Method and device for predicting production indexes of semiconductor production line

Abstract

The invention provides a method and a device for predicting production indexes of a semiconductor production line. The prediction method comprises the following steps: acquiring encrypted training samples of at least two participants; each training sample comprises a production parameter sample of a participant and a production index sample corresponding to the production parameter sample; training a pre-established initial network model by using the training sample and adopting a federal learning training method to obtain at least one production index prediction model; and the participant inputs the production parameters into the production index prediction model to obtain a prediction result of the production index corresponding to the production parameters. The data privacy of each participant can be protected, the universality is high, popularization and implementation are convenient, and the accuracy of the prediction result of the production index is high.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and in particular to a method for predicting production indicators of a semiconductor production line. Background Art

[0002] Integrated circuits play a very important role in all walks of life and have gradually become the cornerstone of the modern information society. Integrated circuits often include a large number of semiconductor devices of various types. The prediction of production indicators (such as process indicators, yield indicators, equipment performance indicators, etc.) of semiconductor production lines is of vital importance to improving production efficiency and saving resources. However, semiconductor production lines often involve many process flows and equipment, such as wafer processing, oxidation, photomasks, etching, thin film deposition, interconnection, testing, packaging and other processes. "How to solve the prediction problem of production indicators through data models" has long been one of the important research topics of technicians in this field. However, since the factors affecting production indicators are very diverse and complex, so far, there has been no record of relevant prediction schemes for production indicators of semiconductor production lines with high prediction accuracy and relatively general. The most common prediction schemes in the relevant technology have defects of one kind or another, which are listed as follows:

[0003] One of them is a more traditional prediction method, the core idea of ​​which is to use process data to directly predict yield indicators. Although this method is relatively simple and easy to implement, due to the limited data samples, the prediction results are usually not accurate enough.

[0004] Another is a yield prediction method based on machine learning, which uses a variety of production data during the semiconductor manufacturing process to establish multiple independent models, and then combines these models to predict the yield. Although this method can obtain more accurate prediction results, it requires all participants in the semiconductor production line to disclose data, which conflicts with the data privacy protection demands of each participant. Therefore, in practical applications, it is difficult to coordinate (it is impossible to guarantee the data privacy of each participant) the implementation of this solution by each participant.

[0005] Another method is to model the multiple links in the semiconductor production line separately, and then use a unified optimization model to learn and optimize the performance of the model in the production scheduling scenario. This method only involves the optimization of production scheduling, and does not involve the prediction and improvement of yield indicators, production rate indicators, etc., and is only applicable to the vertical optimization of a single production line and cannot be applied to the horizontal optimization across production lines. Therefore, this method has the following defects: This method not only has poor versatility, but also requires all participants to disclose data, which cannot meet the needs of each participant for data privacy protection.

[0006] It should be noted that the information disclosed in the background technology section of the invention is only intended to deepen the understanding of the general background technology of the invention, and should not be regarded as an admission or suggestion in any form that the information constitutes prior art already known to those skilled in the art. Summary of the invention

[0007] The purpose of the present invention is to provide a method and device for predicting production indicators of a semiconductor production line. The present invention can not only obtain more accurate prediction results of production indicators while protecting the data privacy of each participant; but also the present invention has strong versatility and can predict process indicators, yield indicators or equipment performance indicators.

[0008] To achieve the above object, the present invention is implemented by the following technical solution: A method for predicting production indicators of a semiconductor production line, comprising:

[0009] Obtaining encrypted training samples of at least two participants; each training sample includes a production parameter sample of the participant and a production indicator sample corresponding to the production parameter sample;

[0010] Using the training samples, a pre-built initial network model is trained using a federated learning training method to obtain at least one production indicator prediction model;

[0011] The production parameters of the participant are input into the production index prediction model to obtain the prediction results of the production index corresponding to the production parameters.

[0012] Optionally, all the participants belong to the same semiconductor production line, and the prediction method runs on the participants. The prediction method also includes: performing feature extraction on the original production parameter samples and the production indicator samples corresponding to the original production parameter samples, and using the extraction results as the training samples.

[0013] Optionally, the prediction method further includes: encrypting the training sample using a homomorphic encryption method to obtain an encrypted training sample.

[0014] Optionally, before encrypting the training samples, the prediction method further comprises: aligning the training samples according to feature values ​​of the training samples provided by all the participants;

[0015] The adopting homomorphic encryption method to encrypt the training samples includes: adopting homomorphic encryption method to encrypt the aligned training samples.

[0016] Optionally, each of the participants corresponds to a production indicator prediction model; each of the participants obtains the corresponding production indicator prediction model through training through the following steps:

[0017] Generate a homomorphically encrypted private key and public key, and use your own public key to exchange the public key with each other participant to obtain the public key of each of the other participants;

[0018] Using the encrypted training sample of the party, training the pre-built first initial network model until a first preset training end condition is met to obtain a first candidate prediction model; and using the first candidate prediction model to obtain a first intermediate prediction result corresponding to the encrypted training sample;

[0019] Encrypting the first intermediate prediction result using the public key of the party to obtain a second intermediate prediction result; and using the second intermediate prediction result to exchange with the second intermediate prediction result of each of the other participants to obtain third intermediate prediction results of all other participants;

[0020] Encrypting the first intermediate prediction result using the public key of each of the other participants to obtain a fourth intermediate prediction result;

[0021] The first candidate prediction model is trained by respectively using the third intermediate prediction result and the fourth intermediate prediction result encrypted by the public key provided by the same other participant until a second preset training end condition is met, thereby obtaining the first model correction amount of the other participant by the own party, and using the first model correction amount to exchange with the other participant to obtain the corresponding second model correction amount of the own party;

[0022] The private key is used to decrypt all the second model correction values, and the decrypted second model correction values ​​are used to correct the first candidate prediction model to obtain the production indicator prediction model.

[0023] Optionally, the using the first model correction amount to exchange with the other participant includes: encrypting the first model correction amount using random noise, and using the encrypted first model correction amount to exchange with the other participant;

[0024] Before using the private key to decrypt all the second model corrections, the prediction method further includes: filtering the first model corrections using a filtering method corresponding to the random noise to obtain the second model corrections.

[0025] Optionally, the using the decrypted second model correction amount to correct the first candidate prediction model to obtain the production index prediction model includes:

[0026] The accumulated result of all the correction values ​​of the second model is used as the correction value of the first candidate prediction model, and the first candidate prediction model is modified using the correction value to obtain the production index prediction model.

[0027] Optionally, based on all the second model corrections, the correction contribution of each of the other participants to the production indicator prediction model is calculated and recorded.

[0028] Optionally, inputting the production parameters of the participant into the production index prediction model to obtain the prediction results of the production indicators corresponding to the production parameters includes:

[0029] Disassemble each of the production indicator prediction models to obtain the model body and output layer corresponding to each participant;

[0030] According to the output layer corresponding to each of the participants, a joint output layer is obtained;

[0031] Fixing the model parameters of the model body, and continuing to use the federated learning training method to optimize each of the model bodies and the joint output layer, so as to obtain a joint prediction model according to all the optimized output layers;

[0032] The participating party inputs the production parameters into the joint prediction model to obtain the prediction results of the production indicators corresponding to the production parameters.

