A century per wafer: the AFM throughput problem, and how parallel heads change it

Start with a number published for hybrid-bonding metrology: imaging an 80 × 80 µm area with AFM takes about five minutes. That is 0.0064 mm² per five minutes, or roughly 0.08 mm² per hour of pure imaging.

A 300 mm wafer has an area of about 70,700 mm². At that rate, imaging every square millimetre with one tip would take on the order of a hundred years. Nobody does that, of course. Fabs sample: a few sites, on a few wafers, per lot.

Why not just scan faster?

Single-tip speed is bounded by physics: the cantilever’s resonance, the bandwidth of the feedback loop, and the tip-sample forces you can tolerate before you damage the tip or the surface. High-speed AFM research has reached video rate on small scan areas, but large scan ranges and production tip life are much harder to speed up.

Change the scaling model instead

In principle, throughput scales with the number of tips working at once. One barrier to many tips over a wafer is head size: a conventional AFM head is about 100 mm across, so you cannot put many over one wafer.

Shrink the head to 20 × 20 mm and the geometry changes. On a 300 mm wafer, a grid of 20 mm heads leaves room for dozens of independent scanners. Forty heads at conventional speed give about forty times the area rate of one tip, roughly 3 mm² per hour, without asking any single tip to go faster. Our 50–100 mm² per hour target goes further: it also needs each head to scan faster than the five-minute benchmark, about 1.25–2.5 mm² per hour per head.

The hard part

Many heads means many feedback loops sharing one wafer, one stage and one vibration environment. Each head needs its own calibration, and every head must read the same height on the same feature. That is where our engineering effort goes: synchronization, calibration, stability, electronics and control. Our dual-head prototype is the first step: prove parallel with two, then scale to many.