Abstract: In order to determine contour edges within a provided image, a plurality of image cells (e.g., groupings of pixels) are created within the image. For each image cell, a numerical value for each of the pixels is compared to a predetermined threshold value to determine comparison values for each pixel. A total numerical value for each image cell may then be determined utilizing the comparison values and numerical values for each pixel within each image cell. An associated contour cell (indicating present contour edges) is then determined for each image cell by comparing the total numerical value for the image cell to a contour cell index. These operations may be performed in parallel by a graphics processing unit (GPU) for each image cell, which may improve a performance of contour edge determination for the image. The stitching of contour edges may also be performed using the GPU, which may provide additional performance improvements for image contour extraction.

Abstract: A method comprising identifying a set of target features for a plurality of data instances of an input data collection; determining feature values for the set of target features for the plurality of data instances; identifying a plurality of outlier data instances based on the determined feature values; identifying a plurality of noisy data instances from the outlier data instances based on feature values of the plurality of noisy data instances, wherein a noisy data instance is identified based on a determination that noise is present in noisy data instance; and providing an indication of the plurality of noisy data instances.

Abstract: Images are accessed representing a status in a fabrication of a semiconductor chip corresponding to a particular stage in the fabrication. Distortion is removed from the images and actual features of the semiconductor chip are extracted from the images. Synthesized ideal features of the semiconductor chip associated with completion of the particular stage in the fabrication are determined from the one or more images. The actual features are compared to the ideal features to determine whether anomalies associated with the particular stage exist in the semiconductor chip.

Abstract: A method comprising determining a binary classification value for each of a plurality of data instances based on a first threshold value assigned to each of the plurality of data instances; applying at least one clustering model to a first subset of the plurality of data instances to identify one or more dominant clusters of data instances; determining a second threshold value to assign to a second plurality of data instances that are included within the one or more dominant clusters of data instances; and redetermining a binary classification value for each of the plurality of data instances based on the second threshold value assigned to the second plurality of data instances and the first threshold value, wherein the first threshold value is assigned to at least a portion of data instances of the plurality of data instances that are not included in the second plurality of data instances.

Abstract: A semantic pattern extraction system can distill tremendous amounts of silicon wafer manufacturing data to generate a small set of simple sentences (semantic patterns) describing physical design geometries that may explain manufacturing defects. The system can analyze many SEM images for manufacturing defects in areas of interest on a wafer. A tagged continuous itemset is generated from the images, with items comprising physical design feature values corresponding to the areas of interest and tagged with the presence or absence of a manufacturing defect. Entropy-based discretization converts the continuous itemset into a discretized one. Frequent set mining identifies a set of candidate semantic patterns from the discretized itemset. Candidate semantic patterns are reduced using reduction techniques and are scored. A ranked list of final semantic patterns is presented to a user. The final semantic patterns can be used to improve a manufacturing process.

Abstract: An improved random forest model is provided, which has been trained based on silicon data generated from tests of previously fabricated chips. An input is provided to the random forest model, the input including a feature set of a pattern within a particular chip layout, the feature set identifying geometric attributes of polygonal elements within the pattern. A result is generated by the random forest model based on the input, where the result identifies a predicted attribute of the pattern based on the silicon data, and the result is generated based at least in part on determining, within the random forest model, that geometric attributes of the pattern were included in the previously fabricated chips, where the previously fabricated chips have chip layouts are different from the particular chip layout.

Abstract: A machine readable storage medium, a method and an apparatus. The method comprises selecting a candidate set of parameters from a plurality of available parameters comprising variables that affect an outcome of a lithography process; performing a set of optimizations wherein each optimization of the set of optimizations is subject to a plurality of objectives and tolerances and a set of constraints, wherein performance of said each optimization comprises: modifying values of at least a portion of the candidate set of parameters to derive a predicted outcome for said each optimization; and determining whether a difference between the predicted outcome and an intended outcome is within an error threshold; and if the difference exceeds the error threshold, perform a subsequent optimization, and otherwise generate an input file including modified values, corresponding to a last one of the set of optimizations, for the at least a portion of the candidate set of parameters.

Abstract: Reinforcement learning methods are applied to the multi-domain problem of developing photoresist models for advanced semiconductor technologies. In an iterative process, candidate photoresist models are selected or generated, with each model comprising an optical imaging model, one or more analytical chemistry or deformation kernels, and one or more photoresist development model terms. Model parameters to be calibrated in an iteration are selected. The candidate photoresist models are calibrated to best fit photoresist contours extracted from SEM images. Values for the calibration model parameters are determined and the most useful analytical kernels are kept in each model while the others are dropped. A genetic algorithm uses the best calibrated photoresist models from the prior iteration to develop candidate models for the next iteration. The process iterates until no further accuracies can be gained. A residual minimization model can be trained to correct for residual errors in the final model.

Abstract: A machine readable storage medium, a method and an apparatus. The method comprises selecting a candidate set of parameters from a plurality of available parameters comprising variables that affect an outcome of a lithography process; performing a set of optimizations wherein each optimization of the set of optimizations is subject to a plurality of objectives and tolerances and a set of constraints, wherein performance of said each optimization comprises: modifying values of at least a portion of the candidate set of parameters to derive a predicted outcome for said each optimization; and determining whether a difference between the predicted outcome and an intended outcome is within an error threshold; and if the difference exceeds the error threshold, perform a subsequent optimization, and otherwise generate an input file including modified values, corresponding to a last one of the set of optimizations, for the at least a portion of the candidate set of parameters.

