Leakage-Calibrated Counting of Rare Aggregates in Flowing Nanoparticle Streams

When nanomedicines move from a 10 mL vial to continuous production, the failure that matters clinically is often not the average particle. It is the rare one. A clump of hundreds of nanometres occurring once in ten thousand particles can block capillaries or trigger immune clearance. Yet the inline monitors used in automated synthesis report bulk averages over billions of particles, which hide that clump completely. Rare-event physics has spent decades on the mirror-image problem: how to count a handful of true signals inside a flood of look-alike background, and how to prove how often background leaks through. This note proposes to bring that measurement discipline, rather than any single detector trick, to nanoparticle manufacturing.

Concretely, we would divert a slowed side-stream of a continuous synthesis line through a wide inspection chamber watched by an event camera. An event camera is a sensor whose pixels independently report brightness changes with microsecond timing, instead of capturing whole frames. For every particle that crosses the chamber, we compute two independent size estimates. One comes from how brightly it scatters light. The other comes from how much it jitters sideways by Brownian motion once the known flow profile is subtracted. True aggregates are large on both. Bubbles, dust, wall glints and two monomers crossing at once disagree between the two. The discriminator is then calibrated the way DEAP-3600 calibrated its pulse-shape cut: with spiked "signal" samples and background-only samples. The output is an upper limit on the defect fraction in production with a stated confidence level and a measured leakage rate, which no current nanomanufacturing monitor provides.

Core Claim

Requiring an optical-size and a diffusive-size estimate of the same particle to agree, measured in flow with the flow advection removed by a physical model, gives a rare-aggregate discriminator. Its leakage of monomers and optical artifacts can be measured below $10^{-4}$ at useful signal acceptance. That is enough to place 90% confidence upper limits on aggregate fractions that change measurably with mixing conditions, while ensemble absorbance does not change.

The novelty is not "event cameras for nanoparticles." Nanoparticle Diameter Measurements With Event Camera Tracking already showed that event cameras size particles in quiescent liquid. On a 51.4 ± 4.1 nm gold reference they reported a median of 53.1 nm and a robust coefficient of variation of 19.8% at 80× magnification, against 55.7 nm and 21.8% for a frame camera. That work measured average size in still samples and made no attempt to classify rare particles or quantify misclassification. We move the measurement into flow, we target the tail of the distribution rather than its median, and we add the calibration and statistical-limit machinery of rare-event searches.

Why the physics analogy is the right one

The method borrows three ideas from rare-event physics. The first is the two-observable consistency cut. In the MAJORANA germanium detectors, α-event characterization and rejection in point-contact HPGe detectors combined a delayed-charge-recovery variable with an amplitude-over-energy variable. Because surface alphas fail one test that bulk events pass, the combination rejected all surface alphas beyond 3 mm while losing only 0.2% of bulk events. Our brightness-versus-diffusion agreement plays the same role. Scattered intensity depends steeply on size but also on refractive index, so a bright bubble or a metallic dust grain can mimic an aggregate. Brownian spread depends only on hydrodynamic size and fluid viscosity. Each observable alone is fooled by some background; their disagreement is the signature of background.

The second idea is diffusion broadening as a position and size handle. Topological background discrimination in the PandaX-III neutrinoless double beta decay experiment used the diffusion-induced widening of drifting charge to reject uranium surface events, cutting that background to a factor of 0.15 while keeping 75% of signal. That work showed a cheap 2D strip readout could match 3D pixels once diffusion was modelled. We use transverse Brownian spread in the same spirit: a physically modelled broadening that a simple sensor can read out.

The third idea, and the most important, is calibrated leakage. First Results from the DEAP-3600 Dark Matter Search with Argon at SNOLAB reports its pulse-shape cut as a measured electronic-recoil leakage below $1.2 \times 10^{-7}$ at 90% confidence in a stated energy window. That number came from dedicated background calibration, not from training accuracy. Nanomanufacturing QC has no equivalent statement, and that absence is the real gap.

