Search

IBM Research is engaged in a research program in symbiotic cognitive computing to investigate how to embed cognitive computing in physical spaces. This article proposes 5 key principles of symbiotic cognitive computing. We describe how these principles are applied in a particular symbiotic cognitive computing environment and in an illustrative application.

An analytical method based on transmission matrices is used to study the imaging performance of multilayer silver superlenses for near-field optical lithography. The spatial-frequency transfer functions of these systems show complex structure, and the best lens stack to use depends upon the feature sizes to be imaged, the distance from the final silver surface and the image plane, and the required level of transmission of the lens stack. The balance between the DC and the fundamental spatial-frequency component in the transfer function is also crucial in determining the quality of the final image that is formed. Graded-thickness lens stacks are introduced and analyzed for the first time, and it is shown that these systems can have a flat transfer function over a wide range of spatial frequencies, which may make them more versatile for practical systems.

The resolution of a near-field lithography technique that uses a planar silver lens to form a near-field image has been investigated and compared with hard-contact lithography. Sub-diffraction-limited imaging has been observed through a 50 nm silver film, confirming a recent superlensing proposal [J.B. Pendry, Phys. Rev. Lett. 85 (2000) 3966], with grating periods down to 170 nm pitch being patterned into resist using simple broadband UV-illumination.

The near-field imaging properties of planar silver layers are reviewed here, with a view to identifying their potential and pitfalls for sub-100nm photolithography. Initial studies have shown that half-pitch features of 60-70nm can be achieved using simple contact exposure with low-cost mercury lamp sources, which is less than 1/5th of the exposure wavelength. Sub-100nm isolated features and line pairs with sub-100nm spacing have also been resolved, and conventional resolution limits are exceeded in these cases too. Multi-layer silver lenses have also been studied, as it has been predicted that these will improve image resolution for a fixed mask-resist separation. Initial experimental results are inconclusive, as roughness on the additional interfaces tends to mitigate the benefits of going to the laminated structure. Simulations have been used to support the findings of the silver-lens experiments, and it is predicted that sub-50nm half-pitch resolution should be possible.

Near-field imaging through planar silver lenses has been demonstrated using a modified conformal-mask optical lithography arrangement. Dense feature resolution down to 250 nm (on a 500 nm period) has been achieved in 50 nm thick photoresist on silicon using broadband illumination from a mercury lamp. Finite difference time domain simulations have been performed to show the resolution improvements that can be expected for imaging through such silver lenses compared with near-field proximity imaging. The resolution enhancements that are predicted are in good agreement with the experimental results, and the conditions by which sub-diffraction-limited resolution may be achieved are given.

Patterning with sub-diffraction-limited resolution has been demonstrated using a simple photolithography technique. Evanescent fields and surface plasmons are critical to the image formation, which is investigated here using computer simulations and experiments. A regime exists in which surface plasmons are resonantly excited, which we have named Evanescent Interferometric Lithography (EIL); period halving and reduced exposure times characterize this exposure mode. Two other exposure modes have been investigated in which surface plasmons on a planar metallic film beneath the mask are used to improve pattern formation. In the first, Planar Lens Lithography (PLL), a planar silver layer excited near its plasma frequency is used to form a projected near-field image. For a 40-nm thick silver layer, we predict that resolution down to 40 nm should be possible. However, the image is affected by the loss in the silver layer, the mask period, duty cycle and surrounding refractive index. Experimental verification of PLL is presented for 1-micron period structures imaged through 120 nm of silver. Finally, simulations are used to show that surface plasmons on an underlying silver layer can be used to improve process latitude and depth of field. We have named this mode Surface Plasmon Enhanced Contact Lithography (SPECL).

Optical imaging through a thin planar silver layer has been achieved by utilizing near-field lithography techniques. A 120 nm thick silver lens that was placed 60 nm below a patterned mask, imaged the mask’s features onto a photosensitive material located 60 nm below the silver. The entire structure was exposed from above with a mercury lamp. Features sizes as small as 350 nm (at a 700 nm period) were imaged onto the photosensitive material, demonstrating the lensing ability of the planar silver slab.

