With EUV reticle features shrinking and becoming more complicated, conventional 193 nm inspection tools pose significant challenges due to poor signal-to-noise (SNR) ratio, low optical image stability, and limited data processing capability. To meet these challenges, we have implemented machine learning techniques in EUV reticle inspection starting from advanced technology nodes, which effectively improve defect SNR and eliminates false counts. This accomplishment has been made possible through a combination of several innovations addressed in KLA’s third generation Teron™ 640e Series system: 1. Aberration Control Compensation Techniques: These techniques reduce intrinsic optical noise, enhancing accurate defect detection. 2. Focus drift improvement: Controlling focus drift within tens of nanometers through a full mask area scan is achieved by deploying high-frequency focus trajectory calibration and a low thermal expansion stage. 3. X30 Die-to-Database (DB) Inspection Mode: Leveraging Gen-2 deep learning algorithms, this encompasses a comprehensive analysis of layout dimensions and the integration of design elements through to the final pattern generation. The objective is to enhance the modeling process, thereby diminishing noise levels for superior inspection sensitivity. 4. Curvilinear-Friendly Geometry Classification Scheme: KLA-designed Gen-2 feature map for advanced inspection sensitivity control. 5. Enhanced Data Preparation Server: Efficiently handling data sizes of OASIS P49 MULTIGON format four times larger than traditional Manhattan formats, this server ensures comparable data preparation time. The 3rd generation Teron™ 640e Series system has been demonstrated to meet production requirements for N technology node and beyond. The next step will focus on cutting-edge optical and algorithm design to overcome resolution limitations and implementing these advanced technologies in the most suitable areas of EUV mask inspection.
A proper surface treatment, such as O2 plasma, helps to improve particle removal efficiency (PRE) because of the
formation of hydrogen bonding between particles, water and the mask surface after treatment. The effectiveness of
surface treatments cannot be determined only by the static wettability after processes. More key indexes should be
considered. In this paper, we report our findings on the relationship between surface treatments on photomasks and the
resulting wettability. In addition, added defects after the treatment and the cleaning process were inspected with a 193-
nm KLA inspector on 193-nm immersion and EUV photomasks, which consist of SiO2, MoSi, Cr, Ta-based absorber
and Ru. Based on our work, three indexes can be built for determining the effectiveness of surface treatments. The first is
to check whether the surface becomes super-hydrophilic after treatment. The second is to determine the efficiency of
surface treatments on enhancing wettability. The last is to quantify the added watermark count after the surface treatment
and the cleaning process. With a proper surface treatment, watermarks can be greatly eased. These three indexes can
quickly determine possible effective methods for treating the surfaces of different materials.
We present a series of baseline techniques for inspection, cleaning, repair, and native defect mitigation of extreme ultraviolet (EUV) masks. Deep-ultraviolet inspectors are capable of inspecting patterns down to about 45 nm in pitch on wafer. Cleaning methods involving both chemical and physical forces have achieved good particle removal efficiency while minimizing absorber shrinkage and have realized 90% PRE in removing particles from the backside of an EUV mask. In addition, our compensation method for native defect repair has achieved partial success.
We investigated methods to extend the damage-free process window for fragile Sub-Resolution Assist Features (SRAF)
in mask cleaning using MegaSonic and binary spray techniques. Particle removal efficiency (PRE) was found to increase
by 8% and damage reduced from 7 ppm to 0 ppm with the optimization of the spray droplet characteristics through liquid
media control. MegaSonic damage was eliminated completely from 10 ppm to 0 ppm by varying physical and chemical
properties of the cleaning media. Since particles in the deep trenches are very difficult to remove using droplet spray
alone, a combination of MegaSonic and Binary Spray processes was tested. The acoustic effects generated through the
MegaSonic combined with optimized droplet impact showed an improvement of 4% in PRE of hard-to-remove trench
particles. Overall, the improved process points to a promising solution for overcoming the roadblock in mask cleaning
for the advanced mask cleaning.
As semiconductor manufacturing advances to sub-20-nm nodes, specification (size < 50 nm) for extremely fine particles on photomasks is getting more and more stringent. Photomask cleanliness, which seriously impacts manufacturing cycle time and productivity, is a serious challenge in the development of sub-20-nm node mask cleaning process. Several cleaning approaches, including the use of chemical and physical forces, are widely used in mask cleaning. In this study, we focus on the chemical force through zeta potential (ZP). ZP indicates the degree of repulsion between the particles and the mask surface (mostly quartz). In the nano-scale, stronger repulsion means easier removal of particles from mask surfaces. By controlling ZP of different chemicals from -10 mV to -150 mV in the cleaning process, the particle removal efficiency (PRE) is further improved by about 10%, especially for extremely fine particles. The ZP measurement methodology for different cleaning chemicals on quartz surface is also carried out. ZP is a helpful index in evaluating the performance of new chemicals for mask cleaning. To enhance photomask cleaning for sub-20-nm nodes, the chemical force needs to be increased because the physical force has been constrained to avoid pattern damages, especially when much smaller assistant features are commonly used to gain a greater lithography process window. How to choose a suitable cleaning approach for the next generation mask cleaning is very critical.
