Significance High-precision digital maskless lithography is essential for advanced integrated circuit fabrication, but existing mask-design methods struggle to simultaneously achieve superior pattern fidelity and computational efficiency, especially for complex nested structures. Aim We introduce a multistrategy collaborative optimization framework that integrates genetic algorithm, biogeography-based optimization, and tent map-based chaotic mutation to enhance both accuracy and speed in digital photomask design. Approach Building on the Hopkins imaging model and a tailored multiobjective fitness function, the proposed method merges GA's local search with BBO's global migration strategy. To accelerate convergence and avoid premature stagnation, adaptive parameter tuning, chaotic mutation via a tent map, and an early stopping criterion are incorporated, ensuring a balance between exploration and exploitation without excessive computation. Results For complex nested geometries, the proposed Tent-MP-GA-BBO framework achieved a PE of 1 pixel, maintained SSIM values above 0.99, and reduced computational time by up to 40% compared with the MP-GA-BBO baseline, demonstrating superior contour fidelity and optimization efficiency. Conclusions The integrated mapping strategy effectively harmonizes global exploration and local refinement, delivering a markedly more precise and efficient solution for high-precision maskless lithography mask optimization.
Sub-10-nm spacings on a wafer can be obtained using phase reversal on the photomask in conjunction with the 0.55 numerical aperture extreme-ultraviolet lithography. These spacings are a result of the minimum aerial-image intensity that occurs at the boundary of a pi-phase shifter, which can be fabricated by depositing additional molybdenum-silicon bilayers on top of the standard 40 pairs on the mask blank and selectively removing the added bilayers by masking and reactive-ion etching. Our calculated results are confirmed by rigorous simulations.
Background: Lateral gate-all-around (LGAA) device architectures-including nanosheet and forksheet structures-are becoming the dominant transistor platforms for sub-3 nm technology nodes. These architectures introduce multiple buried features such as silicon-germanium (SiGe) layers, dielectric-wall voids, and channel-release interfaces, many of which lie beneath hard-mask or dummy-gate stacks in actual process flows. As a result, the need for subsurface metrology that can access these features prior to or without cross-sectioning has grown substantially. Aim: The aim is to isolate and evaluate the intrinsic subsurface-visibility capability of high-voltage critical-dimension scanning electron microscopy (HV CD-SEM). This study investigates its feasibility using nanosheet test structures intentionally prepared without upper hard-mask or dummy-gate layers. Approach: Three LGAA-relevant inspection targets were examined: (1) voids within dielectric walls, (2) SiGe lateral recess amounts, and (3) multilayer SiGe residues after channel release. HV CD-SEM images were acquired at elevated landing energies to enhance backscattered-electron (BSE) sensitivity to embedded features. Monte Carlo simulations were performed to interpret voltage-dependent BSE signatures, and selected features were verified by transmission electron microscopy (TEM) or energy-dispersive X-ray spectroscopy (EDX). Results: HV CD-SEM successfully visualized dielectric-wall voids, extracted SiGe recess amounts using SE/BSE paired contrast, and detected SiGe residues originating from distinct depths within the nanosheet stack. Recess extraction was applied at 163 sampling locations, enabling wafer-level evaluation of layout-dependent micro-loading effects, whereas TEM cross-sections provided limited but essential reference values. Simulations correctly predicted the voltage-dependent detectability of multilayer SiGe residues, and EDX confirmed Ge-related contrast in the HV CD-SEM images. Conclusions: These results show that HV CD-SEM possesses intrinsic subsurface-visibility capability when upper hard-mask and dummy-gate layers are absent, demonstrating feasibility for detecting voids, SiGe recess profiles, and multilayer SiGe residues in nanosheet structures. Extending this approach to fully processed LGAA devices-where upper gate-stack layers remain in place-represents an essentially next step toward establishing HV CD-SEM as a comprehensive inline subsurface metrology technique.