Source-mask optimization (SMO) in optical lithography has in recent years been the subject of increased exploration as an enabler of 22/20nm and beyond technology nodes [1-6]. It has been shown that intensive optimization of the fundamental degrees of freedom in the optical system allows for the creation of non-intuitive solutions in both the source and mask, which yields improved lithographic performance. This paper will demonstrate the value of SMO software in resolution enhancement techniques (RETs). Major benefits of SMO include improved through-pitch performance, the possibility of avoiding double exposure, and superior performance on two dimensional (2D) features. The benefits from only optimized source, only optimized mask, and both source and mask optimized together will be demonstrated. Furthermore, we leverage the benefits from intensively optimized masks to solve large array problems in memory use models (MUMs). Mask synthesis and data prep flows were developed to incorporate the usage of SMO, including both RETs and MUMs, in several critical layers during 22/20nm technology node development. Experimental assessment will be presented to demonstrate the benefits achieved by using SMO during 22/20nm node development.

Mask simulations are made consistent with the brightness theorem if the source map is rescaled by pixel-solid-angle. Standard radiometry factors preserve consistency during propagation, and are derivable from rigorous vector diffraction integrals.

In recent years the potential of Source-Mask Optimization (SMO) as an enabling technology for 22nm-and-beyond lithography has been explored and documented in the literature.1-5 It has been shown that intensive optimization of the fundamental degrees of freedom in the optical system allows for the creation of non-intuitive solutions in both the mask and the source, which leads to improved lithographic performance. These efforts have driven the need for improved controllability in illumination5-7 and have pushed the required optimization performance of mask design.8, 9 This paper will present recent experimental evidence of the performance advantage gained by intensive optimization, and enabling technologies like pixelated illumination. Controllable pixelated illumination opens up new regimes in control of proximity effects,1, 6, 7 and we will show corresponding examples of improved through-pitch performance in 22nm Resolution Enhancement Technique (RET). Simulation results will back-up the experimental results and detail the ability of SMO to drive exposure-count reduction, as well as a reduction in process variation due to critical factors such as Line Edge Roughness (LER), Mask Error Enhancement Factor (MEEF), and the Electromagnetic Field (EMF) effect. The benefits of running intensive optimization with both source and mask variables jointly has been previously discussed.1-3 This paper will build on these results by demonstrating large-scale jointly-optimized source/mask solutions and their impact on design-rule enumerated designs.

We demonstrate experimentally for the first time the feasibility of applying SMO technology using pixelated illumination. Wafer images of SRAM contact holes were obtained to confirm the feasibility of using SMO for 22nm node lithography. There are still challenges in other areas of SMO integration such as mask build, mask inspection and repair, process modeling, full chip design issues and pixelated illumination, which is the emphasis in this paper. In this first attempt we successfully designed a manufacturable pixelated source and had it fabricated and installed in an exposure tool. The printing result is satisfactory, although there are still some deviations of the wafer image from simulation prediction. Further experiment and modeling of the impact of errors in source design and manufacturing will proceed in more detail. We believe that by tightening all kind of specification and optimizing all procedures will make pixelated illumination a viable technology for 22nm or beyond. Publisher's Note: The author listing for this paper has been updated to include Carsten Russ. The PDF has been updated to reflect this change.

Near-field interference lithography is a promising variant of multiple patterning in semiconductor device fabrication that can potentially extend lithographic resolution beyond the current materials-based restrictions on the Rayleigh resolution of projection systems. With H2O as the immersion medium, non-evanescent propagation and optical design margins limit achievable pitch to approximately 0.53λ/nH2O = 0.37λ. Non-evanescent images are constrained only by the comparatively large resist indices (typically1.7) to a pitch resolution of 0.5/nresist (typically 0.29). Near-field patterning can potentially exploit evanescent waves and thus achieve higher spatial resolutions. Customized near-field images can be achieved through the modulation of an incoming wavefront by what is essentially an in-situ hologram that has been formed in an upper layer during an initial patterned exposure. Contrast Enhancement Layer (CEL) techniques and Talbot near-field interferometry can be considered special cases of this approach. Since the technique relies on near-field interference effects to produce the required pattern on the resist, the shape of the grating and the design of the film stack play a significant role on the outcome. As a result, it is necessary to resort to full diffraction computations to properly simulate and optimize this process. The next logical advance for this technology is to systematically design the hologram and the incident wavefront which is generated from a reduction mask. This task is naturally posed as an optimization problem, where the goal is to find the set of geometric and incident wavefront parameters that yields the closest fit to a desired pattern in the resist. As the pattern becomes more complex, the number of design parameters grows, and the computational problem becomes intractable (particularly in three-dimensions) without the use of advanced numerical techniques. To treat this problem effectively, specialized numerical methods have been developed. First, gradient-based optimization techniques are used to accelerate convergence to an optimal design. To compute derivatives of the parameters, an adjoint-based method was developed. Using the adjoint technique, only two electromagnetic problems need to be solved per iteration to evaluate the cost function and all the components of the gradient vector, independent of the number of parameters in the design.

