradeoffs. Push the frame rate up to freeze motion, and the data volume balloons while exposure time shrinks and signal quality drops. Keep the frame rate modest, and fast particles blur into uselessness.
Engineers have tried many workarounds. Stroboscopic illumination freezes motion with short, intense light pulses and has reached throughputs above fifty thousand cells per second, though it demands precise timing and strong light sources. Microlens arrays and multichannel optics widen the field of view at the cost of system complexity. More recently, deep learning has been used to sharpen or rebuild low-resolution images, yet that approach adds heavy computation and can distort the very spectral and shape information the instrument was meant to preserve. In every case, the underlying architecture stays frame based, and the frame remains the bottleneck.
How Event-Based Sensing Rewrites the Rules
Event-based sensing, also called neuromorphic vision, takes a fundamentally different stance. Instead of capturing whole frames on a clock, each pixel acts on its own. A pixel watches the light landing on it, encodes that brightness on a logarithmic scale, and stays quiet until the change crosses a set threshold. Only then does it emit an event, defined by its location, a microsecond-resolution timestamp, and a single bit of polarity that records whether the light grew brighter or dimmer.
The consequences are considerable. Static background produces no output at all, so the sensor stops wasting bandwidth on parts of the scene that are not moving. Regions of rapid motion, by contrast, generate dense bursts of data exactly where the interesting physics is happening. This gives event cameras microsecond temporal resolution, a wide dynamic range, and a much lighter memory and computing load than frame-based rivals. Prior work had already shown event cameras tracking fluid flows at velocities up to 1.7 meters per second and sizing particles with neural networks. What had not been shown was the extraction of color, or spectral identity, from opaque flowing particles inside a microfluidic channel. That gap is what this work fills.
Encoding Color in Space Rather Than in Software
Here is the clever move. Event cameras report change, not color, so on their own they cannot tell a red fragment from a blue one. The researchers solved this by placing a spatially multiplexed RGB filter mask directly in front of the event sensor. The mask divides the sensor into three adjacent regions, each covered by a different off-the-shelf Lee photographic filter, namely Tokyo Blue, Primary Green, and Marius Red, chosen because their transmission bands barely overlap. White LEDs light the channel from above, and the system collects the backscattered light, so each particle responds according to its own surface reflectance rather than merely casting a shadow.
The effect is that color becomes position. When a fragment whose reflectance matches the green band passes beneath the green region of the mask, it triggers a dense cluster of events there and far fewer events elsewhere. Spectral identity and motion are therefore encoded at the moment of sensing, written straight into the geography of the event stream. Because the color information lives in where the events land, no learned model is needed to decode it. This is the heart of the zero-inference idea.
Best-of-Frames: Classification Without a Network
To turn raw events into a color call, the authors first slice the asynchronous stream into short frame-like windows, testing rates of 500, 1000, and 1500 hertz, then sum events into 10 by 10 spatial bins to lift the signal above the noise. For each frame, the algorithm sums the cleaned event activity falling within the red, green, and blue mask regions, producing three running scores.
Crucially, a fast, tumbling particle does not peak in all three regions at the same instant, since it crosses different parts of the physical sensor at different moments. So rather than judging a single frame or averaging over the whole transit, the method performs an independent temporal search for each color, recording the highest score that band ever reaches during the passage. The final call is simply the band with the tallest peak. The authors name this the Best-of-Frames strategy, and its appeal is that it always grades each color on its single best moment, regardless of when that moment arrived. Better still, keeping the runner-up scores opens a second-place analysis, which becomes useful for particles that show two strong spectral peaks at once.
A 3 by 3 Kernel Instead of a GPU
Cleaning the signal is where the no-network philosophy shows its teeth. Real event streams are speckled with stochastic noise, single isolated firings that mean nothing. The researchers exploit a simple physical truth, which is that genuine particle activity clusters in space while noise tends to stand alone. To separate the two, they convolve a binary map of active pixels with a small 3 by 3 kernel that counts each pixel’s eight neighbors.
K =
1 1 1
1 0 1
1 1 1
A pixel survives only if its own intensity clears a threshold and it has at least one active neighbor. Anything more isolated is treated as noise and zeroed out. This single convolution, cheap enough to run on the most modest hardware, scrubs the event trajectories clean so that color decisions rest on dense, physically real clusters rather than background flicker. There is no training set, no inference pass, and no graphics processor in the loop. A few arithmetic operations per pixel do the work that an industry reflex would have handed to a deep model.
Reading Size and Speed From the Same Stream
The platform extracts kinematics from the same events. For tracking, the method computes an intensity-weighted center of mass for each frame, then links those centroids across frames using a nearest-neighbor rule that tolerates a few missed detections when a particle dims or drifts out of focus. The track that survives the most frames wins, since noise tracks rarely last more than a frame or two. Velocity then follows from the displacement between consecutive centroids, scaled by the calibrated pixel pitch and the frame integration time.
Sizing leans on both positive and negative events to capture the full boundary of each fragment. After a morphological closing step fills small internal gaps, the software fits an equivalent ellipse using a weighted spatial covariance matrix, with a Gaussian term that favors the dense core over noisy edges. Because particles tumble and the major axis can stretch along the direction of motion, the authors take the minor axis as the more stable estimate of true diameter. Across the experiments, most fragments measured between 0.05 and 0.30 millimeters and traveled between roughly 3 and 80 millimeters per second.
What the Numbers Show
The researchers deliberately worked under messy, application-relevant conditions. They abraded colored thermoplastics into a population of irregular fragments spanning a wide size range, suspended them in water, and pushed them through the channel with a syringe pump while air bubbles and 3D-printed couplers stirred up non-laminar flow. These are exactly the conditions that defeat frame-based instruments.
