An end-to-end knowledge graph generation method

Abstract

The application relates to an end-to-end knowledge graph generation method, which comprises the following steps: acquiring at least one target data in a non-text form output in semiconductor manufacturing, each target data comprising a numerical sequence or a numerical matrix; determining a target knowledge point according to the type of each target data by using a corresponding anomaly detection model, wherein the target knowledge point comprises an abnormal type related to semiconductor manufacturing contained in corresponding target data; and generating a knowledge graph in the field of semiconductor manufacturing according to the target knowledge point.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing, and more particularly to an end-to-end knowledge graph generation method. Background Technology

[0002] The semiconductor manufacturing process generates various types of data, which can be broadly categorized into two types based on their source: one type comes from various equipment, such as equipment operating status, equipment parameters, equipment manuals, and operating logs; the other type comes from the testing data of the produced chips, such as the chip's electrical characteristics and defects. In current production processes, often only a small portion of this data is analyzed to guide production, and the majority of the data remains underutilized.

[0003] In knowledge graph generation, the conventional approach involves extracting nodes and their causal / sequential relationships from open-domain text documents using Natural Language Processing (NLP) to construct knowledge points, which are then used to generate the corresponding knowledge graph. However, due to the large amount of structured data in semiconductor manufacturing, using conventional knowledge graph generation methods requires first generating event description text from the data using a language generation model, and then extracting knowledge points from the event description text to construct the knowledge graph. This data processing involves language generation and deconstruction, which is cumbersome and inefficient. Summary of the Invention

[0004] This specification describes one or more embodiments of an end-to-end knowledge graph generation method, aiming to process various data in the semiconductor manufacturing field using an end-to-end approach to generate a knowledge graph, which can then guide the entire semiconductor production process. Furthermore, this invention focuses on knowledge points, avoiding the generation of event description text from data, but instead directly generating knowledge points from the data, and then generating a knowledge graph based on these knowledge points, thereby simplifying the knowledge graph generation process throughout semiconductor manufacturing.

[0005] Firstly, an end-to-end knowledge graph generation method is provided, including:

[0006] Acquire at least one type of target data in non-textual form generated during semiconductor manufacturing, including numerical sequences or numerical matrices;

[0007] Based on the type of each target data, the corresponding anomaly detection model is used to determine the target knowledge points, which include anomaly types related to semiconductor manufacturing contained in the corresponding target data;

[0008] Based on the target knowledge points, a knowledge graph for the semiconductor manufacturing field is generated.

[0009] In one possible implementation, the at least one target data includes: statistical process control (SPC) data, wafer map data, and fault detection and classification (FDC) data.

[0010] In one possible implementation, determining the target knowledge points using the corresponding anomaly detection model based on the type of each target data item includes:

[0011] Based on the type of each target data item, the target data is encoded using the corresponding encoding method to obtain encoded data;

[0012] The encoded data is input into the corresponding anomaly detection model, and the target knowledge point is output.

[0013] In one possible implementation, the target knowledge point is in the form of a triple, wherein the triple includes the exception type as an element.

[0014] In one possible implementation, the at least one target data item includes Statistical Process Control (SPC) data, which forms a numerical sequence; the step of determining the target knowledge points using a corresponding anomaly detection model based on the type of each target data item includes:

[0015] Feature extraction and position embedding are performed on the sequence segments of the numerical sequence to obtain a feature vector sequence;

[0016] The feature vector sequence is input into the Transformer model to obtain the anomaly type and the relative position of the anomaly;

[0017] The start and end positions of the anomaly are determined based on the position of the sequence fragment in the numerical sequence.

[0018] The triplet consisting of the exception type, the start position of the exception, and the end position of the exception is taken as the corresponding target knowledge point.

[0019] In one possible implementation, the at least one type of target data includes wafer image data; the step of determining target knowledge points using a corresponding anomaly detection model based on the type of each type of target data includes:

[0020] Based on the pixel matrix corresponding to the wafer image, the wafer image is segmented to obtain several defect sub-images containing a single wafer defect;

[0021] The target defect subgraph in the aforementioned defect subgraphs is input into the defect classification model to obtain the corresponding defect type;

[0022] The defect type is treated as an anomaly type, and the corresponding target knowledge point is determined by combining the target defect subgraph.

[0023] In one possible implementation, before inputting the target defect subgraph from the plurality of defect subgraphs into the defect classification model, the method further includes:

[0024] Feature extraction is performed on any defect subgraph to obtain subgraph features;

[0025] Whether the defect subgraph belongs to a known defect type is determined by searching and comparing within a predetermined feature set; the predetermined feature set consists of graph features of known defect types.

