S-Genius is a new universal scatterometry platform, which gathers all the LTM-CNRS know-how regarding the rigorous
electromagnetic computation and several inverse problem solver solutions. This software platform is built to be a userfriendly,
light, swift, accurate, user-oriented scatterometry tool, compatible with any ellipsometric measurements to fit
and any types of pattern. It aims to combine a set of inverse problem solver capabilities — via adapted Levenberg-
Marquard optimization, Kriging, Neural Network solutions — that greatly improve the reliability and the velocity of the
solution determination. Furthermore, as the model solution is mainly vulnerable to materials optical properties, S-Genius
may be coupled with an innovative material refractive indices determination. This paper will a little bit more focuses on
the modified Levenberg-Marquardt optimization, one of the indirect method solver built up in parallel with the total SGenius
software coding by yours truly. This modified Levenberg-Marquardt optimization corresponds to a Newton
algorithm with an adapted damping parameter regarding the definition domains of the optimized parameters.
Currently, S-Genius is technically ready for scientific collaboration, python-powered, multi-platform
(windows/linux/macOS), multi-core, ready for 2D- (infinite features along the direction perpendicular to the incident
plane), conical, and 3D-features computation, compatible with all kinds of input data from any possible ellipsometers
(angle or wavelength resolved) or reflectometers, and widely used in our laboratory for resist trimming studies, etching
features characterization (such as complex stack) or nano-imprint lithography measurements for instance. The work
about kriging solver, neural network solver and material refractive indices determination is done (or about to) by other
LTM members and about to be integrated on S-Genius platform.
Proc. SPIE. 7638, Metrology, Inspection, and Process Control for Microlithography XXIV
KEYWORDS: Metrology, Optical lithography, Data modeling, Calibration, Inspection, Atomic force microscopy, Scanning electron microscopy, Process control, Optical proximity correction, Current controlled current source
Mask and metrology errors such as SEM (Scanning Electron Microscopy) measurement errors are currently not accounted for when calibrating OPC models. Nevertheless, they can lead to erroneous model parameters therefore causing inaccuracies in the model prediction if these errors are of the same order of magnitude than targeted modeling accuracy. In this study, we used a dedicated design of hundres of features exposed through a Focus Exposure Matrix for the metrology error, we compared the SEM measurements to AFM measurements for as much as 105 features exposed in various process conditions of does and defocus. These data have then been used in a OPC model calibration procedure. We show that the impact of the metrology error is not negligible and demonstrate the importance of taking into account these errors in order to improve the reliability of the OPC models.
In this paper, an ill-posed inverse ellipsometric problem for thin film characterization is studied. The aim is to determine the thickness, the refractive index and the coefficient of extinction of homogeneous films deposited on a substrate without assuming any a priori knowledge of the dispersion law. Different methods are implemented for the benchmark. The first method considers the spectroscopic ellipsometer as an addition of single wavelength ellipsometers coupled only via the film thickness. The second is an improvement of the first one and uses Tikhonov regularization in order to smooth out the parameter curve. Cross-validation technique is used to determine the best regularization coefficient. The third method consists in a library searching. The aim is to choose the best combination of parameters inside a pre-computed library. In order to be more accurate, we also used multi-angle and multi-thickness measurements combined with the Tikhonov regularization method. This complementary approach is also part of the benchmark. The same polymer resist material is used as the thin film under test, with two different thicknesses and three angles of measurement. The paper discloses the results obtained with these different methods and provides elements for the choice of the most efficient strategy.
Optical Proximity Correction (OPC) is used in lithography to increase the achievable resolution and pattern transfer
fidelity for IC manufacturing. Nowadays, immersion lithography scanners are reaching the limits of optical resolution
leading to more and more constraints on OPC models in terms of simulation reliability. The detection of outliers coming
from SEM measurements is key in OPC . Indeed, the model reliability is based in a large part on those measurements
accuracy and reliability as they belong to the set of data used to calibrate the model. Many approaches were developed
for outlier detection by studying the data and their residual errors, using linear or nonlinear regression and standard
deviation as a metric .
