// Showcase

Measuring the wedge error with autofocus

When wafer and tool come together, they may be tilted against each other by no more than 20 µm.

Semiconductors & Microfabrication Sensors & Metrology

Measuring the wedge error with autofocus

Semiconductors & Microfabrication Sensors & Metrology

01Starting point

When wafer and tool come together, they may be tilted against each other by no more than 20 µm. There was no distance sensor in the concept for measuring that tilt – the height was meant to come from the same camera that finds the alignment marks, via the point at which the image is sharpest. Measure three marks and you have a plane, and with it the wedge. Elegant in principle. The only question was whether it is accurate enough, before the machine reaches the point where only ±15 µm remains correctable.

02Approach

We turned five sharpness measures loose on identical image stacks – four from the textbook list and one we added along the way. And we varied more than the algorithms: one campaign with the wafer re-placed under the camera before each series, two without. That is where the answer was. With the wafer left undisturbed, the best method scattered by 0.010 mm. With the wafer re-placed, by 0.095 to 0.106 mm – a factor of ten, and the algorithm was the same one both times.

03Result

The customer learns before the build where the limit really sits – and that it does not sit in the software. Instead of a promise that code will close the gap, there are four named measures: an objective with less depth of field than the 0.4 mm available, reproducible light, more precise positioning, and a lower-noise camera. Three of those are a purchase. Considerably cheaper than declaring the wedge solved and finding out at the contact point, where every failed attempt costs a wafer and a tool.

What's inside

A comparison of five sharpness measures on the same image data, three measurement campaigns under different conditions, a formal gauge study with three appraisers, three repetitions and ten measurement points per method, plus characterisation over a hundred measurements each at two sampling pitches – on wafer marks and on tool marks.

And what happened next

The most counter-intuitive result made it into our design rules. Sampling more finely does not make you more accurate: at 50 µm per image the winner was the same frame in all one hundred runs; at 10 µm per image it moved almost every time, and the scatter in z went up by a factor of three to six. The reason is simple – the sharpness maximum is as wide as the depth of field of the lens, and inside it camera noise decides. A flat maximum is not fixed by a finer grid, but by a narrower peak.

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.