Access Official Technical PDF

Share Summary:LinkedInEmail

As semiconductor processes pursue tighter line width roughness (LWR) targets, reliable measurement becomes essential. LWR metrology standardization is not just a matter of agreeing on a numerical specification: engineers also need a consistent way to determine whether a process meets it. Without shared acquisition and analysis methods, measurements from different tools or organizations may not be meaningfully comparable.

A multi-company round robin reported a three-sigma spread of approximately ±30% in roughness results for the same wafer, with participants using their own best-known methods. That level of variation can complicate comparisons against specifications and between process-development efforts. The imec roughness protocol (iRP) was proposed to address the problem with more complete guidance for scanning electron microscope (SEM) image acquisition and roughness analysis.

Why LWR metrology standardization matters

A roughness value depends on more than the physical edge being measured. It is also affected by how the image is acquired, how edges are detected, and how the measurement data are processed. Settings such as pixel size, frame averaging, digital filtering, and measurement length can change the reported result.

This distinction matters because a measurement intended for process monitoring may still be useful when settings remain consistent. But comparing results across tools, materials, or organizations requires tighter controls. Otherwise, differences in measurement conditions can be mistaken for actual changes in pattern quality.

How SEM noise affects roughness measurements

SEM image noise can interfere with edge detection and make measured roughness appear higher than the underlying feature roughness. One approach to reducing this effect is to calculate the power spectral density (PSD) of the edge data and subtract the estimated SEM noise floor. Integrating the corrected, or unbiased, PSD then provides a roughness estimate intended to be less sensitive to noise.

This correction relies on an important assumption: at high spatial frequencies, the SEM noise behaves like a flat, white-noise floor. When that assumption holds, noise-floor subtraction can reduce measurement variability. When it does not, the correction may no longer be reliable.

Frame averaging has practical limits

Frame averaging affects the amount of noise in an SEM image. With fewer frames, noise can make edge detection less reliable and increase a biased LWR estimate. In the reported tests, the unbiased estimate remained more stable than the biased one down to about 10 frames. Below that level, the high-frequency PSD became sloped rather than flat, indicating that the white-noise assumption was no longer satisfied.

The implication is not simply to apply noise correction in every case. Acquisition settings must also keep the image within conditions where that correction is valid.

Filtering and pixel size can change the result

Digital filtering may help some edge-detection methods, but it can also remove real high- and mid-frequency information from the PSD and introduce artifacts. Because roughness is calculated from the PSD, filtering can alter the reported value. The study therefore recommends avoiding digital filters for roughness analysis. Where a tool cannot operate without filtering, the protocol limits its use to a specified setting rather than leaving the choice open-ended.

Pixel size is another important variable. Changing it affects the high-frequency noise contribution, so biased measurements made at different pixel sizes may not agree. Noise correction can reduce this effect, but it is less dependable when pixels are too large. The protocol consequently specifies a pixel size of 0.8 nm in both image directions.

By contrast, changing the field of view (FOV) by changing the number of pixels was found to have little effect when pixel size stayed constant. This provides flexibility to adjust image dimensions without introducing the same inconsistency associated with changing pixel resolution.

Relating measurement length to correlation length

The length of the line segment used in a roughness calculation also affects the result. Longer segments include more low-frequency components, while very short segments can produce artificially low roughness estimates. Earlier guidance called for a measurement box length of at least 2 μm, but the analysis argues that a fixed length does not account for the feature’s correlation length.

Instead, the recommended measurement length is about 50 times the correlation length, which is intended to provide accuracy on the order of a few percent. In practice, the number of pixels may be adjusted to achieve the required field of view while preserving the specified pixel size.

What the imec roughness protocol specifies

The iRP 2018 combines acquisition and analysis requirements to make results more comparable. Its principal settings include:

  • Pixel size: 0.8 nm in both image directions.
  • Beam settings: 8 pA, with landing energy of 500 eV for after-develop inspection (ADI) or 800 eV for after-etch inspection (AEI).
  • Image dimensions: 2048 by 2048 pixels as the standard, with pixel counts allowed to change when needed.
  • Frame averaging: 16 frames for ADI and 32 for AEI.
  • Filtering: none as the standard; any permitted option should be reported.
  • Analysis: at least 50 uncompressed TIFF images and mandatory PSD unbiasing.
  • Measurement length: set the field of view in the relevant direction to approximately 50 times the correlation length.

Reporting deviations is part of the protocol’s intent. If an allowed option is used, the measurement data should include that information so that readers can interpret comparisons appropriately.

Practical implications for semiconductor teams

Unbiased metrology is necessary for reducing the influence of SEM noise, but it is not sufficient on its own. Low frame averaging, large pixel sizes, or filtering can undermine the assumptions behind the correction or distort the PSD. Consistent results therefore depend on controlling the full measurement chain, not just applying a noise-subtraction step.

For engineers and technical decision-makers, the practical takeaway is to treat LWR as a method-dependent measurement unless acquisition and analysis conditions are documented and aligned. A defined protocol helps make comparisons between teams, tools, and specifications more defensible. As roughness targets tighten, consistent metrology is a prerequisite for knowing whether measured process performance represents the feature—or the measurement settings.

Access Official Technical PDF

About The Author