Product 02 · Zentari Nano

Carrier design across models, motion and flow.

Candidate manifests move from fast analytical references through Brownian and Stokesian dynamics, optimisation and controlled CFD without losing their evidence lineage.

Analytic Stokes surface traction on an undisturbed carrier and illustrative deformed response — FIG 03 **Analytic Stokes field · illustrative response**`main · 1cacc2d · 14 µm radius · 160 nm slip length` Analytic creeping-flow reference; deformation is not OpenFOAM, solved FSI, hardware or laboratory evidence.
Open evidence →

Engineering stateDeterministic carrier and simulation pipeline

Scientific evidenceL2 in-silico internal benchmark

Next gateMeasured rig data and sim-to-real reconciliation

A score is useful only when its physics and provenance remain inspectable.

Workflow

Fast exploration, controlled escalation.

Baseline methods remain available while challengers earn promotion through real-data benchmark reports.

01Specify
Carrier geometry, payload, actuation and fluid context.
02Evaluate
Analytic Stokes and BD/SD establish fast references.
03Search
NSGA-II, MAP-Elites, ParEGO or qNEHVI behind a common seam.
04Challenge
Selected cases escalate to a governed OpenFOAM runner.
05Validate
Convergence, physics QC and non-convergent demotion.
06Retain
Manifest, fields, Pareto set and benchmark evidence.

Actual results

A field view and an internal reference floor.

The hero is analytic, not CFD. The benchmark is simulated, not measured.

Internal MicroCarrierBench plot comparing magnetic, acoustic and passive reference carriers across four tasks — FIG 04 **MicroCarrierBench v0.1 · internal L2 benchmark**`zen-magnetic 88.56 · zen-acoustic 88.22 · passive floor 2.44` Every entry is unmeasured and in-silico; no external benchmark or sim-to-real gap is available.
Open evidence →

Magnetic reference: 88.56
Acoustic reference: 88.22
Passive floor: 2.44
Reference radius: 14 µm
Slip length: 160 nm
Measured entries: 0

Methods

A validation-gated learning stack.

Each stage has a default baseline and a challenger that cannot become default on name or novelty alone.

Forward models

Analytic Stokes, physics-aware test doubles, real OpenFOAM runners and uncertainty-aware surrogate interfaces.

Figure 03 uses the analytic path.

Stochastic motion

Brownian and Stokesian dynamics support trajectory and mobility evidence.

Model assumptions remain in-silico until a rig closes the gap.

Optimisation

Multi-objective, quality-diversity and Bayesian searchers share benchmark and promotion records.

Search finds candidates; it does not validate them.

Active learning

Random, uncertainty and diverse-uncertainty acquisition strategies can target expensive labels.

Real OpenFOAM or measured labels are required for promotion.