Intelligent silicon-based OLED yield prediction method and device, and storage medium

Abstract

The present disclosure relates to the technical field of data processing, including an intelligent silicon-based OLED yield prediction method, device and storage medium. Production data of each intelligent silicon-based OLED is obtained; first data types and type values corresponding to the first data types in the production data are deleted, and target data obtained includes multiple second data types and type values corresponding to the second data types; a first data format of the target data is transposed into a second data format, and data in a row direction of transposed data obtained includes type values of multiple second data types of the same intelligent silicon-based OLED; an unqualified product prediction model is trained based on the transposed data; Shapley values of each feature type in the multiple second data types are determined based on the transposed data and the unqualified product prediction model; and multi-dimensional explanation information is generated based on the Shapley values of each feature type. The training efficiency of the unqualified product prediction model and the explainability of yield prediction can be improved.

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to a smart silicon-based OLED yield prediction method, device and storage medium. Background Technology

[0002] Organic Light-Emitting Diodes (OLEDs) on silicon are a new type of display device that combines semiconductor technology and OLED display technology. Intelligent silicon OLEDs are fabricated using monocrystalline silicon as the active driving backplane. To improve the product yield and production efficiency of intelligent silicon OLEDs, yield management is necessary. Yield management for intelligent silicon OLEDs refers to a series of measures taken during the production process to improve the yield and production efficiency, reduce scrap and rework, and ultimately improve product quality.

[0003] Traditional semiconductor yield management typically relies on manual experience to adjust production processes and parameters. Simultaneously, data-driven methods are used to analyze and predict yield issues during production. These data-driven methods collect and analyze large amounts of production data to identify the root causes of defects, predict future yield trends, and implement corresponding improvement measures.

[0004] However, the production data of smart silicon-based OLEDs collected through traditional data-driven methods are complex, leading to low efficiency in the analysis and prediction of yield problems. Summary of the Invention

[0005] In view of this, this disclosure proposes a smart silicon-based OLED yield prediction method, device, and storage medium, which can solve the problem of low efficiency in traditional data-driven methods for yield prediction, improve the training efficiency of defect prediction models, and thus improve yield prediction efficiency. Simultaneously, it can generate multi-dimensional explanatory information based on the Shapley value of each feature type, thereby intuitively reflecting the root causes of defect types and improving the interpretability of yield prediction.

[0006] According to one aspect of this disclosure, a method for predicting the yield of smart silicon-based OLEDs is provided, the method comprising:

[0007] Acquire production data for each smart silicon-based OLED; wherein, the production data includes multiple data types and a type value corresponding to each data type;

[0008] Delete the first data type and the corresponding type value of the first data type from the production data to obtain the target data, wherein the target data includes multiple second data types and the corresponding type values ​​of the second data types;

[0009] The first data format of the target data is transposed to the second data format to obtain transposed data; the transposed data is N×M structured data, and the data in the row direction of the transposed data includes the type values ​​of multiple second data types of the same smart silicon-based OLED; N and M are both integers greater than 1;

[0010] A poor prediction model is trained based on the first part of the transposed data;

[0011] Based on the second part of the transposed data and the adverse prediction model, determine the Shapley value of each feature type in the multiple second data types;

[0012] Multidimensional explanatory information is generated based on the Shapley value for each feature type. This multidimensional explanatory information is used to describe the impact of each feature type on yield prediction from different dimensions.

[0013] In one possible implementation, the target data is structured data represented by multiple first data tables, where each first data table includes, along with its row-wise data, a type identifier for a second data type of the same smart silicon-based OLED and a type value corresponding to that second data type; accordingly,

[0014] The step of transposing the first data format of the target data to the second data format to obtain transposed data includes:

[0015] Create a second data table corresponding to the second data format. The first column of the second data table represents the product identifier of each smart silicon-based OLED, and the first row represents the type identifier of each second data type. The type identifiers are arranged in a preset order.

[0016] For each product identifier and type value in the same row of each first data table, the type value is converted to the position indicated by the product identifier and the type identifier in the second data table to obtain the transposed data.

[0017] In one possible implementation, before training the poor prediction model based on the first portion of the transposed data, the method further includes:

[0018] The transposed data is filtered at least once to obtain filtered transposed data;

[0019] The first part of the data and the second part of the data are determined based on the filtered transposed data.

[0020] In one possible implementation, the step of performing at least one level of filtering on the transposed data to obtain filtered transposed data includes:

[0021] Determine the variance of the type values ​​corresponding to each second data type;

[0022] Based on the comparison between the variance and the preset variance threshold, the transposed data is subjected to first-level filtering.

[0023] For any two second data types after primary filtering, the correlation between the two second data types is determined based on the type values ​​corresponding to the two second data types.

[0024] Based on the comparison result between the correlation and the preset correlation threshold, the transposed data is subjected to secondary filtering to obtain the filtered transposed data.

[0025] In one possible implementation, the multiple second data types include the feature type and the output type, wherein the feature type is the second data type input to the adverse prediction model, and the output type is the second data type corresponding to the prediction result of the adverse prediction model; the multi-dimensional explanatory information includes a Shapley overall ranking graph, which is used to indicate the feature type causing the adverse result and the output type that the adverse prediction model can predict; accordingly,

[0026] The generation of multi-dimensional explanatory information based on the Shapley value for each feature type includes:

[0027] Determine the average absolute value of the Shapley value for each feature type;

[0028] For each output type, determine the proportion of that output type in the average absolute value of all feature types;

[0029] The Shapley overall sorting graph displays each feature type via the first coordinate axis, the average absolute value of each feature type via the second coordinate axis, and the percentage of each output type via a color legend; wherein the first coordinate axis is perpendicular to the second coordinate axis.

[0030] In one possible implementation, the multiple second data types include the feature type and the output type, wherein the feature type is the second data type input to the adverse prediction model, and the output type is the second data type corresponding to the prediction result of the adverse prediction model; the multi-dimensional interpretation information includes a Shapley heatmap, which is used to indicate the ranking of the Shapley values ​​of each feature type predicting the output type for each output type to be analyzed by the adverse prediction model, and to indicate whether each feature type has a positive or negative impact on the output type; accordingly,

[0031] The generation of multi-dimensional explanatory information based on the Shapley value for each feature type includes:

[0032] For each output type to be analyzed, determine the first Shapley absolute value when predicting the output type based on each feature type;

[0033] The second part of the data in each group is clustered according to explanatory similarity to obtain multiple sample sets; where explanatory similarity refers to multiple groups of data with the same output type and the same feature type.

[0034] Determine the second Shapley absolute value when predicting the output type based on each sample set;

[0035] Based on the aforementioned adverse prediction model, the prediction results corresponding to each sample set are determined;

[0036] The Shapley heatmap displays each feature type on a first axis, the first Shapley absolute value corresponding to each feature type on a second axis, and each sample set corresponding to each feature type on a third axis. A color representation of the second Shapley absolute value is displayed between the first and third axes, indicating the magnitude of the second Shapley absolute value. The first and second axes are perpendicular to the third axis.