Source optimization in optical lithography has been the subject of increased exploration in recent years [1-4], resulting in the development of multiple techniques including global optimization of process window [4]. The performance advantages of source optimization have been demonstrated through theory, simulation, and experiment. This paper will emphasize global optimization of sources over multiple patterns, e.g. co-optimization of critical SRAM cells and the critical pitches of random logic, and implement global source optimization into current resolution enhancement techniques (RETs). The effect on optimal source due to considering multiple patterns is investigated. We demonstrate that optimal source for limited patterns does work for a large clip of layout. Through theoretical analysis and simulations, we explain that only critical patterns and/or critical combinations of patterns determine the final optimal source; for example those patterns that contain constraints which are active in the solution. Furthermore, we illustrate, through theory and simulation, that pixelated sources have better performance than generic sources and that in general it is impossible for generic sources to construct a truly optimal solution. Sensitivity, tool matching, and lens heating issues for pixelated sources are also discussed in this paper. Finally, we use a RETs example with wafer data to demonstrate the benefits of global source optimization.

Traditional OPC is essentially an iterated feedback process, in which the position of each target edge is corrected by adjusting a controlling mask edge. However, true optimization adjusts the mask variables collectively, and in so-called SMO approaches (for Source Mask Optimization) the source variables are adjusted as well. Optimized masks often have high edge density if synthesis methods are used in an effort to obtain a more global solution, and the correspondence between individual mask edges and printed target edges becomes less clearcut than in traditionally OPC'd masks. Restrictions on phase shift and MEEF tend to reduce this departure from traditional solutions, but they trade off the theoretical performance advantage in dose and focus latitude that phase shift provides for a reduced sensitivity to thick mask topography and to manufacturing error. Mask variables couple across long distances only in the indirect sense of stitched connection across chains of neighbor-to-neighbor interactions, but source variables interact directly across entire masks. Source+mask optimization of large areas therefore involves long-range communication across the parts of the calculation, though the number of source variables involved is small. Tradeoffs between source structure and pattern diversity are illustrated, taking into account the limited (but unknown) number of binding features in a large layout. SMO's exploitation of complex source designs is shown to provide superior solutions to those obtained by mask optimization alone. Moreover, in development work the ability to adjust the source opens up new options in process engineering, and these will become particularly valuable when future exposure tools provide greater flexibility in programmable source control. Such capabilities can be explored in a preliminary way by using programmed multi-scans to compose optimized compound sources with e.g. multiple poles or annular elements.

In this paper, we will outline the approach for optimizing the illumination conditions to print three-dimensional images in resist stacks of varying sensitivity in a single exposure. The algorithmic approach for acheiving both optimal common and weakest window is presented. Results will be presented which demonstrate the ability of the technique to create threedimensional structures. The performance of the common and weakest window formulation will be explored using this approach. Additionally, due to physical restrictions there are limitations to the type of patterns that can be printed with a single exposure in this manner, thus the abilities of such a technique will be explored.

