BRENNEN

Industries / Technology / Electronics and semiconductors / Semiconductor equipment

Inspection and measurement

Inspection and measurement tools find defects on wafers and masks and measure what was made, such as line widths, film thickness and how well layers line up, so that a bad step is caught before it spoils many wafers.

5

Milestones achieved, from 1984 to 2019.

3

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1984

    Electron microscopes measure line widths

    By the 1980s chip lines were close to the smallest size an optical microscope can resolve, so measuring them with light stopped working. Hitachi developed the S-6000 CD-SEM, a scanning electron microscope that measures the "critical dimension", the width of the smallest lines, at a low voltage of about 1 kV so the beam would not charge or damage the circuit; the Semiconductor History Museum of Japan says this made control of line widths below one micron possible.

    Lessonwhen features shrink past what light can resolve, measurement moves to electrons, and the tool must not harm what it measures.

    shmj.or.jpchiphistory.org

    Credit

    Hitachi (now Hitachi High-Tech)

  2. 2015

    KLA scales back mask tool

    KLA-Tencor had been building an actinic EUV mask inspection tool, one that inspects masks with the same EUV light that prints from them. In April 2015 it said it expected no customer funding and was scaling the programme back sharply; chief executive Rick Wallace said "there is not a market for it, there is no other application for it".

    Lessona tool with one use needs customers who commit to paying for it; the first such inspector for patterned masks came four years later from Lasertec.

    semiengineering.comoptics.org

    Credit

    KLA-Tencor

  3. 2016

    Lam and KLA merger collapses

    Lam Research had agreed to buy KLA-Tencor, the inspection and measurement toolmaker, in a deal Fortune valued at $10.6 billion. On 5 October 2016 the two ended it, after the US Justice Department said it would not continue with a consent decree (an agreed settlement) they had been negotiating; Fortune reported the pair would have held 42% of the wafer fab equipment market.

    Lessona merger of two of the largest toolmakers could not be made acceptable to US antitrust regulators, even with negotiated conditions.

    ir.kla.comfortune.com

    Credit

    Lam Research; KLA-Tencor; US Department of Justice

  4. 2016

    ASML buys an e-beam inspector

    In June 2016 ASML agreed to buy Taiwan's Hermes Microvision (HMI) for about €2.75 billion in cash. HMI made e-beam inspection tools, which scan a wafer with an electron beam to find defects too small for light to show, and ASML said the deal fit its plan to feed measurement data back into its lithography machines; Singapore's competition regulator cleared it. ASML later launched the eScan 1000, with nine beams in a 3 by 3 grid, which it says inspects up to 600% faster than earlier e-beam tools.

    Lessona single electron beam sees small defects but is slow, so the next step was many beams in one tool.

    sec.govsemiengineering.comccs.gov.sgasml.com

    Credit

    ASML; Hermes Microvision

  5. 2019

    EUV light inspects EUV masks

    Lasertec announced the ACTIS A150, which it calls the world's first actinic patterned EUV mask inspection system, able to find phase defects: tiny bumps or pits buried in a mask's mirror layers that change the reflected light and print onto wafers. Semiconductor Engineering reported in January 2020 that it had shipped to Intel, whose engineer Mark Phillips said optical and e-beam tools "cannot detect all printable defects".

    Lessoninspecting with a different light from the one that prints can miss defects that print.

    lasertec.co.jpsemiengineering.com

    Credit

    Lasertec; Intel (first customer)

Watch for

  1. Faster EUV mask inspection

    Lasertec says its ACTIS A200HiT inspects EUV masks three times as fast as the A150, and aims it at chip fabs checking masks when they arrive and at set intervals, rather than at mask makers.

    Achieved whena chipmaker or an independent party reports it in routine use for incoming masks, or publishes its measured speed against the A150.

    lasertec.co.jp

    Claimed by

    Lasertec, October 2025

  2. Ninety-one beams in one tool

    Hitachi High-Tech and ZEISS SMT are developing in-line inspection tools built around ZEISS's 91-beam electron column. They say prototypes tested on site inspect several dozen times faster than single-beam tools and find small defects that optical inspection misses, and that they will test on mass-production wafers aiming for "process of record" status, a fab's approval for production use.

