Sharpness search algorithms for automatic focusing in the scanning electron microscope C.F. BATTEN, D.M. HOLBURN, B.C. BRETON, N.H.M. CALDWELL Department of Engineering, Cambridge University, Cambridge, UK
The scanning electron microscope’s transition from a re-search device to a common industrial tool has increased the need for instrument automation, both in conventional and re-mote microscopy.1Traditionally, autofocusing research has concentrated on finding an optimal sharpness measure, which is then applied over a range of focal lengths and the focal length with the maximum sharpness is chosen as the best focus.2The present work investigates the use of more so-phisticated sharpness search algorithms that decrease search time without sacrificing the sharpness of the final image.
Four sharpness measures were evaluated based on their ro-bustness to noise, applicability to different specimens, im-plementation cost, and adherence to the strict unimodality property. A strictly unimodal sharpness measure has a sin-gle peak at the best focus and is strictly decreasing away from this peak. Strict unimodality is particularly important to the success of more sophisticated search algorithms. The sharp-ness measures considered were based on the image gradient, sum of specific frequency domain components, image auto-correlation, and image variance.
The gradient measure was found to be the most suscepti-ble to noise, while the variance method was largely insensi-tive to noise. The auto-correlation measure was usually strictly unimodal but had poor reproducibility, and while the frequency domain measures performed well, the imple-mentation cost of performing frequency domain transforms in software was significant. The variance measure was cho-sen as the primary sharpness measure for this work because of its strict unimodality regardless of noise, as well as its sim-ple imsim-plementation.
Signal-to-noise (SNR) ratio and objective lens hysteresis ef-fects are important concerns when developing sharpness search
algorithms. In this work, SNR issues were addressed by choos-ing a moderate level of hardware noise reduction and uschoos-ing a robust sharpness measure. Hysteresis was compensated for by resetting the focal length between iterative sweeps and by di-rect calculation of a hysteresis offset when returning to the best focus following a focus sweep. This hysteresis offset was cal-culated using a new technique based on the relative sharpness of images obtained during and after the sweep.
The fixed step-size search required the most images for a given accuracy. Stopping the search after detecting a sharp-ness peak improved performance but is applicable only under high SNR conditions. The iterative search (using mul-tiple sweeps at decreasing step-sizes) required fewer image captures but additional hysteresis corrections. Following these searches, interpolation can be used to further improve the estimated best focus. A new interpolation technique was developed which uses a model of the image variance as a function of defocus and was found to be reasonably suc-cessful under most test conditions, but adds significant over-head to the search.
The variable step-size search attempts to decrease the step-size during the sweep as the sharpness increases. This technique requires careful initial step-size choice, and is less effective in low SNR conditions. The Fibonacci search partitions the search space into segments and iteratively re-duces the search space using additional, carefully placed sharpness measurements. This technique requires the fewest images, but like the variable step-size search, it relies heav-ily on the strict unimodality property.
The search algorithms were tested on various standard and non-biological samples, using both stored focus series and live images on a LEO 440. Their results were verified against maximal sharpness and human assessments of “best focus”.
This work has investigated the potential of more sophis-ticated sharpness search algorithms which require slightly higher SNR requirements but significantly fewer image cap-tures, and introduced novel hysteresis correction and inter-polation techniques.3The Fibonacci search using a variance sharpness measure is robust and efficient for rapid fine fo-cusing, and combined with the other methods could create a complete full focusing solution.
Acknowledgments: The authors would like to thank the Winston Churchill Scholarship Foundation, LEO Electron Microscopy Ltd., and the Isaac Newton Trust for their support of this research.