There is a surprising lack of clarity about the exact quantity that a lithographic source map should specify. Under the plausible interpretation that input source maps should tabulate radiance, one will find with standard imaging codes that simulated wafer plane source intensities appear to violate the brightness theorem. The apparent deviation (a cosine factor in the illumination pupil) represents one of many obliquity/inclination factors involved in propagation through the imaging system whose interpretation in the literature is often somewhat obscure, but which have become numerically significant in today's hyper-NA OPC applications. We show that the seeming brightness distortion in the illumination pupil arises because the customary direction-cosine gridding of this aperture yields non-uniform solid-angle subtense in the source pixels. Once the appropriate solid angle factor is included, each entry in the source map becomes proportional to the total |E|^2 that the associated pixel produces on the mask. This quantitative definition of lithographic source distributions is consistent with the plane-wave spectrum approach adopted by litho simulators, in that these simulators essentially propagate |E|^2 along the interfering diffraction orders from the mask input to the resist film. It can be shown using either the rigorous Franz formulation of vector diffraction theory, or an angular spectrum approach, that such an |E|^2 plane-wave weighting will provide the standard inclination factor if the source elements are incoherent and the mask model is accurate. This inclination factor is usually derived from a classical Rayleigh-Sommerfeld diffraction integral, and we show that the nominally discrepant inclination factors used by the various diffraction integrals of this class can all be made to yield the same result as the Franz formula when rigorous mask simulation is employed, and further that these cosine factors have a simple geometrical interpretation. On this basis one can then obtain for the lens as a whole the standard mask-to-wafer obliquity factor used by litho simulators. This obliquity factor is shown to express the brightness invariance theorem, making the simulator's output consistent with the brightness theorem if the source map tabulates the product of radiance and pixel solid angle, as our source definition specifies. We show by experiment that dose-to-clear data can be modeled more accurately when the correct obliquity factor is used.

Near-field imaging through plasmonic 'superlensing' layers can offer advantages of improved working distance (i.e. introducing the equivalent of a focal length) and control over image intensity compared to simple near-field imaging. In a photolithographic environment at ultra-violet (UV) wavelengths the imaging performance of single- and multi-layer silver plasmonic superlenses has been studied both experimentally and via computer simulations. Super-resolution imaging has been demonstrated experimentally, with the sub-100 nm resolution currently being limited by issues of roughness in the silver layers and the ability to deposit high-quality silver-dielectric multilayers. The simulation studies have shown that super-resolved imaging should be possible using surprisingly thick silver layers (>100 nm), with the cost of much reduced image intensity, which is something that is yet to be shown experimentally. The use of multilayer plasmonic superlenses also introduces richness to the imaging behaviour, with very high transmission possible for certain spatial frequency components in the image. This has been widely touted as a means for improving image resolution, but the complexity of the spatial-frequency transfer functions for these systems does not make this a universal fact for all classes of objects. Examples of imaging situations are given where multi-layer superlenses are actually detrimental to the image quality, such as the case of closely-separated dark-line objects on an otherwise bright background.

We provide an expanded description of the global algorithm for mask optimization introduced in our earlier papers, and discuss auxiliary optimizations that can be carried out in the problem constraints and film stack. Mask optimization tends inherently to be a problem with non-convex quadratic constraints, but for small problems we can mitigate this difficulty by exploiting specialized knowledge that applies in the lithography context. If exposure latitude is approximated as maximization of edge slope between image regions whose intensities must print with opposite polarity, we show that the solution space can be approximately divided into regions that contain at most one local minimum. Though the survey of parameter space to identify these regions requires an exhaustive grid search, this search can be accelerated using heuristics, and is not the rate-limiting step at SRAM scale or below. We recover a degree of generality by using a less simplified objective function when we actually assess the local minima. The quasi-binary specialization of lithographic targets is further exploited by searching only in the subspace formed by the dominant joint eigenvectors for dark region intensity and bright region intensity, typically reducing problem dimensionality to less than half that of the full set of frequency-domain variables (i.e. collected diffraction orders). Contrast in this subspace across the bright/dark edge will approximately reflect exposure latitude when we apply the standard fixed edge-placement constraints of lithography. However, during an exploratory stage of optimization we can define preliminary tolerances which more explicitly reflect constraints on devices, e.g. as is done with compactor codes for design migration. Our algorithm can handle vector imaging in a general way, but for the special case of unpolarized illumination and a lens having radial symmetry (but arbitrary source shape) we show that the bilinear function which describes vector interference within the film stack can be expressed in terms of three generic radial functions, enabling rapid numerical evaluation of the Hopkins kernel. By inspection these functions show that one can in principle recover classical scalar-like imaging even at high NA by exposing a very thin layer spaced above a reflective substack. The reflected image largely restores destructive interference in TM polarized fringes, if proper phasing is achieved. With an ideal reflector, the first-order azimuthal contrast loss term vanishes in all TCC components, and complete equivalence to scalar imaging is obtained in classical two-beam imaging.