In this paper, we will present a statistical approach for detection of outlier measurements. This approach consists of
scanning Critical Dimension (CD) measurements by process conditions using a statistical method based on fuzzy CMean
clustering and the used of a covariant distance for checking aberrant values cluster by cluster. We propose to use
the Mahalanobis distance  in order to improve the discrimination of the outliers when quantifying the similarity within
each cluster of the data set.
This fuzzy classification method was applied on the SEM CD data collected for the Active layer of a 65 nm half pitch
technology. The measurements were acquired through a process window of 25 (dose, defocus) conditions. We were able
to detect automatically 15 potential outliers in a data distribution as large as 1500 different CD measurement. We will
discuss about these results as well as the advantages and drawbacks of this technique as automatic outliers detection for
large data distribution cleaning.
At present, the question of the move from 193 to 157nm lithography is under discussion. There are still several major issues such as the development of 157nm photo-resists and pellicles, as well as calcium-fluoride lens material availability. The extension of the 193nm lithography down to the 65- and 45-nm half pitch technologies is now considered as a serious alternative. This requires several technical challenges with the use of phase shift masks (PSM), optical proximity effects corrections or liquid immersion. Simulation gives information on expected process latitudes and is an important tool to help this technical choice. Previous works have shown that the "Diffused Aerial Image Model" (DAIM) is accurate for CD prediction. Reliable process latitudes can be extracted from the simulated focus-exposure matrices (FEM). The model is used for the process latitudes evaluation of the different lithography approaches possibly used to print the 65- and 45-nm half pitches. 193nm illumination in addition to PSM is compared to 157nm lithography associated with conventional or optimized illumination schemes. This work shows that PSM at 193nm gives generally better exposure latitude for all pitches and CD, and confirms that 193nm lithography is a possible alternative to achieve 45nm and 70nm half pitches patterning. The process windows are nevertheless very small, and huge mask error factors (MEEF) are another sign that printing such small features is close to the physical limit (k1 factor close to the quarter).
Lithography modeling is a very attractive way to predict the critical dimensions of patterned features after lithographic processing. In a previous paper, we have presented the assessment of three different simplified resist models (aerial image model, aerial image convolved with fixed gaussian noise and aerial image convolved with variable gaussian noise) by using a systematic comparison between experimental and simulated data. It has been shown that the aerial image convolved with fixed gaussian noise, or "diffused aerial image model" (DAIM), exhibits surprisingly good results of CD prediction for lines @ 193nm. Using these datasets, the DAIM appeared as an accurate model for CD prediction. This approach allows also an easy run, and because it needs only four adjustable parameters, it avoids the difficult task of resist parameters extraction associated to full resist models.
In this paper, we enlarge the datasets used for the assessment of the DAIM by considering both lines and contact holes of various sizes printed at different wavelengths. The reference wafers have been printed at 248nm, 193nm and 157 nm. The procedure used to extract the model parameters has been improved and now needs less data to provide acceptable values. We will show that the validity of the DAIM extends well outside the results presented in Ref. 1. Experimental data printed using various wavelengths, resists and exposure tools can be simulated accurately with CD prediction error ranging within few percents. It is to be noted that the results that will be presented on contact holes data indicate that the model is valid for 2D features. Finally, a comparison with full resist models shows that the accuracy of DAIM is comparable to more sophisticated and heavier models.
Resist modeling is an attractive way to predict the critical dimensions of patterned features after lithographic processing. Unfortunately, previous works have shown that model parameters are very difficult to determine and have often a poor range of validity outside the dataset that have been used to generate them. The goal of this work is to assess different simplified resist models using a systematic method. We have studied the accuracy of aerial image model and aerial image plus Gaussian noise convolution model. The approach is based on the comparison between simulated and experimental data for periodic lines of various dimensions at various illumination conditions. We also propose a reliable expression for Bossung curves fitting. Using simple physical considerations, the expression has been made very simple and efficient. After a proper setting of the model parameters to the experimental data, mean CD discrepancies between simulation and experiment are as small as 5% and can be 3% for certain feature types. Moreover, we show that simple Gaussian noise convolution models can be predictive with the same accuracy. The method for CD prediction is fully described in this paper. Significant improvements have been made in resists modeling over the last several years, but simplified resist models such as 'aerial image + Gaussian noise' seems to be an effective tool for CD prediction, which remains the major demand of IC manufacturers.