Failure rate-aware prediction models for hotspot prediction in circuit designs

Abstract

A method for hotspot prediction for circuit designs through failure rate-aware prediction models may include training a failure rate-aware prediction model. Training the failure rate-aware prediction model may include correlating confirmed hotspot locations and confirmed non-hotspot locations to circuit elements of a training circuit design, extracting hotspot feature vectors and non-hotspot feature vectors from the correlated circuit elements, clustering the hotspot feature vectors and the non-hotspot feature vectors into feature vector clusters in a partitioned feature space. Training may also include determining a failure rate value for the given feature space partition that the given feature vector cluster is located within, the failure rate value computed based at least on the number of hotspot feature vectors included in the given feature vector cluster. The method may further include analyzing an input circuit design through the trained failure rate-aware prediction model to predict hotspot locations the input circuit design.

Description

202414041FAILURE RATE-AWARE PREDICTION MODELS FOR HOTSPOT PREDICTION IN CIRCUIT DESIGNSBACKGROUND

[0001] Electronic circuits, such as integrated circuits, are used in nearly every facet of modern society, from automobiles to microwaves to personal computers. Design of circuits may involve many steps, known as a "design flow." The particular steps of a design flow are often dependent upon the type of circuit being designed, its complexity, the design team, and the circuit fabricator or foundry that will manufacture the circuit. Electronic design automation (EDA) applications support the design and verification of circuits prior to fabrication. EDA applications may implement various EDA procedures, e.g., functions, tools, or features to analyze, test, or verify a circuit design at various stages of the design flow.BRIEF DESCRIPTION OF THE DRAWINGS

[0002] Certain examples are described in the following detailed description and in reference to the drawings.

[0003] Figure 1 shows an example of a computing system that supports hotspot prediction for circuit designs through failure rate-aware prediction models.

[0004] Figure 2 shows an example training of a failure rate-aware prediction model according to the present disclosure.

[0005] Figure 3 shows an example determination of failure rate values from feature vector clusters for training of a failure rate-aware prediction model.

[0006] Figure 4 shows an example of a trained failure rate-aware prediction model according to the present disclosure.

[0007] Figure 5 shows an example determination of predicted hotspots for an input circuit design through a failure rate-aware prediction model.202414041

[0008] Figure 6 shows an example of logic that a system may implement to support hotspot prediction for circuit designs through failure rate-aware prediction models.

[0009] Figure 7 shows an example of a computing system that supports hotspot prediction for circuit designs through failure rate-aware prediction models.DETAILED DESCRIPTION

[0010] Electronic circuits, such as integrated circuits (ICs), are used in nearly every facet of modem society, from automobiles to microwaves to personal computers. The design, verification, and physical manufacture of circuit devices often involve several steps, sometimes referred to as a "design flow." The particular steps of a design flow are dependent upon various factors, such as the type of integrated circuit being designed, its complexity, the design team, and the integrated circuit fabricator (e.g., foundry) that will manufacture the physical circuit. Typically, software and hardware tools can verify the circuit designs at various stages of the design flow, for example through complex rule checks, software-based simulations, hardware-based emulation, and various other techniques supported by modern EDA technology. These steps of a design flow aid in the discovery of errors in circuit designs, and allow design teams and engineers to correct or otherwise improve the designs prior to, during, or after physical manufacture.

[0011] Modern circuit design technologies can provide various mechanisms to detect potential or actual defects that can occur in manufactured circuits. Physical inspection of manufactured chips is one way to detect actual-occurring circuit defects and circuit hotspots. As used herein, a “hotspot” may refer to any area in a circuit that is defective (e.g., an improperly manufactured circuit component unintended by a circuit design or a flaw in the circuit design itself). On a manufactured circuit wafer, circuit hotspots are commonly complex in nature, as variations in circuit design characteristics and manufacture process parameters can cause circuit defects to arise from a combination of multiple factors. Hotspot confirmations on manufactured circuits are possible today. Some hotspot detection processes may involve use of high precision imaging techniques, such via as scanning electron beam microscopy (“SEM”) techniques. High precision inspection techniques can be used to inspect specific windows of manufactured circuits to detect hotspots or otherwise confirm that no hotspots exist in the window, e.g., a non-hotspot region of the manufactured circuit.202414041

[0012] Confirmed hotspots and non-hotspot chip locations in manufactured chips can be leveraged in various hotspot prediction technologies. Supervised or semisupervised machine learning (ML) models can be trained based on detected hotspot locations on physical chips to predict hotspot locations in other circuit designs. In that regard, ML models can extract circuit data from (e.g., proximate to) confirmed hotspot locations and non-hotspot locations and label such extracted circuit data accordingly, supporting ML-based prediction of hotspots in circuit designs. As chip designs continue to increase in complexity and fabrication technologies support manufacture of circuit structures merely a few atoms wide, the sheer number of circuit elements that can be analyzed in a circuit design make hotspot predictions based on singular circuit element characteristics increasingly challenging. Clustering algorithms and techniques can be used to group circuit elements with similar characteristics (e.g., similar feature vectors) together into a single cluster, and predictive models can learn, model, or characterize circuit behavior based on clustered circuit elements.

[0013] Limitations exist in conventional clustering and ML feature labeling techniques. In many conventional clustering techniques, clusters are labeled in binary label scheme, e.g., a “yes” or “no” label, a “pass” or “fail” label, and the like. However, the same cluster of feature vector data can have both hotspot and non-hotspot feature vectors due to stochastic nature of circuit failures. Due to randomness in lithographical process steps and photoresist materials of circuit fabrication processes, defects in circuit structures may only sometimes appear, even for the same or very similar underlying layout structures. In the context of hotspot detections, the same feature vector group (e.g., same cluster) can include circuit elements that pass on one inspected circuit location (e.g., confirmed non-hotspot location), but fail at another location (e.g., confirmed hotspot location). In ML-based circuit analysis clustering techniques, conventional processes may label cluster with simply a “hotspot” or “non- hotspot” that applies to all circuit elements grouped into the single cluster. For example, if a given cluster of circuit elements (e.g., extracted feature vectors thereof) include both hot-spot and non-hotspot feature vectors, then conventional clustering processes may assign a “fail” or “hotspot” label to the entire cluster. This may be the case even though the cluster includes a combination of hotspot and non-hotspot circuit elements, e.g., due to stochastic failures.202414041

[0014] Attributing an entire cluster in a binary labeling scheme with only a “pass” or “fail” label cannot adequately represent stochastic failures that can occur for layout structures of a circuit design. For such stochastic failures, defects and hotspots may sometimes arise for circuit structures with very similar characteristics, and thus hotspot feature vectors and non-hotspot feature vectors may have similar feature values. Accordingly, these hotspot and non-hotspot feature vectors may be clustered together, and definitively labeled as “hotspot” and “non-hotspot” in conventional labeling and prediction schemes. The limitations of such conventional binary labeling schemes can cause issues in hotspot prediction and limit the accuracy of prediction models trained with such binary-labeled training data. If a cluster comprised of hotspot and non- hotspot feature vectors is wholly-labeled as “fail” due to the presence of even a single hotspot feature vector, then all other feature vectors also grouped into this same cluster will be predicted as failure data points. Such an all-or-nothing cluster labeling scheme can lead to high false alarm rates for hotspot prediction, reducing the accuracy and usefulness of such prediction models trained with conventional binary cluster labeling techniques.