Metal oxide resists (MORs) have shown great promise for high-resolution patterning in extreme ultraviolet (EUV) lithography, with potential for integration into high-volume manufacturing. However, MORs have recently been shown to exhibit sensitivity to process conditions and environment, leading to critical dimension (CD) variation. Although this variation can be reduced with proper process control, there is a current lack of fundamental knowledge on how these aspects affect the pattern formation mechanism. Moreover, the diverse composition of atmospheric environments makes it difficult to disentangle the role of individual atmospheric components on the lithographic performance of this promising class of EUV photoresists. To bridge these knowledge gaps, we deploy a coordinated, fundamentals-focused approach to yield deep insights into MOR exposure and process chemistry. Our results on a model MOR, an n-butyl Sn-Ox system, reveal how parameters such as exposure dose, post-exposure bake (PEB) temperature, and atmospheric environment influence the EUV exposure and post-exposure delay (PED) and PEB mechanisms. Using an advanced toolset, we show that EUV-induced ligand cleavage likely occurs via homolytic Sn-C bond breaking, resulting in a Sn-based radical "active site," which serves as a reactive center that endows MOR materials with their sensitivity to atmospheric components (e.g., H2O and O-2 ) during PED and PEB. We resolve the roles of H2O during PED and PEB, and in particular, we show that PEB environments containing O-2 exhibit increased litho performance (reduced dose and improved development contrast), suggesting that O-2 plays a critical role in the exposure and thermal mechanisms of MOR materials. Our results, and the coordinated approach using correlative spectroscopies, provide a strong foundation for understanding the critical EUV exposure and PED and PEB mechanisms in MOR materials, provide insights into potential optimization routes via environmental control during the process, and finally offer the potential to link mechanistic aspects and MOR lithographic performance and stability.
Background Dimensional scaling of line features can lead to line wiggling after development and after the etching process step. The line wiggle (WGL) phenomenon is strongly influenced by the process conditions and by the film stack, due to both thickness and stiffness of the materials involved. Aim We aim to understand the relationship between critical dimensional (CD) and line wiggle, which is essential to identifying material, process, and stack requirements that can minimize the WGL as the CD scales down. Approach The impact of resist properties, development process, and film stack properties online wiggle is investigated by applying the Euler-Bernoulli bending theory to sub-10 nm lines, after development and after etch. Results WGL is mainly driven by mechanical properties of the patterned line features, such as (i) aspect ratio, (ii) material stiffness (Young's modulus), and (iii) etching-induced forces. Conclusions Investigating and optimizing material mechanical properties, stack engineering, and etching is crucial to minimizing the WGL phenomenon in high NA EUV patterning as severe line wiggles can lead to line collapse with consequently low electrical device yield.
Background: Inverse lithography technology (ILT) improves mask printability, but classical optimizers are computationally intensive. Learning-based ILT reduces runtime yet often relies on heavy transformer backbones. Aim: We seek to lower resist L2 error while preserving a compact model that delivers fast turnaround time (TAT) at inference. Approach: The framework follows a U-Net encoder-decoder with skip connections, where the bottleneck incorporates TransMamba blocks that combine multihead attention and Mamba state space modules to capture both global dependencies and local inductive bias. The entire pipeline is trained end-to-end with differentiable lithography simulation and resist modeling, directly supervising mask printability and process variation balance. Results: On ICCAD-2013 benchmarks, HeteroM-ILT reduces resist L2 by 35.47% and 32.95% relative to Neural-ILT and A2-ILT, respectively. In terms of process variation band, HeteroM-ILT achieves a comparable level to prior learning-based baselines, a 0.97% reduction versus Neural-ILT, with a modest trade-off versus A2-ILT. Thanks to the compact Mamba backbone and streamlined decoding, the method achieves a TAT approaching 0.1 s during deployment, providing 42.9 & times; to 106.3 & times; speedup over recent learning-based ILT baselines. Conclusions: We contribute to the semiconductor industry by enhancing manufacturability while significantly reducing computational cost, making large-scale ILT more feasible in advanced technology nodes.