[0037] The Shapley heatmap displays the change curves of the prediction results for each sample set at designated locations.

[0038] In one possible implementation, the multiple second data types include the feature type and the output type, wherein the feature type is the second data type input to the adverse prediction model, and the output type is the second data type corresponding to the prediction result of the adverse prediction model; the multi-dimensional explanatory information includes a Shapley decision graph, which is used to indicate the change process of the prediction result after gradually adding each feature type for each output type that the adverse prediction model can predict; accordingly,

[0039] The generation of multi-dimensional explanatory information based on the Shapley value for each feature type includes:

[0040] For each set of second part data, the type values ​​corresponding to each feature type in the second part data are added to the input data one by one according to the preset input order. The prediction results corresponding to various input data are predicted by the bad prediction model, and the change process of the prediction results corresponding to the second part data is obtained.

[0041] The change process of different prediction results that meet the preset conditions is clustered to obtain the change process of the clustered prediction results.

[0042] Obtain the mean of the prediction results corresponding to the second part of the data in each group;

[0043] The Shapley decision graph displays each feature type in the input order via the first coordinate axis and each prediction result via the second coordinate axis. Based on the change process of the prediction results after clustering, the prediction value change curve is extended along the first coordinate axis from the mean of the prediction results on the second coordinate axis; wherein, the first coordinate axis is perpendicular to the second coordinate axis.

[0044] In one possible implementation, the multiple second data types include the feature type and the output type, wherein the feature type is the second data type input to the poor prediction model, and the output type is the second data type corresponding to the prediction result of the poor prediction model; the multi-dimensional interpretation information includes a Shapley dependency graph, which indicates the influence of the type value of the feature type to be analyzed on the type value of another feature type for each output type that the poor prediction model can predict; accordingly,

[0045] The generation of multi-dimensional explanatory information based on the Shapley value for each feature type includes:

[0046] For the output type and feature type to be analyzed, the first coordinate axis of the Shapley dependency graph displays each Shapley value, and the second coordinate axis displays the different type values ​​corresponding to the feature type. Color points are displayed between the first and second coordinate axes. The position of the color points is determined by the Shapley value corresponding to the type value of the feature type to be analyzed when predicting the output type to be analyzed. The color of the color point indicates the magnitude of the type value of the other feature type.

[0047] According to another aspect of this disclosure, a smart silicon-based OLED yield prediction device is provided, comprising: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the above method when executing the instructions stored in the memory.

[0048] According to another aspect of this disclosure, a non-volatile computer-readable storage medium is provided that stores computer program instructions thereon, wherein the computer program instructions, when executed by a processor, implement the above-described method.

[0049] According to another aspect of this disclosure, a computer program product is provided, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is run in a processor of an electronic device, the processor in the electronic device performs the above-described method.

[0050] By acquiring production data for each smart silicon-based OLED; deleting the first data type and its corresponding type value from the production data to obtain target data, which includes multiple second data types and their corresponding type values; transposing the first data format of the target data to the second data format to obtain transposed data; the transposed data is N×M structured data, and the row data in the transposed data includes the type values ​​of multiple second data types for the same smart silicon-based OLED; training a defect prediction model based on the first part of the transposed data; determining the Shapley value of each feature type in the multiple second data types based on the second part of the transposed data and the defect prediction model; generating multi-dimensional explanatory information based on the Shapley value of each feature type; since the first data format of the target data is transposed to the second data format, the type values ​​of the same row can be directly used for training the defect prediction model without having to call the target data in the first data format to find the second data type corresponding to each type value during training, thus solving the problem of low efficiency in traditional data-driven methods for predicting yield, improving the training efficiency of the defect prediction model, and thus improving the yield prediction efficiency.

[0051] In addition, by generating multi-dimensional explanatory information based on Shapley values ​​for each feature type, the root causes of defective types can be intuitively reflected, improving the interpretability of yield prediction.

[0052] Other features and aspects of this disclosure will become clear from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0053] The accompanying drawings, which are included in and form part of this specification, illustrate exemplary embodiments, features, and aspects of this disclosure together with the specification and serve to explain the principles of this disclosure.

[0054] Figure 1 A flowchart is shown for a smart silicon-based OLED yield prediction method according to an embodiment of the present disclosure;

[0055] Figure 2 A schematic diagram illustrating a data transposition process according to an embodiment of the present disclosure is shown;

[0056] Figure 3 A schematic diagram showing an overall Shapley arrangement diagram according to an embodiment of the present disclosure;

[0057] Figure 4 A schematic diagram of a Shapley heat map according to an embodiment of the present disclosure is shown;

[0058] Figure 5 and Figure 6 A partial schematic diagram of a Shapley decision graph with an output type of Class 12 according to an embodiment of the present disclosure is shown.

[0059] Figure 7 A Shapley dependency graph is shown when the feature type to be analyzed is VTH_LIN_value_rangeb and the other feature type is MOB_SAT_value_std, according to an embodiment of the present disclosure.

[0060] Figure 8 A schematic diagram illustrating an output processing procedure according to an embodiment of the present disclosure is shown;

[0061] Figure 9 A block diagram of a smart silicon-based OLED yield prediction device according to an embodiment of the present disclosure is shown.

[0062] Figure 10 A block diagram of a smart silicon-based OLED yield prediction device according to another embodiment of the present disclosure is shown. Detailed Implementation

[0063] Various exemplary embodiments, features, and aspects of this disclosure will now be described in detail with reference to the accompanying drawings. The same reference numerals in the drawings denote elements that have the same or similar functions. Although various aspects of the embodiments are shown in the drawings, they are not necessarily drawn to scale unless specifically indicated otherwise.

[0064] The term “exemplary” as used herein means “serving as an example, embodiment, or illustration.” Any embodiment illustrated herein as “exemplary” is not necessarily to be construed as superior to or better than other embodiments.

[0065] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can be practiced without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0066] Figure 1A flowchart illustrating a smart silicon-based OLED yield prediction method according to an embodiment of this disclosure is provided. In this embodiment, the method is a smart silicon-based OLED yield prediction method based on the Shapley model, and the example given is its application in a computing-capable electronic device. This electronic device can be a user terminal or a server, such as a computer, tablet computer, or mobile phone. This embodiment does not limit the implementation method of the electronic device. Figure 1 As shown, the method includes:

[0067] Step 101: Obtain production data for each smart silicon-based OLED.

[0068] The production data includes multiple data types and the corresponding type values ​​for each data type.

[0069] Production data is used to indicate the production process and results of smart silicon-based OLEDs (or samples).

[0070] Indicatively, the data indicating the production process includes, but is not limited to: key process parameters, electrical parameters, and production parameters generated during the production process.

