A model updating method, system and product
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN EXX IND AUTOMATION CO LTD
- Filing Date
- 2026-03-25
- Publication Date
- 2026-07-14
Smart Images
Figure CN121902645B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of machine tool model updating technology, specifically to a model updating method, system, and product. Background Technology
[0002] Semiconductor manufacturing involves multiple processes, each of which can be divided into several steps and stages. Different processes can be completed by different machines. Each machine is equipped with multiple sensors, and high-precision monitoring relies on high-frequency data acquisition with a wide range of values. Therefore, each machine collects massive amounts of data.
[0003] In response, traditional technologies have proposed some model training methods for semiconductor equipment.
[0004] For example, patent application CN121637971A discloses a model training method, a silicon wafer processing method, an apparatus, a storage medium, and a device. The model training method includes: acquiring initial state information and target state information of the silicon wafer to be processed; generating processing action information for a processing machine based on the initial and target state information using an action decision model; determining reward parameters corresponding to the processing action information, whereby the reward parameters characterize the processing quality and efficiency of the processing machine when processing the silicon wafer based on the processing action information; updating the initial action decision model based on the reward parameters if the reward parameters are not satisfied with preset conditions, thus obtaining an updated action decision model; and returning to the step of generating processing action information based on the initial and target state information using the action decision model, based on the updated action decision model, until the reward parameters satisfy the preset conditions, thus obtaining the target action decision model.
[0005] For example, patent application CN118521855A discloses a supercomputing model training and deployment system, which relates to the field of model training. It includes a model training center node and several terminal edge nodes. The terminal edge nodes include a data server and several production line machines. The production line machines detect wafer material boards based on a defect detection model and collect defect images and confirm defect information. The data server generates defect samples based on the defect images and defect information, uploads the defect samples, and receives push-down instructions and trained AI models, deploying the AI models to the target production line. The model training center node includes an image annotation server and a model training server. The image annotation server classifies and annotates the uploaded defect samples. The model training server performs iterative training based on the classified and annotated defect samples.
[0006] For example, patent application CN116108642A discloses a parameter debugging model training method, debugging method, and equipment for a welding machine. The parameter debugging model training method for the welding machine includes: acquiring welding data of multiple historical products welded by the target welding machine; the welding data of each historical product includes the material information of the historical product, the good product information of the historical product, and the welding configuration parameters of the target welding machine when welding the historical product; the good product information of the historical product is used to characterize whether the historical product is good; constructing a training sample set based on the welding data of multiple historical products welded by the target welding machine; and training a preset debugging model based on the training sample set to update the debugging model to the parameter debugging model of the target welding machine.
[0007] However, due to the complexity of semiconductor manufacturing processes (such as adjustments to process details or updates to production line equipment), traditional AI models face high maintenance costs after deployment. Summary of the Invention
[0008] The purpose of this invention is to provide a model update method, system, and product that partially solves or alleviates the above-mentioned shortcomings in the prior art and can improve the scene adaptability of machine models.
[0009] To solve the aforementioned technical problems, the present invention specifically adopts the following technical solution:
[0010] A first aspect of the present invention is to provide a model update method, comprising:
[0011] S201, construct a model pool for the machine group, the machine group includes: at least two wafer machines, and the model pool includes: (1) at least one primary model, and the primary model is created corresponding to at least one of the wafer machines; (2) at least one secondary model, and the secondary model is created uniformly corresponding to at least two wafer machines;
[0012] S202, identify the scene change type of the target machine, wherein the scene change type includes at least one of the following: process parameter change, model output change, machine object change, component update change, model performance change; and the target machine is one of the wafer machines in the machine group;
[0013] S203, Select a reference pool for model update according to the scenario change type, wherein the reference pool is used to define the source range of reference data for updating the target model, the target model is the first-level model or the second-level model corresponding to the target machine, and the source range includes at least one of the following: (A) training data source; (B) model architecture source;
[0014] The training data sources include at least one of the following: (1) the first operating data of the target machine; (2) the second operating data of a similar machine; and the similar machine is the wafer machine that has the same working scenario as the target machine; the model architecture sources include: (1) the first-level model of the similar machine; (2) the second-level model corresponding to the target machine;
[0015] S204, Update the target model according to the reference pool.