[0015] The disclosure herein may provide systems, methods, devices, and logic for hotspot prediction for circuit designs through failure rate-aware prediction models. Any technical aspect described herein may be a part of the failure rate-aware prediction technology of the present disclosure. As described herein, failure rates can be computed on a per-cluster basis, allowing for characterization and labeling of partitions in a feature space with computed failure rate values instead of a binary “pass” or “fail” label. Accordingly, the failure rate-aware prediction technology described herein can enable failure rate-based rankings during cluster generation as well as hotspot prediction stages, increasing the flexibility and accuracy of detecting stochastic failures and hotspots thereof. Failure rate-aware prediction models as described herein can differentiate feature space partitions that clusters are located within based on computed failure rates from training data, allowing for finer and more accurate predictions from input circuit designs. As compared to conventional binary labeling techniques, limited to “fail” or “not fail’ labels, ranking of circuit elements can happen based on failure rate values, allowing for more comprehensive, efficient, and accurate hotspot prediction processes.202414041

[0016] These and other technical features and technical benefits of the failure rate- aware prediction technology are presented herein.

[0017] Figure 1 shows an example of a computing system that supports hotspot prediction for circuit designs through failure rate-aware prediction models. The computing system 100 may take the form of a single or multiple computing devices such as application servers, compute nodes, desktop or laptop computers, smart phones or other mobile devices, tablet devices, embedded controllers, and more. In some implementations, the computing system 100 hosts, instantiates, executes, supports, or implements an EDA application or EDA system that supports circuit design and analysis, and may accordingly provide or implement any of the failure rate- aware prediction technology described herein.

[0018] As an example implementation to support any combination of the failure rate- aware prediction technology described herein, the computing system 100 shown in Figure 1 includes a failure rate-aware prediction engine 110. The computing system 100 may implement the failure rate-aware prediction engine 110 (including components thereof) in various ways, for example as hardware and programming. The programming for the failure rate-aware prediction engine 110 may take the form of processor-executable instructions stored on a non-transitory machine-readable storage medium and the hardware for the failure rate-aware prediction engine 110 may include a processor to execute those instructions. A processor may take the form of single processor or multi-processor systems, and in some examples, the computing system 100 implements multiple engines using the same computing system features or hardware components (e.g., a common processor or a common storage medium).

[0019] In operation, the failure rate-aware prediction engine 110 may access a training dataset comprised of confirmed hotspot locations and confirmed non-hotspot locations on manufactured circuits of a training circuit design and train a failure rate- aware prediction model through the training dataset. Training of the failure rate-aware prediction model by the failure rate-aware prediction engine 110 may include correlating confirmed hotspot locations and the confirmed non-hotspot locations to circuit elements of the training circuit design, extracting hotspot feature vectors for the confirmed hotspot locations and non-hotspot feature vectors for the confirmed non- hotspot locations from the correlated circuit elements of the training circuit design, and clustering the hotspot feature vectors and the non-hotspot feature vectors into feature202414041 vector clusters in a partitioned feature space based on feature values of the hotspot feature vectors and the non-hotspot feature vectors.

[0020] Training of the failure rate-aware prediction model by the failure rate-aware prediction engine 110 may further include determining failure rate values for feature space partitions that the feature vector clusters are located within in the partitioned feature space. For a given feature space partition, the failure rate-aware prediction engine 110 may compute a failure rate value based on a number of hotspot feature vectors included in a given feature vector cluster located within the given feature space partition. In operation, the failure rate-aware prediction engine 110 may further analyze an input circuit design through the failure rate-aware prediction model to predict hotspot locations in the input circuit design, doing so in any of the ways described herein.

[0021] These and other technical features and technical benefits of the failure rate- aware prediction technology are described in greater detail next.

[0022] Figure 2 shows an example training of a failure rate-aware prediction model according to the present disclosure. In the example of Figure 2, the failure rate-aware prediction engine 110 may access a training dataset 210. The training dataset 210 may take the form of any suitable data format by which the failure rate-aware prediction engine 110 may extract training data. The training dataset 210 may include confirmed hotspot locations, confirmed non-hotspot locations, or a combination of both. The hotspot and non-hotspot locations in the training dataset 210 may be confirmed in the sense that the determined locations are verified on physically manufactured circuits. In the example of Figure 2, the training dataset 210 includes confirmed hotspot locations (shown as confirmed hotspot locationi and more) and confirmed non-hotspot locations (shown as confirmed non-hotspot location and more), which may be physically verified from the manufactured circuits 220. The manufactured circuits 220 may be physically manufactured ICs or wafers of a training circuit design, which may refer to any circuit design by which the failure rate-aware prediction engine 110 can use hotspot or non-hotspot location data in support of failure rate-aware predictions.

[0023] The training dataset 210 may be represented in any number of forms and can vary based on a particular imaging or detection technique used to identify the circuit defects on the manufactured circuits 220. In some examples, the manufactured circuits 220 may include one or more circuit wafers or circuit lots manufactured with202414041 specific process parameters. Defects on the manufactured circuits 220 may be physically detected using any number of circuit imaging techniques, such as SEM techniques, or any other circuit inspection process. In such cases, hotspot locations may be represented in the training dataset 210 as positional identifiers, circuit coordinates, captured image data (e.g., centered around a detected hotspot location), and the like.

[0024] In some implementations, the training dataset 210 accessed by the failure rate-aware prediction engine 1 10 may take the form of image data. SEM images are used as a continuing illustrative example, but any suitable image format or imaging technique is contemplated herein. High-precision images, like SEM images, can capture circuit features of manufactured circuits in order to verify or otherwise detect circuit defects, failures, or other hotspot attributes (e.g., bridging effects pinches, or other circuit defects). Each SEM image may capture a field of view of a circuit (e.g., a 10 micrometer by 10 micrometer view, or windows of any other configurable dimensions), and confirmed hotspot locations in the field of view may be pinpointed or annotated in the image, or otherwise specified in any other suitable manner. A confirmed non-hotspot location may take the form of a SEM image without any hotspot locations in the field of view (and may thus represent an entire circuit section without hotspots, instead of merely a singular point). As such, hotspot and non-hotspot locations confirmations through SEM images may be one example format by which confirmed hotspot locations or confirmed non-hotspot locations can be represented in a training dataset 210. The failure rate-aware prediction engine 110 may access the training dataset 210 in any suitable manner. For example, the failure rate-aware prediction engine 110 may load the training dataset 210 from a memory, receive the training dataset 210 over a communication network, or obtain the training dataset 210 via user input via an EDA application.

[0025] The failure rate-aware prediction engine 110 may support failure rate-aware model training through the training dataset 210. In the example of Figure 2, the failure rate-aware prediction engine 110 trains a model shown as the failure rate-aware prediction model 230. The failure rate-aware prediction model 230 may take the form of any supervised or semi-supervised machine learning (ML) model, and may utilize, incorporate, implement, or otherwise access any type of ML technology in support of hotspot predictions in circuit designs. For example, the failure rate-aware prediction202414041 model 230 may include, implement, or utilize any relevant clustering technology, allowing the failure rate-aware prediction model 230 to cluster circuit elements correlated to hotspot and non-hotspot locations in the training dataset 210 or extracted feature vectors thereof. The failure rate-aware prediction engine 110 may train, construct, or maintain the failure rate-aware prediction model 230. In some implementations, the failure rate-aware prediction engine 110 may itself implement the failure rate-aware prediction model 230.