Solid-state electrochemical switches are new devices that may find application in low power logic or memory circuits. The reversible electrochemical oxidation and reduction of a sub 10nm-thick silver sulphide layer is the critical process for these devices' operation, which determines the switching speed and the OFF/ON resistance ratio. A new process has been developed for fabricating QCAS devices, including a novel SF6 plasma sulphidation technique for forming the critical silver sulphide switchable insulator layer. QCAS devices were successfully created with OFF/ON resistance ratios of up to 106 , with switching speeds comparable to previous devices and far less problematic yields

A technique for permanently capturing a replica impression of biological cells has been developed to facilitate analysis using nanometer resolution imaging tools, namely the atomic force microscope (AFM). The method, termed Bioimprint™, creates a permanent cell 'footprint' in a non-biohazardous Poly (dimethylsiloxane) (PDMS) polymer composite. The transfer of nanometer scale biological information is presented as an alternative imaging technique at a resolution beyond that of optical microscopy. By transferring cell topology into a rigid medium more suited for AFM imaging, many of the limitations associated with scanning of biological specimens can be overcome. Potential for this technique is demonstrated by analyzing Bioimprint™ replicas created from human endometrial cancer cells. The high resolution transfer of this process is further detailed by imaging membrane morphological structures consistent with exocytosis. The integration of soft lithography to replicate biological materials presents an enhanced method for the study of biological systems at the nanoscale.

A performance enhancement to planar lens lithography (PLL) through the use of i-line narrowband exposures has been investigated. Experimental results show that for a 50nm silver layer the image fidelity of narrowband exposures out performs broadband exposures. This is due to the removal of off-plasmonic-resonance wavelengths, which cause unwanted background exposure and a loss of image fidelity. Dense gratings have been resolved down to 145nm periods, as well as line-pairs down to separation distances of 117nm. These results out perform the diffraction-limits that restrict traditional optical-system resolution limits.

In the recent past, scaling of semiconductor fabrication systems has been dominated by wavelength and numerical aperture modifications. This is now no longer the case for 193-nm immersion projection lithography (193i) systems as there are no technical paths for continued benefit from the in these areas. Instead, a range of techniques including patterning processes and system optimization are being used to push the limits of the system. This paper will review the elements that are now driving scaling for a system of fixed wavelength and numerical aperture.

An experimental comparison of the performance of single- and double-layer planar lens lithography has been carried out. A direct comparison is made with a single 50 nm silver lens and a double silver lens with two 30 nm layers. Sub-diffraction-limited features have been imaged in both cases, with dense grating periods down to 145 and 170 nm for the single- and double-layered stacks, respectively. For the same total thickness of silver, the resolution limit is qualitatively better for a double-layer stack. However, pattern fidelity is reduced in the double-layer experiments, owing to increased surface roughness. Finite-difference time-domain simulations are also presented to back up the experimental results.

It has been proposed that a planar silver layer could be used to project a super-resolution image in the near field when illuminated near its plasma frequency [J. B. Pendry, Phys. Rev. Lett. 86, 3966 (2000)]. This has been investigated experimentally using a modified form of conformal-mask photolithography, where dielectric spacers and silver layers are coated onto a tungsten-on-glass mask. We report here on the experimental confirmation that super-resolution imaging can be achieved using a 50-nm thick planar silver layer as a near-field lens at wavelengths around 365 nm. Gratings with periods down to 145 nm have been resolved, which agrees well with our finite-difference time domain (FDTD) simulations.

A variation of proximity lithography has been proposed using a planar metallic film in a near-field configuration, where metal acts as a near-field “superlens.” We report here on experimental evidence of such optical lithography with a planar silver lens. Silver layers of varying thickness (85–120nm) placed at specific distances (40–60nm) below a patterned mask were able to image the mask’s features onto a photoresist located after a gap (26–60nm) below the silver. The entire structure was exposed from above with a mercury lamp. Feature sizes as small as 250nm (at a 500nm period) were imaged, demonstrating the lensing ability of the planar silver slab.