[0016] In some embodiments, S203 includes the step of:
[0017] When the scenario change type is a type of change, the source range of the reference pool includes: the training data source; and the type of change includes: process parameter changes and model performance changes;
[0018] When the scenario change type is a type II change, the source range of the reference pool includes: the model architecture source; and the type II change includes: model output change, machine object change, and component update change.
[0019] In some embodiments, when the source range of the reference pool includes: model architecture source, correspondingly, S204 includes the step:
[0020] Obtain the model score of the model, wherein the model includes: a first-level model of the similar machine and / or a second-level model corresponding to the target machine;
[0021] The model whose score is greater than the set score is used as the reference model;
[0022] The model architecture of the target model is updated using the reference model.
[0023] In some embodiments, when the reference model includes: a primary model of the similar machine, updating the model architecture of the target model using the reference model includes the following steps:
[0024] Identify the historical events experienced by the similar machines during a set time period, and the corresponding times of occurrence;
[0025] When the historical event is of the same type as the scene change, the historical event is identified as a matching event; otherwise, the historical event is identified as a non-matching event.
[0026] The reference weights of the reference model are determined based on the matching event and the occurrence time.
[0027] The target model is updated based on the reference weights and the reference model.
[0028] In some embodiments, the later the occurrence time, the higher the reference weight.
[0029] In some embodiments, before updating the target model according to the reference weights and the reference model, the method further includes the following steps:
[0030] Identify non-matching events that occur after the match has occurred;
[0031] Compare the correlation between the non-matching events and the matching events;
[0032] If the correlation degree is greater than the preset correlation degree, the reference weight is maintained or reduced.
[0033] If the correlation degree is less than or equal to the correlation degree, the corresponding reference model will be removed from the reference pool.
[0034] In some embodiments, S204 includes:
[0035] Updated training data is obtained from the reference pool;
[0036] The first-level model is updated using the updated training data.
[0037] In some embodiments, the input of the primary model is process parameters, and the output of the primary model is business rules, which include at least one of the following categories: yield indicators, stability indicators, capacity indicators, and cost indicators.
[0038] And / or, the update objective in S204 is to require the output of the first-level model to meet the business rules;
[0039] And / or,
[0040] Before applying the updated primary model to the target machine, the following steps are also included:
[0041] Obtain process parameters based on the updated primary model;
[0042] Test production was carried out using sample wafers under the stated process parameters;
[0043] The updated primary model is verified based on the test production results, and after successful verification, the updated primary model is allowed to be applied to the target machine.
[0044] A second aspect of the present invention is to provide a model update system, comprising:
[0045] A model pool construction module is used to construct a model pool for a group of wafer rigs, the group of wafer rigs including at least two wafer rigs, and the model pool including: (1) at least one primary model, and the primary model is created corresponding to at least one of the wafer rigs; (2) at least one secondary model, and the secondary model is created uniformly corresponding to at least two wafer rigs.
[0046] The scene change recognition module is used to identify the scene change type of the target machine. The scene change type includes at least one of the following: process parameter change, model output change, machine object change, component update change, and model performance change; and the target machine is one of the wafer machines in the machine group.
[0047] The reference pool selection module is used to select a reference pool for model updating according to the scene change type. The reference pool is used to define the source range of reference data for updating the target model. The target model is the first-level model or the second-level model corresponding to the target machine. The source range includes at least one of the following: (A) training data source; (B) model architecture source.
[0048] The training data sources include at least one of the following: (1) the first operating data of the target machine; (2) the second operating data of a similar machine; and the similar machine is the wafer machine that has the same working scenario as the target machine. The model architecture sources include: (1) the first-level model of the similar machine; (2) the second-level model corresponding to the target machine.
[0049] The update module is used to update the target model based on the reference pool.
[0050] A third aspect of the present invention is to provide a computer program product comprising a computer program that, when executed by a processor, implements the model update method as described in any embodiment of the present invention.