[0033] Optionally, all the participants are equipment belonging to the same process flow of different semiconductor production lines or process control systems belonging to the same process flow of different semiconductor production lines; all the participants correspond to one production index prediction model;

[0034] The prediction method is run on a central server, and the training sample is used to train a pre-built initial network model using a federated learning training method to obtain at least one production indicator prediction model, including:

[0035] Building a second initial network model, and sending the second initial network model to each of the participants respectively;

[0036] The following steps are iteratively performed until the production index prediction model is obtained:

[0037] Receiving model reference parameters obtained by each participant from training the second initial network model; and modifying the second initial network model according to the model reference parameters to obtain a second candidate prediction model;

[0038] Sending the second candidate prediction model to each of the participants respectively, and receiving evaluation results of the second candidate prediction model by each of the participants;

[0039] Based on the evaluation results of all the participants, determine whether the second candidate prediction model meets the third prediction training end condition. If so, notify each of the participants that the training is completed so that each of the participants uses the second candidate prediction model as the production indicator prediction model; if not, notify each of the participants to continue training so that each of the participants uses the second candidate prediction model to replace the second initial network model, and repeat the above steps.

[0040] Optionally, after receiving the model reference parameters obtained by each participant from training the second initial network model, the prediction method further includes:

[0041] Based on all the model reference parameters, calculate and record the revised contribution of each participant to the production indicator prediction model.

[0042] Optionally, all of the participants are equipment belonging to the same process flow of different semiconductor production lines or process control systems belonging to the same process flow of different semiconductor production lines; all of the participants correspond to one production index prediction model; the prediction method is run on the participants, and the training samples are used to train the pre-built initial network model using the federated learning training method to obtain at least one production index prediction model, including:

[0043] Receiving a second initial network model sent by the central server;

[0044] And iteratively perform the following steps until the production index prediction model is obtained:

[0045] Using the encrypted training samples of the party, training the second initial network model until a fourth preset training end condition is met, obtaining model reference parameters after training the second initial network model; and sending the model reference parameters to the central server;

[0046] receiving a second candidate prediction model obtained by the central processor according to the model reference parameters, and evaluating the second candidate prediction model to obtain an evaluation result;

[0047] Uploading the evaluation results to the central server;

[0048] Receive the training result sent by the central server. If the training is completed, use the received second candidate prediction model as the production indicator prediction model; if the training continues, use the second candidate prediction model to replace the second initial network model, and repeat the above steps.

[0049] Optionally, the model reference parameters include model parameters or gradient parameters after homomorphic encryption.

[0050] Optionally, all the participants belong to equipment or process control systems of similar process flows, and the training samples are used to train a pre-built initial network model using a federated learning training method to obtain at least one production indicator prediction model, including:

[0051] Using the homomorphically encrypted private training samples of all the participants, a third initial network model built in advance is trained using a federated learning training method to obtain a global production indicator prediction model;

[0052] For each of the participants, the following steps are performed to obtain the corresponding production indicator prediction model:

[0053] Using the private training sample of the participant to train the pre-built fourth initial network model to obtain a private production indicator prediction model; and disassembling each of the private production indicator prediction models to obtain a private model body and a private output layer;

[0054] Integrate the private model body with the global production index prediction model to obtain an optimized output layer;

[0055] According to the usage model body, the global production index prediction model and the optimized output layer, the production index prediction model corresponding to the participant is obtained.

[0056] In order to achieve the above object, the present invention further provides a prediction device for production indicators of a semiconductor production line, the prediction device comprising:

[0057] A training sample acquisition unit, configured to acquire encrypted training samples of at least two participants; wherein the participants include equipment or process control systems of the semiconductor production line, and all the participants belong to the same semiconductor production line, equipment of the same process flow of different semiconductor production lines, or process control systems of the same process flow of different semiconductor production lines; each training sample includes a production parameter sample of the participant and a production indicator sample corresponding to the production parameter sample;

[0058] A production index prediction model acquisition unit is configured to use the training samples to train a pre-built initial network model using a federated learning training method to obtain at least one production index prediction model;

[0059] The production index prediction result acquisition unit is configured to input the production parameter into the production index prediction model to obtain the prediction result of the production index corresponding to the production parameter.

[0060] Compared with the prior art, the method and device for predicting production indicators of a semiconductor production line provided by the present invention have the following advantages:

[0061] The method for predicting production indicators provided by the present invention first obtains encrypted training samples of at least two participants (such as lithography machines, measurement equipment, etc.), then uses the training samples to train the pre-built initial network model using the federated learning training method to obtain at least one production indicator prediction model; finally, the production parameters of the participants are input into the production indicator prediction model to obtain the prediction results of the production indicators corresponding to the production parameters. It can be seen that the method for predicting production indicators provided by the present invention does not impose any restrictions on the participants and production samples. All the participants can belong to the same semiconductor production line, and all the participants can also belong to the equipment of the same process flow of different semiconductor production lines or the process control system of the same process flow of different semiconductor production lines; and the production indicators (which can be yield indicators, process indicators, production rate indicators, etc.) are not limited, and the versatility is strong; further, the prediction method provided by the present invention uses the encrypted training samples of each participant, so it can meet the needs of data privacy protection of each participant and is more convenient for promotion and implementation. Furthermore, the production indicator prediction model is trained using the federated learning training method, which can achieve joint modeling and optimization among all participants without the participants sharing training samples, thereby improving the prediction accuracy of the production indicator prediction model.

[0062] Since the device for predicting production indicators of a semiconductor production line provided by the present invention and the method for predicting production indicators of a semiconductor production line provided by the present invention belong to the same inventive concept, the device for predicting production indicators of a semiconductor production line provided by the present invention at least has all the advantages of the method for predicting production indicators of a semiconductor production line provided by the present invention. For details, please refer to the above description of the method for predicting production indicators of a semiconductor production line, which will not be elaborated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 A schematic diagram of the overall process of the method for predicting production indicators of a semiconductor production line provided by the present invention;

[0064] Figure 2 This is a schematic diagram of a specific process of obtaining a corresponding production index prediction model through training by each participant in one implementation mode of the first embodiment of the present invention;

[0065] Figure 3 A schematic diagram of the disassembly and assembly process of a production index prediction model provided as a specific example of the implementation of the present invention;

[0066] Figure 4A schematic diagram of the packaging process of a production index prediction model provided as another specific example of the first embodiment of the present invention; Figure 5 This is a schematic diagram of a specific process of obtaining a production index prediction model by training a central server according to one implementation of the second embodiment of the present invention;

[0067] Figure 6 This is a schematic diagram of a specific process of obtaining a production index prediction model through training by a participant in another implementation manner of the second embodiment of the present invention;

[0068] Figure 7 A schematic diagram of a specific process of obtaining a production index prediction model according to one implementation method of the third embodiment of the present invention;

[0069] Figure 8 A schematic diagram of the disassembly and packaging process of a production index prediction model provided as a specific example of Embodiment 3 of the present invention;

[0070] Fig. 9 A schematic diagram of the structure of a device for predicting production indicators of a semiconductor production line provided in a fourth embodiment of the present invention;

[0071] The reference numerals are as follows:

[0072] Training sample acquisition unit-100, production index prediction model acquisition unit-200, production index prediction result acquisition unit-300. DETAILED DESCRIPTION

[0073] The following is a further detailed description of the method and device for predicting production indicators of a semiconductor production line proposed by the present invention in conjunction with the accompanying drawings. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the drawings are in a very simplified form and use non-precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. In order to make the purposes, features and advantages of the present invention more obvious and easy to understand, please refer to the accompanying drawings. It should be noted that the structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification for people familiar with this technology to understand and read, and are not used to limit the limiting conditions for the implementation of the present invention. Any modification of the structure, change of the proportional relationship or adjustment of the size, in the case of the same or similar effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical content disclosed by the present invention. The specific design features of the present invention disclosed herein include, for example, specific dimensions, directions, positions and shapes, which will be determined in part by the specific environment to be applied and used. And, in the embodiments described below, sometimes the same figure mark is used in common between different drawings to represent the same part or part with the same function, and its repeated description is omitted. In this specification, similar reference numerals and letters are used to refer to similar items, and therefore, once an item is defined in one figure, it need not be further discussed in subsequent figures. In addition, if the method described herein includes a series of steps, the order of these steps presented herein is not necessarily the only order in which these steps can be performed, and some of the steps described may be omitted and / or some other steps not described herein may be added to the method.