References
1. Caldwell NHM, Breton BC, Holburn DM: WebXpertEze: In-telligent instruments via the internet. In 12thEuropean Con-gress on Electron Microscopy, Czechoslovak Society for Elec-tron Microscopy, (Brno) 3, 411–412 (2000)
2. Tee WJ, Smith KCA, Holburn DM: An automatic focusing and stigmating system for the SEM. Journal of Physics E: Scientific Instrumentation, 12, 35–38 (1979)
3. Batten CF: Autofocusing and Astigmatism Correction in the Scanning Electron Microscope, Master’s Thesis, University of Cambridge (2000) (unpublished)
Effect of pseudo-random scan parameters on nega-tive specimen charging and beam landing errors in the scanning electron microscope
J.T.L. THONG, W.K. WONG, A. ZAINAL
Centre for IC Failure Analysis and Reliability (CICFAR), Faculty of Engineering, National University of Singapore
The use of a pseudo-random scanning algorithm1has previously been shown to reduce image degradation due to negative charging effects when specimens with insulating areas are viewed at high-beam energies in the scanning electron microscopy (SEM). On insulating specimens with slight conductivity, random scanning reduces the occur-rence of surface breakdown and Malter discharge. Beam deflection errors and contrast artefacts can also be effec-tively reduced. The principle underlying the reduction of surface charging is to spread out deposition of charge by the beam over the scan field in a manner that consecutive pix-els are well separated.2On a slightly conductive specimen, this allows surface charge to leak away at any point in the scan field before the beam revisits its vicinity again. With a conventional raster scan, consecutive pixels are adjacent to one another, which results in a rapid build-up of charge that can lead to dielectric breakdown and other severe charging artefacts.
The pseudo-random algorithm is based on a pixel inter-lace factor (IF) which is the number of pixels to increment on a one-dimensional raster scan pixel map, and remapped onto a two-dimensional scan map. A judicious choice of IF is found to be important as it affects the beam landing error due to the charge pattern that has been already been estab-lished on the specimen surface by the time sequence of charge deposition. Assuming an initially uncharged surface for a conventional raster scan, the total charge build-up in-creases monotonically in the frame scan direction, and this results in a wider scan field towards the end of the frame.
This effect is acceptable especially if the rate of charge build-up is relatively small. On the other hand, with random scanning, the beam vectors all over the scan field, and de-pending on the extant surface charge, may result in periodic beam landing errors that manifest as “jagged” images. It was observed experimentally that certain IF give images that are practically free of jagged edges.
To understand this phenomenon, computer simulation was carried out to model the effect of evolving surface charge on the beam landing error. In the model, the primary beam is scanned over a dielectric surface according to the pseudo-random scan algorithm. At any irradiated point, the surface charge is assumed to redistribute uniformly by electron beam-induced conductivity (EBIC) within the backscattering radius of the beam. The trajectory of the pri-mary beam is then traced through the electric field result-ing from the surface charge, and this determines the land-ing position of the next point to be irradiated. In the present model, charge decay is assumed to occur through the
sub-strate but does not otherwise redistribute laterally on the surface when the beam is not positioned at the area con-cerned. The charging state of the surface and the surface po-tential can be visualised as a three-dimensional surface plot as scanning progresses.
The beam landing error for a raster scan is as expected, with the scan field expanding towards the end of the frame.
However, at higher charge doses (higher beam currents), the beam landing error interacts with the charge deposition and results in ripples appearing in the charge distribution.
For random scanning, the IF affects the distribution of charge, but for sensible values of IF the difference in the maximum charge is marginal and is invariably lower than that of a raster scan. However, the beam landing error can take on very different forms as the IF is changed. Certain scan sequences result in monotonic distortion of the scan field, while others give rise to seemingly erratic beam po-sitioning errors, supporting the experimental observations of random-scan images. A truly random scan, where the scan coordinates are generated by a pseudo-random num-ber generator, gives the worst possible scenario with un-predictable beam landing error and an irregular surface charge distribution. On the other hand, IF-based pseudo-random scanning with a proper choice of IF can yield the benefits of the technique without the effects of erratic beam positioning errors.
References
1. Lee KW, Thong JTL, Wong WK: Reduction of charging effects using pseudo-random scanning In the scanning electron micro-scope, Scanning 22, 112–113 (2000)
2. Postek MT: Low accelerating voltage inspection and linewidth measurement in the scanning electron microscope, scanning electron microscopy Part III. 1065–1074 (1984)
Acknowledgments: Supported by the Horserace Betting Levy Board.
Specimen charging characterization using com-puter-based image contrast-emission processing for the scanning electron microscope
W.K. WONG, Y.Z. WEI, J.C.H. PHANG, J.T.L. THONG Centre for Integrated Circuit Failure Analysis and Reliability (CICFAR), Faculty of Engineering, National University of Singapore, Singapore
Specimen charging problems are frequently encountered in imaging with the scanning electron microscope (SEM).