[0051] Beneficial technical effects:
[0052] It is worth noting that the restrictive update strategy provided by this invention is not intended to provide the most accurate model in the initial stage, but to limit the range of reference data to a certain extent, so as to quickly ensure that the model meets the basic business rules in new scenarios (e.g., in the updated machine or in the next production stage) with relatively limited update resource investment, and can be quickly enabled on the production line.
[0053] In other words, this invention focuses on providing an executable model solution for the production line in the early stages of model deployment or in scenarios where model performance may decline, that is, balancing between executableness and model training cost.
[0054] To address this, the present invention first provides an update strategy recommendation mechanism that restricts the scale of model updates based on the type of scene change. Specifically, the magnitude of the update is controlled by restricting the source of the data referenced for the update.
[0055] In other words, this invention chooses to switch between updating the model architecture or updating the parameters to match different changing scenarios.
[0056] Specifically, this invention provides a further guided narrowing scheme for model architecture update paths that are difficult and costly to update. This scheme filters eligible models based on overall model performance and adjusts their reference weights based on historical events experienced by the models. Through the combined efforts of overall filtering and personalized parameter tuning, the pressure of updating the model architecture can be further reduced.
[0057] Furthermore, this invention also employs a comprehensive analysis of matching and non-matching events over time to avoid or limit excessive narrowing of the reference pool while appropriately reducing the impact of low-reference-degree models on the target model update. Attached Figure Description
[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. The elements or parts in the drawings are not necessarily drawn to scale. Obviously, the drawings described below are some embodiments of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative effort.
[0059] Figure 1 A schematic flowchart of a model update method provided by the present invention;
[0060] Figure 2 A schematic diagram of the structure of a model update system provided by the present invention;
[0061] Figure 3 This is a schematic block diagram of the structure of a computer device provided by the present invention. Detailed Implementation
[0062] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.
[0063] In this document, suffixes such as "module," "part," or "unit" used to denote elements are used only for the purpose of illustrative purposes and have no specific meaning in themselves. Therefore, "module," "part," or "unit" may be used interchangeably.
[0064] In this document, the terms "upper," "lower," "inner," "outer," "front," "rear," "one end," and "the other end," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the present invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the present invention. Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.
[0065] In this document, unless otherwise explicitly specified and limited, the terms "installed," "equipped with," "connected," etc., should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection, a direct connection, or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0066] In this document, "and / or" includes any and all combinations of one or more of the listed related items.
[0067] In this article, "multiple" means two or more, that is, it includes two, three, four, five, etc.
[0068] As used in this specification, the term "about" typically means + / -5% of the value, more typically + / -4% of the value, more typically + / -3% of the value, more typically + / -2% of the value, even more typically + / -1% of the value, and even more typically + / -0.5% of the value.
[0069] In this specification, certain embodiments may be disclosed in a range-bound format. It should be understood that this "range-bound" description is merely for convenience and brevity and should not be construed as a rigid limitation on the disclosed range. Therefore, the description of a range should be considered as having specifically disclosed all possible subranges and the individual numerical values within those ranges. For example, a description of the range 1-6 should be considered as having specifically disclosed subranges such as from 1 to 3, from 1 to 4, from 1 to 5, from 2 to 4, from 2 to 6, from 3 to 6, etc., and the individual numbers within those ranges, such as 1, 2, 3, 4, 5, and 6. This rule applies regardless of the breadth of the range.
[0070] Definition of the noun:
[0071] A wafer can refer to a silicon chip used to manufacture silicon semiconductor circuits.
[0072] A wafer fabrication unit (or simply unit) can refer to any one or more processing modules or devices on a wafer fabrication line.
[0073] For example, a machine refers to equipment on a production line that performs specific process steps. In a semiconductor wafer fab, this could specifically refer to lithography machines, etching machines, thin film deposition equipment, ion implanters, or chemical mechanical polishing equipment. During the operation of each machine, machine data is collected and recorded (which is used as training data) for real-time monitoring, fault diagnosis, yield analysis, and predictive maintenance.