For the diffusive size, consider a particle with residence time $\tau$ in the chamber. After subtracting the expected Poiseuille trajectory at its lateral position, let $\langle \delta y^2 \rangle$ be its mean squared transverse displacement. With temperature $T$, viscosity $\eta$ and Boltzmann constant $k_B$, the Stokes–Einstein relation gives

$$ d_D = \frac{2 k_B T \, \tau}{3 \pi \eta \, \langle \delta y^2 \rangle}. $$

The per-particle discrimination statistic compares this with the optical size $d_O$ inferred from integrated event count via a calibration curve, $\Delta = \ln d_O - \ln d_D$. We classify a particle as an aggregate when both sizes exceed a threshold and $|\Delta|$ lies inside a band set on calibration data.

A feasibility estimate sets the chamber design. In water, a 50 nm particle has a diffusion coefficient near $9 \times 10^{-12}$ m²/s, so it spreads about 130 nm in 1 ms and about 400 nm in 10 ms. A 500 nm aggregate spreads about three times less. Separating these with dark-field localization at tens of nanometres needs residence times of several milliseconds. This is why we inspect a decelerated side-stream rather than the production channel itself.

Research Program

The work has three stages. The first builds the inspection module: a split-off line from the synthesis outlet feeding a widened, shallow chamber where velocity falls to the millimetre-per-second range. The chamber is imaged in dark field onto an event camera. A beam splitter also sends the light to a frame camera, reproducing the dual-sensor arrangement of Daugherty et al. so the comparison is like-for-like. The flow profile in the chamber is computed from channel geometry and the measured flow rate, and is checked with tracer trajectories. Advection subtraction is therefore a physical model with no learned parameters. This follows the finding of Optimization of microfluidic synthesis of silver nanoparticles: a generic approach using machine learning that describing a reactor by Reynolds and Dean numbers rather than raw pump rates makes models transfer across conditions.

The second stage is calibration, organised as in a dark-matter experiment. "Signal" samples are the 51.4 nm gold reference spiked with larger gold or polystyrene particles in the 200–500 nm range at prepared ratios from $10^{-3}$ down to $10^{-5}$. "Monomer-background" samples are the reference material alone; these measure how often ordinary particles and coincident pairs leak into the aggregate class. "Artifact-background" samples are filtered particle-free buffer with deliberately induced microbubbles, seeded dust, and wall-fouled chips. From these we measure acceptance, the two leakage fractions, and the linearity of measured versus prepared spike ratio. Cut thresholds are frozen on one half of the calibration runs before the other half is examined, the blind-analysis habit of the physics experiments.

The third stage turns to production. We take the T-junction silver nanoparticle synthesis of the Dean-number study and deliberately vary the Dean-to-Reynolds ratio. That study found that secondary Dean vortices narrow the size distribution, so lowering the ratio should broaden the tail. This gives a set of controlled "good" and "worse" runs. In each run we report three things together: the inline absorbance spectrum, the median size, and a 90% confidence upper limit or measured value for the aggregate fraction. As Tingting Zhang et al. and Yue Ke et al. argue, early mixing variations propagate nonlinearly into batch quality, and the tail is where we expect that propagation to show first.

We deliberately do not promise closed-loop sorting of individual particles. The decision latency of the per-particle statistic is reported on a CPU. EAP4EMSIG - Enhancing Event-Driven Microscopy for Microfluidic Single-Cell Analysis shows that closed-loop microfluidic triggers built on lightweight, physically motivated features are workable at the 87–121 ms scale. That scale suits a batch-level alarm or diversion valve, which is the realistic first use.

Baselines

The naive reference is the monitor current platforms already use: inline UV-Vis absorbance, as in Two-step machine learning enables optimized nanoparticle synthesis, where a spectrophotometer steered silver nanoprism synthesis to a target spectrum within 120 sampled conditions. We record it on every calibration and production run. The proposal beats it if spike ratios of $10^{-4}$ produce a detected aggregate excess while the absorbance spectrum stays within its run-to-run noise.

The simplest single-particle competitor is a brightness threshold on the same event stream: call any particle whose integrated event count exceeds the aggregate calibration point an aggregate. It uses identical data, the same flow-subtraction code, and a threshold tuned on the same calibration half. This is the baseline most likely to be "good enough," because scattering grows so steeply with size. Beating it means lower artifact-background leakage at equal acceptance, by at least a factor of ten, which is what the two-variable cut in the MAJORANA work achieved over single-variable cuts.