[0074] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. The singular forms "a", "an" and "the" include plural objects, the term "or" is generally used in a sense including "and / or", the term "several" is generally used in a sense including "at least one", and the term "at least two" is generally used in a sense including "two or more". In addition, the terms "first", "second" and "third" are used for descriptive purposes only and are not to be understood as indicating or suggesting relative importance or implicitly indicating the number of the indicated technical features.

[0075] The core idea of ​​the present invention is to provide a method and device for predicting production indicators of a semiconductor production line, so as to obtain more accurate prediction results of production indicators while protecting customer privacy; at the same time, it has strong versatility and applicability, and can predict process indicators, yield indicators or equipment performance indicators.

[0076] In order to realize the above idea, the method for predicting production indicators of a semiconductor production line provided by the present invention is implemented by the following scheme. Figure 1 , Figure 1 The overall flow chart of the method for predicting production indicators of a semiconductor production line provided by the present invention is as follows. Figure 1 It can be seen that the production index prediction method provided in this embodiment includes:

[0077] S100: Obtaining encrypted training samples from at least two participants; each training sample includes a production parameter sample of the participant and a production indicator sample corresponding to the production parameter sample;

[0078] S200: Using the training samples, a pre-built initial network model is trained using a federated learning training method to obtain at least one production indicator prediction model;

[0079] S300: Inputting the production parameters of the participant into the production index prediction model to obtain the prediction results of the production indicators corresponding to the production parameters.

[0080] The method for predicting production indicators provided by the present invention first obtains encrypted training samples of at least two participants (such as lithography machines, measurement equipment, etc.), then uses the training samples to train the pre-built initial network model using the federated learning training method to obtain at least one production indicator prediction model; finally, the production parameters of the participants are input into the production indicator prediction model to obtain the prediction results of the production indicators corresponding to the production parameters. It can be seen that the method for predicting production indicators provided by the present invention does not impose any restrictions on the participants and production samples. All the participants can belong to the same semiconductor production line, and all the participants can also belong to the equipment of the same process flow of different semiconductor production lines or the process control system of the same process flow of different semiconductor production lines; and the production indicators (which can be yield indicators, process indicators, production rate indicators, etc.) are not limited, and the versatility is strong; further, the prediction method provided by the present invention uses the encrypted training samples of each participant, so it can meet the needs of data privacy protection of each participant and is more convenient for promotion and implementation. Furthermore, the production indicator prediction model is trained using the federated learning training method, which can achieve joint modeling and optimization among all participants without the participants sharing training samples, thereby improving the prediction accuracy of the production indicator prediction model.

[0081] Preferably, in some exemplary embodiments, the participants include but are not limited to photolithography machines, measurement equipment, process control systems, etc. The participants may be at least one of the photolithography machines, measurement equipment, process control systems, etc., or other equipment in the semiconductor production line except the photolithography machines, measurement equipment, process control systems, etc. and the glue coating machines; the production indicators include but are not limited to process indicators, production rate indicators, yield indicators and equipment performance indicators. It should be noted that, as can be understood by those skilled in the art, all the specific contents of the participants (such as photolithography machines, process control systems, etc.) and the specific contents of the production indicators (such as yield indicators, equipment performance indicators, etc.) in this document are exemplary descriptions, not limitations of the present invention, and the present invention does not impose any limitations on the participants and the production indicators.

[0082] Specifically, the federated learning training method was first proposed by Google, with the goal of enabling multiple participants to jointly model and improve the model effect without sharing data. Its technical basis is homomorphic encryption technology and distributed computing technology, that is, after the data is homomorphically encrypted, the model is trained using distributed computing technology to obtain model improvement information, which becomes the private model optimization results of each participant after decryption. Taking the training process of the classic federated learning framework as an example, its training process can be roughly summarized into the following steps: the coordinator establishes a basic model and informs each participant of the basic structure and parameters of the model; each participant uses local data to train the model and returns the results to the coordinator; the coordinator summarizes the models of each participant and builds a more accurate global model to improve the overall model performance and effect. This meets the needs of data privacy protection for each participant. For more detailed information about the federated learning training method, please refer to the relevant technologies known to those skilled in the art, and no further explanation will be given here.

[0083] To facilitate understanding of the present invention, the following Example 1 takes the example of different equipment and / or process control systems of the participants from the same semiconductor production line, Example 2 takes the example of similar equipment and / or similar process control systems of the participants from different semiconductor production lines, and Example 3 takes the example of predicting production indicators of a new semiconductor production line with fewer training samples but similar equipment or process control systems to existing semiconductor production lines, to respectively illustrate the production indicator prediction method provided by the present invention.

[0084] In addition, it should be noted that although Embodiment 1, Embodiment 2 and Embodiment 3 are all described using a photolithography machine as an example, this is not a limitation of the present invention. The method for predicting production indicators of a semiconductor production line provided by the present invention does not impose any restrictions on the process flow and equipment of the semiconductor production line.

[0085] Embodiment 1

[0086] First, the core idea of ​​the method for predicting production indicators of a semiconductor production line provided in this embodiment is to improve the prediction accuracy of the production indicator prediction model of different equipment or process control systems of the same semiconductor production line through vertical federated learning training. In other words, the method for predicting production indicators provided in this embodiment is preferably applied in scenarios where each participant has a lot of the same sample data, but the data features between these sample data rarely overlap.

[0087] Specifically, in some exemplary embodiments, all of the participants belong to the same semiconductor production line, and the prediction method runs on the participants. Correspondingly, the prediction method also includes: extracting features from the original production parameter samples and the production index samples corresponding to the original production parameter samples, and using the extraction results as the training samples. Therefore, since the prediction method runs on the participants, the data privacy needs of each participant can be met; and since each participant comes from the same semiconductor production line, joint modeling of different equipment or process control systems of the same semiconductor production line is achieved under the premise of protecting data privacy.

[0088] For ease of understanding and description, the following is an example of three participants A1, B1 and C1 from the same semiconductor production line. For example, taking semiconductor advanced process control software (assuming that A1, B1 and C1 are photolithography machines, measurement equipment and process optimization software, respectively) as an example, its optimization target is process indicators, which include but are not limited to overlay accuracy, image quality and defects. In the related art, the photolithography machine is exposed according to the process recipe set by the process engineer, and the measurement equipment measures the actual results after exposure and etching. Finally, the process optimization software is modeled according to the key data submitted in plain text by each participant, and the optimization direction is given to form a sub-prescription for process correction. However, by adopting the production index prediction method provided in this embodiment, each participant A1, B1 and C1 does not need to submit key data in plain text, but only needs to submit encrypted training data to optimize the production index prediction model. After that, the prediction result of the production index can be obtained through the production index prediction model, and the optimization direction can be obtained according to the prediction result to form a sub-prescription for process correction. For more detailed information on how to obtain a sub-prescription for process correction based on the prediction results of the production indicators, please refer to the relevant technologies known to those skilled in the art, which will not be elaborated here.