[0071] Critical process parameters refer to those process parameters whose values ​​change during production, potentially affecting the production results (such as yield) of smart silicon-based OLEDs. These process parameters need to be monitored and controlled to ensure that the produced smart silicon-based OLEDs meet predetermined quality standards. Critical process parameters include, but are not limited to, production temperature, evaporation rate, pressure, and / or production time. This embodiment does not limit the specific content of the critical process parameters.

[0072] Electrical parameter data refers to data obtained through electrical testing during semiconductor manufacturing. Electrical parameter data includes, but is not limited to: the gate voltage corresponding to the transistor when it first starts conducting current when the transistor is operating in the linear region, the gate voltage corresponding to the transistor when it first starts conducting current when the transistor is operating in the saturation region, the mobility corresponding to the transistor operating in the saturation region, current-voltage (IV) characteristics, etc. This embodiment does not limit the parameter content of the electrical parameter data.

[0073] Production parameter data is used to characterize parameter data generated in real time during the production process. Specifically, it can be data obtained after feature processing of parameter data generated in real time during production. Because real-time generated parameter data changes over time, the amount of data is large. Therefore, feature processing can preserve the characteristics of the parameter data without consuming significant storage and processing resources. For example, the parameter data could include parameters such as the dwell time and production temperature of a smart silicon-based OLED on a certain production equipment.

[0074] Optionally, the method of obtaining production parameter data by feature processing of parameter data includes, but is not limited to: obtaining the maximum value of parameter data, obtaining the minimum value of parameter data, or calculating the mean and / or variance of parameter data. Different types of parameter data can have different feature processing methods. This embodiment does not limit the implementation method of feature processing.

[0075] Data indicating production results includes, but is not limited to: defect types and the defect rate corresponding to each defect type.

[0076] During the production of smart silicon-based OLEDs, data is collected through various acquisition components to obtain key process parameters, electrical parameters, and production parameters. After production, the smart silicon-based OLED undergoes defect detection to obtain the production result. If the production result indicates that the smart silicon-based OLED is defective (or unqualified), the defect type is further investigated. By statistically analyzing the defect types of each smart silicon-based OLED, the corresponding defect rate for each type is obtained. This yields a set of production data for each smart silicon-based OLED, which includes multiple data types.

[0077] Schematic illustration: The production data is a tree-structured data structure. For example, the factory workshop is the root node of the tree structure, and multiple production devices in the factory workshop are the first-level child nodes of the root node. The data generated by each production device is the second-level child node of the corresponding first-level child node. In this case, production data of different data types can be queried between each other. For example, the factory workshop and production devices can be obtained through the data in the second-level child nodes. In other embodiments, the production data can be other data structures; this embodiment does not limit the data structure of the production data.

[0078] Optionally, in this embodiment, production data indicating the defects of the smart silicon-based OLED is used as an example for explanation. In other embodiments, both qualified and defective production data indicating the smart silicon-based OLED can be analyzed. This embodiment does not limit the type of production data.

[0079] Step 102: Delete the first data type and its corresponding type value from the production data to obtain the target data; the target data includes multiple second data types and their corresponding type values.

[0080] Among them, the prediction results of smart silicon-based OLED remain basically unchanged when the type value corresponding to the first data type changes, while the second data type is different from the first data type.

[0081] "Remains basically unchanged" means that the value remains unchanged or the change is less than the preset fluctuation range. Since the first data type has little impact on the prediction results, it is unnecessary to analyze its influence. The first data type and its corresponding value can be deleted, thus saving computational resources. For example, if the factory workshop has little impact on the prediction results, then the factory workshop is the first data type.

[0082] Deletion methods include, but are not limited to: obtaining the whitelisted data types; comparing the data types in the production data with the whitelisted data types; and deleting the first data type that matches the whitelisted data types and the type value corresponding to the first data type.

[0083] Alternatively, display production data for each data type; upon receiving a deletion instruction to delete the first data type, delete the first data type indicated by the deletion instruction and the type value corresponding to the first data type.

[0084] The deletion command can be generated through a human-computer interaction interface or sent by other devices. This embodiment does not limit the method of obtaining the deletion command.

[0085] Step 103: Transpose the first data format of the target data to the second data format to obtain transposed data; the transposed data is N×M structured data, and the data in the row direction of the transposed data includes the type values ​​of multiple second data types of the same smart silicon-based OLED; N and M are both integers greater than 1.

[0086] In one example, the target data is structured data represented by multiple first data tables. Each first data table includes, along with its row-wise data, a type identifier and a corresponding type value for a second data type specific to the same smart silicon-based OLED. The second data type and type value of the same smart silicon-based OLED stored in different first data tables are associated through the product identifier of that smart silicon-based OLED.

[0087] Optionally, the product identifier can be the primary key of the production data of the smart silicon-based OLED, or a product number, etc. This embodiment does not limit the implementation method of the product identifier.

[0088] The type identifier can be the type name of the second data type, or the type code, etc. This embodiment does not limit the implementation method of the type identifier.

[0089] Accordingly, the first data format of the target data is transposed to the second data format to obtain transposed data, including: creating a second data table corresponding to the second data format, wherein the first column of the second data table represents the product identifier of each smart silicon-based OLED, the first row represents the type identifier of each second data type, and the type identifiers are arranged in a preset order; for the type identifier and type value of the same row in each first data table corresponding to each product identifier, the type value is converted to the position indicated by the product identifier and type identifier in the second data table to obtain transposed data.

[0090] For example: Reference Figure 2 The data transposition process shown involves two first data tables, 201 and 202. The first row of both tables includes attributes corresponding to a single column of data. The second row of table 201 includes a type identifier (class0) for a second data type corresponding to the smart silicon-based OLED 213a, and a type value of 0.0017. The third row of table 201 includes a type identifier (class1) for another second data type corresponding to the smart silicon-based OLED 213a, and a type value of 0.0317. Correspondingly, the attributes of the first row of table 201 include attributes for class0 and class1 (defect type) and attributes for 0.0017 and 0.0317 (defect rate).

[0091] The second row of the first data table 202 includes a type identifier VFB1_max for one type of second data type corresponding to the smart silicon-based OLED 213a, and a type value of 0.13 corresponding to VFB1_max; the third row of the first data table 202 includes a type identifier VFB1_avg for another type of second data type corresponding to the smart silicon-based OLED 213a, and a type value of 0.18 corresponding to VFB1_avg. Correspondingly, the attributes of the first row of the first data table 202 include attributes for VFB1_max and VFB1_avg, which are parameters, and attributes for 0.13 and 0.18, which are parameter values.