[0026] In contrast to conventional binary labeling techniques, the failure rate-aware prediction engine 110 may train the failure rate-aware prediction model 230 with cluster labels indicative of failure rate. Such cluster labels may apply to feature space partitions in which a given cluster is located within or otherwise defined by. In support of failure rate-aware predictions, the failure rate-aware prediction engine 110 may correlate hotspot and / or non-hotspot locations to circuit elements of a circuit design, and group together circuit elements with similar characteristics into a given cluster. Failure rate determinations can be computed for clustered feature vectors, and in any of the ways described herein. Example features of failure rate determinations for training of the failure rate-aware prediction model 230 are described next with reference to Figure 3.

[0027] Figure 3 shows an example determination of failure rate values from feature vector clusters for training of a failure rate-aware prediction model. In the example of Figure 3, the failure rate-aware prediction engine 110 correlates confirmed hotspot locations and confirmed non-hotspot locations (e.g., from the training dataset 210) to circuit elements specifically in the form of OPC fragments. The failure rate-aware prediction engine 110 may do so in any suitable or relevant manner. Example technical aspects of circuit element correlations and feature vector extractions are described in International Application Number PCT / US2020 / 041141 titled “MACHINE LEARNING-BASED DOWN SELECTION OF CANDIDATE HOTSPOT LOCATIONS OF CIRCUIT DESIGNS” (the ‘141 application), the entirety of which is incorporated by reference herein. The failure rate-aware prediction engine 110 may implement any of the techniques and technical aspects of the ‘141 application in support of failure rate- aware predictions according to the present disclosure.

[0028] For example, the failure rate-aware prediction engine 110 may correlate confirmed hotspot locations and confirmed non-hotspot locations of the training202414041 dataset 210 detected on physically manufactured circuits to layout data for a training circuit design of the physically manufactured circuits. Layout data may refer to or include any circuit data for a given circuit design, e.g., at a polygon-level of a circuit design. As such, layout data may refer to a physical circuit design that includes, describes, or represents specific geometric elements (e.g., polygons) that define the shapes and circuit components that will be created in various circuit materials in order to physically manufacture the circuit. Through layout data, physical layers of a physical circuit can have a corresponding layer representation in the layout design, and the geometric elements described in a layer representation can define the relative locations of the circuit device components that will make up a physically-manufactured circuit.

[0029] In the example of Figure 3, the failure rate-aware prediction engine 110 correlates confirmed hotspot locationi (e.g., as included in a SEM image of the training dataset 210) to the layout data 310. The layout data 310 may include a specific portion of a layout design that includes the position of confirmed hotspot locationi in a given layer of the layout design as well as surrounding geometric elements, e.g., in a polygon of the layout design surrounding confirmed hotspot locationi. In some examples, the failure rate-aware prediction engine 110 may determine the layout data 310 as that corresponding to the circuit portion depicted in a SEM image that includes confirmed hotspot locationi. To correlate a given hotspot location to layout data, the failure rate- aware prediction engine 110 may identify a position in a layout design at which the given hotspot location occurs. For instance, the training dataset 210 may depict confirmed hotspot locationi at a particular location and physical layer in a physically- manufactured chip, whether specified via circuit coordinates, physical measures or offsets from a particular circuit position, and the like. The failure rate-aware prediction engine 110 may then identify this particular location in a layout design of the training circuit design, which may include positional coordinates as well as a target layer corresponding to the physical chip location at which confirmed hotspot locationi occurs.

[0030] The failure rate-aware prediction engine 110 may access layout data 310 that includes decomposed OPC fragments from EDA applications, as OPC or other EDA- based resolution enhancement techniques (RET) may generate fragmented layout data. Delineations of polygon-level data (e.g., decomposed into OPC fragments) may202414041 provide a mechanism through which the failure rate-aware prediction engine 110 may perform feature vector extraction for different portions of an extraction window or circuit design, thus allowing for failure rate-based labeling of clusters to support subsequent ML modeling and failure rate-aware hotspot predictions. While many of the feature vector extraction examples described herein are presented with reference to OPC fragments, any other delineated sub-portions of circuit designs or layout designs are contemplated herein for feature vector extraction, data characterization, failure ratebased cluster labeling, or any form of failure rate-aware hotspot predictions as described herein.

[0031] For a given confirmed hotspot location, the failure rate-aware prediction engine 110 may correlate the confirmed hotspot location to circuit elements proximate to the confirmed hotspot location, e.g., within a threshold distance. An example of proximity is shown in Figure 3 through the interaction zone 312 that surrounds confirmed hotspot locationi of the training dataset 210. The failure rate-aware prediction engine 110 may configure the shape, dimensions, or any other relevant parameter of the interaction zone 312, e.g., as a predetermined shape configuration, via user input, or combinations thereof. The failure rate-aware prediction engine 110 may identify any OPC fragments within or intersecting the interaction zone 312 as hotspot fragments. In Figure 3, the failure rate-aware prediction engine 110 identifies six OPC fragments in the layout data 310 as hotspot fragments 314 for confirmed hotspot locationi of the training dataset 210. In a consistent manner, the failure rate- aware prediction engine 110 may identify hotspot fragments for other confirmed hotspot locations of the training dataset 210.

[0032] After identification of hotspot fragments for confirmed hotspot locations of a training dataset 210, the failure rate-aware prediction engine 110 may extract hotspot feature vectors for the hotspot fragments. The failure rate-aware prediction engine 110 may extract a feature vector from a given OPC fragment in any number forms, and each feature vector extracted from layout data of a circuit design may represent a discrete OPC fragment (or other circuit portion) of a circuit design. The failure rate- aware prediction engine 110 may extract feature vectors of any type or format, and extracted fragment feature vectors may track any number of characteristics of OPC fragments of a circuit design. In some implementations, extracted feature vectors may take the form of n-dimensional vectors of numerical parameters values captured for202414041OPC fragments. As an illustrative example, a feature vector extracted by the failure rate-aware prediction engine 110 may represent a given OPC fragment in a layout design and example parameter values of the extracted feature vector may numerically represent OPC fragment and geometry data, lithographical data, intensity or density values, fragment lengths, simulation or convolution-based geometry data of the OPC fragment and neighboring fragments, contour data, fragment positional data, neighboring geometry, or any number of additional or alternative characteristics specific to the OPC fragment. In some implementations, the failure rate-aware prediction engine 110 may extract feature vectors that represent micro-level characteristics of OPC fragments (whether for the OPC fragment itself or neighboring OPC fragments, but not for entire chip parameters or characteristics). In the example of Figure 3, the failure rate-aware prediction engine 110 extracts the hotspot feature vectors 320 for the hotspot fragments 314, including the six hotspot fragments identified for confirmed hotspot locationi.

