Semiconductor & Industrial Imaging

Every structure, inspected.
Every result, proven.

OptiFlow is a domain-aware image-analytics platform for semiconductor and industrial imaging — from SEM, FIB-SEM and TEM to X-ray, optical and infrared inspection. Build the analysis visually, find what manual review misses, and run it from a single image to a full batch or a live feed — exporting results that carry the exact method that made them. No code, air-gapped, 100% local processing.

In production at ETH Zürich Air-gapped & 100% local processing 20+ peer-reviewed publications Ongoing trials at NTU
The inspection gap

The defect is already in your images.
Catching every one is the hard part.

Semiconductor and industrial imaging has outpaced the tools used to inspect it. Datasets from electron microscopes and high-resolution inspection lines are far beyond what anyone can review by hand; off-the-shelf AI was never trained on this kind of imagery; and an inspection decision that can't be reproduced or audited can't be trusted where it matters most.

Defects beyond manual review

Production-resolution EM and inspection datasets are too large to examine exhaustively, yet too critical to sample — the one anomaly that slips through is the one that reaches a customer.

Generic AI breaks on domain images

The noise statistics of an electron micrograph differ fundamentally from those of a natural photograph. Models trained on everyday images mis-segment, over-flag, and miss the subtle, domain-specific signatures that matter.

Decisions you can't reproduce

When results shift between operators, stations and shifts, they can't be traced, defended, or relied upon — a critical gap in regulated manufacturing and security-sensitive work.

Who we serve

One platform.
Built around your inspection.

Semiconductor and industrial imaging spans many different teams — from hardware-assurance labs to process engineering and R&D — yet they share one need: inspection they can trust, repeat, and defend. OptiFlow adapts to each, rather than forcing your work into a generic mould.

Hardware assurance & security

Verify integrity at component resolution — with or without a trusted reference.
  • Reference-free anomaly & hardware-Trojan detection
  • Golden-reference registration & comparison
  • Counterfeit & supply-chain screening

Failure analysis & defect review

Characterise and classify defects across electron-microscopy modalities.
  • Defect characterisation & classification
  • SEM / FIB-SEM / TEM workflows
  • Reproducible, shareable FA reports

Process & inline defect detection

Catch known defects and surface novel failure modes at production throughput.
  • Supervised detection on your defect library
  • Unsupervised discovery — no library needed
  • Drift & excursion monitoring

Metrology & dimensional inspection

Turn pixels into calibrated measurements in real, traceable units.
  • Instance segmentation & boundary delineation
  • Feature & critical-dimension measurement
  • Per-feature statistics & distributions

Research & development

Device and materials imaging, analysed reproducibly — without scripting.
  • EM & X-ray analytics, no code required
  • Methods you can publish and re-run
  • Models trained on your own data

A different problem?

Industrial machine vision and beyond — if your imaging challenge isn't here, it's likely still a fit.
Signature capability

Catch the component that doesn't belong.
No golden reference required.

OptiFlow flags anomalous and tampered components at single-component resolution using self- and unsupervised learning — it learns what normal looks like from your own imagery, so it surfaces the unexpected even when no trusted reference device exists. Where a golden reference is available, multi-resolution registration and comparison add a second, complementary line of detection.

Preparing live demo…
From your own field — every component of interest located, characterised, then gated to a handful of suspects. No golden reference; no labelled defects.
  • No golden reference needed

    Detect anomalies without a known-good device — built for new designs, legacy parts, and counterfeit screening where no trusted original exists.

  • Single-component resolution

    Localise the specific component or region that deviates — not just a whole-image pass or fail — so review goes straight to what matters.

  • Self- and unsupervised — no labels

    Learns normal appearance directly from your own images and flags novel deviations that no labelled training set could anticipate.

  • Golden-reference comparison, when you have one

    Multi-resolution registration aligns a device to its reference and surfaces differences — a complementary check whenever a trusted reference is available.

  • Triage hundreds to a handful

    Every component is scored and ranked, so expert review goes to the few that matter — not the entire field.

Ongoing trials at NTU
The semiconductor toolkit

Every step, from raw frame
to the verdict.

Compose exactly the pipeline your inspection needs from a deep library of domain-aware operations — purpose-built for industrial imaging. Preview any of them on your own images, then license only what your work demands.

