Machine vision solutions for detecting bubbles in optical epoxy layers
Air bubbles trapped inside cured epoxy encapsulants are one of the most persistent causes of optical failure in precision assemblies. Even a void smaller than 50 micrometres can scatter enough light to distort a sensor reading or scramble a measurement taken through the bonded interface. As lens stacks grow thinner and adhesive layers become part of the optical path rather than a mechanical afterthought, the tolerance for invisible inclusions has effectively shrunk to zero. Machine vision offers a way to find these voids before the part leaves the production cell, provided the imaging chain is designed around the way light actually behaves in a transparent resin.
Australian manufacturers working in medical devices, photonics, and mining instrumentation have already absorbed the cost of miscaught defects through warranty claims and field returns. Sydney and Melbourne both host clusters of firms that build optical sub-assemblies for export, and their quality teams routinely deal with the kind of high-mix, low-volume runs where manual inspection simply cannot keep up with the surface area being produced each shift. Brisbane and Adelaide add defence and aerospace work that demands traceable inspection records.
Regulatory pressure reinforces this technical need. The Therapeutic Goods Administration scrutinises optical components used in diagnostic instruments, and consumer-grade products must satisfy the guarantees of the Australian Consumer Law. Both regimes require documented evidence that a part has been inspected against a measurable criterion. A vision system that records every bonded layer against an automated void threshold provides exactly the kind of traceable record that auditors now expect during a routine visit.
The core engineering problem is deceptively simple. An epoxy layer cured against glass, metal, or another polymer will trap gas at three predictable points in the process: during dispensing, when the resin wets the substrate, and during the exothermic rise that drives cure shrinkage. Each of these entry points produces voids with a slightly different shape, location, and optical signature, which is why a single lighting recipe rarely catches everything that matters in the field.
How air bubbles form in optical epoxy encapsulants
Dispense dynamics create the first family of voids. When a needle deposits a resin bead onto a cool substrate, the air that was sitting on the surface cannot always escape ahead of the advancing fluid front. The result is a thin chain of flattened bubbles along the wetting line, often smaller than the resin film itself and almost impossible to catch with the naked eye.
The second family forms during cure. Epoxy systems undergo a small volumetric contraction as they cross from liquid to gel, and if the rheology of the bond has already started to lock up, the contracting mass will pull away from internal stress concentrators and leave behind lens-shaped voids. These voids tend to sit near the centre of the bond line, where the material cooled last and where shrinkage had the greatest distance over which to act.
Outgassing from the substrate produces the third family. Many optical cements and some anti-reflective coatings release small molecules during temperature ramp-up, and if the cure profile runs hot, those molecules nucleate inside the still-fluid resin. The bubbles they form tend to cluster near the substrate boundary, where the gas has the shortest path to travel before it is trapped, and they often appear as a faint haze rather than as discrete circles.
Each family needs its own detection strategy. A vision system calibrated only for surface bubbles will miss the central voids that form during cure. A system tuned only for the largest inclusions will miss the chains of micro-bubbles that follow the dispense bead. Reasonable coverage means pairing bright-field imaging with at least one cross-sectional or sub-surface technique, ideally on the same part and within the same cycle time, so that no defect family slips past unrecorded.
Imaging setups for sub-millimetre bubble detection
The first decision in any optical assembly cell is whether to inspect the part from above or from the side. Top-down cameras read the projected silhouette of the encapsulant layer and rely on illumination that penetrates the resin. Side cameras, mounted on a telecentric probe or on a low-angle mirror, read the bond line directly and see voids that hide beneath a glossy surface finish. The two geometries answer different questions and rarely substitute for each other on a critical part.
For thin bond lines below 200 micrometres, top-down imaging dominates. A 5-megapixel camera with a telecentric lens can resolve voids down to roughly 15 micrometres in a 50-millimetre field of view, provided the working distance is held within a few hundred micrometres. The bond line is in focus, the surrounding fixture is in focus, and software can subtract the fixture cleanly. The same camera can be reused across part families if a simple calibration routine is built into the changeover script.
Where bond lines are thicker, or where the part geometry blocks a straight top-down path, angled illumination does most of the work. A ring light offset from the optical axis will pick up the meniscus of any bubble that breaks the surface, and a dark-field ring placed further off-axis will show the side wall of any void that sits just below the surface. The two ring positions together cover both the visible and the just-buried voids without much extra cycle time.
For the deepest voids, time-domain approaches enter the picture. Structured light patterns projected onto the part can be analysed for distortion caused by subsurface refraction, and spectral techniques can be used to read the way each pixel of resin absorbs and scatters light. Confocal line-scan cameras, which reject out-of-focus light optically rather than by software, also perform well here. These methods are slower than simple bright-field imaging and demand more careful calibration, but they recover voids that other optical routes simply cannot reach.
Lighting strategies that expose subsurface voids
Bright-field transmission is the workhorse of bubble detection in transparent resins. A collimated backlight placed behind the part sends parallel rays through the bond line, and any void refracts a measurable fraction of those rays out of the camera aperture. The contrast is high and the setup is forgiving, which is why many lines in Brisbane and Adelaide begin with transmission and only add complexity when the part mix demands it.
Dark-field illumination does the opposite. Light enters the part from grazing angles that cannot reach the camera directly, so a clean bond line renders dark. A bubble, however, acts like a tiny lens and steers grazing light toward the camera, producing a bright point wherever a void sits. Dark field is exceptional at finding small, near-surface inclusions that bright-field would average away, and it pairs naturally with bright field on a single inspection station.