[0033] As another example shown in Figure 3, the failure rate-aware prediction engine 110 may extract non-hotspot feature vectors, including for confirmed nonhotspot locations of the training dataset 210. The failure rate-aware prediction engine 110 may do so by correlating a given confirmed non-hotspot location to circuit elements, such as OPC fragments of the layout data 330 in Figure 3. As confirmed non-hotspot locations may be in the form of SEM images without any hotspots, the failure rate-aware prediction engine 110 may identify layout data in a layout design that corresponds to the circuit portion depicted by a SEM image for a confirmed non- hotspot location. Each circuit element in the circuit image may identified by the failure rate-aware prediction engine 110 as a non-hotspot circuit element. Thus, the failure rate-aware prediction engine 110 may identify each OPC fragment in the layout data 330 for confirmed non-hotspot locationi as non-hotspot fragments 334 for confirmed non-hotspot locationi of the training dataset 210. Then, the failure rate-aware prediction engine 110 may extract feature vectors for the non-hotspot fragments 334, shown as the non-hotspot feature vectors 340 in Figure 3.

[0034] In any such manner, the failure rate-aware prediction engine 110 may extract hotspot feature vectors and non-hotspot feature vectors from a training dataset of confirmed hotspot locations and confirmed non-hotspot locations. Note that the failure rate-aware prediction engine 110 can extract non-hotspot feature vectors from circuit202414041 elements within a depicted circuit section of a confirmed hotspot location (e.g. within a SEM image), but outside of the interaction zone 312 or at least a threshold distance from the interaction zone 312 (e.g., in any of the ways described in the ‘141 application). Note that the failure rate-aware prediction engine 110 may extract hotspot feature vectors and non-hotspot feature vectors in a consistent or identical manner and extraction process, e.g., with the same feature vector format or with identical features. Doing so may allow for characterization, and thus clustering, of hotspot and non-hotspot feature vectors together. The failure rate-aware prediction engine 110 may thus extract hotspot feature vectors for confirmed hotspot locations and non-hotspot feature vectors for confirmed non-hotspot locations of a training dataset 210 from correlated circuit elements of a training circuit design.

[0035] After extraction, the failure rate-aware prediction engine 110 may cluster hotspot feature vectors 320 and non-hotspot feature vectors 340 into feature vector clusters. The clustering capabilities used to group feature vectors may be implemented or provided by the failure rate-aware prediction model 230, which the failure rate-aware prediction engine 110 may use, implement, or access to provide any of the clustering capabilities described herein. As used herein, “clustering” by the failure rate-aware prediction engine 110 may be interchangeable terms with “clustering” by the failure rate-aware prediction model 230. In Figure 3, the failure rate-aware prediction model 230 clusters hotspot feature vectors 320 and non-hotspot feature vectors 340 into feature vector clusters 350. The failure rate-aware prediction model 230 may cluster extracted feature vectors in a partitioned feature space based on feature values of the hotspot feature vectors and the non-hotspot feature vectors. A partitioned feature space may refer to any space that is delineated into different sections (e.g., partitions). The failure rate-aware prediction model 230 may apply, implement, or use any relevant or suitable clustering algorithm, technique, or process to determine the feature vector clusters 350. In particular, the failure rate-aware prediction model 230 may partition a feature space of the hotspot and non-hotspot feature vectors, and group the feature vectors located in (e.g., mapped into) each feature space partition in the partitioned feature space into the same feature vector cluster.

[0036] To illustrate, the extracted hotspot and non-hotspot feature vectors may be ‘n’-dimensional vectors with ‘n’ number of different feature types, and thus ‘n’ number202414041 of different feature values for each feature vector. The failure rate-aware prediction engine 110 may configure the failure rate-aware prediction model 230 to determine a feature space of up to ‘n’ dimensions, with each dimension in the feature space for a respective feature value of the ‘n’ feature value types. Any suitable feature vector processing, feature vector transformations, feature dimension selections and the like may be performed by the failure rate-aware prediction engine 110 in support of clustering of feature vectors and to tune the failure rate-aware prediction model 230. Partitioning of dimension ranges in each dimension of a feature space may specify distinct value ranges in the dimension, and the partitioned value ranges across each dimension may together form a partitioned feature space. The failure rate-aware prediction model 230 may then map the hotspot and non-hotspot feature vectors into the partitioned feature space, and the mapped feature vectors in a given feature space partition may form a feature vector cluster. As noted herein, the failure rate-aware prediction engine 110 and the failure rate-aware prediction model 230 may employ any suitable clustering process, which may include any suitable feature space generation and partitioning techniques, algorithms, or processes. Example feature space generation and partitioning techniques are described in International Application Number PCT / US2020 / 041153, titled “HYPERSPACE-BASED PROCESSING OF DATASETS FOR ELECTRONIC DESIGN AUTOMATION (EDA) APPLICATIONS” (the ‘153 application), the entirety of which is incorporated by reference herein. The failure rate-aware prediction engine 110 or the failure rate- aware prediction model 230 may implement any of the techniques and technical aspects of the ‘153 application in support of failure rate-aware predictions according to the present disclosure.

[0037] Thus, the failure rate-aware prediction engine 110 may generate a partitioned feature space through the failure rate-aware prediction model 230 in any suitable manner and map the hotspot and non-hotspot feature vectors extracted for the training dataset 210 into the partitioned feature space. Each different feature space partition in the partitioned feature space may be specific to a different feature vector cluster. Understood in a different way, each feature space partition may define a separate cluster in the feature space, such that any feature vectors with feature values that are mapped to a given feature space partition are assigned or grouped into the same feature vector cluster. In that regard, a feature space partition may define similarity202414041 requirements between feature vectors to be grouped into the same cluster. Thus, the failure rate-aware prediction model 230 may group any feature vectors mapped into the same feature space partition of the partitioned feature space into the same feature space cluster. In such a manner, the failure rate-aware prediction engine 110 may determine the feature vector clusters 350 through the failure rate-aware prediction model 230 for the hotspot and non-hotspot feature vectors extracted for a training dataset 210.

[0038] As the failure rate-aware prediction model 230 may perform a clustering process to map both hotspot and non-hotspot feature vectors into the same partitioned feature space, any given feature space partition in the partitioned feature space may include a combination of hotspot and non-hotspot feature vectors. A given feature vector cluster among the feature vector clusters 350 may thus include a combination of hotspot and non-hotspot feature vectors. Understood in another way, a given feature vector cluster may include one or more hotspot feature vectors, one or more non-hotspot feature vectors, or a combination thereof, located within a given feature space partition in the partitioned feature space. As each feature vector may be extracted from, and thus represent, a given circuit element (e.g., OPC fragment), the clusters determined by the failure rate-aware prediction engine 110 may be understood as feature vector clusters, circuit element clusters (e.g., fragment clusters), or both.

[0039] The failure rate-aware prediction engine 110 may determine labels from individual clusters among the feature vector clusters 350. In contrast to conventional binary labeling schemes, the failure rate-aware prediction engine 110 may instead determine labels from the feature vector clusters 350 with a value indicative of a relative degree or rate of failure among the elements that form a given cluster, e.g., through a failure rate value. In the context of hotspot and non-hotspot feature vectors, the failure rate-aware prediction engine 110 may measure or express failure values as a function of a number of hotspot feature vectors in a feature vector clusters. That is, the failure rate-aware prediction engine 110 may determine failure rate values for a given feature vector cluster as a function of the number of hotspot feature vectors included in the given feature vector cluster. Thus, instead of a simple “pass” or “fail” labeled determined based on the presence of any hotspot feature vector in a given feature vector cluster, a failure rate value computed by the failure rate-aware202414041 prediction engine 110 may instead provide insight into the rate or proportion of which hotspot feature vectors occur in a given feature vector cluster.