[0074] For example, machine data may include engineering data on the equipment's own health status (such as time-series data such as temperature, pressure, gas flow rate, motor speed, and vibration spectrum), process parameters of the processing (such as recipe settings, actual execution values, alarm logs, and event records), and contextual information related to production results (such as wafer ID, batch number, process step identifier, process sequence label, and timestamp), etc., without limitation. In this article, the primary model can be a dedicated model created for a specific wafer machine in a machine group, adapted to the machine's own operating characteristics (such as process type, operating status, and hardware attributes), also known as a personalized model.
[0075] A secondary model can be a general model created by extracting the common features of at least two wafer rigs in a group of rigs, which can be adapted to the common working scenarios of these rigs. It is also called a balanced model.
[0076] In this paper, the first-level model and the second-level model are preferably AI models.
[0077] A rig group can refer to a collection that contains at least two wafer rigs. For example, a rig group refers to a collection of rigs that have the same or similar attributes (such as rig type, hardware configuration, or process formulation).
[0078] For example, multiple etching machines used to perform the same process can be grouped into a single machine group.
[0079] For example, they can also be subdivided according to model number, such as grouping machines that perform the same process and are from the same batch (or have the same batch of core components) into a group of machines.
[0080] In the complex manufacturing environment of wafer fabrication lines, configuring AI models for equipment to generate process suggestions is a key means to improve yield and efficiency. However, the application of AI models faces high configuration and maintenance costs.
[0081] For example, the state of a machine tool can shift over time. This shift can stem from various factors, such as wear and tear on the electrostatic chuck, accumulation of polymer within the cavity, or aging of consumables. When the machine tool shifts, the data distribution upon which the AI model trained on historical data relies also changes, leading to a decrease in the model's predictive accuracy and even generating process recommendations that do not match actual operating conditions. Therefore, to ensure the effectiveness of the AI model, it must be regularly updated and maintained to enable it to track and adapt to the latest state of the machine tool.
[0082] However, the complexity of production line changes places enormous cost pressure on this maintenance work. In the initial setup phase, massive amounts of data need to be collected for different models and configurations of machines to train a reliable initial model. In the later maintenance phase, due to the large number of machines and their significant individual differences, regularly maintaining or updating the model for each machine will also consume substantial manpower and computing resources.
[0083] In other words, AI models face significant technical challenges in both the application deployment and maintenance phases.
[0084] Example 1:
[0085] This embodiment focuses on providing a recommended mechanism for model maintenance strategies, so as to avoid excessive model maintenance costs (i.e., complete model update and deployment with limited update resources) while updating the machine offset problem in a timely manner.
[0086] Please see Figure 1 This invention proposes a model update method, comprising:
[0087] S201, construct a model pool for the machine group, the machine group includes: at least two wafer machines, and the model pool includes: (1) at least one primary model, and the primary model is created corresponding to at least one of the wafer machines; (2) at least one secondary model, and the secondary model is created uniformly corresponding to at least two wafer machines;
[0088] S202, identify the scene change type of the target machine, wherein the scene change type includes at least one of the following: process parameter change, model output change, machine object change, component update change, model performance change; and the target machine is one of the wafer machines in the machine group;
[0089] S203, Select a reference pool for model update according to the scenario change type, wherein the reference pool is used to define the source range of reference data for updating the target model, the target model is the first-level model or the second-level model corresponding to the target machine, and the source range includes at least one of the following: (A) training data source; (B) model architecture source;
[0090] The training data sources include at least one of the following: (1) the first operating data of the target machine; (2) the second operating data of a similar machine; and the similar machine is the wafer machine that has the same working scenario as the target machine; the model architecture sources include: (1) the first-level model of the similar machine; (2) the second-level model corresponding to the target machine.
[0091] In some embodiments, the target model is the primary model or the secondary model corresponding to the target machine.
[0092] In this embodiment, the reference pool for model updates is differentiated and its range is limited based on the differences in the working environment of the machine, thereby reducing or limiting excessive model updates to a certain extent.
[0093] In some embodiments, machines located in the same group of machines as the target machine are considered similar machines to the target machine.
[0094] In other words, this invention creates a rich reference pool containing various types of training data and reference models to provide ample data support for model updates and maintenance. Furthermore, during actual operation, the reference direction for model updates can be strategically selected based on specific operational needs, striking a balance between update cost and reliability.