[0026] If the defect belongs to a known defect type, the defect subgraph is identified as the target defect subgraph.

[0027] In one possible implementation, the at least one target data includes a sequence of measurement data from a target sensor; the step of determining target knowledge points using a corresponding anomaly detection model based on the type of each target data item includes:

[0028] The target subsequence ordered first in the measurement data sequence is encoded using the encoder of a variational autoencoder (VAE) to obtain latent space variables; the temporal information of the target subsequence is extracted using a gated recurrent unit (GRU) model; the latent space variables and the temporal information are input into the decoder of the variational autoencoder (VAE) for decoding to obtain prediction data.

[0029] Based on the predicted data and the measurement data of the target subsequence in the measurement data sequence, the anomaly type related to the target sensor is determined;

[0030] Generate target knowledge points, including the anomaly type, the production equipment number to which the measurement data sequence is targeted, and the wafer number being processed by the production equipment.

[0031] In one possible implementation, before generating the knowledge graph based on the target knowledge points, the method further includes:

[0032] The knowledge points generated from the data of each objective are deduplicated to remove duplicate knowledge points.

[0033] In one possible implementation, a knowledge graph in the field of semiconductor manufacturing is generated based on the target knowledge points, including:

[0034] The anomaly types in the target knowledge points are added as entity nodes to the knowledge graph.

[0035] Secondly, an end-to-end knowledge graph generation device is provided, including:

[0036] The data acquisition unit is configured to acquire at least one type of non-textual target data produced in semiconductor manufacturing, wherein each target data includes a numerical sequence or a numerical matrix; the knowledge point determination unit is configured to determine target knowledge points based on the type of each target data using a corresponding anomaly detection model, wherein the target knowledge points include anomaly types related to semiconductor manufacturing contained in the corresponding target data; and the graph generation unit is configured to generate a knowledge graph in the field of semiconductor manufacturing based on the target knowledge points.

[0037] In one possible implementation, the device further includes:

[0038] The deduplication unit is configured to perform deduplication operations on the knowledge points generated from each target data, removing duplicate knowledge points.

[0039] This invention proposes an end-to-end knowledge graph generation method. Using an end-to-end approach, it processes various data in the semiconductor manufacturing field to generate a knowledge graph. With knowledge points as the core, it does not generate event description text from the data, but directly generates knowledge points from the data, and then generates a knowledge graph based on the knowledge points, thereby simplifying the process of generating a knowledge graph in the entire semiconductor manufacturing process. Attached Figure Description

[0040] To more clearly illustrate the technical solutions of the various embodiments disclosed in this specification, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only a few embodiments disclosed in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0041] Figure 1 This is a flowchart of an end-to-end knowledge graph generation method disclosed in an embodiment of the present invention;

[0042] Figure 2 This is a flowchart of a method for generating knowledge points based on SPC data, as disclosed in an embodiment of the present invention.

[0043] Figure 3 This is a flowchart of a method for generating knowledge points based on wafer image data, as disclosed in an embodiment of the present invention.

[0044] Figure 4 This is a flowchart of a method for generating knowledge points based on univariate time series data, as disclosed in an embodiment of the present invention.

[0045] Figure 5 This is a schematic block diagram of an end-to-end knowledge graph generation device disclosed in an embodiment of the present invention. Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0047] Before providing a detailed description of the embodiments of this application, the terms and symbols used in the embodiments of this application will be explained first.

[0048] SPC: Statistical Process Control, is a process control tool that uses mathematical statistics. It can analyze and pre-control production processes, keeping them in a controlled state where they are only affected by random factors, in order to control quality, improve quality, and achieve continuous improvement.

[0049] A wafer, also known as a silicon wafer, is a silicon chip used to fabricate silicon semiconductor circuits. Its raw material is silicon. High-purity polycrystalline silicon is dissolved, doped with silicon seed crystals, and then slowly pulled out to form a cylindrical single-crystal silicon ingot. After grinding, polishing, and slicing, the silicon ingot is formed into a silicon wafer, or crystal.

[0050] FDC: Fault Detection and Classification, is a systematic approach that replaces manual monitoring of machine processes to prevent large-scale scrapping and performs differential analysis when anomalies occur, enabling users to quickly troubleshoot the anomalies.