Background Metrology techniques with high accuracy and throughput are needed to address critical challenges in the characterization of periodic nanostructured materials. Copper interconnects are structures for which precise topography characterization is needed. Although common methods include atomic force and scanning electron microscopy, extreme-ultraviolet (EUV) scatterometry has unique capabilities for the characterization of structure parameters and material properties. Aim We study the effect of experimental design parameters (wavelength and incidence angle) on EUV scatterometry. We explore the fundamental sensitivity capabilities of EUV scatterometry for interconnect metrology when data are collected such that the information content in a single EUV diffraction pattern is maximized by tuning the wavelength and incidence angle.Approach We applied experimental design methods based on the Fisher information matrix to EUV scatterometry of an interconnect to obtain the optimized EUV wavelength and angles at which to obtain a single diffraction pattern. Using rigorous coupled wave analysis simulations and Monte Carlo uncertainty quantification, the accuracy of the parameter reconstructions under an optimal design is quantified. Results The uncertainty of the sample parameters depends strongly on wavelength and incidence angle, necessitating experimental design methods to achieve maximal sensitivity. When data are collected under an optimal experimental design, simulations indicate that sub-angstrom accuracy can be obtained from only a single diffraction pattern for the dishing depth of interconnect pads (0.02 nm standard deviation), and the substrate density can additionally be extracted (0.015 g/cm3 standard deviation). Decorrelation of the sample parameters is observed with optimal design, and 13.8 nm EUV light provides better sensitivity than longer wavelengths such as 30 nm. Conclusions Experimental design methods are an essential tool for ensuring that experimental data are collected under an optimal design that maximizes the information in the data. Implementing these methods opens further capabilities for fast and precise interconnect metrology with EUV scatterometry.
Background Image-based overlay (IBO) target is a type of overlay metrology method in semiconductor manufacturing. Focus optimization is a critical factor that determines the quality of the IBO target signals in IBO target measurement. To calculate the overlay, two target images are required. As the semiconductor process becomes complex, optimizing the IBO target focus for both targets becomes increasingly difficult. Aim We target to develop a deep-learning-based method that tackles two issues in IBO target focus optimization. Instead of choosing a compromised focus offset for both targets or measuring different targets sequentially, we train a generative model that outputs IBO target images with different focus offsets. We use the model for the IBO target focus optimization. Approach The proposed method consists of two steps. The first step is to train a model; conditional deblurring network (cDN) generates IBO target images with different camera focus offsets. Our model is based on the deep image deblurring methods using the generative models with auxiliary focus offsets. The second step is to optimize IBO target images with the trained model. We can generate an optimal target image that consists of IBO target images with the best focus for each layer. Results In experiments, we have compared different baseline models for cDN. The best model, which adopted DeblurGAN-v2, achieved a peak signal-to-noise-ratio of 49.04, a Fr'echet inception distance of 1.11, a mean squared error (MSE) of contrast precision of 0.0010, and an MSE of center-of-symmetry of 0.0672. The model can generate focus-adjusted IBO target images comparable to the ground truth. With the trained model, we optimized IBO target images. We achieved the improved target quality compared with the compromised best focus images. Conclusions We proposed a deep-learning-based methodology to solve the focus optimization problem for IBO target images. We demonstrated that our method can solve the IBO target optimization problem more efficiently than existing methods. In addition, we proved the applicability of deep learning in semiconductor manufacturing.
Laser-based patterning in semiconductor manufacturing can suffer critical dimension (CD) distortion when heat diffuses beyond the intended pattern during exposure. We seek beam shape design settings that confine the thermal field within a target CD boundary for shaped beam pattern projection. A continuum mechanics computational framework solves transient heat diffusion under projected spatially uniform and nonuniform beam profiles. We then formulate cost functions that penalize CD deviation for above-threshold heating. A genetic algorithm optimizes beam intensity and pulse durations. For the nonuniform and uniform beams, the optimized minimum costs were 8.87 & centerdot;10(-5) and 2.75 & centerdot;10(-3), with above temperature threshold areas of 1.11 & centerdot;10(-12) m(2) and 1.33 & centerdot;10-12 m(2), respectively. With 2% intensity fluctuations applied to the optimal uniform beam, critical dimension uniformity (CDU) improved by 57.1% when the fluctuations were skewed toward the lower bound of the GA-optimal intensity. Optimizing spatiotemporal beam delivery can reduce CD distortion and overheating, and the sensitivity to upward power drift motivates robust beam-control strategies for thermal-constrained patterning.