[0092] In the second data table 203, the first column represents the product identifier 213a of each smart silicon-based OLED, and the first row represents the type identifiers VFB1_max, VFB1_avg, class0, and class1 of each second data type. For each product identifier 213a, the type identifier class0 and type value 0.0017 in the same row of each first data table 201 and 202 are converted to the position indicated by the product identifier 213a and type identifier class0 in the second data table 203. Similarly, the type identifier class1 and type value 0.0317 in the same row are converted to the position indicated by the product identifier 213a and type identifier class1 in the second data table 203. Other data in the first data table are transposed in the same way to obtain transposed data.

[0093] As shown in the second data table 203, a set of data for each smart silicon-based OLED is located in the same row. This row of data constitutes a multi-input multi-output data structure corresponding to the poor prediction model, including both feature type values ​​and output type values. In this case, the model can determine the corresponding second data type based on the position of each data point, without needing to search for the type value of each second data type in the first data table during training, thus improving model training efficiency. Here, the feature type refers to the second data type to be input into the poor prediction model, and the output type refers to the output type corresponding to the prediction result of the poor prediction model.

[0094] Step 104: Train a poor prediction model based on the first part of the transposed data.

[0095] Optionally, before training a poor prediction model based on the first part of the transposed data, the transposed data can be filtered, specifically including: performing at least one level of filtering on the transposed data to obtain filtered transposed data; and determining the first part of the data and the second part of the data described below based on the filtered transposed data.

[0096] In one example, the transposed data is filtered at least once to obtain filtered transposed data, including: determining the variance of the type values ​​corresponding to each second data type; performing first-level filtering on the transposed data based on the comparison result between the variance and a preset variance threshold; for any two second data types after first-level filtering, determining the correlation between the two second data types based on the type values ​​corresponding to the two second data types; and performing second-level filtering on the transposed data based on the comparison result between the correlation and a preset correlation threshold to obtain filtered transposed data.

[0097] Specifically, based on the comparison between the variance and a preset variance threshold, the transposed data undergoes a first-level filtering process, including: if the variance corresponding to the second data type is less than the variance threshold, then the second data type and its corresponding type value are deleted. The variance threshold can be 10. -10 Alternatively, other values ​​can be set according to the filtering requirements. This embodiment does not limit the value of the variance threshold.

[0098] Optionally, the correlation between two second data types can be determined by calculating the Pearson correlation coefficient between them. Correspondingly, based on the comparison between the correlation coefficient and a preset correlation threshold, a secondary filtering process is performed on the transposed data, including: if the correlation threshold corresponding to the second data type is greater than the correlation threshold, then the second data type and its corresponding type value are deleted. The correlation threshold can be 0.99, or it can be other values ​​set according to the filtering requirements. This embodiment does not limit the value of the correlation threshold.

[0099] In other embodiments, the correlation between two second data types can be determined in other ways, such as determining the correlation based on Kendall's rank correlation. This embodiment does not limit the way the correlation is determined.

[0100] In other embodiments, the transposed data may be filtered at only one level or at more levels. This embodiment does not limit the filtering method.

[0101] In this embodiment, by filtering the transposed data, the problems of information redundancy or insufficient information in the transposed data can be avoided, thereby realizing feature engineering dimensionality reduction of the transposed data and improving the model calculation efficiency.

[0102] Optionally, before filtering the data or training a poor prediction model, the transposed data can be preprocessed, including but not limited to: missing value handling, outlier handling, and / or standardization.

[0103] In this context, a missing value refers to a second data type of a certain smart silicon-based OLED that does not have a type value, but other smart silicon-based OLEDs with the same second data type do have a type value. In this case, the type value that the smart silicon-based OLED does not have is the missing value. Missing value handling methods include, but are not limited to: deleting a set of data with missing values; or filling the missing value with the average value of the second data type corresponding to the missing value, etc. This embodiment does not limit the method of handling missing values.

[0104] An outlier refers to a value of a second data type that exceeds the corresponding preset value range, or a change in the value that exceeds the corresponding preset change range. Outlier handling methods include, but are not limited to: deleting a set of data with outliers; or correcting the outlier to a normal value using a preset correction algorithm. This embodiment does not limit the methods of outlier handling.

[0105] Standardization refers to scaling type values ​​to a standard range, typically to a range with a mean of 0 and a standard deviation of 1. Standardization methods include, but are not limited to, normalization and logarithmic transformation. This embodiment does not limit the specific standardization method used.

[0106] After filtering and / or preprocessing the transposed data, the processed transposed data is divided according to a preset ratio to obtain a training set, a validation set, and a test set. The pre-created mathematical model is then trained using the training and validation sets to obtain a poor prediction model. The training set is the primary source for the model to learn data features and establish prediction rules, enabling the model to learn patterns and relationships within the data. The validation set is used to adjust model parameters and determine the model structure during training, evaluating the model's performance on unseen data and preventing overfitting. The test set is used to evaluate the performance of the poor prediction model.

[0107] In this embodiment, the training set and validation set constitute the first part of the data, and the test set constitutes the second part of the data hereinafter. In other embodiments, the first part of the data may only include the training set, and correspondingly, the second part of the data may include the validation set and the test set. This embodiment does not limit the method of determining the first part of the data and the second part of the data.

[0108] Optionally, the mathematical models used to train the poor prediction model include, but are not limited to: random forest models, elastic networks, partial least squares regression, extreme gradient boosting (XGBoost), and support vector regression (SVR). This embodiment does not limit the implementation method of the mathematical model. Since the random forest model has good prediction performance in the yield prediction of smart silicon-based OLEDs and has strong interpretability, this embodiment uses the poor prediction model trained based on the random forest model as an example for illustration.

[0109] In this embodiment, the various second data types include feature types used as input to the adverse prediction model and output types corresponding to the prediction results of the adverse prediction model. Optionally, the output types in the second data types are at least two, for example: the output types include multiple adverse types, and correspondingly, the prediction results include the adverse rate corresponding to each adverse type. In this case, it is necessary to ensure that the output of the adverse prediction model is a multi-dimensional output. In one example, the adverse prediction model is a multi-output model, in which case the output dimensions of the multi-output model correspond one-to-one with the output types.

[0110] In another example, if a single poor prediction model cannot achieve multi-dimensional output, the poor prediction model can be configured to include a single-output regression model for each output type, with different single-output regression models for different output types. A single-output regression model is a regression model that outputs only one type of output. This regression model indicates the relationship between the type value of the feature type and the type value of the output type, predicting the type value of the output type based on the type value of the feature type.

[0111] Since the correlation between different defect types is relatively small during the production of smart silicon-based OLEDs, different single-output regression models corresponding to different output types will not affect the accuracy of the defect prediction model. Optionally, the model types of the single-output regression models corresponding to different output types can be the same or different. For example, the single-output regression model corresponding to output type 1 is a random forest model, and the single-output regression model corresponding to output type 2 is an XGBoost model. In this case, the model types of the single-output regression models corresponding to different output types are different.