[0095] For example, in some scenarios, this invention chooses to use the original architecture of the first-level model, focusing on updating the model through updates to the training data. In other scenarios, this invention focuses on using other models (such as other first-level or second-level models) for multi-model fusion updates.
[0096] Among them, changes in process parameters can refer to changes in working conditions (such as changes in chamber temperature).
[0097] Among them, changes in model output can refer to changes in the business requirements of the model, such as changing the focus from yield rate to cost control.
[0098] Among these, "change of machine object" can refer to replacing the current machine with another machine. For example, when the current machine experiences aging issues, a new machine will be used to replace it.
[0099] Among these, component updates and changes can refer to scenarios involving the replacement of parts. For example, if a core component in the current machine tool has exceeded its service life and may need to be replaced, then a model update may also need to be considered.
[0100] Among them, model performance change can refer to the model / device experiencing performance drift due to long-term operation, resulting in a decrease in model adaptability.
[0101] In some embodiments, a work scenario can be defined by at least one of the following work indicators: process parameters (such as process steps, process formulas, process cycles, etc.), business requirements (such as yield standards, capacity requirements, product specifications, etc.), hardware performance (such as machine models, core component configurations, etc.), and operating environment requirements (such as temperature and humidity, power supply stability, etc.).
[0102] In some embodiments, "same working scenario" can mean that the number of identical working indicators is greater than or equal to a preset threshold (e.g., 3 items). For example, if target machine a and machine b have the same three working indicators—process parameters, business requirements, and hardware performance—then machine b can be considered a similar machine to target machine a.
[0103] Preferably, "same" in this embodiment does not necessarily mean absolute sameness. For example, taking process parameters as an example, when the process parameter is process temperature, "same process parameters" means that the two process temperatures are within the same temperature range, or that the difference between the two temperatures is relatively small, and they can be regarded as being in the same or similar temperature environment in the actual wafer processing process.
[0104] Alternatively, in some embodiments, machines with the same working scenario are usually machines that are classified into the same machine group, that is, similar machines are machines located in the same machine group as the target machine.
[0105] In some embodiments, the first-level model may also be referred to as the personalized model. The second-level model may also be referred to as the equilibrium model or the group model.
[0106] In some embodiments, S203 includes the step of:
[0107] When the scenario change type is a type of change, the source range of the reference pool includes: training data sources; and the type of change includes: process parameter changes and model performance changes.
[0108] When the scenario change type is a type II change, the source range of the reference pool includes: model architecture source; and the type II change includes: model output change, machine object change, and component update change.
[0109] For example, if a machine needs to fine-tune the chamber temperature when producing a certain type of wafer, but the machine, product, and hardware remain unchanged, only one process parameter needs adjustment, the original model framework can be used. The model can be trained using machine operation data, allowing it to relearn the patterns in the operation data.
[0110] In other words, in this embodiment, when dealing with a type of change that requires fine-tuning of details, the training data source can be selected as the reference pool for model updates. That is, for a type of change, this invention chooses to use historical machine operation data to calibrate or update the model, thereby fine-tuning the model at a lower cost to quickly adapt to minor changes.
[0111] For example, if the core components of the equipment are replaced (such as the radio frequency module), the operating characteristics of the equipment (such as parameter fluctuation patterns and response speed) may change significantly. In this case, a model architecture adapted to the new hardware can be used to learn and match the operating patterns of the equipment with the new hardware.
[0112] For cases involving relatively large variations, such as type II changes, this invention selects the model architecture source as the source range of the reference pool, that is, it selects to switch to other models to better adapt to the current scenario, thereby improving the accuracy of model calculation.
[0113] Furthermore, by combining actual needs to recommend different model update strategies, this invention can avoid excessive model updates, thereby reducing the update pressure on the model.
[0114] It should be noted that the present invention addresses one type of update scenario:
[0115] After a wafer fabrication machine has been running for a period of time, it needs to be maintained and repaired. During the maintenance process, it was found that a certain core component did not meet the working requirements, so the core component was replaced.
[0116] At this point, due to slight deviations in different components due to batches, these slight deviations may cause process fluctuations during the high-precision processing of wafers, thereby affecting wafer yield or production capacity.
