Martin A. Hunt

Direct to digital holography for semiconductor wafer defect detection and review

A method for recording true holograms directly to a digital video medium in a single image has been invented. This technology makes the amplitude and phase for every pixel of the target object wave available. Since phase is proportional wavelength, this makes high-resolution metrology an implicit part of the holographic recording. Measurements of phase can be made to one hundredth or even one thousandth of a wavelength, so the technology is attractive for dining defects on semiconductor wafers, where feature sizes are now smaller than the wavelength of even deep UV light.

Direct To Digital Holography For High Aspect Ratio Inspection of Semiconductor Wafers

Direct to Digital Holography (DDH) has been developed as a semiconductor wafer inspection tool and in particular as a tool for seeing defects in high aspect ratio (HAR) structures on semiconductor wafers and also for seeing partial‐height defects. While the tool works very well for general wafer inspection, it has unusual capabilities for high aspect ratio inspection (HARI) and for detecting thin residual film defects (partial height defects). Inspection of HAR structures is rated as one of the highest unmet priorities of the member companies of International SEMATECH, and finding residual thin film defects (in some cases called “stringers”) is also a very difficult challenge. The capabilities that make DDH unusually sensitive include: 1) the capture of the whole wave—both the classical amplitude captured by traditional optical systems, and the phase of the wave, with phase potentially measured to ∼1/1000’th of a wavelength or ∼2 to 3 Angstroms for a deep ultra‐violet (DUV) laser; 2) heterodyne detection—...

Imaging tristimulus colorimeter for the evaluation of color in printed textiles

The high-speed production of textiles with complicated printed patterns presents a difficult problem for a colorimetric measurement system. Accurate assessment of product quality requires a repeatable measurement using a standard color space, such as CIELAB, and the use of a perceptually based color difference formula, e.g. (Delta) ECMC color difference formula. Image based color sensors used for on-line measurement are not colorimetric by nature and require a non-linear transformation of the component colors based on the spectral properties of the incident illumination, imaging sensor, and the actual textile color. This research and development effort describes a benchtop, proof-of-principle system that implements a projection onto convex sets (POCS) algorithm for mapping component color measurements to standard tristimulus values and incorporates structural and color based segmentation for improved precision and accuracy. The POCS algorithm consists of determining the closed convex sets that describe the constraints on the reconstruction of the true tristimulus values based on the measured imperfect values. We show that using a simulated D65 standard illuminant, commercial filters and a CCD camera, accurate (under perceptibility limits) per-region based (Delta) ECMC values can be measured on real textile samples.

Subpixel measurement of image features based on paraboloid surface fit

A digiuil image processing inspection system is under development at Oak Ridge National Laboratory that will locate image features on printed material and measure distances between them to accuracies of 0. 001 in. An algorithm has been developed for this system that can locate unique image features to subpixel accuracies. It is based on a least-squares fit of a paraboloid function to the surface generated by correlating a reference image feature against a test image search area. Normalizing the correlation surface makes the algorithm robust in the presence of illumination variations and local flaws. Subpixel accuracies of better than 1/16 of a pixel have been achieved using a variety of different reference image features.

Moment invariants for automated inspection of printed material

The use of moment invariants for the detection of flaws in automated image processing inspection of printed graphic material is investigated. Prior work with moment invariants has concentrated on twodimensional image pattern recognition. A major limitation in pattern recognition applications has been the segmentation of the image from its background. Automated image processing inspection of printed material does not suffer from this limitation because a standard image background exists. The potential for separating flawed and unflawed printed material using moment invariants is demonstrated with formal statistical experiments.

Nonlinear filter derived from topological image features

A digital machine-inspection system is being developed at Oak Ridge National Laboratory to detect flaws on printed graphic images. The inspection is based on subtraction of a digitized test image from a reference image to determine the location, number, extent, and contrast of potential flaws. When performing subtractive analysis on the digitized information, two sources of errors in the amplitude of the difference image can develop: (1) spatial misregistration of the reference and test sample, or (2) random fluctuations in the printing process. Variations in printing and registration between samples will generate topological artifacts related to surface structure, which is referred to as edge noise in the difference image. Most feature extraction routines require that the difference image be relatively free of noise to perform properly. A novel algorithm has been developed to filter edge noise from the difference images. The algorithm relies on the a priori assumption that edge noise will be located near locations having a strong intensity gradient in the reference image. The filter is based on the structure of the reference image and is used to attenuate edge features in the difference image. The filtering algorithm, consisting of an image multiplication, a global intensity threshold, and an erosion/dilation, has reduced edge noise by 98% over the unfiltered image and can be implemented using off-the-shelf hardware.