The strongest existing sizing method is event-camera tracking as in Nanoparticle Diameter Measurements With Event Camera Tracking, applied offline to grab samples in a static chamber. Its reported figure on the reference material is a 53.1 nm median and 19.8% robust CV at 80×. For rare-aggregate detection in flow or at stated leakage, there is no reported result on Lacuna, so we must measure it ourselves. We will reproduce their static numbers on our hardware first. The in-flow system should match them within the reference's stated uncertainty, and should reach a lower detectable aggregate fraction per minute of operation, because it inspects continuously rather than from grab samples.

The frame-camera arm runs the same two-observable pipeline on the frame sensor. It is cropped to the frame rate needed for millisecond residence times, so we report the field of view, and hence particles inspected per second, as part of the result. A learned classifier comes from Deep learning based pulse shape discrimination for germanium detectors, which showed networks can replace hand-built pulse-shape variables. Here it is a small network on per-particle event-stream patches, trained on the same calibration half under a matched tuning budget. Two ablations isolate the key ingredients: removing the diffusive size (leaving brightness plus transit shape), and replacing model-based flow subtraction with simple per-track detrending.

Why not just…?

Why not just threshold on brightness, since aggregates scatter far more light? This is a real threat for pure gold or silver product, where an aggregate is unambiguously brighter. It weakens in two cases. Bright artifacts such as bubbles and metallic dust share the high-brightness region. And for low-contrast carriers like lipid particles, an aggregate's brightness can overlap a dense contaminant's. The artifact-background calibration settles this directly. If brightness-only leakage on bubble and dust samples is already below $10^{-4}$, the diffusion observable is unnecessary for metal particles, and we would say so and restrict the claim to low-contrast materials.

Why not just use a high-speed frame camera, which is mature and cheap? Daugherty et al. showed the frame sensor must be cropped to reach kilohertz rates, and that it lost tracks of particles moving through the focal plane. In our setting, cropping shrinks the inspected volume, and rare-event sensitivity scales directly with particles inspected. The frame-camera arm runs the identical pipeline. If it reaches the same leakage and limit per minute, the event sensor is not needed, and that would be a useful negative result.

Why not just train a neural classifier on the event stream and skip the physics? A learned classifier may match the physics cut on the conditions it was trained on. The concern is transfer. EAP4EMSIG reports that its models degraded across chip geometries and illumination. The Dean-number study found that physically parameterised inputs raised predictive R² from 0.45 to 0.62–0.66. We therefore train both on one flow rate and chip and test both on a different flow rate and chamber depth. Leakage drift under that shift is the deciding number.

Why not just sample product offline into a commercial-style static tracking chamber? Offline sampling sees a small volume at one moment and cannot follow a run whose mixing drifts. The production experiment varies Dean number within a run. Grab-sample static tracking with the same sensor is run in parallel, so we can report whether it detects the tail change, and how late.

Proposed Experiments

The first decisive check is static agreement. On the 51.4 nm reference in a still chamber, our pipeline must reproduce a median and robust CV comparable to the published 53.1 nm and 19.8%. It must then hold those values in flow after advection subtraction. If the in-flow median shifts by more than the reference's stated uncertainty, the flow model is wrong, and nothing downstream is trustworthy.

The second check is the calibrated leakage curve. Using the frozen cuts on the held-out calibration half, we report acceptance for spiked aggregates and leakage for monomers, coincident pairs, bubbles and dust. The comparison covers the full method, brightness-only, the frame arm, the learned classifier and both ablations. Success means measured spike ratios track prepared ratios down to $10^{-4}$ with artifact leakage at least ten times below brightness-only at equal acceptance. Failure on either count would be reported as the central result.

The third check is the production test. Across T-junction syntheses at several Dean-to-Reynolds ratios, the aggregate fraction or its upper limit should rise as Dean mixing weakens, while absorbance spectra and median size stay statistically indistinguishable. If the tail moves together with the ensemble signal, single-particle monitoring adds nothing for this chemistry. If it moves alone, we will have shown that the defect that matters is visible only to a rare-event measurement.

The final check is transfer: freeze all cuts and models, change flow rate and chamber depth, and re-measure leakage. This is where the physics-based and learned discriminators are expected to diverge.