[0112] The process of training a maladaptive prediction model based on the first part of the data includes: inputting the type values ​​of the feature types in the first part of the data into a pre-created mathematical model to obtain the model result; comparing the model result with the type values ​​of the same set of output types and the model result; iteratively updating the model parameters of the mathematical model based on the comparison result to obtain the maladaptive prediction model.

[0113] Because during training, reading the same row of data in the first part of the data directly yields the type values ​​corresponding to multiple output types, which can then be compared sequentially with the corresponding model results, without needing to return to the first data table to query the type value corresponding to each output type, the model training efficiency can be improved.

[0114] Step 105: Based on the second part of the transposed data and the poor prediction model, determine the Shapley value for each feature type in various second data types.

[0115] The Shapley value for each feature type indicates the contribution of that feature type to the prediction result.

[0116] The second part of the data consists of multiple sets, each containing type values ​​for various second-type data corresponding to a specific smart silicon-based OLED. For example: [reference needed] Figure 2 The second set of data corresponding to the smart silicon-based OLED213a includes type values ​​of 0.13 and 0.18 for two feature types, and type values ​​of 0.0017 and 0.0317 for two output types.

[0117] Suppose that the data set P consists of p feature types corresponding to a certain smart silicon-based OLED in the second part of the data, i.e., the data dimension of the data set P is p. The defect prediction model v can obtain the prediction results (such as the defect rate corresponding to each defect type) based on any subset S of P, and v(φ) = 0, where φ is an empty set. For a subset S of P, v(S) represents the average prediction result obtained after the joint action of the features in set S, i.e., the prediction value of these features. The Shapley value of the i-th feature type is... The following formula is used to calculate:

[0118]

[0119] Where |S| represents the number of elements in subset S; v(S∪{i})-v(S) represents the contribution of the type value of the i-th feature type to the prediction result; P\{i} represents the set formed after removing the type value corresponding to feature type i from set P; S belongs to P\{i}; and SU{i} represents the union of set S and the type value corresponding to feature type i. The larger the absolute value of the Shapley for the i-th feature type, the higher the importance (or contribution) of feature type i to the prediction. A positive Shapley value indicates that feature type i will have a positive impact on the prediction result, making the prediction result higher; a negative Shapley value indicates that feature type i will have a negative impact on the prediction result, making the prediction result lower. By comparing the absolute values ​​and positive / negative values ​​of Shapley for different feature types i, the importance of each feature type in the poor prediction model and its impact on the prediction result can be explained.

[0120] Step 106: Generate multi-dimensional explanatory information based on the Shapley value for each feature type. This multi-dimensional explanatory information is used to describe the impact of each feature type on yield prediction from different dimensions.

[0121] Optionally, the multi-dimensional explanatory information can be textual information and / or graphical information. This embodiment does not limit the implementation method of the multi-dimensional explanatory information.

[0122] To improve the intuitiveness of obtaining multi-dimensional explanatory information, this embodiment uses the generation of multi-dimensional explanatory information in the form of charts as an example for illustration.

[0123] In one example, the multidimensional explanatory information includes a Shapley overall ordination diagram, which indicates the types of features that cause adverse outcomes and the types of outputs that the adverse prediction model can predict.

[0124] Accordingly, multi-dimensional explanatory information is generated based on the Shapley value for each feature type, including: determining the average absolute value of the Shapley value for each feature type; for each output type, determining the proportion of that output type in the average absolute value of each feature type; displaying each feature type on the first coordinate axis of the overall Shapley ranking graph, displaying the average absolute value corresponding to each feature type on the second coordinate axis, and displaying the proportion of each output type on the third coordinate axis; wherein the first and third coordinate axes are perpendicular to the second coordinate axis.

[0125] Since each feature type may have different type values ​​in different groups of the second part of the data, and the Shapley values ​​corresponding to different type values ​​may be different, the absolute values ​​of each Shapley value corresponding to the same feature type are taken and averaged to obtain the average absolute value.

[0126] For each feature type, each type value corresponds to a prediction result, which includes type values ​​(i.e., predicted values) for multiple output types. Correspondingly, for each output type, the type values ​​of multiple feature types may be predicted by a poor prediction model to obtain the type value for that output type. Therefore, by statistically analyzing the Shapley values ​​corresponding to the type values ​​of each feature type that can yield predicted values ​​for that output type, and then averaging the absolute values ​​of these Shapley values, the quotient of this average value and the average absolute value is determined to obtain the proportion of that output type in the average absolute value of all feature types.

[0127] refer to Figure 3 The Shapley overall ranking chart shown uses the first axis as the left vertical axis and the second axis as the horizontal axis, with color legends located on the right side of the axes. The first axis displays all feature types, specifically indicated by their type identifiers. The average absolute value of each feature type is represented by a bar chart, with the bar height indicated by the corresponding value on the second axis. The feature types displayed on the first axis are arranged in descending order of the bar chart's height; a higher bar indicates a greater impact of that feature type on the prediction result. The proportion of each output type within the average absolute value of all feature types is represented by a rectangle, with the proportions of each output type arranged in descending order, and different colored rectangles indicating the proportion: blue represents a large proportion, and purple represents a small proportion. For example: Figure 3In the diagram, the Shapley values ​​for each feature type primarily influence the output variables in the blue section (e.g., defect type Class 12), while the Shapley values ​​for the output variables in the purple section corresponding to each feature type are very small. The performance of the defect prediction model is mainly determined by the output variables with a relatively high proportion, such as... Figure 3 In the diagram, the blue output types (Class 12 and Class 0) account for a large proportion. The performance of the poor prediction model is mainly determined by the prediction results corresponding to the output types Class 12 and Class 0. The Shapley mean absolute value of the feature type VFB1_value_max is the largest, indicating that it has the greatest impact on the prediction results of the poor prediction model.

[0128] In other embodiments, the first coordinate axis may also be the horizontal axis, and correspondingly, the second coordinate axis may be the vertical axis; the average absolute value of each feature type may also be represented by a line chart or other methods instead of a bar chart; the proportion of each output type may also be represented by a color legend of different colors. This embodiment does not limit the display method of the Shapley overall sort chart.

[0129] In another example, the multidimensional explanatory information includes a Shapley heatmap, which is used for each output type to be analyzed for the poor prediction model, indicating the ranking of the Shapley values ​​of the various feature types that predict that output type, and indicating whether each feature type has a positive or negative impact on the output type.

[0130] Optionally, the output type to be analyzed can be any output type, or it can be the output type with the highest proportion determined by the Shapley overall ranking chart, where n is a positive integer. This embodiment does not limit the method of determining the output type to be analyzed.