Such problems are especially common during the observa-tion of poorly conducting specimens using high-beam en-ergies and at high-current densities associated with high magnifications.1Therefore, in situ imaging analysis in the form of charging detection and quantification to derive ad-ditional information on the sample charge state is needed if
consistent performance from the SEM is desired. One use-ful way of representing the information on the charging at-tributes of the sample is in the form of a 2-dimensional charging map of the sample that can be derived from a com-bination of both physical and image-based measurements.
The basis for an image-based charging detection tech-nique can be constructed based on the fundamental emis-sion behavior of insulators under an electron-beam probe as described by the electron emission coefficient, σ which represents the ratio of the total emission current over the primary beam current. For most typical semiconducting and insulating materials but excluding metals, σ can span ranges both within and in excess of unity depending on the effective electron landing energy.2This property allows different charging-characteristic zones to be investigated at different beam energies. For example, in the σ > 1 region, in which more electrons are emitted than injected into the sample, the resultant net positive charging situation results in a relatively darker image intensity of the final secondary electron (SE) image of the sample due to suppression of electron emission from the now positively charged sur-face.3Conversely, operation in regions where σ < 1 results in a net higher image intensity due to negative charging. It is assumed that all primary electron-optical parameters re-main constant and charge proximity or neighborhood ef-fects4are negligible. A means of mapping out regions that exhibit this contrast fluctuation with beam voltage can then be constructed to form a charging map technique for charg-ing detection.
In the first case of moderate charging in which the charg-ing artifact is manifested as relatively subtle contrast aber-rations in the resultant image without gross image distor-tions or drifts in the image field-of-view, a method employing voltage-by-voltage differential beam energy analysis (BEA) has been developed. Using beam energy al-teration as the primary stimuli, the charging condition changes according to the total yield curve behavior of the regions of interest. In the subsequent processing of the image data obtained, two methods, the maximum distance distribution (MDD) and the least misclassified distribu-tion (LMD) techniques, have been implemented. Compar-ing the contrast on each combination of image frames ob-tained under different beam energies, the combination that gives maximum discrimination of the feature-of-interest from the noncharging background is selected. It was found that the MDD method generally requires less computation at the expense of the higher charging region misclassifica-tion error rate. The LMD algorithm, on the other hand, re-quires larger amounts of computation and hence results in longer charging detection turnaround times, but with bet-ter charging detection accuracy.
A secondary algorithm has also been developed to detect relatively severe charging, where charging artifacts are at-tributed to beam deflection effects on the primary electron beam by strong surface electric field build-up.5In this case, an image drift detection technique is implemented as the primary detection parameter. Using a combination of
frame-by-frame comparison, drift detection, and thresh-olding techniques, a charging map is also generated to de-scribe the zones of severe charging.
The charging map generation techniques described above provide the critical information required for the de-tection and eventual compensation of specimen charging.
The above-mentioned modules are part of an overall charge control architecture for the SEM which is presently under development at CICFAR. The successful development of charge control techniques would ultimately result in self-optimising SEMs where the accuracy of analysis in the face of charging can be more reliably ascertained com-pared with present conventional SEMs.
References
1. Wong WK, Phang JCH, Thong JTL: Charging control using pulsed scanning electron microscopy. Scanning 17, 312–315 (1995)
2. Joy DC: A model for calculating secondary and backscattered electron yields. J Microsc 147, 51–64 (1987)
3. Joy DC: Control of charging in low-voltage SEM. Scanning 11,1–4 (1989)
4. Postek MT: Low accelerating voltage inspection and linewidth measurement in the scanning electron microscope. Scan Elec-tron Microsc 1984, Part III, 1065–1074
5. Shaffner TJ, Hearle JWS: Recent Advances in Understanding Specimen Charging. Proc Ninth Annual Scan Electron Microsc Symposium. IIT Research Institute, Chicago (Ed. Johari O). I, 61–70 (1976)
Visualization of the energy band contrast in SEM scanning electron microscopy through low-energy electron reflectance
I. MÜLLEROVÁ, L. FRANK, O. HUTAR
Institute of Scientific Instruments AS CR, Brno, Czech Republic
The trend towards lowering the primary beam energy in the scanning electron microscope (SEM), motivated by the aim of observing non-conductors with suppressed charging-up, and at of improving the surface relief visual-ization, reaches its limits (in conventional instruments) slightly below 1 keV, as a result of growing aberrations and decreasing current. Further energy lowering is possible when non-constant beam energy along the column is em-ployed and the final deceleration is performed close to the specimen surface. In cases in which the decelerating field is applied directly to the specimen, i.e.that is, when a “cath-ode lens” is introduced, the beam energy can be decreased to units of eV without any significant sacrifice to the reso-lution.1The specimen is then exposed to a strong electric field of, say, 2 to 3 kV/mm. This does not, however, create a gradient exceeding 10 volts per micrometer in the silicon oxide, so most semiconductor devices should not be en-dangered.