[0089] Preferably, in some exemplary implementations, in step S100, each participant encrypts the training sample using homomorphic encryption to obtain an encrypted training sample. Thus, the use of homomorphic encryption allows each participant to use the encrypted training sample for model training when the training sample is encrypted and to obtain the same effect as the training model training using the training sample before encryption. While ensuring the data privacy of each participant, the amount of training sample data is effectively increased, laying a solid foundation for improving the prediction accuracy of production indicators.

[0090] Specifically, homomorphic encryption is an algorithm that satisfies the homomorphic operation property of ciphertext, that is, the plaintext data D is encrypted by the homomorphic encryption algorithm A to obtain the ciphertext S, and the ciphertext S is subjected to a specific calculation F. The result R1 obtained after the homomorphic decryption of the ciphertext calculation result SR is completely consistent with the result R2 obtained by directly performing the specific calculation F on the plaintext data D.

[0091] Preferably, in some exemplary implementations, before encrypting the training samples in step S100, the prediction method further includes: aligning the training samples according to the characteristic values ​​of the training samples provided by all the participants. Correspondingly, the encrypting the training samples using the homomorphic encryption method includes: encrypting the aligned training samples using the homomorphic encryption method. Therefore, the prediction method for production indicators provided in this embodiment, by aligning the training samples, enables each participant to clearly know what kind of training samples to provide for joint training, avoids exposing unique training samples that do not overlap with other participants, and further meets the needs of data privacy protection.

[0092] Specifically, each participant can use a feature extraction method to convert their own training samples into feature values. Each participant contributes the feature values ​​of the training samples. The intersection of these training samples is calculated based on these feature values. Then, each participant can know which training samples they should take out to participate in the joint training through the feature values ​​contained in the intersection, thereby avoiding exposing unique non-overlapping training samples. It should be noted that the present invention does not limit the above-mentioned feature extraction method. The feature extraction method includes but is not limited to the use of a one-way hash algorithm or an RSA-based blind signature algorithm. Furthermore, as can be understood by those skilled in the art, the feature extraction process is irreversible, so other participants cannot restore the original training samples of other participants only through feature values.

[0093] Preferably, in some exemplary implementations, each of the participants corresponds to a production indicator prediction model. Figure 2 , Figure 2 This is a schematic diagram of a specific process of training each participant to obtain a corresponding production index prediction model in one implementation of this embodiment. Figure 2 It can be seen that each participant obtains the corresponding production index prediction model through training through the following steps:

[0094] S211: Generate a homomorphically encrypted private key and public key, and use the public key of the party to exchange the public key with each other participant to obtain the public key of each other participant;

[0095] S212: using the encrypted training sample of the party, training the pre-built first initial network model until a first preset training end condition is met, and obtaining a first candidate prediction model; and using the first candidate prediction model, obtaining a first intermediate prediction result corresponding to the encrypted training sample;

[0096] S213: using the public key of the party concerned to encrypt the first intermediate prediction result to obtain a second intermediate prediction result; and using the second intermediate prediction result to exchange with the second intermediate prediction result of each of the other participants to obtain third intermediate prediction results of all other participants;

[0097] S214: using the public key of each of the other participants to encrypt the first intermediate prediction result to obtain a fourth intermediate prediction result;

[0098] S215: respectively using the third intermediate prediction result and the fourth intermediate prediction result encrypted by the public key provided by the same other participant to train the first candidate prediction model until a second preset training end condition is met, obtaining the first model correction amount of the other participant by the own party, and using the first model correction amount to exchange with the other participant to obtain the second model correction amount corresponding to the own party;

[0099] S216: using the private key to decrypt all the second model correction values, and using the decrypted second model correction values ​​to correct the first candidate prediction model to obtain the production index prediction model.

[0100] Therefore, in the production indicator prediction method provided by this embodiment, each participant realizes the joint training of their respective production indicator prediction models through the exchange of training samples and encrypted prediction results, which not only ensures the needs of each party's data privacy protection, but also because each participant comes from the same semiconductor production line, it can improve the prediction accuracy of each production indicator prediction model.

[0101] In order to further assist in understanding the present invention, the above steps are described as follows, taking the joint training of the model by three participants A1, B1 and C1 from the same semiconductor production line as an example:

[0102] First, each participant generates a homomorphic public-private key pair (public key PK and private key SK) and exchanges the public key (PK A1 ,PK B1 ,PK C1 ). Next, each participant uses the encrypted training sample (Participant A1 uses D A1 , Party B1 uses D B1 , Participant C1 uses D C1) trains and optimizes the first initial network model until the first candidate prediction model is obtained (Party A1 obtains model M A1 , Participant B1 obtains model M B1 , Participant C1 obtains model M C1 ), and use the first candidate prediction model (Party A1 uses model M A1 , Participant B1 uses model M B1 、Participant C1 uses model M C1 ) to obtain the first intermediate prediction result (R A1 , R B1 , R C1 ). Then, each participant uses his own public key (Participant A1 uses PK A1 , Party B1 uses PK B1 , Participant C1 uses PK C1 ) encrypts the first intermediate prediction result (Participant A1 encrypts R A1 Encryption, Party B1 to R B1 Encryption, Participant C1 to R C1 Encrypted) to obtain the second intermediate prediction result (Participant A1 obtains PK A1 R A1 , Participant B1 gets PK B1 R B1 , Participant C1 gets PK C1 R C1 ) and exchange. For participant A1, the exchange obtains the second intermediate prediction results PK of participants B1 and C1 B1 R B1 and PK C1 R C That is to say, each participant has the same third intermediate prediction result PK A1 R A1 ,PK B1 R B1 and PK C1 R C Next, the first intermediate prediction result is encrypted using the public key of each of the other participants to obtain a fourth intermediate prediction result. Taking participant A1 as an example, the first intermediate prediction result R is encrypted using the public keys of participants B1 and C1 respectively. A1 Encryption is performed to obtain the fourth intermediate prediction result PK B1 R A1 and PK C1 R A1Then, the third intermediate prediction result and the fourth intermediate prediction result encrypted with the same public key are used to train the first candidate prediction model, and the first model correction amount of the model of the other participant corresponding to the public key is obtained. Taking participant A1 as the participant and participant B1 as the other participant, the third intermediate prediction result PK B1 R B1 and the fourth intermediate prediction result PK B1 R A1 The first candidate prediction model of participant A1 is trained to obtain the encrypted first model correction value PK B1 L A1B1 , and use the data of participant A1 to optimize the first candidate prediction model corresponding to participant B1. Similarly, the third intermediate prediction result PKc1R held by participant A1 C1 and the fourth intermediate prediction result PK C1 R A1 The first candidate prediction model of participant A1 is trained to obtain the encrypted first model correction value PK C1 L A1C1 , used to optimize the first candidate prediction model corresponding to participant C1. At this time, through exchanges with other participants, the second model correction amount corresponding to participant A1 is PKA1L B1A1 and PKA1L C1A1 Finally, use your own private key SK A1 Decrypting the second model correction PKA1L B1A1 and PKA1L C1A1 , and obtain the second model correction value L after decryption B1A1 and L C1A1 By correcting the first candidate prediction model corresponding to participant A1, the production indicator prediction model corresponding to participant A1 can be obtained.

