Semiconductor Technology from A to Z

Particles are among the most critical sources of defects in semiconductor manufacturing. A single sub-micrometer particle, if it lands on an active structure, can already cause a short circuit, an open line, or another electrical defect, rendering an entire chip non-functional. Since modern feature sizes are now in the single-digit nanometer range, particle contamination on the wafer must be monitored continuously.

Particles can originate from a wide variety of sources: from the ambient air of the cleanroom, from equipment components (e.g. flaking coatings, seal abrasion), from process gases and chemicals, or even from the wafer itself (e.g. crystal defects that emerge at the surface). To narrow down the cause and ensure cleanroom quality as well as equipment performance, wafers are checked for particle contamination after critical process steps.

The most common method for particle detection uses the interaction of light with the wafer surface, similar to some film-thickness measurement techniques. However, since particles do not form large, homogeneous layers but rather represent localized point defects, this method does not rely on the interference or polarization of reflected light, but instead on the scattering of light at small objects.

In laser scattering measurement, the wafer surface is systematically scanned with a focused laser beam. When the beam hits a flat, defect-free area, the light is reflected in a largely directional manner. If, however, a particle, a scratch, or another topographic irregularity is present at that location, the light is scattered in almost all directions.

Principle of scattered-light measurement at a particle

One or more detectors positioned outside the direct reflection path capture this scattered light. The intensity of the scattered light can be used to approximate the particle size – however, this is typically not reported as the actual physical dimension, but as the so-called Latex Sphere Equivalent (LSE) size, i.e. the size of a reference sphere (usually polystyrene latex) that would produce the same scattered-light intensity.

Because different materials (metal, photoresist, crystal fragments, etc.) scatter light to differing degrees, the real particle size often deviates from the measured LSE size. Modern systems therefore frequently use several wavelengths and detection angles simultaneously (multi-channel or multi-angle scattering) to allow initial conclusions about the particle type in addition to its size.

Modern inspection systems achieve a detection limit of roughly 20–30 nm (LSE), although actual sensitivity depends heavily on the wafer surface (film structure, roughness, reflectivity) and the wavelength used. Patterned wafers with existing device structures are inherently more difficult to inspect than unstructured (blanket) wafers, since structure edges themselves scatter light and can therefore be falsely identified as particles.

The result of a particle measurement is usually displayed as a so-called wafer map: a graphical top-down view of the wafer on which every detected defect is marked as a point at its exact x/y coordinate. In addition to the position, the LSE size and a preliminary classification code are stored for each defect.

Example of a wafer map with particle distribution

The spatial distribution of the defects already provides important clues about the root cause. Typical patterns include, for example:

• Edge-heavy concentration: often indicates handling problems or uneven process conditions at the wafer edge (e.g. during spin-coating or CMP processes).
• Comet-shaped patterns: often result from a single, larger particle that is dragged across the wafer during a rotational process (e.g. resist spinning), leaving a trail of further, smaller defects behind it.
• Random, uniform distribution: usually points to contamination from the ambient air of the cleanroom.
• Localized clusters at specific coordinates: can indicate a faulty piece of equipment that touches the wafer at the same location every time (e.g. a wafer handler or chuck).

The inspection software automatically assigns a preliminary class to each defect based on its scattered-light signature, size, and shape (so-called Automatic Defect Classification, ADC). However, this automatic pre-classification is not always reliable, especially for new or unknown defect types, which is why critical or conspicuous defects must be examined more closely in a further step.

Since optical scattered-light measurement only provides position and an approximate size, but no reliable information about the defect's shape and material, selected particles are subsequently examined more closely in a so-called defect review. For this purpose, the wafer map is imported into a scanning electron microscope (SEM), which automatically drives to the previously determined coordinates (Automatic Defect Review, ADR).

Workflow of the automated SEM review

Upon reaching the target location, the SEM captures a high-resolution image of the defect. Since modern inspection systems can detect several thousand defects per wafer, it is not practically possible to review every single defect. Instead, a representative sample is usually driven to automatically (e.g. the largest 20–50 defects, or a random selection per class).

In addition to plain image capture, many SEM systems feature an energy-dispersive X-ray detector (Energy Dispersive X-ray Spectroscopy, EDX). This allows the chemical composition to be determined directly at the defect location, which enables conclusions about its origin: if, for example, a metal is detected that is not used in any of the current process steps, this points to contamination from an equipment component (e.g. worn seals, chamber coating).

Based on shape, size, and material composition, the defect is then finally classified (e.g. particle, scratch, crystal defect, residue). This classification is usually compared against historical reference images (a so-called defect library), either manually by a technician or automatically via image recognition.

The results of particle measurement are not only considered for the individual wafer, but are evaluated statistically across multiple wafers, lots, and a longer time period. A key metric here is the total defect count per wafer (Total Defect Count) above a defined size threshold, which is compared against control limits (Upper Control Limit, UCL). If this limit is exceeded, the process is considered “out of control” and must be investigated.

Pareto evaluation of the defect classes across multiple wafers

A common method for prioritization is Pareto analysis: the various defect classes are displayed sorted by frequency, so that the class with the largest share is immediately apparent. Since experience shows that the majority of defects can usually be traced back to a few main causes, troubleshooting can be focused specifically on the most frequent classes instead of tracking down every individual cause separately.

Of particular importance is also the correlation between defect density and the later electrical yield of the chips. Not every defect actually leads to a device failure – a particle lying on an uncritical area (e.g. in the scribe line between the chips) has no effect on functionality. By comparing defect positions with the later test results of the individual chips (yield map), it is possible to statistically determine which defect classes and sizes are actually yield-relevant. This so-called defect-to-yield correlation is one of the most important tasks of failure analysis in semiconductor manufacturing, as it allows resources to be focused specifically on the truly critical defect sources instead of pursuing every detected particle class with equal intensity.

In the long term, the insights gained feed into a continuous improvement program: equipment with elevated particle loads is specifically maintained or requalified, process parameters are adjusted, and cleanroom conditions are improved as needed, in order to reduce defect density and, ultimately, manufacturing cost per functional chip.