This paper studies the repeatability and the reliability of CDUs from a mask inspector and their correlation with CD
SEM measurements on various pattern attributes such as feature sizes, tones, and orientations. Full-mask image
analysis with a mask inspector is one of potential solutions for overcoming the sampling rate limitation of a mask
CD SEM. By comparing the design database with the inspected dimension, the complete CDU behavior of specific
patterns can be obtained without extra work and tool time. These measurements can be mapped and averaged over
various spatial lengths to determine changes in relative CDU across the mask. Eventually, success of this
methodology relies on the optical system of the inspector being highly stable.
A programmed-defect mask consisting of both bump- and pit-type defects on the LTEM mask substrate has been
successfully fabricated. It is seen that pit-type defects are less printable because they are more smoothed out by the
employed MLM deposition process. Specifically, all bump-type defects print even at the smallest height split of 1.7 nm
whereas pit-type defects print only at the largest depth split of 5.7 nm. At this depth, the largest nonprintable 1D and 2D
defect widths are about 23 nm and 64 nm, respectively.
We report inspection results of early 22-nm logic reticles designed with both conventional and computational
lithography methods. Inspection is performed using a state-of-the-art 193-nm reticle inspection system in the reticleplane
inspection mode (RPI) where both rule-based sensitivity control (RSC) and a newer modelbased
sensitivity control (MSC) method are tested.
The evaluation includes defect detection performance using several special test reticles designed with both conventional
and computational lithography methods; the reticles contain a variety of programmed critical defects which are
measured based on wafer print impact. Also included are inspection results from several full-field product reticles
designed with both conventional and computational lithography methods to determine if low nuisance-defect counts can
be achieved. These early reticles are largely single-die and all inspections are performed in the die-to-database
inspection mode only.
The fundamentals of droplet-based cleaning of photomasks are investigated and performance regimes that enable the use
of binary spray technologies in advanced mask cleaning are identified. Using phase Doppler anemometry techniques, the
effect of key performance parameters such as liquid and gas flow rates and temperature, nozzle design, and surface
distance on droplet size, velocity, and distributions were studied. The data are correlated to particle removal efficiency
(PRE) and feature damage results obtained on advanced photomasks for 193-nm immersion lithography.
A new DUV high-resolution reticle defect inspection platform has been developed. This platform is designed to meet the reticle qualification requirements of the 65-nm node and beyond. In this system, the transmitted and reflected inspection lights are collected simultaneously to produce reticle images at high speed. Transmitted and reflected inspections in the die-to-die (DD) and the die-to-database (DB) modes can be executed concurrently. Both images can be gathered at full synchronization with low noise. Basically, both inspection modes are needed to detect as many types of hard and soft defects as possible. Concurrent inspection saves time from using transmitted and reflected lights sequentially. In this presentation, results of DD and DB inspection using standard programmed defect test reticles as well as advanced 65-nm production reticles, are given, showing high-sensitivity and low-false-count detections being achieved with low operating cost.
As the lithography design rule of IC manufacturing continues to migrate toward more advanced technology nodes, the mask error enhancement factor (MEEF) increases and necessitates the use of aggressive OPC features. These aggressive OPC features pose challenges to reticle inspection due to high false detection, which is time-consuming for defect classification and impacts the throughput of mask manufacturing. Moreover, higher MEEF leads to stricter mask defect capture criteria so that new generation reticle inspection tool is equipped with better detection capability. Hence, mask process induced defects, which were once undetectable, are now detected and results in the increase of total defect count. Therefore, how to review and characterize reticle defects efficiently is becoming more significant.
A new defect review system called ReviewSmart has been developed based on the concept of defect grouping disposition. The review system intelligently bins repeating or similar defects into defect groups and thus allows operators to review massive defects more efficiently. Compared to the conventional defect review method, ReviewSmart not only reduces defect classification time and human judgment error, but also eliminates desensitization that is formerly inevitable. In this study, we attempt to explore the most efficient use of ReviewSmart by evaluating various defect binning conditions. The optimal binning conditions are obtained and have been verified for fidelity qualification through inspection reports (IRs) of production masks. The experiment results help to achieve the best defect classification efficiency when using ReviewSmart in the mask manufacturing and development.