[0103] Preferably, in some exemplary embodiments, the use of the first model correction amount to exchange with the other participants includes: encrypting the first model correction amount using random noise, and using the encrypted first model correction amount to exchange with the other participants. Correspondingly, before using the private key to decrypt all the second model correction amounts, the prediction method also includes: filtering the first model correction amount using a filtering method corresponding to the random noise to obtain the second model correction amount. Therefore, the prediction method of the production index provided in this embodiment further improves the security of the data transmission process by encrypting the first model correction amount using random noise, and can better meet the needs of data privacy protection.

[0104] For example, for participant A1, by modifying the first model correction amount PK B1 LA1B1 and PK C1 L A1C1 By performing random noise encryption, the first model correction value PK after random noise encryption can be obtained B1 (L A1B1 +RN A1 ) and PK C1 (L A1C1 +RN A1 ), after exchanging with participants B1 and C1, participant A1 will receive PK A1 L B1A1 +PK A1 RN B1 and PK A1 L C1A1 +PK A1 RN C1 , after decryption, we get L B1A1 +RN B1 and L C1A1 +RN C1 , use the filtering algorithm to remove random noise and obtain the second model correction value L after decryption B1A1 and L C1A1 .

[0105] Preferably, in some exemplary embodiments, the use of the decrypted second model correction amount to correct the first candidate prediction model to obtain the production index prediction model includes:

[0106] The cumulative result of all the second model corrections is used as the correction of the first candidate prediction model, and the correction is used to modify the first candidate prediction model to obtain the production index prediction model. Therefore, the production index prediction method provided by this embodiment uses the cumulative result of all the second model corrections as the correction of the first candidate prediction model, and the calculation method is simple and easy to implement. Specifically, taking participant A1 as an example, the decrypted second model corrections obtained by participant A1 have two L B1A1 and L C1A1 , can be used for L B1A1 and L C1A1 The sum is added and the result of the addition is used as the correction amount of the first candidate prediction model corresponding to the participant A1. It should be noted that, as can be understood by those skilled in the art, the above is only an exemplary description and not a limitation of the present invention. In other embodiments, different weights may be applied to different other participants before obtaining the correction amount of the first candidate prediction model corresponding to the participant A1.

[0107] It should be particularly noted that, as those skilled in the art can understand, each participant can choose whether to use the second model correction amount to optimize its corresponding first candidate prediction model, that is, the number of participants using the second model correction amount can be equal to or less than the number of initial participants.

[0108] Preferably, in some exemplary implementations, the production index prediction method further comprises: calculating and recording the correction contribution of each of the other participants to the production index prediction model according to all the second model correction amounts. Thus, by recording the correction contribution of each other participant to the production index prediction model, a solid foundation is laid for more effectively improving the production indicators of the same production line from a global perspective while meeting the data privacy protection needs of each participant.

[0109] It should be noted that, as can be understood by those skilled in the art, the present invention does not limit the specific method of recording the modified contribution of each other participant. In some embodiments, a database may be used for recording; in other embodiments, a blockchain may be used for recording; in still other embodiments, a text may be used for recording. I will not elaborate on them one by one.

[0110] Preferably, in some exemplary implementations, step S300 inputs the production parameters of the participant into the production index prediction model to obtain the prediction results of the production index corresponding to the production parameters, including:

[0111] S310: Disassemble each of the production indicator prediction models to obtain a model body and an output layer corresponding to each participant;

[0112] S320: fixing the model parameters of the model body, and continuing to use the federated learning training method to optimize the output layer corresponding to each of the participants, so as to obtain a joint output layer according to all the optimized output layers;

[0113] S330: Obtaining a joint prediction model according to each of the model bodies and the joint output layer;

[0114] S340: The participant inputs the production parameters into the joint prediction model to obtain the prediction results of the production indicators corresponding to the production parameters.

[0115] Therefore, the method for predicting the production index of the semiconductor production line provided by this embodiment not only allows each participant to independently use their corresponding production index prediction model, but also obtains a joint prediction model by disassembling the production index prediction model of each participant and combining the output layer, thereby satisfying the data privacy protection needs of each participant, simplifying the difficulty of building the production index prediction model of the semiconductor production line and the complexity of training, and realizing the joint prediction of each participant of the semiconductor production line. Specifically, in step S320, the basic principle of continuing to use the federated learning training method to optimize the output layer corresponding to each of the participants is similar to the above steps S211 to S216, but it should be noted that, as mentioned above, when optimizing the output layer corresponding to each participant, the model parameters of the model body of the production index prediction model corresponding to each participant need to be fixed. Due to space limitations, it will not be repeated. For more detailed content, please refer to the adaptive understanding of steps S211 to S216 above. Those skilled in the art should be able to understand that step S320 is to further optimize (i.e., Finetune) the production index prediction model to obtain a joint output layer.

[0116] For scenarios where the production index prediction models corresponding to multiple participants can be disassembled, it is preferred to first disassemble the production index prediction models corresponding to each participant in the manner of steps S310 to S320, and then integrate them to obtain a joint prediction model. For example, see Figure 3 , which schematically shows a schematic diagram of the disassembly and assembly process of the production indicator prediction model given in one of the specific examples of the implementation of the present invention. Assume that in one of the application scenarios, there are three participants, and the production indicator prediction models trained by the three participants are M1O1, M2O2 and M3O3 respectively. Among them, the model body of the production indicator prediction model M1O1 is M1, and its output layer is O1; the model body of the production indicator prediction model M2O2 is M2, and its output layer is O2; the model body of the production indicator prediction model M3O3 is M3, and its output layer is O3. After disassembly in the above step S310, the model bodies M1, M2 and M3 are obtained; the output layers O1, O2 and O3. After the above step S320 optimizes the output layers O1, O2 and O3 and step S330 integrates the model bodies M1, M2 and M3 and the optimized output layers O1, O2 and O3, a joint prediction model is obtained, which includes the model bodies M1, M2, M3 and the joint output layer Ox. For scenarios where the generation indicator prediction models corresponding to multiple participants need to be repackaged, they can be combined and packaged directly through the above step S320. For details, please refer to Figure 4 , Figure 4 FIG. 1 is a schematic diagram of the packaging process of a production index prediction model provided as another specific example of the first embodiment of the present invention. Figure 4It can be seen that this method directly encapsulates the original production indicator prediction models corresponding to multiple participants (ie, without going through step S310) to obtain a joint prediction model.

[0117] It should be noted that, as those skilled in the art can understand, the present invention does not limit the specific implementation method of obtaining a unified output layer based on the output layers optimized by each participant. For example, in one embodiment, the average value of the model parameters of each optimized output layer can be used as the model parameters of the unified output layer; in other embodiments, different weighting factors can be applied to the model parameters of each optimized output layer to obtain the model parameters of the same output layer. In some other embodiments, other methods besides the above two methods can also be adopted. For more details, please refer to the relevant technologies known to those skilled in the art, which will not be described here one by one.

[0118] Taking a semiconductor production line with a photolithography machine as an example, the photolithography machine can optimize the process control algorithm of the photolithography machine and improve the performance index of the photolithography machine according to the prediction results of its optimized production index prediction model (model body and optimized output layer) (for example, if the input is historical data on how to optimize the process control algorithm, then the prediction result is also parameter data on how to optimize the process control algorithm). The joint prediction model can directly optimize the process indicators of the semiconductor production line. Exemplarily, if the input data is yield / yield / equipment performance data, the prediction result of the production index finally output by the joint prediction model is yield / yield / equipment performance data. If the input data is historical data on how to optimize yield / yield / equipment performance, the prediction result of the production index finally output by the joint prediction model is the parameters on how to optimize yield / yield / equipment performance.