[0051] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0052] Figure 1 This is a flowchart of an end-to-end knowledge graph generation method disclosed in an embodiment of the present invention. Figure 1 As shown, the method includes at least the following steps: Step 101, acquiring at least one type of non-textual target data produced in semiconductor manufacturing, wherein each type of target data includes a numerical sequence or a numerical matrix; Step 102, determining target knowledge points using a corresponding anomaly detection model based on the type of each type of target data, wherein the target knowledge points include anomaly types related to semiconductor manufacturing contained in the corresponding target data; Step 104, generating a knowledge graph in the field of semiconductor manufacturing based on the target knowledge points.

[0053] In step 101, at least one type of target data in non-textual form generated during semiconductor manufacturing is acquired, including numerical sequences or numerical matrices.

[0054] Various types of data are generated during semiconductor manufacturing. These data can be broadly categorized into two types based on their source: data from various equipment, including equipment operating status, parameters, manuals, and logs; and test data from the manufactured chips, including their electrical characteristics and defects. These data can be further classified by data type: structured data, semi-structured data, and unstructured data. Structured data refers to chip data and equipment operating status data; semi-structured data refers to log data and work order data; and unstructured data refers to equipment documents. The method described in this invention primarily involves directly generating knowledge points from the non-textual data produced during semiconductor manufacturing.

[0055] In step 101, the at least one type of target data includes: statistical process control (SPC) data, wafer map data, and fault detection and classification (FDC) data.

[0056] In step 102, based on the type of each target data item, the corresponding anomaly detection model is used to determine target knowledge points. These target knowledge points include anomaly types related to semiconductor manufacturing contained within the corresponding target data. In some possible implementations, the target knowledge points are in the form of triples, where the anomaly type is included as an element.

[0057] Specifically, based on the type of each target data, the target data is encoded using the corresponding encoding method to obtain encoded data; the encoded data is input into the corresponding anomaly detection model to output the target knowledge points.

[0058] In one embodiment, the target data is encoded using a corresponding encoding method according to its type to obtain encoded data. Then, a teacher-student model is used to detect the encoded data and determine the anomaly type. After encoding the data, the encoded data is input into the teacher-student model for detection to determine the anomaly type. The input to the teacher model is a complete input without masking, equivalent to a reference answer. Features output by the teacher model in the masked region are used as soft labels to train the student model. The structure of the teacher model and the student model are exactly the same, only the parameters differ.

[0059] In another embodiment, the at least one target data includes Statistical Process Control (SPC) data, which forms a numerical sequence. The SPC data is a wafer manufacturing-related numerical sequence, which may include numerical sequences related to wafer manufacturing equipment. For example, data on the status of the wafer manufacturing equipment (also referred to as a machine) obtained by sensors during production, including but not limited to: temperature, humidity, pressure, voltage, current, etc.; and the utilization rate of the wafer manufacturing equipment; and / or wafer-related data, such as data obtained through defect detection during the wafer production process (e.g., wafer defect data); data obtained through electrical testing during the wafer production process (e.g., wafer failure type data, including CPU zone failure, GPU zone failure, memory zone failure, etc.); and purely numerical data such as wafer yield data.

[0060] In this embodiment, step 102 is implemented as follows: Figure 2 As shown.

[0061] In step 201, feature extraction and position embedding are performed on the sequence segments of the numerical sequence to obtain a feature vector sequence. The feature extraction can be implemented using a convolutional neural network (CNN) or a recurrent neural network (RNN).

[0062] In step 202, the feature vector sequence is input into the Transformer model to obtain the anomaly type and the relative position of the anomaly.

[0063] In step 203, the start position and end position of the anomaly are determined based on the position of the sequence fragment in the numerical sequence.

[0064] In step 204, the triplet consisting of the exception type, the exception start position, and the exception end position is taken as the corresponding target knowledge point. For example, the triplet can be in the form of (exception type, exception start position, exception end position).

[0065] In yet another embodiment, the at least one target data includes wafer pattern data.

[0066] In this embodiment, step 102 is implemented as follows: Figure 3 As shown. The method includes at least:

[0067] In step 301, the wafer image is segmented according to the pixel matrix corresponding to the wafer image to obtain several defect sub-images containing individual wafer defects. This image segmentation can be implemented using the U-Net image segmentation algorithm.

[0068] In step 305, the target defect sub-image in the plurality of defect sub-images is input into the defect classification model to obtain the corresponding defect type. The defect classification model can be a Detection Transformer model.

