// Showcase

Machine vision tested against ±1 µm – in the plane and along the axis

The machine replicates glass wafers: wafer and tool are positioned relative to one another, brought together and pre-cured.

Semiconductors & Microfabrication Sensors & Metrology

Machine vision tested against ±1 µm – in the plane and along the axis

Semiconductors & Microfabrication Sensors & Metrology

01Starting point

The machine replicates glass wafers: wafer and tool are positioned relative to one another, brought together and pre-cured. The requirement is a placement accuracy below ±1 µm. Corrections are driven in x, y, z and in three angles – two of which are tilts, that is, the wedge between wafer and tool. A tilt can only be corrected if the height at a minimum of three fiducials is known. No separate height sensor was foreseen: z was to come from the same camera, via focus. On top of that sits a point of no return – below a gap of 100 to 200 µm, only corrections of ±15 µm remain permitted. Whatever the vision system is going to contribute has to be right before that line.

02Approach

We did not debate it, we measured it. In the plane, two methods were set against each other: finding edges and intersecting them, versus teaching a pattern and finding it again, each over hundreds to a thousand fiducials. Along the axis, four sharpness measures were compared on identical stored image stacks – first derivative, second derivative, FFT and depth from focus – supplemented by 30 series of 21 images each and a formal gauge R&R across all the methods.

03Result

In the plane the answer is comfortable: edge finding delivers a standard deviation of 0.04 µm in X – except that depending on the illumination direction it finds 100 fiducials out of 100 one time and 37 out of 100 the next. The taught-pattern method found 1000 out of 1000, every time, at 0.18 to 0.64 µm. For a production machine the detection rate wins, not the last decimal place. Along the way it emerged that illumination direction outweighs the choice of algorithm – shallow incident light at about 10° against steep at about 80° makes a factor of three and a half. Along the axis the answer is honestly no: the accuracy is not yet sufficient for wedge compensation, and what is missing is written down as a hardware list rather than as a software promise.

What's inside

A requirement specification with the process requirements written out, an autofocus evaluation comparing four methods, accuracy summaries for x, y, angle and z, and a gauge R&R evaluation that fixes the ranking in numbers: standard deviation of the second derivative 15.1 %, FFT 17.8 %, mean of the second derivative 33.8 %, first derivative 38.6 %, depth from focus 42.6 %. Worth noting: the addendum contradicts the recommendation in the main body – the method favoured there comes fourth of five.

And what happened next

The finding to take away is the uncomfortable one. When a measurement is not accurate enough, the instinct is to sample it more finely – here that instinct was exactly wrong. At 50 µm per image the sharpest image fell on the same index in all 100 measurements, with a scatter of 0.67 µm. At 10 µm per image the index wandered by one or two positions, the scatter rose to 1.8–3.9 µm and the range opened to as much as 18 µm. At a flat maximum, camera noise decides which of several nearly equally sharp images wins. The remedy is not more samples along a flat curve but a lens with less depth of field.

All showcases

Get in touch

We are ready for new challenges

You bring the challenge. We develop the solution. From the first idea through concept and development to the validated machine, we guide your project.