Coaxial or polarised setups add another dimension. Polarised light passing through a strained resin layer will rotate its polarisation in proportion to the local stress, and a void alters that stress field in a way that the polariser can reveal. Cross-polar imaging in particular suppresses glare from the upper surface and leaves only the buried structure visible. This approach is slower than the others but produces maps of mechanical integrity as well as void location, which can be useful when cure uniformity is a concern in its own right.
For finding paper-thin chains of dispense-bead bubbles, diffraction-based illumination has proven its worth. A laser line projected across the bond line at a low angle shows each chain as a bright streak against an otherwise uniform field. The method requires careful tuning of laser power to avoid heating the resin, but for high-precision optics, it remains the most sensitive way to read what happened at the dispense moment.
| Inspection method | Resolution typical | Cycle time per part | Best suited to | Main limitation |
|---|---|---|---|---|
| Bright-field transmission | 10–20 μm | Under 1 s | Large bond areas, flat parts | Loses deep voids |
| Dark-field ring | 5–15 μm | Under 1 s | Surface and near-surface voids | Sensitive to glare on glossy parts |
| Polarised cross-polar | 20–50 μm | 1–3 s | Stress mapping, mid-layer voids | Slower, needs careful alignment |
| Confocal line scan | 1–5 μm | 3–10 s | Thin bond lines, deep voids | Higher equipment cost |
| Structured-light 3D | 10–30 μm | 2–5 s | Curved surfaces, hidden cavities | Demands accurate part fixturing |
| X-ray or CT imaging | 5–20 μm | 30 s to minutes | Critical assemblies, validation | Slow, limited throughput |
For glare-heavy parts such as coated lenses and polished metal mounts, additional guidance is worth reviewing before commissioning any of the optical techniques above. A practical walk-through of the overcoming glare and reflection issues that show up once the lighting has been chosen and the line is live can save weeks of trial-and-error once the cell has been installed.
Software pipelines for segmentation and classification
A bubble's signature in the image is rarely just a circle. It can be a flattened oval, a chain of pinpoints, or a curved arc that hugs the substrate. Classical thresholding handles the simpler cases, but it tends to over-report when the bond line carries residue or when the lighting was not perfectly even. Modern pipelines lean on trained models that have seen thousands of real and synthetic voids, and they treat every defect as a piece of data rather than a binary verdict.
Effective pipelines share a few habits. They start with a flat-field correction that normalises the illumination across the field of view, then apply a noise-reducing pass that preserves edges. The segmentation step uses either a learned model or a sequence of morphological operators, depending on whether the application is fixed or evolving. Finally, every detected void is classified by size, aspect ratio, and proximity to the part edge, and the resulting record is written to a line-side database with a timestamp and a part serial number.
Classification matters because not every void has the same impact on the part. A 30-micrometre bubble in the centre of a sensor window may be invisible to the final application, while a 10-micrometre bubble in a laser collimator can scatter enough energy to burn an upstream filter. Vision systems that report only a pass or fail lose the rich data that engineers need to drive process improvements back at the dispense station.
Integration with line control is the next layer. The detection result feeds the programmable logic controller that governs dispense pressure, stage temperature, and cure ramp. A persistent cluster of mid-layer voids will slow the dispense, advance the cure profile, or trigger a maintenance request. Turning the inspection system into a feedback loop rather than a passive gate is what separates a workable solution from a reliable one, and the data collected along the way is what makes continuous improvement possible on lines that change part every shift.
Integrating vision systems with Australian production lines
Australian lines run smaller batch sizes than the large Asian and European plants where most optical resins are formulated, but they tend to carry higher product variety. Adelaide's defence photonics suppliers, for instance, build run quantities that range from a few dozen to a few thousand, with frequent changeovers. A vision system installed in such a cell must therefore be reconfigured quickly and must survive years of intermittent use without losing calibration.
Ambient conditions are another local factor. Plants in Western Australia that serve the Pilbara resource sector see temperature swings, airborne dust, and occasional exposure to corrosive atmospheres. Enclosures for cameras, lights, and controllers need to be sealed to a higher standard than the indoor laboratory where the system was first commissioned. Seal ratings, cable gland choices, and positive-pressure cabinet design all matter once the hardware leaves the controlled-environment test bench, and cabling should be specified for UV exposure if any part of the cell sits near a cure lamp.
Workforce skills are local too. Australian technicians are well trained in instrumentation, and vision vendors that provide clear remote diagnostics fare better than those that demand frequent site visits. Modules that expose diagnostics over a secure web page, along with a library of canned reports, let a small local team supervise multiple lines without leaving the office. A useful overview of how this kind of supplier support fits into a broader recognition programme can be seen on the industry recognition page.
Local compliance paperwork has to be considered from the start. Any system sold into a medical-device line will need evidence of validation against an Australian or international standard, and the Australian Consumer Law requires that product warranties be supported by reasonable inspection records. The budget for documentation does not have to be large, but it does have to be planned for. Designers who leave paperwork to the end of a project often find that some of the data they need has not been logged, and that gap is harder to fix later than it is to anticipate early.
Once a method has been chosen, the rest of the project is familiar to anyone who has installed a vision cell before. Cabling, staging, calibration, and operator training follow the same logic as any other capital purchase. The Australian wrinkle is the small batch size combined with strict regulatory scrutiny, which rewards systems that are easy to reconfigure and produce audit-ready records from the first part onwards. For more on the manufacturer's full capability and product line, the Dalian Optical Technology portal collects the relevant datasheets, application notes, and contact details in one place, and serves as a reasonable starting point when scoping a new inspection cell from scratch.