Precision vision

The industrys never-ending quest to scale down is throwing up some tough challenges for metrology. Neal Sullivan and Paul Knutrud of Solurus explain how one metrology method is measuring up to sub-100nm processes.

As the industry approaches the 65nm technology node, improved performance in critical dimension (CD) measurements is becoming increasingly valuable and, at the same time, increasingly difficult to achieve. The ITRS roadmap requires printed gate CD control to drop from 3.3nm at the 90nm node to 2.2nm at the 65nm node, with a maximum bias of no more than 10%. Gate CD control directly impacts device performance and therefore yields and process profitability.

At the same time, CD-scanning electron microscope (CD-SEM) metrology is confronting a number of fundamental technical challenges related to its effective resolution, measurement algorithms and new process materials, all of which make improved performance more of a challenge. In many ways, these challenges can be compared to those faced by photolithography as it pushed into the sub-wavelength regime.
The easy improvements (shorter wavelengths) have become more difficult, and next generation solutions will be based on detailed understanding and manipulation of the fundamental physical processes that limit performance (as was the case with such innovations as resolution enhancement technology and phase shift masks).

Effective resolution
On a flat sample, SEM resolution is largely determined by the size of the spot formed by the beam on the sample surface. Secondary electrons (SE) – which are sample electrons scattered by interactions with the high energy electrons of the scanning beam – have low energies and can escape to be detected only from a very shallow region. After entering the sample, electrons continue to penetrate, creating additional SE or perhaps being scattered back out of the sample by a collision with the nucleus of a sample atom. Most of these deep secondary electrons are reabsorbed before escaping, however, a high energy back-scattered electron (BSE) that exits the surface can create detectable SE as it escapes. The situation becomes more complicated when the sample is not flat.

In fact it is one of the strengths of the SE signal that it is very sensitive to local topography. A local peak will emit a stronger SE signal, appearing brighter in the image. This is because more of the subsurface volume, within which the beam electron can scatter and create secondary electrons, is close to the surface. The CD-SEM exploits this phenomenon when it measures line width as the distance between the lines bright edges.
As shown in figure 1, the effective resolution of a CD-SEM is not determined by the spot size as much as the size of the subsurface interaction volume. The spot itself may be some distance from the edge while scattered beam electrons are already creating SE as the beam electrons escape through the sidewall of the line.
The size of the interaction volume is a function of material composition and beam energy, with higher atomic number samples and lower beam energies allowing smaller volumes. Clearly, the effective resolution of a CD-SEM is improved at lower energies.
This improvement is usually offset by the increased difficulty of focusing low energy electrons into a sufficiently small spot. The CD-SEM used to obtain the results discussed here (Soluris Yosemite) is specifically designed for ultra low voltage (ULV) operation and can acquire fully automated measurements with beam landing energies less than 200eV while maintaining better than 3nm imaging resolution. Below 100eV, the interaction volume becomes smaller than the spot size so there is less benefit to be gained from further reductions in landing energy.

Typically, CD-SEM measurements are acquired at a landing energy of 500ev to 800eV. Between 175eV and 800eV, the electron range, and therefore the radius of the interaction volume, increases from 3nm to 30nm. A recent experiment compared measurements made on 180nm etched polysilicon lines over a range of beam energies.

The results demonstrated a statistically significant increase in measured line width of around 3nm between 175eV and 800eV. Such differences may be attributed to the influence of electron range and interaction volume [1,2,3]. Reducing these influences by reducing the beam energy increases the systems effective resolution.

New materials
Another issue that affects the accuracy of CD-SEM metrology is the use of materials that are sensitive to electron beam exposure. Advanced ArF resists exhibit significant slimming upon exposure to the beam. Slimming degrades the accuracy and the repeatability of measurements on these materials [4]. Slimming also reduces line edge roughness (LER), leading to underestimates of LER in process control.

Slimming (figure 2) is difficult to quantify since most of it appears to occur on the first exposure to the beam, eliminating the possibility of obtaining an unexposed measurement for comparison. Attempts to characterise slimming by extrapolating curves formed by sequential measurements back to some original value have yielded inconclusive results. Curves taken at different beam energies do not extrapolate back to the same point. Likewise, efforts to evaluate slimming based on the difference between the first measurement and the second have not been successful.