[0040] The failure rate-aware prediction engine 110 may determine the failure rate value for feature vector clusters in various ways. As an illustrative example, the failure rate-aware prediction engine 110 may compute the failure rate value for a given feature vector cluster as a ratio of a number of hotspot feature vectors included in the given feature vector cluster to a total number of feature vectors in the given feature vector cluster. Such an implementation of the failure rate value may indicate or specify a percentage of the feature vectors in a given feature vector cluster that are hotspot feature vectors. Accordingly, such a failure rate value implementation may specify or express a rate of hotspot circuit failures for circuit elements (and extracted feature vectors thereof) for a given cluster. As another implementation example, the failure rate-aware prediction engine 110 may compute the failure rate value of given feature vector cluster as a ratio of the number of hotspot feature vectors included in the given feature vector cluster to the number of non-hotspot feature vectors included the given feature vector cluster. This implementation may similarly specify a rate at which hotspots occur for similarly characterized circuit elements, albeit as a ratio or comparison between hotspot and non-hotspot occurrences. While some examples are presented herein, any suitable function based on the number of hotspot feature vectors in a given feature vector cluster is contemplated herein for failure rate value computations by the failure rate-aware prediction engine 110.

[0041] Note that, in the present disclosure, any failure rate value computations based on a number of non-hotspot feature vectors in a feature vector cluster (e.g., percentage of non-hotspot feature vectors in a given feature vector cluster) can be understood as a suitable function that is (at least indirectly) based on the number of hotspot feature vectors in a given feature vector cluster, and thus within the scope of a failure rate value computation as described herein. This may be case since the number of non-hotspot feature vectors may be expressed as the total number of feature vectors in a given feature vector cluster minus the number of hotspot feature vectors in the given feature vector cluster. Thus, any function of the number of non- hotspot feature vectors in a given feature vector cluster can be understood also as a function of the number of hotspot feature vectors in a given feature vector cluster.202414041

[0042] The failure rate-aware prediction engine 110 may compute failure rate values for each feature vector cluster of the feature vector clusters 350 in Figure 3, the computed values shown as the failure rate values 360. A given feature vector cluster (for a which a given failure rate value is computed) may be comprised of feature vectors located within (e.g., mapped to) a given feature space partition in the partitioned feature space. Thus, the failure rate value computed for the given feature vector cluster from the training dataset 210 can be used to characterize the given feature space partition in the partitioned feature space in which given feature vector cluster is located within, e.g., the parameters or space that define the dimensional limits of particular feature vector cluster. Thus, the failure rate-aware prediction engine 110 may utilize the computed failure rate values 360 from the feature vector clusters 350 to characterize or label specific feature space partitions of a feature space. Such labeling may be useful in order to characterize feature vectors extracted from other circuit data (e.g., outside of the training dataset 210). Accordingly, the training dataset 210 may allow the failure rate-aware prediction engine 110 to determine failure rates for different feature vector clusters located in different feature space partitions of a partitioned feature space, and labeling of these different feature vector clusters may be implemented by labeling the feature space partitions that each of these different feature vector clusters are located within. Doing so may support subsequent characterization and analysis of feature vectors extracted from input circuit designs.

[0043] Thus, in Figure 3, the failure rate-aware prediction engine 110 may train the failure rate-aware prediction model 230 with the determined failure rate values 360. Example aspects of a trained failure rate-aware prediction model according to the present disclosure are described next with reference to Figure 4.

[0044] Figure 4 shows an example of a trained failure rate-aware prediction model according to the present disclosure. In the example of Figure 4, the failure rate-aware prediction engine 110 trains the failure rate-aware prediction model 230 with failure rate values 360 computed for feature vector clusters 350 of a training dataset 210. The failure rate-aware prediction engine 110 may assign determined failure rate values to the various feature space partitions of a partitioned feature space used to cluster the hotspot and non-hotspot feature vectors extracted for the training dataset 210. Example feature space partitions of a partitioned feature space for feature vectors are shown in Figure 4, labeled as partition , partition, partitions, and so forth.202414041Each of these partitions may represent a different partition in the partitioned feature space (e.g., different section in a ‘n’-dimensional space) used by the failure rate-aware prediction model 230 to cluster extracted feature vectors. The failure rate-aware prediction engine 110 may assign a computed failure rate value for each partition, such as the failure rate value of “0.25” assigned to partitioni , as shown in Figure 4.

[0045] In Figure 4, the failure rate-aware prediction engine 110 assigns a failure rate value for a feature space partition as a percentage of the feature vectors located within the feature space partition that are hotspot feature vectors. To illustrate, partitioni of the partitioned feature space may include or define a cluster of feature vectors comprised of two (2) hotspot feature vectors and six (6) non-hotspot feature vectors extracted from OPC fragments correlated through the training dataset 210. Thus, in this illustrative example, the failure rate-aware prediction engine 110 may compute and assign the failure rate value of “0.25” to partitioni based on the feature vector cluster located within and defined by partitioni of a partitioned feature space. In a similar manner, the failure rate-aware prediction engine 110 may compute and assign failure value rates for other partitions of the partitioned feature space used by the failure rate-aware prediction model 230 to cluster extracted feature vectors from the training dataset 210.

[0046] In some implementations, training of the failure rate-aware prediction model 230 may include tuning clustering parameters employed by the failure rate-aware prediction model 230. A clustering parameter may refer to any configurable value that affects how a clustering process employed by the failure rate-aware prediction model 230 clusters circuit elements, e.g., impacts generation or partitioning of a feature space for clustering of extracted feature vectors. The clustering parameters tuned by the failure rate-aware prediction engine 110 may be different from the feature values of extracted feature vectors, and may instead refer to tuning variables that, for example, control how strictly or loosely a clustering process partitions feature space dimensions, the number of partitions in a given dimension, the number of feature dimensions of the extracted feature vectors to include in the feature space, and so forth. Thus, clustering parameters tuned by the failure rate-aware prediction engine 110 may include any parameter or clustering process configuration that affects the number or size of partitions within a partitioned feature space generated by the failure rate-aware prediction model 230.202414041

[0047] To tune clustering parameter values, the failure rate-aware prediction engine 110 may utilize a validation dataset to assess performance of the failure rate-aware prediction model 230 trained using different values of the clustering parameter. The failure rate-aware prediction engine 110 may configure or train the failure rate-aware prediction model 230 with the value for the clustering parameter that results in the best performance, as quantifiably assessed through one or more performance metrics. A validation dataset used by the failure rate-aware prediction engine 1 10 may be in the same form as the training dataset 210 used to train the failure rate-aware prediction model 230. Thus, a validation dataset may include confirmed hotspot and non-hotspot locations (e.g., in the form of SEM images) through which the failure rate-aware prediction engine 110 can validate performance of the failure rate-aware prediction model 230 in clustering or prediction of hotspots for circuit elements. In some implementations, the failure rate-aware prediction engine 1 10 may obtain the training dataset 210 and validation dataset from a common input dataset. The input dataset may be a set of circuit images or other circuit data comprised of confirmed hotspot and non-hotspot locations of manufactured circuits, e.g., verified via SEM inspection. The failure rate-aware prediction engine 110 may split the input dataset into a training dataset portion and a validation dataset portion, e.g., 70% of the confirmed hotspot and non-hotspot locations of the input dataset as part of the training dataset and the remaining 30% of the confirmed hotspot and non-hotspot locations as part of the validation dataset.