Under that stress, the frame construction rate proved central. At 500 hertz, the long integration window let particles drift far within a single frame, mixing colors and dragging global accuracy down to 60 percent. At 1000 hertz the system struck the best balance between temporal resolution and event density, reaching 74 percent global color accuracy, while 1500 hertz edged slightly lower at 71 percent because shorter windows gathered fewer events per frame. Performance was strongest on smaller fragments, where color accuracy reached about 82 percent in the 0.08 to 0.18 millimeter band, then fell to 66 percent for the 0.18 to 0.25 millimeter range as larger footprints began to straddle two filter regions at once.
The data story is even more striking. A 1280 by 1024 CMOS camera running at 1000 frames per second would spew roughly 3.84 gigabytes per second, most of it redundant background. The event sensor, recording a continuous 100-second stream, produced about 16 megabytes per second, a reduction of more than two hundred forty times, and it did so under a deliberately high-sensitivity bias tuned to catch even faint, out-of-focus particles. On monitoring capacity, the platform held an area throughput of 460 square millimeters per second at full resolution, against 13.8 for a 30 frames-per-second CMOS camera and 32 for a 50-kilohertz line-scan sensor once the cost of rebuilding a two-dimensional image is paid. Snapshot multispectral cameras, meanwhile, stall near 100 hertz and record every voxel of a spectral cube whether or not a particle is present.
Where the Method Bends, Honestly
The authors are candid about the soft spots. Larger and faster particles smear across filter boundaries and confuse the classifier, since a wide footprint can light up two color masks at once. The reflection spectra of the red and blue fragments also overlap at the wavelengths near the seam between bands, which seeds some residual red-and-blue confusion no matter how clean the events are. For these harder cases the group offers two routes. At the hardware level, multi-wavelength LEDs matched to the target analytes would sharpen the spectral separation. At the algorithm level, a multi-rank logic that scores the top two color bands against each other can resolve genuine dual-peak signatures, a refinement the authors explore in a dedicated experiment on particles carrying both green and blue peaks.
Read together, these points reinforce the proof-of-concept framing. The reported numbers came from off-the-shelf photographic filters, broad white light, and irregular particles in turbulent flow, so they should be read as a conservative floor rather than a ceiling.
Where This Could Go: Four Concrete Deployments
Because the architecture is lean on data, light on computation, and built from commodity parts, it suits settings where conventional optics or cloud-scale compute cannot follow. Four applications stand out.
Field-Deployable Microplastic Screening
Today, characterizing microplastics in water usually means shipping samples back to a lab for slow Fourier-transform infrared spectroscopy or energy-intensive high-speed microscopy. This platform points toward a different model. Since it runs on a low-power event sensor, a cheap photographic filter array, and ordinary white LEDs, it could fit inside a ruggedized buoy or a hand-held probe. Such a device could count, size, and sort plastic fragments by color and reflectance right on a riverbank, then stream compact summaries over standard IoT networks without ever touching cloud compute.
Label-Free Point-of-Care Hematology and Flow Cytometry
Conventional high-speed flow cytometry leans on fluorescent labels or costly high-bandwidth cameras that pour out gigabytes per second. By swapping the standard RGB mask for a custom narrow-band multispectral set, a clinical instrument could sort cellular variants or hematological components by their intrinsic backscattered signatures alone. The sub-millisecond timing and the steep data reduction together suggest a compact bedside unit that screens large blood volumes in minutes, which matters for tasks such as flagging circulating tumor cells or pathogens at low cost.
Industrial Slurry and Process Monitoring
In milling, crystallization, and polymer production, holding particle size distributions within spec is what keeps a line from clogging or failing. The fact that this platform held its tracking and sizing together under turbulent flow laced with bubbles is the relevant proof here. An inline microfluidic bypass built on the same neuromorphic core could watch a process stream continuously, then flag unexpected aggregation or contamination with sub-millisecond latency, alerting operators before a fault propagates downstream.
Smart Sorting for Agriculture and Mining
Scaling machine vision down to micro-grains, seeds, or pulverized mineral fragments runs straight into frame-rate bottlenecks. The multi-rank prediction logic offers a way through. By scoring the top two spectral bands to catch dual-peak signatures, a microfluidic or conveyor-based sorter could pull high-value targets out of a waste stream on the fly, judging each fragment on both its optical and its structural fingerprint at high throughput.
A Conservative Baseline for Physical AI
What makes this work matter to a broad audience is the principle underneath it. The industry reflex is to treat every classification problem as a software problem and to grow the model until it fits the data. Teixeira and colleagues went the other way, encoding the decision into the optics so that a 3 by 3 kernel and a maximum-likelihood comparison could finish the job with no inference at all. The result handles polydisperse, irregular, real-world particles under non-ideal flow, which is precisely the regime where frame-based cameras give up.
There is room to grow, and the authors say so plainly, pointing toward purpose-built filter sets, tailored illumination, and higher-dimensional spectral encoding. Yet the proof already stands. By moving computation to the sensing stage and letting physics do the sorting, this neuromorphic platform shows that smarter optics can sometimes replace bigger models, and it does so in a form small enough, frugal enough, and fast enough to leave the lab and go to work in the field.
Reference
Teixeira, J., Lopes, T., Ferreira, T. D., Monteiro, C. S., Jorge, P. A. S., and Silva, N. A. (2026). Towards Neuromorphic Event-Based Sensing for High-Speed Multi-Spectral Classification and Tracking of Microparticles. arXiv:2605.31038v1 [physics.optics]. Centre for Applied Photonics, INESC TEC, and Departamento de Física e Astronomia, Faculdade de Ciências, Universidade do Porto.