[0117] At this point, the original model of the wafer foundry can be updated in advance, taking into account the characteristics of component replacement, so that when the wafer foundry is restarted, the wafer processing will not be easily affected by the detailed offset of the components.
[0118] Another update scenario addressed by this invention is:
[0119] When a wafer foundry has been running for a period of time and a serious, intractable malfunction is discovered during maintenance, the entire wafer foundry will be replaced. However, due to differences in manufacturing batches and specifications of the foundry or its components, the original model may be unable to accurately predict the process parameters of the new foundry, affecting wafer processing. Therefore, this scenario also requires pre-updating the original model.
[0120] Alternatively, when a factory needs to add a new production line, the models deployed in the existing production line may not be usable directly and will also need to be adapted.
[0121] However, deploying AI models is extremely difficult (due to costs and technical challenges such as training and updating), and failure to maintain them in a timely manner may lead to model failure, thus delaying wafer production. However, timely maintenance, on the other hand, places extremely high demands on application costs.
[0122] In response, this invention provides scenario-based analysis guidance to restrict update strategies (such as update mechanisms and update reference data) during model maintenance. This allows for rapid optimization and updates of the model with limited and reasonable resource investment, ensuring that the updated model meets the activation criteria.
[0123] In other words, this restrictive update strategy can reduce the time delay that model updates may cause to a certain extent. Unlike the conventional thinking in traditional wafer model applications, this embodiment does not pursue absolute model accuracy, but adopts a restrictive update strategy to enable the model to quickly reach the activation standards for new scenarios and achieve rapid boot-up.
[0124] In some embodiments, the method further includes the step of:
[0125] S204, Update the target model according to the reference pool.
[0126] In some embodiments, when the source of the reference pool includes a model architecture source, correspondingly, S204 includes the step:
[0127] Select a reference model from the reference pool and update the architecture of the target model based on the reference model.
[0128] The reference model can be a primary model and / or a secondary model of the similar machine; the primary model and the secondary model can also be collectively referred to as the model.
[0129] Preferably, the step of selecting a reference model from the reference pool includes:
[0130] Obtain the model score;
[0131] Models whose scores are greater than the set score are used as reference models.
[0132] In this embodiment, the model is used to evaluate the performance of machine learning models on specific tasks.
[0133] Model scores can be calculated comprehensively based on different dimensions (such as prediction error, stability, generalization ability, robustness, etc.). In the context of wafer production lines, model scores can comprehensively consider the model's fitting accuracy to process parameters, the volatility of prediction errors, its adaptability to edge cases, and its trend over time, thus providing an intuitive and comparable basis for model selection, updates, and performance monitoring.
[0134] For example, in some embodiments, taking a business rule of yield greater than 90% as an example, if the actual wafer yield is 85% after running the process recommendations given by the model for a period of time, it indicates that the model has not met the requirements of the business rule, and the model score is low. Conversely, when the actual wafer yield is 90% or higher, the model score is also higher.
[0135] In some embodiments, when the reference model includes: a primary model of the similar machine, updating the model architecture of the target model using the reference model includes the following steps:
[0136] Identify the historical events experienced by the similar machines during a set time period, and the corresponding times of occurrence;
[0137] When the historical event is of the same type as the scene change, the historical event is identified as a matching event; otherwise, the historical event is identified as a non-matching event.
[0138] The reference weights of the reference model are determined based on the matching event and the occurrence time.
[0139] The target model is updated based on the reference weights and the reference model.
[0140] For example, in some embodiments, if the current target machine is experiencing a scenario change type of component update change, and the similar machine is also experiencing a historical event of component update, then the historical event is considered to be a matching event.
[0141] For example, in some embodiments, the matching rules can be more precise. For instance, if the target machine experiences a scene change of type A (part update), then if the similar machine also experiences a part update of type A, the historical event is considered a matching event. Of course, the specific matching rules can be flexibly adjusted by engineers, and will not be elaborated upon here.
[0142] In some embodiments, the later the occurrence time, the higher the reference weight.
