Cleaning the wafer
Silicon Almanac All explainers
Process control · Explainer 08

Finding the defect

Even a defect a few nanometres across can ruin a chip, as NIST put it in 2013. Somewhere on a 300 mm wafer, a machine has to find it, and fast enough to stop the next lot from getting it too.

~1,000×slower: e-beam inspection against optical, as reported
1 or 2 of 25wafers in a lot that are measured
$18 bn+metrology and inspection tools, 2025
Semiconductor Engineering (2016) reported the 1,000×; Hitachi High-Tech gives the 1 or 2 of 25; Yole Group estimated the market in June 2026. Scene: a wafer moving under the fixed head of an optical inspection tool, opened; what it has looked at lights up, and the defects it passes are marked. The tool is generic, not any company's design; the wafer is drawn to scale. All sources are listed at the end.
Scroll to begin
01 · Zoom

One foot in 3,000 miles.

Rick Silver of NIST compared a 10 nm defect on a 300 mm wafer with "one foot in 3,000 miles" in 2013. That is one part in 15.8 million; the defect is one part in 30.0 million, about 1.9 times finer, by our arithmetic.

Our own: blow the wafer up until the defect is a €2 coin, 25.75 mm across. The wafer is then 772.5 km across and its area about 131% of Germany's total area; the metal lines are 5.4 cm apart. Zoom from the wafer to the defect, every level at its true size against the others.

The same view, if the defect were a €2 coin
-

The whole wafer: 648 whole dies of 10 mm, as drawn, with the one the zoom goes into outlined.

One die, 10 mm square as drawn: blocks of logic, memory and wiring.

A block, 1 mm across as drawn: banks of repeated cells.

Repeated cells, 20 µm across as drawn, with the metal wiring over them.

Metal lines 21 nm apart, centre to centre: the tightest pitch imec gave for its N5 node in 2019.

The defect: a 10 nm speck of foreign material in a gap about as wide as itself.

NIST (2013) and its SI guide for the foot and the mile; the European Central Bank for the coin; Destatis for Germany's total area, 357,683 km² at the end of 2024; imec (2019) for the pitch. The ratios and the coin scale are our arithmetic. The wafer, the pitch and the defect are drawn at their sizes; the die, the block, the cells and the lines' width and height are generic and drawn. Whole dies are those lying wholly inside the 300 mm disc on a 10 mm grid, with no edge margin: our count.
02 · Two ways to look

Fast, or sharp.

Inspection finds where the defects are; metrology measures whether the pattern came out the right size. Applied Materials puts the choice in one sentence: optical defect detection "scans wafers in less time at the sacrifice of resolution", while e-beam review "offers high resolution but takes much longer to scan a wafer".

Keith Wells of KLA-Tencor said in 2015 that its optical tools could reach 10 nm, at "maybe a half a wafer an hour". The scene runs both for the same time: light covers the whole wafer, electrons a square of about 8.4 mm, at the 1,000 to 1 reported in 2016.

Finds defects down to
-
Throughput counted in
-
A full wafer takes
-
Readouts: Semiconductor Engineering, in different years and about different tools: 30 nm and 3 nm from experts it did not name (2015); about an hour on KLA's 3900 and 2930, and e-beam "1,000 times (or more) slower" (2016); hours, if not days, with e-beam (2020). Applied Materials (2023); Keith Wells as quoted (2015). The square is our arithmetic. The tools are generic, the e-beam column opened.
03 · Closing the gap

More beams at once.

One answer is many beams, each scanning its own square. ASML, owner of the e-beam maker HMI since 2016, says its 9-beam HMI eScan 1000 (2020) is "up to 600% faster than previous e-beam wafer inspection tools", and its 25-beam eScan 1100 (2022) will "increase throughput by up to 15 times compared to single e-beam inspection tools".

At Delft University of Technology, a 196-beam research microscope for imaging biological tissue "should have 196 times higher throughput than a single beam SEM if we can make the detection as efficient", its builders wrote in 2016.

Others make one beam faster: Applied Materials says its cold field emission source images up to 10× faster than thermal field emission e-beam; KLA's eSL10 uses "a single, high current-density electron beam".

Area per second, ideal
-
This patch, as drawn
-
ASML (product pages; releases of May 2020 and April 2022), its own claims; Delft University of Technology (2016); Applied Materials (December 2022), a launch claim; KLA (July 2020). Area per second is our arithmetic, an ideal: each beam as fast as one alone. The 9 and 25 beams are drawn in ASML's 3 × 3 and 5 × 5 arrays; the column is generic. One beam covers the patch in 45 s.
04 · You cannot look everywhere

Measure a few.

Fabs do not measure every point on every wafer; they sample. Hitachi High-Tech's primer gives a sampling plan for metrology: 1 to 2 wafers taken from a lot of 25, 5 to 20 dies on each, 10 to 100 points on each die, the number of points varying by maker and device. That is 50 to 4,000 points for a lot, by our arithmetic.

KLA-Tencor's David Price and Douglas Sutherland wrote in 2014: "The best known method for optimizing total cost is usually adjusting the overall lot sampling rate."

Points measured per lot
300
Wafers touched
4%
Hitachi High-Tech, metrology primer, which also gives the lot of 25; KLA-Tencor's Price and Sutherland (2014). Totals are our arithmetic. The carrier is generic, its wafers drawn 10 mm apart.
05 · Why it pays

Every lot until someone looks.

When a tool drifts, every lot through it is at risk until an inspection catches it. Here the tool goes wrong at lot 6 of 24, the lots in the order they left it from the far row on, and the arches are inspections. KLA-Tencor's engineers wrote in 2018: "Finding the excursion sooner limits the lots at risk: a smaller and more clearly defined population of lots are exposed to the higher defect count".

Yole Group estimates that metrology and inspection equipment grew 15% in 2025 to over $18 billion, about 14% of all wafer fab equipment spending.

Lots at risk
3
Wafers at risk
75
KLA-Tencor's Price, Rathert and Sutherland (November 2018); Yole Group (18 June 2026), an analyst's estimate. The example and its counts are our arithmetic, 25 wafers to a lot. The carriers and the track are generic and drawn.
06 · Where to go next

Look first, then fix.

KLA-Tencor's engineers put it first among ten truths of process control in 2014: "You can't fix what you can't find. You can't control what you can't measure."

The 11 companies this record files under process control: Advantest Corporation, Applied Materials, ASML Holding, Camtek, Canon Inc, KLA Corporation, Lam Research, Nikon Corporation, Nova Ltd, Onto Innovation and Screen Holdings.

KLA-Tencor, Price and Sutherland (2014). Scene: the inspected wafer with every defect the scan passed marked; the tool is generic, not any company's design.