    Achieved whena chipmaker names the tool as process of record, or publishes its own capture-rate and speed data.

    hitachi.com

    Claimed by

    Hitachi High-Tech and ZEISS SMT, September 2026

  3. Chinese wafer inspection advances

    The Chinese toolmaker Skyverse says its bright-field patterned-wafer inspection tool (one that looks at light reflected straight back and compares neighbouring chips) now sells in volume for older chip generations, that its advanced version is in trials at leading Chinese logic and memory makers and has won some orders, and that its e-beam line-width tool is in customer sample trials.

    Achieved whena named fab or an independent body confirms production use of the advanced tool and states the chip generation.

    pdf.dfcfw.com

    Claimed by

    Skyverse, in results restated by a Soochow Securities analyst note, November 2026

Open

  1. 01

    Open roughness-measurement benchmark

    A public set of simulated CD-SEM images of rough lines whose true edges are known, at several noise levels, pixel sizes and frame-averaging settings, with a scoring script that reports how far any method's line-width roughness (the wobble in a line's width, quoted as three standard deviations in nm) and roughness spectrum are from the truth. Include baseline scores, and a second track that tests whether AI denoising changes the measured numbers.

    The brief
    Credit · unclaimed
  2. 02

    Open nanosheet scatterometry benchmark

    Scatterometry, also called optical critical dimension (OCD) measurement, shines light on a patch of repeated structures and works out their shape by matching the reflected spectrum to computed ones. Build a set of spectra computed with RCWA (rigorous coupled-wave analysis, the standard method for light reflecting off repeated structures) for a stack of nanosheet transistors, the newest transistor shape, labelled with dimensions such as sheet thickness, width and recess depth, with realistic noise, a frozen test set with hidden labels and a script that scores error per dimension in nm.

    The brief
    Credit · unclaimed

Open 01 · the brief

Open roughness-measurement benchmark

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public set of simulated CD-SEM images of rough lines whose true edges are known, at several noise levels, pixel sizes and frame-averaging settings, with a scoring script that reports how far any method's line-width roughness (the wobble in a line's width, quoted as three standard deviations in nm) and roughness spectrum are from the truth. Include baseline scores, and a second track that tests whether AI denoising changes the measured numbers.

Why it is still open

In a 2018 comparison led by the research centre imec, labs across the industry disagreed on roughness by ±30% (three standard deviations); imec answered with a measurement protocol, a procedure rather than a public image set with known answers. Searches of Zenodo and Figshare found no shared test set, and Fractilia's MetroLER, the best-known tool for this measurement, is commercial.

fractilia.com

Built with what exists

ARTIMAGEN, NIST's public-domain generator of SEM images, or Nebula, an open-source simulator of how electrons scatter in matter, which runs on graphics cards; scikit-image for baseline edge finding; PyTorch for learned methods; Zenodo for hosting.

arxiv.orgnebula-simulator.github.io

  1. First step

    Generate 100 images of one roughness spectrum at three noise levels, compute the true roughness from the edge arrays, and measure how much a simple threshold edge-finder overstates it as noise rises.

  2. Done looks like

    at least 10,000 images with ground truth under an open licence, at three or more noise levels and two or more pixel sizes; the script reports each method's roughness error in nm and its spectrum error; at least three baseline methods are scored; an outsider who reruns the script gets the same numbers.

Open 02 · the brief

Open nanosheet scatterometry benchmark

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

Scatterometry, also called optical critical dimension (OCD) measurement, shines light on a patch of repeated structures and works out their shape by matching the reflected spectrum to computed ones. Build a set of spectra computed with RCWA (rigorous coupled-wave analysis, the standard method for light reflecting off repeated structures) for a stack of nanosheet transistors, the newest transistor shape, labelled with dimensions such as sheet thickness, width and recess depth, with realistic noise, a frozen test set with hidden labels and a script that scores error per dimension in nm.

Why it is still open

NIST's public study comparing standard and machine-learning OCD methods used a data set and simulation library "that assumes a 2-D geometry"; searches found nanosheet training spectra only in patents and vendor papers, not as public data.

nist.gov

Built with what exists

torcwa, an RCWA solver in PyTorch that runs on graphics cards (LGPL licence), cross-checked against a second open RCWA code; published optical constants for silicon, silicon-germanium and silicon dioxide; scikit-learn and PyTorch for baselines.

github.com

  1. First step

    Model a simple silicon line grating in torcwa, compare its reflectance with a second open RCWA code at about 10 wavelengths, then fix the nanosheet parameter list and ranges.

  2. Done looks like

    at least 20,000 labelled spectra released, with a hidden test set of at least 2,000; error per dimension in nm for at least two baselines; agreement between the two solvers on a simple 2-D grating reported as a number.