[0131] Accordingly, multi-dimensional explanatory information is generated based on the Shapley value for each feature type, including: for each output type to be analyzed, determining the first Shapley absolute value when predicting the output type based on each feature type; clustering the second part of the data in each group according to explanatory similarity to obtain multiple sample sets; determining the second Shapley absolute value when predicting the output type based on each sample set; determining the prediction result corresponding to each sample set based on the poor prediction model; displaying each feature type on the first axis of the Shapley heatmap, displaying the first Shapley absolute value corresponding to each feature type on the second axis, and displaying each sample set corresponding to each feature type on the third axis, with a color representation of the second Shapley absolute value displayed between the first and third axes, the color representing the magnitude of the second Shapley absolute value; wherein the first and second axes are perpendicular to the third axis; and displaying the change curve of the prediction result corresponding to each sample set at a specified position on the Shapley heatmap.

[0132] Explanation similarity refers to multiple sets of data that have the same output type and the same feature type.

[0133] The method for determining the first Shapley absolute value includes: taking the average of the absolute values ​​of the Shapley values ​​corresponding to each feature data in each group of second-part data that can predict the output type to be analyzed, and obtaining the first Shapley absolute value corresponding to each feature type.

[0134] The second Shapley absolute value is determined by taking the average of the absolute values ​​of the Shapley values ​​for each feature type in the sample set that can predict the output type to be analyzed, and then obtaining the second Shapley absolute value for that sample set.

[0135] refer to Figure 4The Shapley heatmap shown is illustrated using the first axis as the left vertical axis, the second axis as the right vertical axis, and the third axis as the horizontal axis as an example. The first axis displays all feature types, specifically indicated by their type identifiers. The first Shapley absolute value for each feature type is represented by a bar chart on the second axis; the higher the bar, the larger the first Shapley absolute value. The third axis displays the sample sets corresponding to each feature type, specifically indicated by their set numbers, such as 0-400. The area between the first and third axes displays the second Shapley absolute value in color. Red indicates that the feature type value in that sample set improves the predicted value, while blue indicates that the feature type value in that sample set decreases the predicted value. The intensity of the color represents the magnitude of the second Shapley absolute value; darker colors indicate a larger second Shapley absolute value, and lighter colors indicate a smaller second Shapley absolute value.

[0136] The curves showing the changes in the prediction results are displayed above the first, second, and third coordinate axes, representing the curves showing the changes in the prediction results for each sample set. The higher the predicted value of the output type in the prediction results, the higher the corresponding function height. The gray horizontal line represents the mean of the prediction results for each sample set.

[0137] In other embodiments, the first and second coordinate axes may also be horizontal axes, and correspondingly, the third coordinate axis may be vertical axes; the first Shapley absolute value may also be represented not by a bar chart, but by a line chart or other methods; this embodiment does not limit the display method of the Shapley heatmap.

[0138] In yet another example, the multidimensional explanatory information includes a Shapley decision graph, which is used to indicate the change in the prediction results as each feature type is gradually added to indicate the type of output to be analyzed that the poor prediction model can predict.

[0139] Accordingly, multi-dimensional explanatory information is generated based on the Shapley value for each feature type, including: for each group of second-part data, the type values ​​corresponding to each feature type in the second-part data are added to the input data in a preset input order, and the prediction results corresponding to various input data are predicted by the bad prediction model, thus obtaining the change process of the prediction results corresponding to the second-part data; the change processes of different prediction results that meet preset conditions are clustered to obtain the change process of the clustered prediction results; the mean of the prediction results corresponding to each group of second-part data is obtained; each feature type is displayed in the input order through the first coordinate axis of the Shapley decision graph, and each prediction result is displayed through the second coordinate axis, and the change curve of the prediction value is extended along the direction of the first coordinate axis from the mean of the prediction results based on the change process of the clustered prediction results; wherein, the first coordinate axis is perpendicular to the second coordinate axis.

[0140] Among them, the preset conditions include, but are not limited to: the similarity of the change process between different prediction results is greater than or equal to the preset similarity threshold.

[0141] refer to Figure 5 and Figure 6 The diagram shown is a partial illustration of a Shapley decision graph with an output type of Class 12. Figure 5 This is the top of the Shapley decision diagram. Figure 6 This represents the bottom of the Shapley decision graph. Since the change in predicted values ​​in the middle section of the Shapley decision graph is very small—that is, the predicted value change curve is essentially vertical—this embodiment uses the top and bottom of the Shapley decision graph as examples for illustrative purposes. Figure 5 Taking the Shapley decision graph as an example, with the first axis as the left vertical axis and the second axis as the horizontal axis, the first axis displays all feature types, specifically indicated by their type identifiers. The second axis displays all prediction results, specifically the predicted value corresponding to the current output type, Class12. Starting from the mean of the prediction results, the curve of the predicted value change extends along the first axis on the second axis. The color of the curve corresponding to the predicted value change of each group of second-part data indicates the magnitude of the final predicted value when the input data includes all feature types; blue indicates a relatively low final predicted value, and red indicates a relatively high final predicted value.

[0142] In other embodiments, the first coordinate axis may also be the horizontal axis, and correspondingly, the second coordinate axis may be the vertical axis; the preset value change curve may also not have color, or the final predicted value may be represented by other colors. This embodiment does not limit the display method of the Shapley decision graph.

[0143] In another example, the multidimensional explanatory information includes a Shapley dependency graph, which indicates how the predicted result and Shapley value corresponding to the type value of each feature type are affected by the type value of another feature type.

[0144] Accordingly, multi-dimensional explanatory information is generated based on the Shapley value for each feature type, including: for the feature type to be analyzed, displaying each Shapley value through the first coordinate axis of the Shapley dependency graph, displaying different type values ​​corresponding to the feature type through the second coordinate axis, displaying color points between the first and second coordinate axes, the position of the color points being determined by the Shapley value corresponding to the type value of the feature type to be analyzed when predicting the output type to be analyzed, and the color of the color point indicating the magnitude of the type value of another feature type.

[0145] The feature type to be analyzed can be any feature type, or it can be the feature type whose average absolute value is among the top m determined by the Shapley overall ordinal chart, where m is a positive integer. This embodiment does not limit the method of determining the feature type to be analyzed.

[0146] refer to Figure 7 The diagram shows the Shapley dependency plots for two feature types: VTH_LIN_value_rangeb and MOB_SAT_value_std. VTH_LIN_value_rangeb represents the gate voltage at the moment the transistor begins to conduct current, operating in the linear region. MOB_SAT_value_std represents the standard deviation of the mobility when the transistor operates in the saturation region. Figure 7Taking a Shapley dependency graph as an example, with the first axis as the left vertical axis, the second axis as the horizontal axis, and the third axis as the right vertical axis, the first axis displays the various Shapley values; the second axis displays the different type values ​​of VTH_LIN_value_rangeb; and the third axis displays the type value of another feature type, MOB_SAT_value_std. Each colored point between the first and second axes represents the Shapley value corresponding to the type value of feature type VTH_LIN_value_rangeb in a set of data from the second part. The color in the color legend of the third axis represents the magnitude of the type value of the other feature type, MOB_SAT_value_std; the redder the color, the larger the type value of MOB_SAT_value_std. If a type value on the second axis corresponds to a single Shapley value, it means that this type value does not interact with the other feature type, MOB_SAT_value_std. However, if the type value of the second coordinate axis corresponds to multiple Shapley values, it means that the Shapley value is affected not only by the feature type VTH_LIN_value_rangeb, but also by another feature type MOB_SAT_value_std.