The rapid advancement of the semiconductor industry, with technologies now reaching single-digit nanometer scale, has significantly increased the complexity of integrated circuit manufacturing. This complexity is further amplified by the use of a wide variety of compounds used across multiple process steps, making yield enhancement increasingly dependent on stringent contamination control. Although contamination control in earlier decades primarily focused on particles and a limited set of gaseous contaminants, today, even trace levels of foreign molecules-regardless of their chemical category (acidic, basic, condensable, etc.)-can pose serious risks to process integrity and device performance. As a result, real-time monitoring of airborne molecular contaminants (AMCs) with single-digit parts-per-trillion volume (pptv) sensitivity has become essential. We present a metrology tool based on chemical ionization time-of-flight mass spectrometry, designed for comprehensive, real-time measurement of AMCs across all chemical categories. We demonstrate its application through representative, real-world use cases within semiconductor manufacturing environments covering stationary applications (manifold monitoring), arising monitoring approaches (mobile and point-of-use), material off-gassing, and emerging contaminants monitoring, showcasing its ability to sensitively detect both well-known and emerging AMC species such as per- and polyfluoroalkyl substances (PFAS), a class of compounds of increasing concern due to their environmental persistence and potential health impacts. Although traditional AMC monitoring primarily supports yield and process reliability, the detection of PFAS highlights the broader relevance of advanced contamination metrology for sustainability and workplace safety initiatives. These results emphasize the tool's dual role in both enhancing yield and supporting responsible manufacturing practices in advanced semiconductor production.
The semiconductor industry relies on developments in extreme ultraviolet lithography (EUVL) to sustain the continued dimensional scaling of integrated circuits. The recent introduction of the ASML EXE:5000 enables such scaling by facilitating high numerical aperture (High NA) EUVL. To achieve this, the scanner utilizes anamorphic optics that compress the exposure field along the y-axis by a factor of 2. Consequently, stitching of multiple images is required to produce large chips with High NA EUVL. Although at-resolution stitching is demonstrated with Ta-based masks, this is not the case for alternative absorber material masks, such as high-reflectivity low-n masks. Due to high absorber reflectivity, these masks require an additional design solution to enable at-resolution stitching, as discussed in this study. Through the local insertion of sub-resolution gratings (SRGs) in the mask design, we show that high absorber reflectivity can be reduced to the reflectivity level of a Ta-based mask. In a wafer exposure study at 0.33NA, we investigated stitching metrics to demonstrate the insertion of SRGs in the mask design as an effective design solution to suppress low-n absorber reflectivity in the stitching region. Experiments are ongoing to evaluate the current findings through the mirrors of the ASML EXE:5000 scanner for High NA EUVL.
Background: Source mask optimization (SMO) is a crucial technique in photolithography, particularly for sub-resolution printing. It enhances the ability to produce complex patterns and significantly increases the process window, often by a factor of 2x to 6x. Traditional SMO methods, however, can be computationally intensive and time-consuming. Aim: This research aims to introduce a machine learning (ML)-based method for enhancing SMO by predicting seed sources using a convolutional neural network (CNN), thereby improving efficiency and integration in photolithography processes. Approach: The proposed method leverages a CNN within parametric SMO simulations to optimize source patterns. This CNN-based approach processes multiple clip inputs essential for source optimization while maintaining independence from aerial image parameters, thereby improving the adaptability and robustness of the optimization process. Results: The experimental implementation of the CNN model resulted in a significant 87.5% increase in speed compared with conventional optimization techniques. Furthermore, the method was seamlessly integrated into existing SMO workflows. Conclusion: The study highlights the significant potential of applying ML, especially CNNs, to the SMO process. This approach leads to faster optimization, improved print fidelity, and a wider process window, making it a valuable addition to advanced photolithography techniques.