[0048] In support of clustering parameter tuning for the failure rate-aware prediction model 230, the failure rate-aware prediction engine 110 may train the failure rate- aware prediction model 230 with the training dataset 210 with multiple different values of the clustering parameter. As such, the failure rate-aware prediction engine 110 may configure the failure rate-aware prediction model 230 to use a first value for a clustering parameter, then to use a second value for the clustering parameter, then to use a third value for the clustering parameter, and so forth. For each different value of the clustering parameter, a different partitioned feature space may be generated and used to cluster the extracted feature vectors of the training dataset, thus resulting in different number of feature vector clusters and different values for computed failure rate values. The failure rate-aware prediction engine 110 may then assess performance of the failure rate-aware prediction model 230, for each of the different202414041 values of the clustering parameter, doing so through the validation dataset. The validation dataset may be used to assess how the failure rate-aware prediction model 230 clusters and characterizes circuit elements for layout data correlated to the confirmed hotspot non-hotspot locations included in the validation dataset.

[0049] Characterizing (e.g., hotspot prediction rankings for) circuit elements at the confirmed hotspot and non-hotspot locations of the validation dataset may include extracting feature vectors for the circuit elements and clustering the extracted feature vectors in the partitioned feature space generated by the failure rate-aware prediction model 230 for each value of the clustering parameter. Through the clustering, the different correlated circuit elements can be characterized by the failure rate-aware prediction model 230 with a failure rate value based on the feature space partition of the partitioned feature space that the failure rate-aware prediction model 230 maps each feature vector into. Thus, a failure rate can be assigned to the different feature vectors (and, thus, circuit elements) for circuit locations of the validation dataset. Then, the predictive performance of the failure rate-aware prediction model 230 for correlated circuit elements can be compared against the actual confirmed hotspot locations and non-hotspot locations specified in the validation dataset. Performance of the failure rate-aware prediction model 230 with differing values of a clustering parameter can be assessed accordingly.

[0050] Any suitable performance metric is contemplated herein. Example performance metrics for tuning of clustering parameters include a number of feature space partitions that are needed by the failure rate-aware prediction model 230 to identify all (or a threshold portion) of the confirmed hotspot locations in the validation dataset. Another example performance metric is the number of confirmed hotspot locations included in the top one hundred (or any other configurable metric number) of feature space partitions with the assigned highest failure rate values. Such a metric may indicate how relatively efficient and accurate a value of a clustering parameter. The higher the efficiency in accuracy of a clustering parameter value, the better the performance. Thus, the failure rate-aware prediction engine 110 may assess performance and select a value for a clustering parameter for tuning the failure rate- aware prediction model 230 that results in the best performance according to a performance metric.202414041

[0051] In any of the ways described herein, the failure rate-aware prediction engine 110 may tune clustering parameters used by the failure rate-aware prediction model 230 through a validation dataset. As such, the failure rate-aware prediction engine 110 may access a validation dataset of confirmed hotspot locations and confirmed non-hotspot locations on manufactured circuits of a training circuit design, and the validation dataset may include different confirmed hotspot locations and confirmed non-hotspot locations on the manufactured circuits of the training circuit design from a training dataset. The failure rate-aware prediction engine 110 may then train the failure rate-aware prediction model 230 through both the training dataset and the validation dataset, including by tuning a clustering parameter (e.g., partition sizes) in the partitioned feature space based on a performance metric of the failure rate-aware prediction model 230 applied to the validation dataset.

[0052] Through a trained failure rate-aware prediction model 230, analysis of input circuit designs can be performed for hotspot prediction, example features of which are described next with reference to Figure 5.

[0053] Figure 5 shows an example determination of predicted hotspots for an input circuit design through a failure rate-aware prediction model. The failure rate-aware prediction engine 110 may analyze an input circuit design 510 through a trained failure rate-aware prediction model 230, including with assigned failure rate values for feature space partitions. The input circuit design 510 may be in any circuit design format. In some implementations, the input circuit design 510 may be for the same circuit layer or be designed for manufacture with the same process parameters as the training dataset 210 used to train the failure rate-aware prediction model 230.

[0054] To analyze the input circuit design 510, the failure rate-aware prediction engine 110 may extract feature vectors for the input circuit design 510, doing so for circuit elements of the input circuit design 510. For training of the failure rate-aware prediction model 230 through OPC fragments, the failure rate-aware prediction engine 110 may likewise identify OPC fragments in the input circuit design 510 and extract feature vectors for the OPC fragments in the input circuit design 510. The failure rate- aware prediction engine 110 may do so in a consistent manner or feature vector extraction process as that used to train the failure rate-aware prediction model 230. Then, the failure rate-aware prediction engine 110 may provide to the failure rate- aware prediction model 230, as inputs, the extracted feature vectors for the input202414041 circuit design 510. The failure rate-aware prediction model 230 may cluster the feature vectors extracted for the input circuit design 510 in the same partitioned feature space used to determine the failure rate values for the training dataset 210. Doing so may include mapping individual feature vectors the partitioned feature space generated and used by the failure rate-aware prediction model 230, e.g., into feature space partitions thereof. Then, the failure rate-aware prediction engine 110 may assess the feature vectors (and corresponding circuit elements) based on the failure rate values assigned for the feature space partitions that each of the feature vectors is located within.

[0055] Through feature space partitions labeled with failure rate values, the failure rate-aware prediction engine 110 (including through the failure rate-aware prediction model 230) may characterize feature vectors and correlated circuit elements. Correlated circuit elements with higher failure rate values may be assessed by the failure rate-aware prediction engine 110 as more likely to be characterized as a hotspot feature vector, and thus more likely to include a hotspot. As such, the failure rate-aware prediction engine 110 may predict hotspot locations for the input circuit design 510 based on the correlated circuit elements of the input circuit design 510 that have a highest failure rate value, based on mapping of feature vectors into the partitioned feature space used by the failure rate-aware prediction model 230. In Figure 5, the failure rate-aware prediction engine 110 may determine the predicted hotspots 520 for the input circuit design 510 through the trained failure rate-aware prediction model 230.

[0056] In some implementations, the failure rate-aware prediction engine 110 may analyze the input circuit design 510 through the failure rate-aware prediction model 230 to predict hotspot locations in the input circuit design 510 by generating field of view windows in the input circuit design 510 and ranking the generated field of view windows based on failure rate values of circuit elements of the input circuit design 510 included in the generated field of view windows. Any suitable application of predicted hotspots 520 from the trained failure rate-aware prediction model 230 is contemplated herein. Predicted hotspots 520 (e.g., in highest-ranked generated fields of view) may be provided by the failure rate-aware prediction engine 110 for subsequent analysis. For example, the failure rate-aware prediction engine 110 may provide predicted hotspot locations in the input circuit design 510 for performance of an image analysis at the predicted hotspot locations on manufactured circuits of the input circuit design202414041510, e.g., via SEM image analyses. The failure rate-aware prediction engine 110 may flag or annotate predicted hotspot locations in the input circuit design 510, and EDA applications can support subsequent circuit modifications or engineer analysis for such locations in the input circuit design 510 to reduce likelihood of defects or hotspots. In some implementations, the failure rate-aware prediction engine 110 may itself support manufacture of the input circuit design 510, e.g., after modification of selected locations of predicted hotspots or in support of further SEM image analyses at the predicted hotspot locations.