Facet model and mathematical morphology for surface characterization

This paper describes an algorithm for the automatic segmentation and representation of surface structures and non-uniformities in an industrial setting. The automatic image processing and analysis algorithm is developed as part of a complete on-line web characterization system of a paper making process at the wet end. The goal is to: (1) link certain types of structures on the surface of the web to known machine parameter values, and (2) find the connection between detected structures at the beginning of the line and defects seen on the final product. Images of the pulp mixture, carried by a fast moving table, are obtained using a stroboscopic light and a CCD camera. This characterization algorithm succeeded where conventional contrast and edge detection techniques failed due to a poorly controlled environment. The images obtained have poor contrast and contain noise caused by a variety of sources.

Accommodating multiple illumination sources in an imaging colorimetry environment

Researchers at the Oak Ridge National Laboratory have been developing a method for measuring color quality in textile products using a tri-stimulus color camera system. Initial results of the Imaging Tristimulus Colorimeter (ITC) were reported during 1999. These results showed that the projection onto convex sets (POCS) approach to color estimation could be applied to complex printed patterns on textile products with high accuracy and repeatability. Image-based color sensors used for on-line measurement are not colorimetric by nature and require a non-linear transformation of the component colors based on the spectral properties of the incident illumination, imaging sensor, and the actual textile color. Our earlier work reports these results for a broad-band, smoothly varying D65 standard illuminant. To move the measurement to the on-line environment with continuously manufactured textile webs, the illumination source becomes problematic. The spectral content of these light sources varies substantially from the D65 standard illuminant and can greatly impact the measurement performance of the POCS system. Although absolute color measurements are difficult to make under different illumination, referential measurements to monitor color drift provide a useful indication of product quality. Modifications to the ITC system have been implemented to enable the study of different light sources. These results and the subsequent analysis of relative color measurements will be reported for textile products.

Low-power, real-time digital video stabilization using the HyperX parallel processor

Coherent Logix has implemented a digital video stabilization algorithm for use in soldier systems and small unmanned air / ground vehicles that focuses on significantly reducing the size, weight, and power as compared to current implementations. The stabilization application was implemented on the HyperX architecture using a dataflow programming methodology and the ANSI C programming language. The initial implementation is capable of stabilizing an 800 x 600, 30 fps, full color video stream with a 53ms frame latency using a single 100 DSP core HyperX hx3100TM processor running at less than 3 W power draw. By comparison an Intel Core2 Duo processor running the same base algorithm on a 320x240, 15 fps stream consumes on the order of 18W. The HyperX implementation is an overall 100x improvement in performance (processing bandwidth increase times power improvement) over the GPP based platform. In addition the implementation only requires a minimal number of components to interface directly to the imaging sensor and helmet mounted display or the same computing architecture can be used to generate software defined radio waveforms for communications links. In this application, the global motion due to the camera is measured using a feature based algorithm (11 x 11 Difference of Gaussian filter and Features from Accelerated Segment Test) and model fitting (Random Sample Consensus). Features are matched in consecutive frames and a control system determines the affine transform to apply to the captured frame that will remove or dampen the camera / platform motion on a frame-by-frame basis.

Paradigm for selecting the optimum classifier in semiconductor automatic defect classification applications

The automatic classification of defects found on semiconductor wafers using a scanning electron microscope (SEM) is a complex task that involves many steps. The process includes re- detecting the defect, measuring attributes of the defect, and automatically assigning a classification. In many cases, especially during product ramp-up, and when multiple products are manufactured in the same line, there are few training examples for an automatic defect classification (ADC) system. This condition presents a problem for traditional supervised parametric and nonparametric learning techniques. In this paper we investigate the attributes of several approaches to ADC and compare their performance under a variety of available training data scenarios. We have selected to characterize the attributes and performance of a traditional K-nearest neighbor classifier, probabilistic neural network (PNN), and rule-based classifier in the context of SEM ADC. The PNN classifier is a nonparametric supervised classifier that is built around a radial basis function (RBF) neural network architecture. A basic summary of the PNN will be presented along with the generic strengths and weakness described in the literature and observed with actual semiconductor defect data. The PNN classifier is able to manage conditions such as non-convex class distributions and single class multiple clusters in feature space. A rule-based classifier producing built-in core classes provided by the Applied Materials SEMVision tool will be characterized in the context of both few examples and no examples. An extensive set of fab generated data is used to characterize the performance of these ADC approaches. Typical data sets contain from 30 to greater than 200. The number of classes in the data set range from 4 to more than 12. The conclusions reached from this analysis indicate that the strengths of each method are evident under specific conditions that are related to different stages within the VLSI yield curve, and to the number of different products in the line.