Abstract: A method comprising determining a binary classification value for each of a plurality of data instances based on a first threshold value assigned to each of the plurality of data instances; applying at least one clustering model to a first subset of the plurality of data instances to identify one or more dominant clusters of data instances; determining a second threshold value to assign to a second plurality of data instances that are included within the one or more dominant clusters of data instances; and redetermining a binary classification value for each of the plurality of data instances based on the second threshold value assigned to the second plurality of data instances and the first threshold value, wherein the first threshold value is assigned to at least a portion of data instances of the plurality of data instances that are not included in the second plurality of data instances.

Abstract: Images are accessed representing a status in a fabrication of a semiconductor chip corresponding to a particular stage in the fabrication. Distortion is removed from the images and actual features of the semiconductor chip are extracted from the images. Synthesized ideal features of the semiconductor chip associated with completion of the particular stage in the fabrication are determined from the one or more images.

Abstract: An improved random forest model is provided, which has been trained based on silicon data generated from tests of previously fabricated chips. An input is provided to the random forest model, the input including a feature set of a pattern within a particular chip layout, the feature set identifying geometric attributes of polygonal elements within the pattern. A result is generated by the random forest model based on the input, where the result identifies a predicted attribute of the pattern based on the silicon data, and the result is generated based at least in part on determining, within the random forest model, that geometric attributes of the pattern were included in the previously fabricated chips, where the previously fabricated chips have chip layouts are different from the particular chip layout.

Abstract: A method comprising identifying a set of target features for a plurality of data instances of an input data collection; determining feature values for the set of target features for the plurality of data instances; identifying a plurality of outlier data instances based on the determined feature values; identifying a plurality of noisy data instances from the outlier data instances based on feature values of the plurality of noisy data instances, wherein a noisy data instance is identified based on a determination that noise is present in noisy data instance; and providing an indication of the plurality of noisy data instances.

Abstract: A semantic pattern extraction system can distill tremendous amounts of silicon wafer manufacturing data to generate a small set of simple sentences (semantic patterns) describing physical design geometries that may explain manufacturing defects. The system can analyze many SEM images for manufacturing defects in areas of interest on a wafer. A tagged continuous itemset is generated from the images, with items comprising physical design feature values corresponding to the areas of interest and tagged with the presence or absence of a manufacturing defect. Entropy-based discretization converts the continuous itemset into a discretized one. Frequent set mining identifies a set of candidate semantic patterns from the discretized itemset. Candidate semantic patterns are reduced using reduction techniques and are scored. A ranked list of final semantic patterns is presented to a user. The final semantic patterns can be used to improve a manufacturing process.

Abstract: Embodiments of the present invention provide, among other things, methods, apparatus, and systems for tuning a semiempirical process for predicting energy for different molecular configurations. In an example method, an energy value and an energy gradient are determined for each of a plurality of molecular configurations using an accurate method. A functional form of the semiempirical process is optimized using the determined energy values and energy gradients via multiobjective optimization. The functional form relates one or more parameters to energy values and energy gradients.

Abstract: Methods and systems for optimizing a solution set. A solution set is generated, and solutions in the solution set are evaluated. Desirable solutions from the solution set are selected. A structural model is created using the desirable solutions, and a surrogate fitness model is created based on the structural model and the desirable solutions. A new solution set may be generated and/or evaluated, based on analyzing at least one of the structural model and the surrogate fitness model, and determining a method for generating a new solution set and/or evaluating the new solution set based at least in part on the analyzing.

Abstract: Methods and systems for creating a synthetic surrogate fitness function. User decisions are received representing fitness for a plurality of solutions. A partial ordering of the plurality of solutions is provided based on the received user decisions, wherein at least some of the plurality of solutions are represented to have a greater relative fitness than other of the plurality of solutions. A complete order of at least the plurality of solutions is induced based on the normalized partial ordering. A synthetic surrogate fitness function is generated using the induced complete order.

Abstract: Embodiments of the present invention provide, among other things, methods, apparatus, and systems for tuning a semiempirical process for predicting energy for different molecular configurations. In an example method, an energy value and an energy gradient are determined for each of a plurality of molecular configurations using an accurate method. A functional form of the semiempirical process is optimized using the determined energy values and energy gradients via multiobjective optimization. The functional form relates one or more parameters to energy values and energy gradients.

Abstract: Methods and systems for creating a synthetic surrogate fitness function. User decisions are received representing fitness for a plurality of solutions. A partial ordering of the plurality of solutions is provided based on the received user decisions, wherein at least some of the plurality of solutions are represented to have a greater relative fitness than other of the plurality of solutions. A complete order of at least the plurality of solutions is induced based on the normalized partial ordering. A synthetic surrogate fitness function is generated using the induced complete order.

Abstract: Methods and systems for optimizing a solution set. A solution set is generated, and solutions in the solution set are evaluated. Desirable solutions from the solution set are selected. A structural model is created using the desirable solutions, and a surrogate fitness model is created based on the structural model and the desirable solutions. A new solution set may be generated and/or evaluated, based on analyzing at least one of the structural model and the surrogate fitness model, and determining a method for generating a new solution set and/or evaluating the new solution set based at least in part on the analyzing.