[0057] Through any of the ways described herein, the failure rate-aware prediction engine 110 may support failure rate-aware predictions of hotspots through trained failure rate-aware prediction models. As compared to conventional binary cluster labeling schemes, determined failure rate values may indicate a relative rate of failure (e.g., relative rate of hotspot vectors in a feature vector cluster). Failure rate values may be applied as labels to feature space partitions used to cluster extracted feature vectors of a training dataset. Subsequent feature vector extraction and mapping into the partitioned feature space via trained failure rate-aware prediction models can support characterization of circuit elements of input circuit designs based on failure rate values. Compared to binary pass / fail labels, such characterizations may provide increased accuracy in terms of characterizing circuit elements (e.g., OPC fragments). Thus, false positives can be reduced, and the specific circuit elements with greater likelihood of exhibiting stochastic failures can be more readily identified and predicted as hotspot locations.

[0058] Through any combination of the features herein, the failure rate-aware prediction technology of the present disclosure may provide technical improvements to EDA computing systems and processes. The failure rate-aware prediction technology described herein may support hotspot predictions with increased accuracy, doing so by accounting for a hotspot rate among extracted feature vectors. Trained prediction models with increased accuracy may allow for circuit verifications with increased efficiency, especially as high-precision imaging techniques can be cost and resource intensive. Efficient selection of candidate hotspot locations for further image analysis can improve circuit verification processes accordingly.

[0059] Figure 6 shows an example of logic 600 that a system may implement to support hotspot prediction for circuit designs through failure rate-aware prediction202414041 models. For example, the computing system 100 may implement the logic 600 as hardware, executable instructions stored on a machine-readable medium, or as a combination of both. The computing system 100 may implement the logic 600 via the failure rate-aware prediction engine 110, through which the computing system 100 may perform or execute the logic 600 as a method to support hotspot predictions through failure rate-aware prediction models. The following description of the logic 600 is provided using the failure rate-aware prediction engine 110 as an example implementation. However, other implementation options by computing systems are possible.

[0060] In implementing the logic 600, the failure rate-aware prediction engine 110 may access a training dataset comprised of confirmed hotspot locations and confirmed non-hotspot locations on manufactured circuits of a training circuit design (602) and train a failure rate-aware prediction model through the training dataset (604). Training of the failure rate-aware prediction model may include correlating the confirmed hotspot locations and the confirmed non-hotspot locations to circuit elements of the training circuit design (606), extracting hotspot feature vectors for the confirmed hotspot locations and non-hotspot feature vectors for the confirmed non-hotspot locations from the correlated circuit elements of the training circuit design (608), and clustering the hotspot feature vectors and the non-hotspot feature vectors into feature vector clusters in a partitioned feature space based on feature values of the hotspot feature vectors and the non-hotspot feature vectors (610). A given feature vector cluster may include one or more hotspot feature vectors, one or more non-hotspot feature vectors, or a combination thereof, located within a given feature space partition in the partitioned feature space.

[0061] Training of the failure rate-aware prediction model may also include determining failure rate values for feature space partitions that the feature vector clusters are located within in the partitioned feature space (612). For a given feature space partition, the failure rate-aware prediction engine 110 may compute a failure rate value based on a number of hotspot feature vectors included in a given feature vector cluster located within the given feature space partition. In implementing the logic 600, the failure rate-aware prediction engine 110 may further analyze an input circuit design through the failure rate-aware prediction model to predict hotspot locations in the input circuit design (614), doing so in any of the ways described herein.202414041

[0062] The logic 600 shown in Figure 6 provides an illustrative example by which a computing system 100 may support or implement various features of the failure rate- aware prediction technology described herein. Additional or alternative steps in the logic 600 are contemplated herein, including according to any of the various features described herein for the failure rate-aware prediction engine 110.

[0063] Figure 7 shows an example of a computing system 700 that supports hotspot prediction for circuit designs through failure rate-aware prediction models. The computing system 700 may include a processor 710, which may take the form of a single or multiple processors. The processor(s) 710 may include a central processing unit (CPU), microprocessor, or any hardware device suitable for executing instructions stored on a machine-readable medium. The computing system 700 may include a machine-readable medium 720. The machine-readable medium 720 may take the form of any non-transitory electronic, magnetic, optical, or other physical storage device that stores executable instructions, such as the failure rate-aware prediction instructions 722 shown in Figure 7. As such, the machine-readable medium 720 may be, for example, Random Access Memory (RAM) such as a dynamic RAM (DRAM), flash memory, spin-transfer torque memory, an Electrically-Erasable Programmable Read-Only Memory (EEPROM), a storage drive, an optical disk, and the like.

[0064] The computing system 700 may execute instructions stored on the machine- readable medium 720 through the processor 710. Executing the instructions (e.g., the failure rate-aware prediction instructions 722) may cause the computing system 700 to perform or implement any of the failure rate-aware prediction technology described herein, including according to any aspect of the failure rate-aware prediction engine 110.

[0065] For example, execution of the failure rate-aware prediction instructions 722 by the processor 710 may cause the computing system 700 to access a training dataset comprised of confirmed hotspot locations and confirmed non-hotspot locations on manufactured circuits of a training circuit design and train a failure rate-aware prediction model through the training dataset. Training of the failure rate-aware prediction model by the computing system 700 may include correlating the confirmed hotspot locations and the confirmed non-hotspot locations to circuit elements of the training circuit design, extracting hotspot feature vectors for the confirmed hotspot locations and non-hotspot feature vectors for the confirmed non-hotspot locations from202414041 the correlated circuit elements of the training circuit design, and clustering the hotspot feature vectors and the non-hotspot feature vectors into feature vector clusters in a partitioned feature space based on feature values of the hotspot feature vectors and the non-hotspot feature vectors.

[0066] Training of the failure rate-aware prediction model by the computing system 700 may also include determining failure rate values for feature space partitions that the feature vector clusters are located within in the partitioned feature space. For a given feature space partition, the failure rate-aware prediction instructions 722, when executed, may cause the computing system 700 to compute a failure rate value based on a number of hotspot feature vectors included in a given feature vector cluster located within the given feature space partition. Execution of the failure rate-aware prediction instructions 722 may further cause the computing system 700 to analyze an input circuit design through the failure rate-aware prediction model to predict hotspot locations in the input circuit design, doing so in any of the ways described herein.

[0067] Any combination of the failure rate-aware prediction technology as described herein may be implemented via the failure rate-aware prediction instructions 722.

[0068] The systems, methods, devices, and logic described above, including the failure rate-aware prediction engine 110, may be implemented in many different ways in many different combinations of hardware, logic, circuitry, and executable instructions stored on a machine-readable medium. For example, the failure rate- aware prediction engine 110, may include circuitry in a controller, a microprocessor, or an application specific integrated circuit (ASIC), or may be implemented with discrete logic or components, or a combination of other types of analog or digital circuitry, combined on a single integrated circuit or distributed among multiple integrated circuits. A product, such as a computer program product, may include a storage medium and machine-readable instructions stored on the medium, which when executed in an endpoint, computer system, or other device, cause the device to perform operations according to any of the description above, including according to any features of the failure rate-aware prediction engine 110.