[0143] Specifically, this invention provides a further guided narrowing scheme for model architecture update paths that are difficult and costly to update. This scheme filters eligible models based on overall model performance and adjusts their reference weights based on historical events experienced by the models. Through the combined efforts of overall filtering and personalized parameter tuning, the pressure of updating the model architecture can be further reduced.
[0144] In some embodiments, before updating the target model according to the reference weights and the reference model, the method further includes the following steps:
[0145] Identify non-matching events that occur after the match has occurred;
[0146] Compare the correlation between the non-matching events and the matching events;
[0147] If the correlation degree is greater than the preset correlation degree threshold, the reference weight is maintained or reduced.
[0148] If the correlation degree is less than or equal to the correlation degree threshold, the corresponding reference model will be removed from the reference pool.
[0149] In some embodiments, the degree of correlation between different events can be preset by engineers (or users) based on their work experience. For example, there can be a high degree of correlation between changes in machine objects and changes in component updates, while there can be a low degree of correlation between changes in machine objects and changes in model output.
[0150] For example, in some embodiments, taking component update changes as an example, there is a high degree of correlation when both involve the same type of component change, and a low degree of correlation when both involve the same type of component change.
[0151] It is understandable that this embodiment also uses matching events and non-matching events for comprehensive analysis in the time dimension, so as to avoid or limit the excessive narrowing of the reference pool while appropriately reducing the impact of low reference models on the target model update.
[0152] Specifically, in this embodiment, when there are non-matching events between a matching event and the current time (i.e., the update time of the target model), instead of directly excluding the model, the non-matching events are inserted to accommodate them to a certain extent. This moderate tolerance for non-matching events also avoids excessively narrowing the reference pool, which would reduce the value of the reference data.
[0153] In other words, by evaluating the correlation between non-matching events and matching events, it is possible to maintain the data comprehensiveness of the reference pool while limiting its data range.
[0154] In some embodiments, the correlation threshold can also be dynamically adjusted.
[0155] For example, the steps to determine the correlation threshold include:
[0156] Obtain the number of machine groups corresponding to the target machine;
[0157] The correlation threshold is determined based on the quantity, wherein the higher the quantity, the higher the correlation threshold.
[0158] In other words, in this embodiment, when there are a large number of similar models to be selected, the narrowing range of the reference pool can be appropriately increased.
[0159] It should be noted that this control over the reference pool is particularly suitable for model updates in the early stages of production lines. For example, when a wafer fab has just introduced a new process node (such as transitioning from 28nm to 14nm) or a new process module (such as a new atomic layer deposition method), it typically only configures a small number of experimental or verification equipment initially. At this time, by controlling the narrowing of the reference pool, it is possible to ensure that a relatively sufficient number of reference models are selected.
[0160] Alternatively, for some special or niche process nodes, such as special processes (e.g., compound semiconductors, MEMS, power devices) or special processes such as ultra-high aspect ratio etching, they are often only configured on some production lines, and the number of processing equipment configured is also relatively limited. In this case, by adjusting the shrinkage range of the reference pool, it is also possible to avoid over-screening of the reference model, resulting in a small reference surface.
[0161] For example, in some embodiments, updating the model architecture of the target model using a reference model refers to:
[0162] A model fusion method is used to fuse the reference model and the target model to obtain a new target model.
[0163] For example, the model fusion method can be a weight-based fusion method, a knowledge distillation method, or a domain adaptation / domain generalization-based fusion method, etc., and this invention does not limit it.
[0164] At this point, the model architecture of the new target model is affected by the reference model during the fusion process, thereby enabling the reference model to update the model architecture of the target model.
[0165] For example, in some embodiments, updating the model architecture can be done by updating the computation graph structure of the model.
[0166] For example, in some embodiments, updating the model architecture can be done by adjusting the type of layers (such as fully connected, convolutional, recurrent, attention, etc.), the connection method between layers (such as sequential, skip connections, branching, recurrent, etc.), the dimension of layers (such as the number of neurons, the size of convolutional kernels, etc.), or the overall topology (such as encoder-decoder, multi-expert hybrid, residual network, etc.).
[0167] For example, in some embodiments, by limiting the reference data for model updates to the training data, the focus is on guiding parameter updates of the model, such as making fine adjustments to the weight matrix, bias vector, or scaling factor in normalization.