[0147] according to Figure 7 It can be seen that when the type value of VTH_LIN_value_rangeb is between 0.2 and 1, the Shapley value decreases as the type value increases. When the range is large, the change in the Shapley value has little impact, and the impact on poor prediction results is negative. When the type value of VTH_LIN_value_rangeb is near zero, its Shapley value is affected by another feature type, MOB_SAT_value_std. When the type value range of VTH_LIN_value_rangeb is 5 and the type value of MOB_SAT_value_std is large, the Shapley value is relatively low.

[0148] In summary, the intelligent silicon-based OLED yield prediction method provided in this embodiment obtains production data for each intelligent silicon-based OLED; deletes the first data type and its corresponding type value from the production data to obtain target data, which includes multiple second data types and their corresponding type values; transposes the first data format of the target data to a second data format to obtain transposed data; the transposed data is N×M structured data, and the row data in the transposed data includes the type values ​​of multiple second data types for the same intelligent silicon-based OLED; a defect prediction model is trained based on the first part of the transposed data; based on the second part of the transposed data and the defect prediction model, the Shapley value of each feature type in the multiple second data types is determined; multi-dimensional explanatory information is generated based on the Shapley value of each feature type; since the first data format of the target data is transposed to the second data format, the type values ​​of the same row can be directly used for training the defect prediction model without having to call the target data in the first data format to find the second data type corresponding to each type value during the training process. Therefore, it can solve the problem of low efficiency in traditional data-driven methods for yield prediction, improve the training efficiency of the defect prediction model, and thus improve the yield prediction efficiency.

[0149] In addition, by generating multi-dimensional explanatory information based on Shapley values ​​for each feature type, the root causes of defective types can be intuitively reflected, improving the interpretability of yield prediction.

[0150] To better understand the production data processing procedure in this application, an example is provided below. (Reference) Figure 8 The processing procedure shown involves obtaining production data through data collection 801 from various data sources. Figure 8 The production data includes production data with defective test results, and the production data with defective test results includes multiple sets. Each set of data includes: key process parameters, electrical parameter data, production parameter data, and defective test data. The defective test data includes, but is not limited to, defect type and defect rate.

[0151] After collecting the production data, data integration is performed (802). Specifically, data irrelevant to adverse prediction is deleted from the production data; that is, the first data type and its corresponding type value are deleted. The resulting second data type and its corresponding type value are then associated with the product identifier to obtain the target data.

[0152] After the target data is obtained through data integration, the first data format of the target data is transposed to the second data format, and the resulting structured transposed data is preprocessed (803). Specifically, the transposed data undergoes missing value processing, outlier processing, and standardization to obtain preprocessed data. The transposition method and the implementation methods of each preprocessing step are described in the above embodiment, and will not be repeated here.

[0153] After obtaining the preprocessed data, feature engineering is performed to reduce the dimensionality by 804. Specifically, a first-level screening (i.e., pre-screening) is performed based on the variance of the type values ​​corresponding to each second data type; after the first-level screening is completed, a second-level screening (i.e., fine screening) is performed based on the correlation between any two second data types to obtain the filtered transposed data.

[0154] Finally, yield prediction is performed based on the filtered transposed data. Specifically, a defect prediction model is trained using the first part of the filtered transposed data. This model is then used to predict the feature data in the second part of the data. Based on the prediction results and the second part of the data, Shapley values ​​for each feature type are determined, and multi-dimensional explanatory information is generated based on these Shapley values.

[0155] Figure 9 A block diagram of a smart silicon-based OLED yield prediction device according to an embodiment of the present disclosure is shown. Figure 9 As shown, the device includes: a data acquisition module 910, a data deletion module 920, a data transposition module 930, a model training module 940, a Shapley value calculation module 950, and a model interpretation module 960.

[0156] The data acquisition module 910 is used to acquire production data for each smart silicon-based OLED; wherein, the production data includes multiple data types and a type value corresponding to each data type;

[0157] The data deletion module 920 is used to delete the first data type and the type value corresponding to the first data type from the production data to obtain target data. The target data includes multiple second data types and the type value corresponding to the second data types.

[0158] The data transpose module 930 is used to transpose the first data format of the target data to a second data format to obtain transposed data; the transposed data is N×M structured data, and the data in the row direction of the transposed data includes the type values ​​of multiple second data types of the same smart silicon-based OLED; N and M are both integers greater than 1;

[0159] The model training module 940 is used to train a poor prediction model based on the first part of the transposed data.

[0160] Shapley value calculation module 950 is used to determine the Shapley value of each feature type in the multiple second data types based on the second part of the transposed data and the bad prediction model;

[0161] The model interpretation module 960 is used to generate multi-dimensional interpretation information based on the Shapley value of each feature type. The multi-dimensional interpretation information is used to describe the impact of each feature type on yield prediction from different dimensions.

[0162] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0163] This disclosure also proposes a computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the above-described method. The computer-readable storage medium can be volatile or non-volatile.

[0164] This disclosure also proposes an electronic device, including: a processor; and a memory for storing processor-executable instructions; wherein the processor is configured to implement the above method when executing the instructions stored in the memory.

[0165] This disclosure also provides a computer program product, including computer-readable code, or a non-volatile computer-readable storage medium carrying computer-readable code, wherein when the computer-readable code is run in a processor of an electronic device, the processor in the electronic device performs the above-described method.

[0166] Figure 10 This is a block diagram illustrating a smart silicon-based OLED yield prediction device 1900 according to an exemplary embodiment. For example, device 1900 can be provided as a terminal device. (Refer to...) Figure 10 The apparatus 1900 includes a processing component 1922, which further includes one or more processors, and memory resources represented by memory 1932 for storing instructions, such as application programs, that can be executed by the processing component 1922. The application programs stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Furthermore, the processing component 1922 is configured to execute instructions to perform the methods described above.

[0167] Device 1900 may also include a power supply component 1926 configured to perform power management of device 1900, a wired or wireless network interface 1950 configured to connect device 1900 to a network, and an input / output interface 1958 (I / O interface). Device 1900 can operate on an operating system, such as Windows Server, stored in memory 1932. TM macOS X TM Unix TM Linux TM FreeBSD TM Or similar.

[0168] In an exemplary embodiment, a non-volatile computer-readable storage medium is also provided, such as a memory 1932 including computer program instructions that can be executed by a processing component 1922 of the device 1900 to perform the above-described method.