Background: Due to the significant advantages of curvilinear masks, a fundamental transition from traditional Manhattan masks to curvilinear masks is being actively pursued in extreme ultraviolet lithography. However, thick-mask models capable of handling arbitrary mask geometries remain understudied in current research. Aim: As the mask pattern undergoes a fundamental transition, there is an urgent need to develop a fast and accurate thick-mask model capable of handling arbitrary mask geometries, particularly for curvilinear masks. Approach: Based on the physical mechanisms of the mask 3D effect, we propose a slope chain code with direction (SCCD) method for accurate boundary characterization, which enables the segmentation of the mask into distinct feature regions based on the extracted features. Then, the diffraction near field (DNF) for distinct feature regions is computed by convolving region-specific mask patterns with their corresponding kernels. These kernels are derived from preestablished training datasets that employ distinct data augmentation strategies for different mask types. Results: Compared with the DNF results obtained from rigorous electromagnetic field (EMF) methods, the proposed model yields a root mean square error of only 1% to 2% across various test mask geometries (including both curvilinear and Manhattan masks), whereas the thin-mask approximation produces significantly larger errors (almost 15%). Moreover, critical dimension (CD) measurements show that the model differs from the EMF results by only 0.14 nm while delivering a computational speed nearly 300 times faster than EMF methods. Conclusions: Simulation results demonstrate the model's consistent accuracy across diverse mask geometries, particularly for curvilinear masks. Given its exceptional computational efficiency, the proposed model shows strong potential for full-mask lithography simulation applications involving arbitrary-shaped thick masks.
Hybrid bonding is a critical part of advanced packaging development, enabling wafer-to-wafer bonding and die-to-wafer bonding for 3D stacking of chips. This process, which is carried out post-individual wafer fabrication, can suffer from void formation either between copper PADs or between the dielectric, leading to failed interconnections in the former case, affecting performance at the die level and rupture of the dielectric layers in the latter case, leading to significant yield loss at the wafer level. Control of critical parameters such as surface roughness, degree of cleanliness, flatness, recess level of the copper PADs, and dielectric surface is necessary at the wafer level, essentially calling for massive metrology. We present such a massive metrology measurement performed on production wafers in the fab using a high-throughput in-line Atomic Force Microscopy-based metrology tool. Measurements were performed on three wafers, which were designed to be the "top" wafers during the bonding step, where a split condition was applied on these wafers, whereas an additional three wafers were measured, which were to be used as the "bottom" wafers during bonding and were all in POR condition. The analysis of the results obtained is also presented for understanding the expected impact on the final performance of the bonding step.
Background:Nanoscale measurements of critical dimensions in semiconductor manufacturing rely on scanning electron microscopy (SEM) and SEM image analyses. The acquisition of SEM images requires a low primary electron beam current and a low dose of the SEM imaging microscope to avoid integrated circuit (IC) sample charging and inflicted damage to sensitive IC structures. These requirements inevitably result in noisy, low-contrast images, which can make traditional SEM image analyses no longer viable. Aim:With the advancement in computational hardware and artificial intelligence (AI) models, IC structure detection via AI-based SEM image segmentation can extend the viability of these measurements from noisy, low-contrast images. However, the use of AI models raises questions about the detection limits of extracted measurements and the confidence in such measurements. Approach:Our approach is to relate SEM image quality characteristics with AI-based object segmentation accuracy to establish detection limits of AI-based models and their relationships to human detection limits. Using SEM image simulation software, we create six image sets of quasi-circular objects on a substrate with varying noise and contrast characteristics. These sets of SEM images are characterized by 25 image quality metrics and then used to train and evaluate three AI models. The 25 SEM image quality characteristics and three AI model accuracy metrics per SEM image define the mapping between the quality of input SEM images and the performance of the trained AI models. Results:We used the mapping to establish the detection limits of trained AI models with respect to a required confidence and then relate the human detection limits to the trained AI model detection limits. The human detection limit was established by Rose as the minimal signal-to-noise ratio (SNR = 5) to reliably delineate the shape and size of objects in an image. We matched the signal-to-noise ratio (SNR) defined by Rose to image quality characteristics and demonstrated the upper and lower SNR bounds for three AI models with respect to the human detection limit and for a specified confidence. Conclusion:We establish a method to determine the detection limits of AI-based SEM dimensional metrology. The study is relevant to semiconductor vendors and consumers of AI models because critical dimension measurements are derived from noisy and low-contrast SEM images using AI models with varying performance characteristics. Given a measured SEM image with estimated noise and contrast characteristics, each AI model will be characterized by unique detection limits that can be trusted in semiconductor production. Our method enables improving the trust in critical dimensions while using advanced AI models.