[0119] Embodiment 2

[0120] Based on the same inventive concept as that of Example 1, this embodiment provides another method for predicting production indicators of a semiconductor production line. The difference from the method for predicting production indicators of a semiconductor production line provided in Example 1 is that: the method for predicting production indicators provided in this embodiment: all the participants belong to the equipment of the same process flow of different semiconductor production lines or the process control systems of the same process flow of different semiconductor production lines; all the participants correspond to one production indicator prediction model. In step S200, the training samples are used to train the pre-built initial network model using the federated learning training method to obtain at least one production indicator prediction model. The prediction method requires the cooperation of the central server and the participants. Due to space limitations, for the parts not mentioned in this embodiment, please refer to the relevant description in Example 1 for adaptive understanding. The following is a description of the differences between this embodiment and Example 1 from the perspectives of the central server and the participants, respectively.

[0121] First, the specific contents of the production index prediction method provided in this embodiment running on the central server are described as follows. Figure 5 , Figure 5 FIG. 1 is a schematic diagram of a specific process of obtaining a production index prediction model by training a central server in one implementation mode of this embodiment. Figure 5 It can be seen that the step S200 uses the training samples to train the pre-built initial network model using the federated learning training method to obtain at least one production indicator prediction model, which specifically includes:

[0122] S221: Building a second initial network model, and sending the second initial network model to each of the participants respectively;

[0123] S222: Iteratively perform the following steps until the production index prediction model is obtained:

[0124] SA1: receiving model reference parameters obtained by each participant from training the second initial network model; and modifying the second initial network model according to the model reference parameters to obtain a second candidate prediction model;

[0125] SA2: Sending the second candidate prediction model to each of the participants respectively, and receiving evaluation results of the second candidate prediction model by each of the participants;

[0126] SA3: Based on the evaluation results of all the participants, determine whether the second candidate prediction model meets the third prediction training end condition. If so, notify each of the participants that the training is finished so that each of the participants uses the second candidate prediction model as the production indicator prediction model; if not, notify each of the participants to continue training so that each of the participants uses the second candidate prediction model to replace the second initial network model and repeat the above steps SA1-SA3.

[0127] It can be seen that the prediction method of production indicators provided in this embodiment is that the central server provides a second initial network model and distributes it to each participant for training, and then the second initial network model is modified based on the model reference parameters received from each participant; and then the third preset training end condition is determined based on the evaluation results received from each participant. Since each participant only uses its own training samples for training, the needs of data privacy protection of each participant are further met, and since data exchange is not required between the participants, the coupling degree between the participants is low.

[0128] It should be noted that, unlike the preferred application scenario of the first embodiment where each participant belongs to the same semiconductor production line, the preferred application scenario of this embodiment is: similar equipment or processes of different existing semiconductor production lines (i.e., each participant has many common data features, but data samples rarely overlap). Specifically, taking the lithography machine as an example, different users A2, B2, and C2 all have different lithography machines of similar models. Each participant hopes to jointly model their own data to improve their own equipment performance or process indicators, etc., but their original data is the core secret of each user and is difficult to share directly.

[0129] Preferably, in some exemplary embodiments, the model reference parameters include model parameters or gradient parameters after homomorphic encryption.

[0130] Preferably, in some exemplary implementations, after receiving the model reference parameters obtained by each participant from training the second initial network model, the prediction method further includes:

[0131] According to all the model reference parameters, the modified contribution of each participant to the production index prediction model is calculated and recorded. For more details about the modified contribution, please refer to the relevant description in the first embodiment. In order to avoid redundancy, it will not be described here.

[0132] Next, the specific content of the production index prediction method provided in this embodiment running on the participants is described as follows. Figure 6 , Figure 6FIG. 1 is a schematic diagram of a specific process of obtaining a production index prediction model by training a participant in another implementation of this embodiment. Figure 6 It can be seen that the step S200 uses the training samples to train the pre-built initial network model using the federated learning training method to obtain at least one production indicator prediction model, which specifically includes:

[0133] S231: receiving a second initial network model sent by the central server;

[0134] S232: Iteratively perform the following steps until the production index prediction model is obtained:

[0135] SB1: using the encrypted training samples of the party, training the second initial network model until a fourth preset training end condition is met, obtaining model reference parameters of the trained second initial network model; and sending the model reference parameters to the central server;

[0136] SB2: receiving a second candidate prediction model obtained by the central processor according to the model reference parameters, and evaluating the second candidate prediction model to obtain an evaluation result;

[0137] SB3: Uploading the evaluation result to the central server;

[0138] SB4: Receive the training results sent by the central server. If the training is completed, use the received second candidate prediction model as the production indicator prediction model; if the training continues, use the second candidate prediction model to replace the second initial network model, and repeat the above steps SB1-SB4.

[0139] Since the production index prediction method provided in this implementation corresponds to the production index prediction method running on the central server provided in the above implementation, it will not be described in detail here. For more detailed content, please refer to the above description for adaptive understanding.

[0140] Preferably, in one exemplary implementation, sending the model reference parameter to the central server includes: encrypting the model parameter or gradient parameter using a homomorphic encryption method, and sending the encrypted model parameter or gradient parameter to the central server. Specifically, for more details about the homomorphic encryption method, please refer to the relevant description in Example 1, which will not be elaborated here.

[0141] It should be noted that, as can be understood by those skilled in the art, the prediction method for the production index of the semiconductor production line provided in this embodiment can be used independently or jointly by each participant after obtaining the production index prediction model, and the present invention does not limit this. Exemplarily, independent use is more suitable for maximizing the overall performance of the equipment or process control system of a single semiconductor production line, such as the performance of at least one of a certain photolithography machine, including but not limited to overlay accuracy and line width, that is, the optimization target is the overall performance of the upstream and downstream equipment belonging to the same semiconductor production line. Correspondingly, joint use is more suitable for the overall performance of a certain type of equipment, such as the overall performance of a certain type of photolithography machine, including but not limited to at least one of the performance of overlay accuracy and line width, that is, the optimization target is the overall performance of a certain type of equipment or a certain type of process control system on multiple semiconductor production lines.

[0142] Embodiment 3

[0143] Basically similar to the application scenario of Example 2, this embodiment provides another method for predicting the production indicators of a semiconductor production line, and all the participants in this embodiment come from different semiconductor production lines. The difference from the method for predicting the production indicators of a semiconductor production line provided in Example 2 is that in the method for predicting the production indicators provided in this embodiment, all the participants belong to equipment or process control systems with similar process flows. Due to space limitations, for the parts not mentioned in this embodiment, please refer to the relevant description in Example 2 for adaptive understanding. The following application scenarios and specific implementation methods are described only in terms of the differences between this embodiment and Example 2. The application scenario of the method for predicting the production indicators of a semiconductor production line provided in this embodiment is different from that of Example 2 in that the method for predicting the production indicators of a semiconductor production line provided in this embodiment is preferably applied in the following scenario: a new semiconductor production line with less data but similar to existing production lines in some equipment or processes (that is, it is not required that each participant has approximately the same data characteristics, but allows completely different characteristics). Specifically, taking the lithography machine as an example, different users A3 and B3 own similar models of lithography machines, and user C3 owns a lithography machine of another model (different from the models of the lithography machines owned by users A3 and B3). User C3 hopes to purchase the data of users A3 and B3 for modeling to improve equipment performance or process indicators, etc., but their original data are the core secrets of each user and are difficult to share directly.