The standard inspection flow typically consists of transmitted light pattern inspection (die-to-die or die-to-database) and STARlightTM (Simultaneous Transmitted And Reflective Light) contamination inspection. The initial introduction of TeraScan (DUV) inspection system was limited to transmitted pattern inspection modes. Hence, complete inspections of critical mask layers required utilizing TeraScan for maximized pattern defect sensitivity and the previous generation TeraStar (UV) for STARlightTM contamination inspection.
Recently, the reflective light die-to-database (dbR) inspection mode was introduced on the DUV tool to compliment transmitted light die-to-database (dbT) inspection. The dbR inspection mode provides not only pattern inspection but also contamination inspection capabilities.
The intent of this evaluation is to characterize the dbR inspection capability on pattern defects and contaminations. A series of standard programmed defect test plates will be used to evaluate pattern inspection capability and a PSL test plate will be used to determine the contamination performance. Inspection results will be compared to the current inspection process of record (dbT + STARlightTM).
Lastly, the learning will be used to develop and implement an optimal dbR inspection flow for selected critical layers of the 65-nm node to meet the inspection criteria and minimize the cycle time.
The paper presents the results of a study to define a production-worthy inspection technique for subresolution solid and hollow scattering features used in 193-nm lithography. Masks are inspected using conventional high-NA and aerial-imaging-based mask inspection tools. Inspection results are compared regarding capture rate and nuisance defect rate.
As the lithography design rule of IC manufacturing industry migrates into sub-130nm nodes, low k1 factor prevails, the mask error enhancement factor (MEEF) increases. Low k1 processing calls for aggressive sub-resolution assist features and the use of attenuated phase shift masks (AttPSMs). The aggressive OPC features pose challenges to reticle inspection due to high false detection, which is time-consuming for defect classification and impacts the throughput of mask manufacturing. Moreover, the high transmission of the shifter material of 193 nm AttPSM also challenges the UV-based reticle inspection tools with high nuisance counts due to undesirable optical diffraction effects. For a given reticle inspection tool, it is necessary to calibrate the system contrast between the clear and opaque regions (quartz/chrome or quartz/MoSi) of the reticles. In this study, we present the influences of various calibration conditions on sensitivity, false and nuisance detection of reticle inspections. Both the STARlight contamination inspections and the die-to-die pattern inspections were carried out using the KLA-Tencor TeraStar inspection tools with production masks and programmed defect test masks including binary intensity masks (BIMs) and AttPSMs. Successful applications with low false detection and adapted sensitivity will be illustrated in terms of optimizing the calibration setup.
Inspection of aggressive OPC represents one of the major challenges for today's mask inspection methodologies. Systems are phased with high-density layouts, containing OPC features far below the resolution limit of conventional inspection systems. This causes large amounts of false and nuisance defects, especially on production applications. The paper presents the use of Aera193, a new inspection system using aerial imaging as inspection methodology.
Following mask inspection, mask-defect classification is a process of reviewing and classifying each captured defect according to prior-defined printability rules. With the current hardware configuration in manufacturing environments, this review and classification process is a mandatory manual task. For cases with a relatively small number of captured defects, defect classification itself does not put too much burden to operators or engineers. With a moderate increase of defects, it would however, become a time-consuming process and prolong the total mask-making cycle time. Should too many nuisance defects be caught under a given detection sensitivity, engineers would generally loosed the detection sensitivity in order to reduce the number of nuisance defects. By doing that however, there exists potential threat of missing real defects. The present study describes a 'progressive self-learning' (PSL) algorithm for defect classification to relieve loading from operators or engineers and further accelerate defect review/classification process. Basically, the PSL algorithm involves with image extraction, digitization, alignment and matching. One key concept of this PSL algorithm is that there is not any pre-stored defect library in the first place of a particular run. In turn, a defect library is 'progressively' built during the initial stage of defect review and classification at each run. The merit of this design can be realized by its flexibility. An additional benefit is that all defect images are stored and suitable for network transfer. The C language is adopted to implement the present algorithm to avoid the porting issue, so as not bound to a particular machine. Assessment of the PSL algorithm is examined in terms of efficiency and the accurate rate.
Recent observations indicate that wafer CD control for the 0.13-micrometers node is sensitive to non-phase defects between 0.1 micrometers and 0.25 micrometers on a 4X reticle, as a function of the location of the sub-killer defect. Since more and more small defects can be detected by today's advanced mask defect inspection tools, it is important to determine whether these detected defects can impact wafer lithography process window. The experimental result is based on a typical 0.13-micrometers process using a pre-designed Defects Sensitivity Monitor reticle to address the printability of these programmed defects.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.