Building a sub-micron inspection strategy

One of the most challenging inspection goals I worked toward was pushing resolution into the sub-micron range. At that scale, everything matters.

What sub-micron actually demands

Start from the physics. To resolve 0.8 um features:

required NA  ~ 0.61 * 0.55 um / 0.8 um  ~ 0.42
depth of field at NA 0.42  ~ 3 um
pixel on object for 3x sampling  ~ 0.27 um

Three micrometers of depth of field means the entire mechanical and thermal system must hold the surface inside a 3 um window while imaging. That is the real specification, and it has nothing to do with the camera.

The problems

• Vibration - 0.5 um of vibration amplitude is invisible to a normal system and fatal here. Footsteps, air handling, and neighboring machines all show up.
• Temperature drift - steel grows about 12 um per meter per degree. A 100 mm structural loop and a 2 degree drift moves focus by more than the entire depth of field.
• Optical aberrations - at NA 0.42 the lens design, not the sensor, sets image quality, especially off-center.
• Focus repeatability - the Z axis must return to within a micron, repeatedly, for hours.
• Contamination - a 1 um dust particle is now a defect-sized object. Cleanliness becomes an imaging parameter.
• Illumination stability - shot-to-shot variation that was always there becomes measurable and matters.

The major realization

The camera was not the bottleneck. Real improvement came from:

• Better optics
• Better illumination
• Better motion control
• Better calibration targets

The camera is often the least interesting component.

How I would start such a project today

1. Write the resolution and field-of-view requirement, then derive NA, sampling, and depth of field on paper.
2. Measure the environment first: vibration spectrum at the mounting point, temperature swing over 24 hours. These numbers veto designs faster than any optics discussion.
3. Get a calibration target that actually has structures at and below the target resolution - you cannot tune what you cannot see.
4. Build the optical path on a rigid breadboard and prove the resolution statically before any automation exists.
5. Only then add motion, and re-verify the resolution after every mechanical change. It will get worse each time; the engineering is in winning it back.

This completely changed how I evaluate high-resolution systems today: I start with the optical path, the mechanics, and the calibration strategy, and only then worry about the sensor.