[0069] The processing capability of the systems, devices, and engines described herein, including the failure rate-aware prediction engine 110, may be distributed among multiple system components, such as among multiple processors and memories, optionally including multiple distributed processing systems or202414041 cloud / network elements. Parameters, databases, and other data structures may be separately stored and managed, may be incorporated into a single memory or database, may be logically and physically organized in many different ways, and may be implemented in many ways, including data structures such as linked lists, hash tables, or implicit storage mechanisms. Programs may be parts (e.g., subroutines) of a single program, separate programs, distributed across several memories and processors, or implemented in many different ways, such as in a library (e.g., a shared library).

[0070] While various examples and features have been described above, many more implementations are possible.

Claims

1. 202414041CLAIMS1 . A method comprising: by a computing system: accessing a training dataset comprised of confirmed hotspot locations and confirmed non-hotspot locations on manufactured circuits of a training circuit design; training a failure rate-aware prediction model through the training dataset, including by; correlating the confirmed hotspot locations and the confirmed non- hotspot locations to circuit elements of the training circuit design; extracting hotspot feature vectors for the confirmed hotspot locations and non-hotspot feature vectors for the confirmed non-hotspot locations from the correlated circuit elements of the training circuit design; clustering the hotspot feature vectors and the non-hotspot feature vectors into feature vector clusters in a partitioned feature space based on feature values of the hotspot feature vectors and the non-hotspot feature vectors, wherein a given feature vector cluster includes one or more hotspot feature vectors, one or more non-hotspot feature vectors, or a combination thereof, located within a given feature space partition in the partitioned feature space; and determining a failure rate value for the given feature space partition that the given feature vector cluster is located within, the failure rate value computed based on a number of the hotspot feature vectors included in the given feature vector cluster; and analyzing an input circuit design through the failure rate-aware prediction model to predict hotspot locations in the input circuit design.

2. The method of claim 1 , wherein the circuit elements comprise optical proximity correction (OPC) fragments of layout data of the training circuit design.

3. The method of claim 1 or 2, wherein analyzing the input circuit design through the failure rate-aware prediction model comprises:202414041 extracting feature vectors for circuit elements of the training circuit design; clustering the feature vectors extracted for the input circuit design in the partitioned feature space used to train the failure rate-aware prediction model; and predicting hotspot locations in the input circuit design based on clusters of the feature vectors extracted for the input circuit design that have a highest assigned failure rate value from the training dataset.

4. The method of any of claims 1-3, comprising computing the failure rate value of the given feature vector cluster as a ratio of the number of the hotspot feature vectors included in the given feature vector cluster to a total number of feature vectors included in the given feature vector cluster.

5. The method of any of claims 1-4, further comprising accessing a validation dataset of confirmed hotspot locations and confirmed non-hotspot locations on manufactured circuits of the training circuit design, wherein the validation dataset comprises different confirmed hotspot locations and confirmed non-hotspot locations on the manufactured circuits of the training circuit design from the training dataset, and comprising training the failure rate-aware prediction model through both the training dataset and the validation dataset, including by tuning partition sizes in the partitioned feature space based on a performance metric of the failure rate-aware prediction model applied to the validation dataset.

6. The method of any of claims 1-5, wherein analyzing the input circuit design through the failure rate-aware prediction model to predict hotspot locations in the input circuit design comprises generating field of view windows in the input circuit design and ranking the generated field of view windows based on failure rate values of circuit elements of the input circuit design included in the generated field of view windows.

7. The method of any of claims 1-6, further comprising providing the predicted hotspot locations in the input circuit design for performance of an image analysis at the predicted hotspot locations on manufactured circuits of the input circuit design.2024140418. A system comprising: a processor; and a non-transitory machine-readable medium comprising instructions that, when executed by the processor, cause a computing system to: access a training dataset comprised of confirmed hotspot locations and confirmed non-hotspot locations on manufactured circuits of a training circuit design; train a failure rate-aware prediction model through the training dataset, including by; correlating the confirmed hotspot locations and the confirmed non-hotspot locations to circuit elements of the training circuit design; extracting hotspot feature vectors for the confirmed hotspot locations and non-hotspot feature vectors for the confirmed non-hotspot locations from the correlated circuit elements of the training circuit design; clustering the hotspot feature vectors and the non-hotspot feature vectors into feature vector clusters in a partitioned feature space based on feature values of the hotspot feature vectors and the non-hotspot feature vectors, wherein a given feature vector cluster includes one or more hotspot feature vectors, one or more non-hotspot feature vectors, or a combination thereof, located within a given feature space partition in the partitioned feature space; and determining a failure rate value for the given feature space partition that the given feature vector cluster is located within, the failure rate value computed based on a number of the hotspot feature vectors included in the given feature vector cluster; and analyze an input circuit design through the failure rate-aware prediction model to predict hotspot locations in the input circuit design.2024140419. The system of claim 8, wherein the circuit elements comprise optical proximity correction (OPC) fragments of layout data of the training circuit design.

10. The system of claim 8 or 9, wherein the instructions, when executed, cause the computing system to analyze the input circuit design through the failure rate- aware prediction model by: extracting feature vectors for circuit elements of the training circuit design; clustering the feature vectors extracted for the input circuit design in the partitioned feature space used to train the failure rate-aware prediction model; and predicting hotspot locations in the input circuit design based on clusters of the feature vectors extracted for the input circuit design that have a highest assigned failure rate value from the training dataset.11 . The system of any of claims 8-10, wherein the instructions, when executed, cause the computing system to compute the failure rate value of the given feature vector cluster as a ratio of the number of the hotspot feature vectors included in the given feature vector cluster to a total number of feature vectors included in the given feature vector cluster.

12. The system of any of claims 8-11 , wherein the instructions, when executed, further cause the computing system to access a validation dataset of confirmed hotspot locations and confirmed non-hotspot locations on manufactured circuits of the training circuit design, wherein the validation dataset comprises different confirmed hotspot locations and confirmed non-hotspot locations on the manufactured circuits of the training circuit design from the training dataset, and wherein the instructions, when executed, cause the computing system to train the failure rate-aware prediction model through both the training dataset and the validation dataset, including by tuning partition sizes in the partitioned feature space based on a performance metric of the failure rate-aware prediction model applied to the validation dataset.20241404113. The system of any of claims 8-12, wherein the instructions, when executed, cause the computing system to analyze the input circuit design through the failure rate-aware prediction model to predict hotspot locations in the input circuit design by generating field of view windows in the input circuit design and ranking the generated field of view windows based on failure rate values of circuit elements of the input circuit design included in the generated field of view windows.

14. The system of any of claims 8-13, wherein the instructions, when executed, further cause the computing system to provide the predicted hotspot locations in the input circuit design for performance of an image analysis at the predicted hotspot locations on manufactured circuits of the input circuit design.

15. A non-transitory machine-readable medium comprising instructions that, when executed by the processor, cause a computing system to perform a method according to any of claims 1-7.

Images