[0144] Specifically, see Figure 7 , Figure 7 This is a schematic diagram of a specific process of obtaining a production index prediction model according to one implementation method of the third embodiment of the present invention. Figure 8 The following is a schematic diagram of the disassembly and assembly process of the production index prediction model given as a specific example of the third embodiment of the present invention. Figure 7It can be seen that the method for predicting production indicators of a semiconductor production line provided in this embodiment uses the training samples in step S200 to train a pre-built initial network model using a federated learning training method to obtain at least one production indicator prediction model, specifically including:

[0145] S241: Using the homomorphically encrypted private training samples of all the participants, a federated learning training method is used to train the pre-built third initial network model to obtain a global production index prediction model;

[0146] S242: For each of the participants, the following steps are performed to obtain the corresponding production index prediction model:

[0147] SC1: using the private training samples of the participant to train the pre-built fourth initial network model to obtain a private production index prediction model; and disassembling each of the private production index prediction models to obtain a private model body and a private output layer;

[0148] SC2: Integrate the private model body with the global production index prediction model to obtain an optimized output layer;

[0149] SC3: According to the usage model subject, the global production index prediction model and the optimized output layer, the production index prediction model corresponding to the participant is obtained.

[0150] Therefore, the method for predicting production indicators of semiconductor production lines provided in this embodiment, by independently superimposing a global production indicator prediction model (i.e., a feature mapping model) to map the space of different features of each participant into a new common feature space, so that completely different semiconductor production lines can be modeled and optimized together as long as some equipment or some process control systems are similar without sharing training samples among the participants, thereby improving the prediction accuracy of the production indicator prediction model, and further laying a solid foundation for optimizing the process, production rate, yield and other indicators and equipment performance indicators of each participant.

[0151] It should be noted that, as can be understood by those skilled in the art, the present invention does not limit the specific implementation method of obtaining the global production index prediction model by using the private training samples homomorphically encrypted by all the participants in S241 and training the pre-built third initial network model by the federated learning training method. For example, in some of the embodiments, the production index prediction model can be obtained by the prediction method described in any embodiment of Example 2. In other embodiments, it can also be obtained by training with relevant technologies known to those skilled in the art. It is only necessary that all the participants correspond to the same production index prediction model. Further, the present invention does not limit the structure of the third initial network model and the fourth initial network model. The structures of the third initial network model and the fourth initial network model can be the same or different. Further, the present invention does not limit the execution order of steps S241 and SC1. In some embodiments, step S241 can be executed first, and then step SC1; in other embodiments, step SC1 can be executed first, and then step S241; in other embodiments, SC1 and S241 can also be executed simultaneously.

[0152] For example, see Figure 8 , Figure 8 A schematic diagram of the disassembly and packaging process of a production index prediction model given in one specific example of this embodiment is schematically given. For ease of understanding, Figure 8Only two participants are shown in . However, it should be noted that, as can be understood by those skilled in the art, assuming that there are N (N≥2) participants in total, each participant will eventually obtain its corresponding production indicator prediction model, that is, there will be N different production indicator prediction models in the end. Specifically, before integration, the private training sample of one of the participants is D1 (i.e. private data, usually not encrypted), and after homomorphic encryption, the encrypted private training sample SD1 (encrypted data, which can be used for joint training) is obtained; after step SC1, the private production indicator prediction model is trained and the model body obtained by disassembling the model is M1, and the output layer is O1; the private training sample of the other participant is D2 (i.e. private data, usually not encrypted), and after homomorphic encryption, the encrypted private training sample SD2 (encrypted data, which can be used for joint training) is obtained; after step SC1, the private production indicator prediction model is trained and the model body obtained by disassembling the model is M2, and the output layer is O2. Using the encrypted training samples SD1 and SD2, the global production index prediction model Mx is obtained through step S241, and the optimized output layer corresponding to one of the participants after optimization is obtained through step SC2, and the optimized output layer corresponding to the other participant is O'1. Then, through step SC2, the private model body is integrated with the global production index prediction model to obtain the optimized output layer, and step SC3 obtains the production index prediction model corresponding to the participant according to the use model body, the global production index prediction model and the optimized output layer. Figure 8 As shown, the production index prediction model M1Mx O'1 corresponding to one of the participants is composed of the model body M1 and Mx and the optimized output layer O'1; the production index prediction model M1MxO'2 corresponding to the other participant is composed of the model body M2 and Mx and the optimized output layer O'2.

[0153] Embodiment 4

[0154] This embodiment provides a device for predicting production indicators of a semiconductor production line. Specifically, see Fig. 9 ,from Fig. 9It can be seen that the production indicator prediction device provided in this embodiment includes a training sample acquisition unit 100, a production indicator prediction model acquisition unit 200, and a production indicator prediction result acquisition unit 300. Specifically, the training sample acquisition unit 100 is configured to obtain encrypted training samples of at least two participants; wherein the participants include the equipment or process control system of the semiconductor production line, and each training sample includes the production parameter sample of the participant and the production indicator sample corresponding to the production parameter sample. The production indicator prediction model acquisition unit 200 is configured to use the training samples to train a pre-built initial network model using a federated learning training method to obtain at least one production indicator prediction model. The production indicator prediction result acquisition unit 300 is configured to input the production parameters into the production indicator prediction model to obtain the prediction result of the production indicator corresponding to the production parameters.

[0155] Since the prediction device for production indicators of the semiconductor production line provided in the present embodiment and the prediction method for production indicators of the semiconductor production line provided in the above embodiments belong to the same inventive concept, the prediction device for production indicators provided in the present embodiment at least has all the advantages of the prediction method for production indicators of the semiconductor production line provided in the above embodiments. For details, please refer to the description related to the prediction method for production indicators of the semiconductor production line above, which will not be repeated here. In addition, those skilled in the art should be able to understand that the basic principle of the prediction device for production indicators of the semiconductor production line provided in the present embodiment is the same as the basic principle of the prediction method for production indicators of the semiconductor production line provided in the above embodiments. Due to space limitations, it will not be described in detail here. For more detailed content about the prediction device provided in the present embodiment, please refer to the adaptive understanding of the relevant content of the prediction method provided in the above embodiments.

[0156] It should be noted that the devices and methods disclosed in the embodiments of this article can also be implemented in other ways. The device implementation described above is only schematic. For example, the flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the devices, methods and computer program products according to multiple embodiments of this article. In this regard, each box in the flowchart or block diagram can represent a part of a module, program or code, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function, and the module, program segment or part of the code contains one or more executable instructions for implementing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of boxes in the block diagram and / or flowchart can be implemented by a dedicated hardware-based system for performing a specified function or action, or can be implemented by a combination of dedicated hardware and computer instructions.

[0157] In addition, the functional modules in the various embodiments of this document may be integrated together to form an independent part, or each module may exist independently, or two or more modules may be integrated to form an independent part.

[0158] The above description is only a description of the preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any changes and modifications made by a person skilled in the art in the field of the present invention based on the above disclosure are within the scope of protection of the present invention. Obviously, a person skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations fall within the scope of the present invention and its equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for predicting production indicators of a semiconductor production line, characterized in that: include: Obtaining encrypted training samples of at least two participants; each training sample includes a production parameter sample of the participant and a production indicator sample corresponding to the production parameter sample; Using the training samples, a pre-built initial network model is trained using a federated learning training method to obtain at least one production indicator prediction model; The production parameters of the participant are input into the production index prediction model to obtain the prediction results of the production index corresponding to the production parameters.