Technological development and detection of defects for modern large integration semiconductor devices require the use of high-resolution visualization of topology of dif-ferently doped areasdoping profiles. In order toTo restrict the radiation damaged area to the shallowest surface, and also to decrease the energy delivered, low- energy electrons should be used for this purpose. Successful imaging of both p- and n-type doping within a broad energy range 2has been explained on the basis of the surface band bending cre-ating a variable height barrier for the secondary electrons (SE), causing the p-type areas to be brighter in the image.
Some recent results 3regarding the low- energy range and contrast between p+and n+areas with respect to n-type sub-strate are also interpreted with the help of the surface bar-riers connected with the carbonaceous contaminants.
True observation of the electronic structure is, never-theless, possible when employing very low- energy electron reflectance, as can be seen from the I-V curves for the I00 spot of the low energy electron diffraction (LEED) pattern.
At energies below approx. 20 to 30 eV, where all mecha-nisms of inelastic electron scattering settle enough so that both the energy and wave vector of the impacting electron conserve over the electron penetration depth, the reflection rate is inversely proportional to the density of states coupled to the incident wave.4, 5. More simply, electrons hitting the forbidden gaps above the vacuum level are reflected sig-nificantly more than when their energy corresponds to al-lowed states. This phenomenon has been demonstrated in LEED5and also in the low- energy electron microscope (LEEM).4. Implementation into SEM is likely to bring a novel diagnostic method for semiconductor technology.
Distribution of the energy band features within semi-conductor devices, caused by local doping, can be visual-ized when tuning the (very low) landing electron energy so that it fits forbidden gaps only in a part of the variously doped areas, while in the remaining areas electrons pene-trate inside as hot electrons.6. Note that SE emission does not play a role here, and that the contrast is due to reflected electrons only. Fine energy tuning could even reveal, at least at higher doping levels, penetration into donor or acceptor bands, depending on the shape and height of the band bending. Demonstration experiments are best performed on a partly processed substrate with the first doping operation performed, but with no further layers grown.
References
1. Frank FL, Müllerová I: Strategies for low- and very-low-energy SEM. J. Electron Microsc. 48, No. 3, 205–219 (1999) 2. Perovic DD, Castell MR, Howie A, Lavoie C, Tiedje T, Cole
JSW: Field-emission SEM imaging of compositional and dop-ing layer semiconductor superlattices. Ultramicroscopy 58, 104–113 (1995)
3. El-Gomati M, Wells TCR, Frank L, Müllerová I: On the imag-ing of semiconductor dopimag-ing usimag-ing low energy electron micro-scopy. Proc. 12thEur. Congr. El. Microsc. (edEd. Frank FL, Ciampor F), ). CSEM, Brno 2000, Vol . II, 2000, 635–636 4. Bauer E: Low energy electron microscopy. Repts Progr. Phys.
57, 895–938 (1994)
5. Bartos I, van Hove MA, Altman MS: Cu(111) electron band structure and channelling by VLEED. Surface Sci, 352, 660–664 (1996)
6. Howie A: Electron microscopy of electronic structure and be-haviour. Proc. Internat. Centennial Symp. on the Electron (edEd.
A Kirkland, PD A, Brown PD), Cambridge (1997), 135–145
Imaging of unstained and uncoated specimens in the scanning electron microscope at optimum electron energy
L. FRANK, I. MÜLLEROVÁ, J. KÁNOVÁ
Institute of Scientific Instruments AS CR, Brno, Czech Republic
The conventional way of suppressing charging in non-conductive specimens due to electron bombardment in the scanning electron microscope (SEM), consisting of stain-ing by metallic salts and conductive coatstain-ings, causes al-teration to the acquired data, and much effort is being made to avoid these steps of preparation. One currently popular
The conventional way of suppressing charging in non-conductive specimens due to electron bombardment in the scanning electron microscope (SEM), consisting of stain-ing by metallic salts and conductive coatstain-ings, causes al-teration to the acquired data, and much effort is being made to avoid these steps of preparation. One currently popular