Modalities covered SEMFIB-SEMTEMSTEMX-ray / CTOptical & brightfieldDark-fieldInfraredMulti-channel

Restore & enhance

Recover signal from difficult images.

Denoising that is classical, AI-guided, and hybrid — NLM, bilateral, wavelet, total-variation and BM3D for low-SNR SEM and TEM data, preserving nanoscale structure. Flat-field, illumination and glare correction, plus learned deconvolution to recover defocused or distorted features.

Expanded

Register & compare

Align to a reference, surface the difference.

Multi-resolution registration aligns a device to a golden reference across fields and scales; difference mapping highlights deviations for review — the reference-based half of hardware assurance.

New

Detect & localise

Find and outline every feature.

Supervised object detection and instance segmentation — precise localisation and boundary delineation of defects and individual components across every imaging modality.

Anomaly & novelty detection

Find what no library contains.

Self- and unsupervised detection learns normal appearance and flags deviations at component resolution — surfacing novel failure modes and tampering with no labelled defects required.

New

Measure & characterise

Quantify in real units.

Calibrated area, distance and feature measurement read from your metadata, plus per-component morphometrics and texture analysis — with batch normalisation built in for comparable results across runs.

Search, monitor & report

From one image to the whole bank.

Search large image banks for a feature or pattern; live monitoring with action triggers — pause acquisition and notify on a flagged event; then statistics, plotting, and automatic export to Excel and auditable PDF reports.

New

Plus the building blocks every pipeline leans on

Channel and colour operations, lookup tables, geometry and alignment, type conversion, and pixel-wise image maths and compositing — the connective tissue that lets any workflow fit together exactly as your inspection requires.

Modular by design — start with what today's inspection needs and add modules as your work grows, instrument by instrument and line by line. See the full OptiFlow platform.

Reproducible & auditable

Inspections you can defend.
Records that travel with the image.

In hardware assurance and regulated manufacturing, an inspection is only as good as the trail behind it. OptiFlow records the full method for you — so every result is reproducible, every method is portable, and your data never leaves your control.

Embedded, tamper-evident provenance

Every result carries its full recipe.

The complete pipeline — every operation, parameter, calibration and operator — is written into the image itself, encrypted and readable only inside OptiFlow. An encrypted, rollback-resistant timestamp guards the record; where it matters, lock it to a password only your team holds.

An exportable method, on demand

Share it, re-run it, hand it over.

Export the complete pipeline as a standalone file — to hand to a partner lab, attach to a report, or re-run across a batch. Anyone you give it to can reproduce the exact inspection, down to the parameter.

Air-gapped & 100% local processing

No image ever leaves the machine.

Processing reads and writes your own filesystem directly. No image data crosses the network — no cloud account, no exceptions — so sensitive designs and security-critical hardware stay exactly where they belong.

See it: the exact method, recovered from any result.

Every result already contains the workflow that made it — so revisiting an inspection, handing it to a colleague, satisfying an auditor, or re-running across a batch needs nothing rebuilt and nothing re-documented. Drop in a finished result below and watch its full method reappear.

Preparing live demo…
From research to production

Built on science.
Engineered for production.

Logic Byte's semiconductor and electron-microscopy capabilities trace directly to world-class research and to systems already in production.

Computer vision · inspection
Wafer inspection & defect detection
Research on automated detection and localisation of defects in semiconductor imaging — the groundwork for the supervised and unsupervised detection modules.
Autonomous imaging
Reinforcement learning for adaptive imaging-device control
Multi-parameter auto-tuning of imaging instruments via reinforcement learning — the basis for diagnostic agents that optimise acquisition and processing for each image class.
Vision generative AI · ETH Zürich
Generative AI on a commercial instrument
Domain-aware generative imaging deployed commercially on a production optical instrument — evidence that the platform performs in real, demanding settings.
How we work

Start on your own images.
Scale when it earns it.

No rip-and-replace, no leap of faith. Prove the value on your own data first, then grow module by module, line by line.

Explore Deploy Expand Embed
Preview before you buy 60-day module guarantee Priority support, included Pay only for what you use

See it on your imaging data.

Bring a semiconductor or industrial imaging problem and your own data — we'll show you exactly what OptiFlow does with them. If your inspection needs something we haven't built yet, we'll build it.

In production at ETH Zürich Ongoing trials at NTU 20+ peer-reviewed publications Featured by CNA & The Straits Times