2. The prediction method according to claim 1, characterized in that: All of the participants belong to the same semiconductor production line, and the prediction method runs on the participants. The prediction method also includes: extracting features from the original production parameter samples and the production indicator samples corresponding to the original production parameter samples, and using the extraction results as the training samples.

3. The prediction method according to claim 2, characterized in that: The prediction method also includes: encrypting the training sample using a homomorphic encryption method to obtain an encrypted training sample.

4. The prediction method according to claim 3, characterized in that: Before encrypting the training samples, the prediction method further includes: aligning the training samples according to the feature values ​​of the training samples provided by all the participants; The adopting homomorphic encryption method to encrypt the training samples includes: adopting homomorphic encryption method to encrypt the aligned training samples.

5. The prediction method according to claim 2, characterized in that: Each of the participants corresponds to a production index prediction model; each of the participants obtains the corresponding production index prediction model through training through the following steps: Generate a homomorphically encrypted private key and public key, and use your own public key to exchange the public key with each other participant to obtain the public key of each of the other participants; Using the encrypted training sample of the party, training the pre-built first initial network model until a first preset training end condition is met to obtain a first candidate prediction model; and using the first candidate prediction model to obtain a first intermediate prediction result corresponding to the encrypted training sample; Encrypting the first intermediate prediction result using the public key of the party to obtain a second intermediate prediction result; and using the second intermediate prediction result to exchange with the second intermediate prediction result of each of the other participants to obtain third intermediate prediction results of all other participants; Encrypting the first intermediate prediction result using the public key of each of the other participants to obtain a fourth intermediate prediction result; The first candidate prediction model is trained by respectively using the third intermediate prediction result and the fourth intermediate prediction result encrypted by the public key provided by the same other participant until a second preset training end condition is met, thereby obtaining the first model correction amount of the other participant by the own party, and using the first model correction amount to exchange with the other participant to obtain the corresponding second model correction amount of the own party; The private key is used to decrypt all the second model correction values, and the decrypted second model correction values ​​are used to correct the first candidate prediction model to obtain the production indicator prediction model.

6. The prediction method according to claim 5, characterized in that: The using the first model correction amount to exchange with the other participant includes: encrypting the first model correction amount using random noise, and using the encrypted first model correction amount to exchange with the other participant; Before using the private key to decrypt all the second model corrections, the prediction method further includes: filtering the first model corrections using a filtering method corresponding to the random noise to obtain the second model corrections.

7. The prediction method according to claim 5, characterized in that: The step of using the decrypted second model correction amount to correct the first candidate prediction model to obtain the production index prediction model includes: The accumulated result of all the correction values ​​of the second model is used as the correction value of the first candidate prediction model, and the first candidate prediction model is modified using the correction value to obtain the production index prediction model.

8. The prediction method according to claim 5, characterized in that: Based on all the second model corrections, calculate and record the correction contribution of each other participant to the production indicator prediction model.

9. The prediction method according to claim 1, characterized in that: Inputting the production parameters of the participant into the production index prediction model to obtain the prediction results of the production index corresponding to the production parameters includes: Disassemble each of the production indicator prediction models to obtain the model body and output layer corresponding to each participant; The model parameters of the model body are fixed, and the output layer corresponding to each of the participants is continuously optimized using the federated learning training method, so as to obtain a joint output layer according to all the optimized output layers; Obtaining a joint prediction model according to each of the model bodies and the joint output layer; The participating party inputs the production parameters into the joint prediction model to obtain the prediction results of the production indicators corresponding to the production parameters.

10. The prediction method according to claim 1, characterized in that: All of the participants are equipment belonging to the same process flow of different semiconductor production lines or process control systems belonging to the same process flow of different semiconductor production lines; all of the participants correspond to one production index prediction model; The prediction method is run on a central server, and the training sample is used to train a pre-built initial network model using a federated learning training method to obtain at least one production indicator prediction model, including: Building a second initial network model, and sending the second initial network model to each of the participants respectively; The following steps are iteratively performed until the production index prediction model is obtained: Receiving model reference parameters obtained by each participant from training the second initial network model; and modifying the second initial network model according to the model reference parameters to obtain a second candidate prediction model; Sending the second candidate prediction model to each of the participants respectively, and receiving evaluation results of the second candidate prediction model by each of the participants; Based on the evaluation results of all the participants, determine whether the second candidate prediction model meets the third prediction training end condition. If so, notify each of the participants that the training is completed so that each of the participants uses the second candidate prediction model as the production indicator prediction model; if not, notify each of the participants to continue training so that each of the participants uses the second candidate prediction model to replace the second initial network model, and repeat the above steps.

11. The prediction method according to claim 10, characterized in that: After receiving the model reference parameters obtained by each participant from training the second initial network model, the prediction method further includes: Based on all the model reference parameters, calculate and record the revised contribution of each participant to the production indicator prediction model.

12. The prediction method according to claim 1, characterized in that: All the participants are equipment belonging to the same process flow of different semiconductor production lines or process control systems belonging to the same process flow of different semiconductor production lines; all the participants correspond to one production index prediction model; the prediction method runs on the participants, and the training samples are used to train the pre-built initial network model using the federated learning training method to obtain at least one production index prediction model, including: Receiving a second initial network model sent by the central server; And iteratively perform the following steps until the production index prediction model is obtained: Using the encrypted training samples of the party, training the second initial network model until a fourth preset training end condition is met, obtaining model reference parameters after training the second initial network model; and sending the model reference parameters to the central server; receiving a second candidate prediction model obtained by the central processor according to the model reference parameters, and evaluating the second candidate prediction model to obtain an evaluation result; Uploading the evaluation results to the central server; Receive the training result sent by the central server. If the training is completed, use the received second candidate prediction model as the production indicator prediction model; if the training continues, use the second candidate prediction model to replace the second initial network model, and repeat the above steps.

13. The prediction method according to claim 12, characterized in that: The model reference parameters include model parameters or gradient parameters after homomorphic encryption.

14. The prediction method according to claim 1, characterized in that: All the participants belong to equipment or process control systems of similar process flows, and the training samples are used to train the pre-built initial network model using the federated learning training method to obtain at least one production indicator prediction model, including: Using the homomorphically encrypted private training samples of all the participants, a federated learning training method is used to train the pre-built third initial network model to obtain a global production indicator prediction model; For each of the participants, the following steps are performed to obtain the corresponding production indicator prediction model: Using the private training sample of the participant to train the pre-built fourth initial network model to obtain a private production indicator prediction model; and disassembling each of the private production indicator prediction models to obtain a private model body and a private output layer; Integrate the private model body with the global production index prediction model to obtain an optimized output layer; According to the usage model body, the global production index prediction model and the optimized output layer, the production index prediction model corresponding to the participant is obtained.

15. A device for predicting production indicators of a semiconductor production line, characterized in that: include: A training sample acquisition unit, configured to acquire encrypted training samples of at least two participants; wherein the participants include equipment or process control systems of the semiconductor production line, and all the participants belong to the same semiconductor production line, equipment of the same process flow of different semiconductor production lines, or process control systems of the same process flow of different semiconductor production lines; each training sample includes a production parameter sample of the participant and a production indicator sample corresponding to the production parameter sample; A production index prediction model acquisition unit is configured to use the training samples to train a pre-built initial network model using a federated learning training method to obtain at least one production index prediction model; The production index prediction result acquisition unit is configured to input the production parameter into the production index prediction model to obtain the prediction result of the production index corresponding to the production parameter.

Images