SIPPER 1) PAPER TITLE Evolving an Automatic Defect Classification Tool 2) AUTHORS Assaf Glazer Applied Materials, Inc., Rehovot, Israel AND Dept. of Computer Science Ben-Gurion University Be'er Sheva 84105, Israel Assaf_Glazer@amat.com Moshe Sipper Dept. of Computer Science Ben-Gurion University Be'er Sheva 84105, Israel sipper@cs.bgu.ac.il 3) CORRESPONDING AUTHOR Moshe Sipper 4) ABSTRACT Automatic Defect Classification (ADC) is a well-developed technology for inspection and measurement of defects on patterned wafers in the semiconductors industry. The poor training data and its high dimensionality in the feature space render the defect-classification task hard to solve. In addition, the continuously changing environment---comprising both new and obsolescent defect types encountered during an imaging machine's lifetime---require constant human intervention, limiting the technology's effectiveness. In this paper we design an evolutionary classification tool, based on genetic algorithms (GAs), to replace the manual bottleneck and the limited human optimization capabilities. We show that our GA-based models attain significantly better classification performance, coupled with lower complexity, with respect to the human-based model and a heavy random search model. 5) CRITERIA (A) The result was patented as an invention in the past, is an improvement over a patented invention, or would qualify today as a patentable new invention. (D) The result is publishable in its own right as a new scientific result, independent of the fact that the result was mechanically created. (E) The result is equal to or better than the most recent human-created solution to a long-standing problem for which there has been a succession of increasingly better human-created solutions. (F) The result is equal to or better than a result that was considered an achievement in its field at the time it was first discovered. (G) The result solves a problem of indisputable difficulty in its field. 6) WHY THE RESULT SATISFIES THE CRITERIA Automatic Defect Classification (ADC) is a well-developed technology for inspection and measurement of defects on patterned wafers in the semiconductors industry, using heavily patented technology (criterion A). The goal of the ADC process is quite simple to state, though arduous to attain: given a wafer image, classify the defect types found in the image (criterion G). During the production process within a fab, which is a customer's wafer fabrication facility, we would like to automatically find and characterize defects, and determine their sources. Poor data, and a deceptive environment in the fab where the classification problem itself varies over time, renders the ADC task hard to solve. The continuously changing environment---comprising both new and obsolescent defect types encountered during an imaging machine's lifetime---require constant human intervention, limiting the technology's effectiveness. In our research we designed an evolutionary classification tool to replace the manual bottleneck and the limited human optimization capabilities. The major breakthrough of our model is its ability to independently fit itself to the changing environment inside the fab---with a deceptive environment of poor and inaccurate information---achieving high classification rate, increased throughput, and better generalization, with respect to the previous human-based system. Obsolete defects can be isolated using our hint array innovation, and the input data set can dynamically change during the machine's lifetime. In addition, our automated model attains significantly better classification performance, coupled with lower complexity, with respect to the human-based model. While humans attained a top accuracy of 87.0%, using our model we obtained a 90.1% accuracy level, with half the complexity of the classification model, achieving higher throughput (wafers produced per hour) and a better generalization for our automated process (criteria D, E, F). This 3% increase in classification performance is highly significant, and translates directly into multi-million dollar savings in wafer production costs. [We also mention that in addition to much improved results, we introduced a novel genetic algorithm, based upon two GAs working in tandem, with a time delay between them (the “hint array”; see paper).] 7) CITATION A. Glazer and M. Sipper, “Evolving an automatic defect classification tool,” Applications of Evolutionary Computing: Proceedings of EvoWorkshops 2008, M. Giacobini et al., Eds. 2008, vol. 4974 of Lecture Notes in Computer Science, pp. 194–203, Springer-Verlag, Heidelberg. 8) STATEMENT OF PRIZE DISTRIBUTION Any prize money, if any, is to be divided equally among the co-authors. 9) COMPARISON TO OTHER HUMAN-COMPETITIVE ENTRIES We in academia often talk about real-world problems, when we usually mean “real-world” problems, i.e., problems that are difficult to solve and *may* have applications somewhere out there in the big, bad, real world… Our first argument herein is that we *really* solved a *real-world* problem (pun intended): the work was carried out in a large, multi-national, high-tech company (Applied Materials, Inc.), and its aim was to improve a multi-million dollar product, namely, an ADC (automatic defect classification) machine. We competed directly with humans. The results we obtained improve significantly upon (hence replacing) human-based optimization in the ADC process. With our automated model, we attained better results both in terms of classification rate and throughput. In addition, by replacing the human bottleneck, we meet the industry's growing demand for robustness and stability in the production process. Note that there is no automated process equivalent to our model in any other product in the industry, worldwide. We put our GA through its paces by proving its worth on the leading product for the Auto Defect Review process in the semiconductors industry. According to formal publications, the Applied Materials SEMVision DR tool, which is the tool used in our work, holds about 60% market share worldwide (a ~0.5 billon USD market), and it serves as one of AMAT’s top-selling products. Finally, one last possible impact of our work: Nowadays, silicon products are used not only for microchips and wafers, but also for solar cells and renewable energy. We hope that our small contribution may help further the cause of solar energy around the world.