Comparisons to AFM measurements show initial slimming of 5nm or more at 500eV while less than 1nm of slimming occurs at ultra low voltages (<200eV), suggesting that ULV measurements are a good reference for evaluating “first strike” slimming at higher beam energies.

Slimming has also been evaluated through its affect on etched polysilicon lines. Again the results showed that initial exposure at higher beam energies caused several nanometres of slimming, while slimming at beam energies below 200eV was negligible.

Measurement algorithms
The first section of this article explains the use of low beam landing energy to reduce the size of the beam/sample interaction volume and improve the effective resolution of CD-SEM measurements. A clear understanding of interaction volume and its impact on the accuracy and precision of CD-SEM measurements requires a detailed examination of the interactions between the beam electrons and the sample. These interactions can be modelled using a Monte Carlo approach that follows each electron through its series of random encounters with sample atoms. Using this approach, it is possible to model the shape of the SE intensity profile that would be generated by a feature with a given physical shape.

Conventional CD-SEM measure-ment algorithms have taken a simpler approach to extracting the measurements from the electron signal. When SEMs supplanted optical microscopes for CD measurements, the width of the intensity changes associated with the line edge were small relative to the line width.

Very reproducible measurements can be achieved by simply defining a reproducible point on the profile. This is usually chosen as a percentage of the maximum or the point of maximum slope. Such measurements proved to be very repeatable and sufficiently accurate for process control of larger CDs. They are not sufficient for CD control in sub-100nm processes. At the 65nm node, the roadmap requires printed gate lengths of 35nm and metrology bias no greater than 10% or 3.5nm. Earlier, we described a 3nm discrepancy between measurements taken at 175eV and 800eV. What is the true value?
Model-based measurements provide a mechanism for understanding the relationship between the intensity profile and the features physical shape. Figure 3 compares a modelled intensity profile with the physical shape for which it was calculated. In this example, the 50% point on the intensity profile overstates the 50% width of the physical feature by about 10nm. When the other side of the line is included, the error grows to 20nm.

The example in figure 2 illustrates another aspect of CD measurements that impacts both accuracy and precision – that is, the need for two and three-dimensional information to completely characterise the feature. Error in measurement is at least partly due to the presence of a foot on the feature and the corresponding bulge in the intensity curve. Conventional one-dimensional CDs, such as line width, do not comprehend the rich detail of the actual physical feature, detail that is largely preserved in the SEM signal. The problem lies in the many-to-one relationship between the feature shape and the conventional measurement: there are many physical shapes that correspond to any given measurement value. This is a source of both bias and variability in the measurement. Variability clearly degrades process control capability, however, a number of recent studies [5-9] also cite inaccuracy (bias) as a significant factor for process control.
For instance, pairs of features may both measure within specification, but one works and the other fails. With conventional measures, shape is essentially uncontrolled and can lead to performance and yield failures. The lack of shape information, and therefore accuracy, can consume a significant part or all of the allowable process window. This cannot be recovered simply by changing the specification. Moreover, it cannot be calibrated out of the conventional measurement process.
A commercially available, model-based measurement algorithm (Critical Shape Metrology, Soluris) was compared to CD-AFM, cross sectional SEM, and FIB9, and showed excellent agreement in all cases. Critical Shape Metrology implements a library approach to modelling.

The user specifies a range of values for various shape parameters including foot length, corner rounding and sidewall angle. The system calculates intensity profiles at specified increments of each parameter and stores the profiles in a library. Each measured profile is then compared to the library and the best match is selected to calculate the features shape and location.

Once the initial model calculations have been stored in the library, measurement throughput is unaffected by the matching process, allowing the system to run at its maximum capability of 1,000 measurements per hour.

Conclusion
Technologies often develop in spurts, with periods of predictable, incremental change punctuated by revolutionary points when the demands of the marketplace force radical changes on the status quo. The CD-SEM is now at such a point. Decreasing feature sizes and narrowing process windows require tighter controls, yet the easy advances have already been made.

For sub-100nm processes, CD-SEM will require ultra low landing energies to improve effective resolution and eliminate resist slimming; and model-based measurement algorithms to improve precision and accuracy and to extract needed three-dimensional information from the SEM signal. Fortunately these solutions already exist and have demonstrated the capability to support semiconductor manufacturing to the 45nm node and beyond.

Fig 1

Fig 2

Fig 3