[0069] In step 306, the defect type is treated as an anomaly type, and the corresponding target knowledge point is determined by combining it with the target defect subgraph. The knowledge point can be constructed as a triple, specifically, the triple is in the form of (defect type, [x1, x2, y1, y2], defect area). Here, (x1, y1) are the coordinates of the upper left corner of the defect range output by the defect classification model, and (x2, y2) are the coordinates of the lower right corner of the defect range output by the defect classification model.

[0070] In a more specific embodiment, prior to step 305, the method further includes:

[0071] Step 302: Extract features from any defective sub-image to obtain sub-image features. Specifically, a convolutional neural network (CNN) or ResNet model can be used to extract features from the defective sub-image.

[0072] Step 303: Determine whether the defect subgraph belongs to a known defect type by searching and comparing within a predetermined feature set; the predetermined feature set consists of graph features of known defect types.

[0073] Step 304: If the defect type is known, determine the defect subgraph as the target defect subgraph.

[0074] Steps 302 to 304 are used to pre-classify the defect subgraphs, classifying the defect types of the subgraphs into known defect types and unknown defect types. For defect subgraphs with known defect types, they are identified as target defect subgraphs and proceed to step 305. For defect subgraphs with unknown defect types, they are directly output for manual processing by engineers and no further steps are executed. After manual processing, the graph features in the predetermined feature set can be updated for defect subgraphs with unknown defect types.

[0075] In another embodiment, the at least one target data includes a sequence of measurement data from a target sensor. Specifically, the measurement data sequence is the raw data generated by sensors in the semiconductor manufacturing process, processed by an FDC system to obtain a data sequence. A semiconductor manufacturing machine is equipped with many sensors, each monitoring a parameter on the machine, such as temperature, humidity, voltage, current, pressure, etc. The value of any parameter output by the sensor, after being processed by the FDC system, will yield a set of data sequences.

[0076] In this embodiment, step 102 is implemented as follows: Figure 4 As shown.

[0077] In step 401, the target subsequence that is ranked first in the measurement data sequence is encoded using the encoder of a variational autoencoder (VAE) to obtain latent space variables.

[0078] In step 402, the time information of the target subsequence is extracted using a gated recurrent unit (GRU) model.

[0079] In step 403, the latent space variables and the time information are input into the decoder of the variational autoencoder (VAE) for decoding to obtain the prediction data.

[0080] In step 404, based on the predicted data and the measurement data of the target subsequence in the measurement data sequence, the anomaly type related to the target sensor is determined.

[0081] In step 405, target knowledge points are generated, including the anomaly type, the production equipment number to which the measurement data sequence is targeted, and the wafer number being processed by the production equipment. The target knowledge point can be in the form of a triple, specifically, the triple form is (anomaly type, equipment number, wafer number).

[0082] Then return Figure 1 In step 104, a knowledge graph for the semiconductor manufacturing field is generated based on the target knowledge points.

[0083] Specifically, the anomaly types in the target knowledge points are added to the knowledge graph as entity nodes. If no existing knowledge graph exists, a new knowledge graph is created based on the knowledge points in the knowledge point set. If an existing knowledge graph already exists, it is updated using the knowledge points in the knowledge point set. It can be understood that during the update of the existing knowledge graph, if a node-relationship edge combination corresponding to a certain knowledge point already exists in the knowledge graph, this knowledge point will not be added to the knowledge graph again to avoid unnecessary duplication.

[0084] To generate a knowledge graph from knowledge points, we first define two types of entities in the knowledge graph: event entities (Entity) and object entities (Event), as well as two types of relationship edges: the relationship edge between object entities and event entities (Entity-Event, En-Ev), and the relationship edge between object entities (Entity-Entity, En-En).

[0085] The process of building a knowledge graph involves retrieving an existing knowledge graph from the knowledge graph database, or creating a new one if none exists. For any given knowledge point, the process first retrieves its corresponding production process information from the dictionary sequence of the original production data. This information includes the production step (step), production equipment (tool_entity), production time (time), and the wafer ID (wafer_id). Then, based on this information, the graph searches for the corresponding SPC event node (SPCEvent node) (creating one if it doesn't exist). Next, based on the tool_entity, the graph searches for the corresponding equipment entity node (ToolEntity node) (creating one if it doesn't exist). An En-Ev relationship edge is then constructed between the SPCEvent node and the ToolEntity node. Finally, other production information, such as step, time, and wafer_id, is filled into the attribute information of the En-Ev relationship edge. This completes the process of adding a knowledge point to the knowledge graph.

[0086] In some possible implementations, prior to step 104, the method further includes: step 103, performing a deduplication operation on the knowledge points generated from each target data to remove duplicate knowledge points.