[0169] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for predicting the yield of intelligent silicon-based OLEDs, characterized in that, The method includes: Acquire production data for each smart silicon-based OLED; wherein, the production data includes multiple data types and a type value corresponding to each data type; Delete the first data type and the corresponding type value of the first data type from the production data to obtain the target data, wherein the target data includes multiple second data types and the corresponding type values ​​of the second data types; The first data format of the target data is transposed to the second data format to obtain transposed data; the transposed data is N×M structured data, and the data in the row direction of the transposed data includes the type values ​​of multiple second data types of the same smart silicon-based OLED; N and M are both integers greater than 1; A poor prediction model is trained based on the first part of the transposed data; Based on the second part of the transposed data and the adverse prediction model, determine the Shapley value of each feature type in the multiple second data types; Multi-dimensional explanatory information is generated based on the Shapley value for each feature type. This multi-dimensional explanatory information is used to describe the impact of each feature type on yield prediction from different dimensions. The target data is structured data represented by multiple first data tables. Each first data table includes, along with its row-wise data, a type identifier for a second data type specific to the same smart silicon-based OLED and a corresponding type value for that second data type. Accordingly, The step of transposing the first data format of the target data to the second data format to obtain transposed data includes: Create a second data table corresponding to the second data format. The first column of the second data table represents the product identifier of each smart silicon-based OLED, and the first row represents the type identifier of each second data type. The type identifiers are arranged in a preset order. For each product identifier, the type identifier and type value of the same row in each first data table are converted to the position indicated by the product identifier and the type identifier in the second data table to obtain the transposed data, so that the data of the same row in the second table constitutes the multi-input multi-output data structure corresponding to the bad prediction model; The multiple second data types include the feature types and output types. The feature types are the second data types input to the adverse prediction model, and the output types are the second data types corresponding to the prediction results of the adverse prediction model. The multi-dimensional interpretation information includes a Shapley heatmap. The Shapley heatmap is used to indicate the ranking of the Shapley values ​​of each feature type predicting the output type for each output type to be analyzed by the adverse prediction model, and to indicate whether each feature type has a positive or negative impact on the output type. Accordingly, The generation of multi-dimensional explanatory information based on the Shapley value for each feature type includes: For each output type to be analyzed, determine the first Shapley absolute value when predicting the output type based on each feature type; The second part of the data in each group is clustered according to explanatory similarity to obtain multiple sample sets; where explanatory similarity refers to multiple groups of data with the same output type and the same feature type. Determine the second Shapley absolute value when predicting the output type based on each sample set; Based on the aforementioned adverse prediction model, the prediction results corresponding to each sample set are determined; The Shapley heatmap displays each feature type on a first axis, the first Shapley absolute value corresponding to each feature type on a second axis, and each sample set corresponding to each feature type on a third axis. A color representation of the second Shapley absolute value is displayed between the first and third axes, indicating the magnitude of the second Shapley absolute value. The first and second axes are perpendicular to the third axis. The Shapley heatmap displays the change curves of the prediction results for each sample set at designated locations.

2. The method according to claim 1, characterized in that, Before training the poor prediction model based on the first portion of the transposed data, the process also includes: The transposed data is filtered at least once to obtain filtered transposed data; The first part of the data and the second part of the data are determined based on the filtered transposed data.

3. The method according to claim 2, characterized in that, The step of performing at least one level of filtering on the transposed data to obtain filtered transposed data includes: Determine the variance of the type values ​​corresponding to each second data type; Based on the comparison between the variance and the preset variance threshold, the transposed data is subjected to first-level filtering. For any two second data types after primary filtering, the correlation between the two second data types is determined based on the type values ​​corresponding to the two second data types. Based on the comparison result between the correlation and the preset correlation threshold, the transposed data is subjected to secondary filtering to obtain the filtered transposed data.

4. The method according to any one of claims 1 to 3, characterized in that, The multiple second data types include the feature type and the output type, wherein the feature type is the second data type input to the adverse prediction model, and the output type is the second data type corresponding to the prediction result of the adverse prediction model; the multi-dimensional explanatory information includes a Shapley overall ranking graph, which is used to indicate the feature type causing the adverse result and the output type that the adverse prediction model can predict; accordingly, The generation of multi-dimensional explanatory information based on the Shapley value for each feature type includes: Determine the average absolute value of the Shapley value for each feature type; For each output type, determine the proportion of that output type in the average absolute value of all feature types; The Shapley overall sorting graph displays each feature type via the first coordinate axis, the average absolute value of each feature type via the second coordinate axis, and the percentage of each output type via a color legend; wherein the first coordinate axis is perpendicular to the second coordinate axis.

5. The method according to claim 1, characterized in that, The multiple second data types include the feature type and the output type. The feature type is the second data type input to the adverse prediction model, and the output type is the second data type corresponding to the prediction result of the adverse prediction model. The multi-dimensional explanatory information includes a Shapley decision graph, which is used to indicate the change process of the prediction result after gradually adding each feature type for each output type that the adverse prediction model can predict. Accordingly, The generation of multi-dimensional explanatory information based on the Shapley value for each feature type includes: For each set of second part data, the type values ​​corresponding to each feature type in the second part data are added to the input data one by one according to the preset input order. The prediction results corresponding to various input data are predicted by the bad prediction model, and the change process of the prediction results corresponding to the second part data is obtained. The change process of different prediction results that meet the preset conditions is clustered to obtain the change process of the clustered prediction results. Obtain the mean of the prediction results corresponding to the second part of the data in each group; The Shapley decision graph displays each feature type in the input order via the first coordinate axis and each prediction result via the second coordinate axis. Based on the change process of the prediction results after clustering, the prediction value change curve is extended along the first coordinate axis from the mean of the prediction results on the second coordinate axis; wherein, the first coordinate axis is perpendicular to the second coordinate axis.

6. The method according to claim 1, characterized in that, The multiple second data types include the feature type and the output type. The feature type is the second data type input to the poor prediction model, and the output type is the second data type corresponding to the prediction result of the poor prediction model. The multi-dimensional interpretation information includes a Shapley dependency graph, which indicates the influence of the type value of the feature type to be analyzed on the type value of another feature type for each output type that the poor prediction model can predict. The generation of multi-dimensional explanatory information based on the Shapley value for each feature type includes: For the output type and feature type to be analyzed, the first coordinate axis of the Shapley dependency graph displays each Shapley value, and the second coordinate axis displays the different type values ​​corresponding to the feature type. Color points are displayed between the first and second coordinate axes. The position of the color points is determined by the Shapley value corresponding to the type value of the feature type to be analyzed when predicting the output type to be analyzed. The color of the color point indicates the magnitude of the type value of the other feature type.

7. A smart silicon-based OLED yield prediction device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to implement the method of any one of claims 1 to 6 when executing instructions stored in the memory.

8. A non-volatile computer-readable storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images