As high-numerical-aperture (NA) extreme ultraviolet (EUV) scanners have a half-field size, two masks are necessary for exposing a chip that is larger than the half-field of 26 mmx16.5 mm. However, scanner throughput simulations have cleared that it takes a long time to expose the second half-field because the second mask should be loaded on the mask stage after exposing the first half-fields on many wafers to improve the throughput. Consequently, this may affect the stitching area due to a long waiting time after the first half-field exposure until the post-process. Therefore, an introduction of a two-mask stage scanner is proposed. This can drastically reduce the time until exposing the second half-field. However, using a conventional scan with two masks on a single mask stage, the two exposed areas are separate. Therefore, a new scan method is proposed, in which the two exposed areas contact. The scanner throughput simulations indicate that the scanner with the two-mask stage has a 20% to 80% higher throughput than the conventional single-mask scanner. The throughput can be even similar to 85% of that of a scanner with a 6 in.x12 in. mask.
Background: Efficient data representation of piecewise linear masks is essential for balancing mask accuracy and data compression. Traditional mask representations rely on piecewise-linear polygons, but parametric representations, such as Bezier and B-spline curves, offer potential advantages in reducing data size while maintaining accuracy. Previous studies have primarily focused on cubic Bezier and B-spline curves, limiting the exploration of higher-degree curves and different parameterization methods. Aim: We aim to investigate a broader range of Bezier and B-spline curve degrees and parameterization methods to identify the optimal representation that minimizes both mask error and data size. Approach: Target designs are corrected using the inverse lithography technique (ILT), and the least squares approximation is employed to fit Bezier and B-spline curves to the mask designs. Various curve degrees and parameterization methods are examined to optimize the approximation. The root mean square error (RMSE) and vertex reduction are assessed in comparison to the piecewise-linear representation. In addition, aerial image simulations are performed to evaluate the lithography errors introduced by mask conversion, specifically analyzing process variation bands and edge placement error distributions. Results: For logic metal designs, converting piecewise-linear polygons to seventh-degree Bezier curves resulted in an 87.5% vertex reduction with an RMSE of 0.5 nm.Similarly, second-degree B-spline conversion achieved an 86.2% vertex reductionwith a lower RMSE of 0.1 nm. For logic VIA designs, the same conversion methodsexhibited comparable trends. The seventh-degree Bezier representation yielded an81.3% vertex reduction with an RMSE of 0.1 nm, whereas the second-degree B-spline achieved a similar reduction of 81.2% with the same RMSE. The aerial imagesimulations for Logic metal indicate that the errors introduced by both Bezier and B-spline representations are negligible. Conclusions: The results demonstrate that seventh-degree Bezier and second-degree B-spline curves provide significant vertex reduction while maintaining accu-racy, making them viable alternatives to piecewise-linear representations. We high-light the potential of parametric representations in optimizing mask data size whilemaintaining comparable lithographic performance.
Background: A low-n attenuated phase shift mask (low-n mask) has been widely used in the extreme ultraviolet (EUV) lithography industry as it aligns the diffraction phases between the zeroth and first orders and enhances exposure latitude, as well as boosting throughput. However, the aligned phase is achieved only for a specific pitch, usually the densest pitch. A pitch-dependent phase offset inevitably remains in a mask layout with various pitches. This leads to the challenge of best focus (BF) variation, which is even more pronounced in the case of a low-n mask with a large refractive index difference from vacuum. The BF variation through pitch is more critical at higher numerical aperture (NA), where the available depth of focus is significantly lower than at low NA. In addition, side-lobe printing caused by high reflectivity from the low-n mask is also a concern. Aim: We aim to enhance imaging performance and demonstrate CD-based overlapping process window (oPW) while minimizing BF variation and unwanted side-lobe printing. Then, we aim to provide a potential mask absorber roadmap in hyper NA EUVL from an imaging perspective. Approach: We proposed a sub-resolution grating (SRG) crossing the main features perpendicularly, which not only suppresses unwanted side-lobe printing but also improves the imaging quality with a better BF alignment over a broader pitch range by effectively aligning diffraction phases. Our approach involves a step-by-step validation of unit cell patterns to assess the SRG impact on oPW, beginning with various L/S pitches, followed by 2-bar and 3-bar for both horizontal and vertical orientations. Ultimately, we verify oPW using real logic use cases applicable to high- and hyper NA EUV lithography. Results: Rigorous simulation demonstrates improved oPW for both vertical and horizontal logic metal layers. As for the vertical L/S patterns as the SRGs can cover the entire mask pattern area, all types of low-n masks can be utilized. As for the horizontal L/S patterns, covering the entire mask pattern area with the same orientation of horizontal SRGs may be challenging from the perspective of mask manufacturability. To overcome this constraint, alternative absorber structures could be considered. A higher k absorber gives a better imaging performance with less pronounced BF variation. By combining the advantages of these two, a thin high-k absorber structure with selective SRG insertion demonstrated an oPW higher than our set criteria and showed more than 25% improvement in productivity compared with a Ni-based high-k absorber. Conclusions: The SRG insertion technique, along with proper wavefront optimization, is expected to offer tangible benefits of imaging and enhanced productivity for logic metal layers in high NA and hyper NA.