[0087] Figure 5 This is a schematic block diagram of an end-to-end knowledge graph generation device disclosed in an embodiment of the present invention. The device 500 includes at least: a data acquisition unit 501, configured to acquire at least one type of non-textual target data produced in semiconductor manufacturing, wherein each target data includes a numerical sequence or a numerical matrix; a knowledge point determination unit 502, configured to determine target knowledge points based on the type of each target data using a corresponding anomaly detection model, wherein the target knowledge points include anomaly types related to semiconductor manufacturing contained in the corresponding target data; and a graph generation unit 504, configured to generate a knowledge graph in the field of semiconductor manufacturing based on the target knowledge points.

[0088] In one possible implementation, the device further includes a deduplication unit 503, configured to perform a deduplication operation on the knowledge points generated from each target data, removing duplicate knowledge points therein.

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0090] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0091] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An end-to-end knowledge graph generation method, comprising: obtaining at least one target data in a non-text form produced in semiconductor manufacturing, each target data comprising a numerical sequence or a numerical matrix; the at least one target data comprises statistical process control (SPC) data, which forms a numerical sequence; determining a target knowledge point according to the type of each target data using a corresponding anomaly detection model, the target knowledge point comprising an anomaly type related to semiconductor manufacturing contained in the corresponding target data; wherein the determining of the target knowledge point according to the type of each target data using the corresponding anomaly detection model comprises: performing feature extraction and position embedding on a sequence segment of the numerical sequence to obtain a feature vector sequence; inputting the feature vector sequence into a Transformer model to obtain an anomaly type and an anomaly relative position; determining an anomaly start position and an anomaly end position according to the position of the sequence segment in the numerical sequence; and taking a triple of the anomaly type, the anomaly start position and the anomaly end position as the corresponding target knowledge point; generating a knowledge graph in the field of semiconductor manufacturing according to the target knowledge point.

2. The method of claim 1, wherein, The at least one target data comprises statistical process control (SPC) data, wafer map data and fault detection and classification (FDC) data.

3. The method of claim 2, wherein, The determining of the target knowledge point according to the type of each target data using the corresponding anomaly detection model comprises: encoding the target data using a corresponding encoding mode according to the type of each target data to obtain encoded data; inputting the encoded data into the corresponding anomaly detection model to output the target knowledge point.

4. The method of claim 1, wherein, The target knowledge point is in the form of a triple, and the triple comprises the anomaly type as an element.

5. The method of claim 1, wherein, The at least one target data comprises wafer map data; The determining of the target knowledge point according to the type of each target data using the corresponding anomaly detection model comprises: segmenting the wafer map according to a pixel matrix corresponding to the wafer map to obtain a plurality of defect sub-maps containing single wafer defects; inputting a target defect sub-map in the plurality of defect sub-maps into a defect classification model to obtain a corresponding defect type; combining the defect type as an anomaly type with the target defect sub-map to determine a corresponding target knowledge point.

6. The method of claim 5, wherein, Before inputting the target defect sub-map in the plurality of defect sub-maps into the defect classification model, the method further comprises: performing feature extraction on any defect sub-map to obtain a sub-map feature; determining whether the defect sub-map belongs to a known defect type by searching and comparing in a predetermined feature set composed of map features of known defect types; and in the case of belonging to a known defect type, determining the defect sub-map as the target defect sub-map.

7. The method of claim 1, wherein, The at least one target data comprises a measurement data sequence of a target sensor; and the determining of the target knowledge point according to the type of each target data using the corresponding anomaly detection model comprises: The target sub-sequence in the front of the measurement data sequence is encoded using an encoder of a variational autoencoder VAE to obtain a latent space variable; time information of the target sub-sequence is extracted using a gated recurrent unit GRU model; and the latent space variable and the time information are input into a decoder of the variational autoencoder VAE for decoding to obtain predicted data; According to the predicted data and measurement data succeeding the target sub-sequence in the measurement data sequence, an abnormal type related to the target sensor is determined; A target knowledge point is generated, which includes the abnormal type, a production equipment number to which the measurement data sequence is directed, and a wafer number being processed by the production equipment.

8. The method of claim 1, wherein, Before the knowledge graph is generated according to the target knowledge point, the method further includes: A de-duplication operation is performed on the knowledge points generated by each target data to remove duplicate knowledge points.

9. The method of claim 1, wherein, According to the target knowledge point, a knowledge graph in the field of semiconductor manufacturing is generated, including: The abnormal type in the target knowledge point is taken as an entity node and added to the knowledge graph.

Images