Background: Line edge roughness (LER) is a critical performance metric in lithography. Current power-spectral-density-based methods, commonly used to evaluate roughness parameters assuming scale-limited self-affine fractal profiles, provide only an average LER value across multiple patterns. Aim: We aim to perform noise-corrected (unbiased) LER measurements on individual patterns, thus enabling detailed characterization of their distribution across multiple patterns. Approach: We applied our proposed height-height correlation function (HHCF)-based noise-correction technique by acquiring profiles from repeated measurements at identical locations and reducing noise via averaging. The average data were then fitted to a theoretical HHCF expression to separate the unbiased HHCF component from noise, and experimental validation was performed on Si line/space patterns with 50 nm linewidths. Results: The noise-reduced profiles exhibited sufficiently low noise to allow the extraction of unbiased HHCF curves and corresponding roughness parameters from each profile. Subsequent analysis of 1000 such profiles successfully characterized the LER distribution. Conclusion: Using the HHCF-based noise correction technique, we determined roughness parameters for each pattern, enabling detailed characterization of LER distributions. The LER distribution offers an efficient approach for evaluating mean LER, which remains a critical performance metric. In future work, applying this distribution could support the identification of outliers with exceptionally large LER values and thereby facilitate yield improvement and process optimization.
Background: One of the primary challenges in semiconductor metrology is high-resolution inspection of lithographic patterns. Although the patterns can be relatively complex, periodic structures are found in many layers of modern devices, and their metrology is fundamental. In the context of hybrid bonding, a periodic array of copper pads within a wafer is used for electrical connections, and metrology methods are deployed to characterize their topography. Although conventional methods include, e.g., atomic force microscopy and scatterometry, we discuss coherent diffractive imaging (CDI). Aim: We aim to study CDI for the metrology of the copper pads used in hybrid bonding. We aim to demonstrate that CDI is a good option for this problem, especially given the phase sensitivity, which may be advantageous when a sample with slight topographic changes is inspected. Approach: CDI is employed for the metrology of a copper pad array. We model the sample as a reflection function. The impact of the copper pad topography on the data is introduced by a phase term that accounts for the optical path length difference induced by the recessions in the copper pad with respect to the surrounding layer. In addition, standard phase retrieval algorithms are modified to incorporate information on the reflectivity value of the substrate/materials that surround the pad array. Results: Simulations show that the proposed algorithms can drastically improve imaging. Features are better resolved, and the loss function associated with the proposed methods can decrease up to 2 orders of magnitude with respect to standard phase retrieval methods. This improvement is more substantial for an Airy spot illumination, and it is milder when a structured illumination is employed. Conclusions: We present a study of phase retrieval algorithms applied to the metrology of copper pad topography for hybrid bonding. We demonstrate that by including a priori information in the update function of the object, a better estimation of the recession of the copper pads can be achieved, leading to an improved metrology by CDI. (c) The Authors. Published by SPIE under a Creative Commons Attribution 4.0 International License.Distribution or reproduction of this work in whole or in part requires full attribution of the originalpublication, including its DOI.