[0168] In other words, updating the model architecture focuses on adjusting the skeleton of the model, while updating the training data focuses on fine-tuning the details of the skeleton while maintaining the original skeleton as much as possible.
[0169] In some embodiments, if the target model is a first-level model, the method further includes the following steps:
[0170] The primary model is updated based on the reference pool.
[0171] In some embodiments, S204 includes:
[0172] Updated training data is obtained from the reference pool;
[0173] The first-level model is updated using the updated training data.
[0174] In this embodiment, when the training data source is selected as the reference pool for model update, the first running data and / or the second running data can be used as the updated training data, and the original first-level model can be trained based on the updated training data.
[0175] For example, a model in the reference pool can be used as a new first-level model.
[0176] In some embodiments, S204 includes:
[0177] A recommended reference model is selected based on the reference pool;
[0178] The model architecture of the first-level model is updated using the reference model.
[0179] In this embodiment, the reference pool of the primary model is typically created based on the operational data of the machine group to which the target machine belongs. In other words, the data in the reference pool originates from the machine group.
[0180] In this embodiment, when a model architecture source is selected as the reference pool for model updates, the model architecture of the selected reference model can be used as the model architecture of the new first-level model of the target machine.
[0181] For example, multiple models in the reference pool can be scored based on their performance. The higher the model performance and the higher the score, the more likely it is to be selected as a recommended reference model.
[0182] Alternatively, models in the reference pool can be scored based on the number of business rules they satisfy; the more business rules satisfied, the higher the score. For example, if model A satisfies 3 business rules and model B satisfies 4 business rules, model B will score higher and can be selected as the reference model.
[0183] Alternatively, operators can select a recommended reference model from the reference pool based on the actual situation.
[0184] In some embodiments, the input of the primary model is process parameters, and the output of the primary model is business rules.
[0185] Among them, process parameters refer to process-related data generated by the machine during the production process, such as process conditions and process formulas.
[0186] In some embodiments, the business rules include at least one of the following categories: yield indicators, stability indicators, production capacity indicators, and cost indicators.
[0187] Among these, business rules refer to the metrics that need to be achieved to realize a specific goal. For example, one business rule for a target machine is that its yield rate must reach 80% or higher.
[0188] For example, yield can refer to the proportion of qualified products to total products when a machine is producing wafers.
[0189] For example, stability indicators can refer to the standard of stability of process parameters and hardware status of machine tools during production.
[0190] For example, a capacity indicator can refer to the effective production capacity standard of a machine within a unit of time. For instance, a capacity indicator can be determined by the number of qualified products produced by a machine within a set time period.
[0191] For example, cost indicators can refer to the control standards for resource consumption and loss during the machine production process.
[0192] It should be understood that the business rules output by the model are operating standards adapted to individual machines and tailored to the actual equipment conditions. For example, machine A may have a higher temperature threshold, while machine B of the same type may have a relatively lower temperature threshold due to component aging.
[0193] In some embodiments, the update objective in S204 is to require the output of the first-level model to satisfy the business rules.
[0194] In some embodiments, adjusting the process parameters of the machine tool according to the updated primary model can enable the machine tool to meet business rules. For example, the yield of the machine tool can reach the yield index that meets the business rules.
[0195] In some embodiments, before applying the updated primary model to the target machine, the method further includes the step of:
[0196] Obtain process parameters based on the updated primary model;
[0197] Test production was carried out using sample wafers under the stated process parameters;
[0198] The updated primary model is verified based on the test production results, and after successful verification, the updated primary model is allowed to be applied to the target machine.
[0199] In some embodiments, before formally applying the updated primary model to the machine, a small-sample test production can be conducted to determine whether the updated primary model can meet the business rules. This invention creates a rich reference pool containing multiple types of training data and multiple types of reference models. During actual operation, the reference direction of the model can be updated according to actual operational needs.
[0200] In other words, for a certain type of change, this invention chooses to use historical operating data of the machine to calibrate or update the model, thereby fine-tuning the model at a lower cost to quickly adapt to minor changes.
[0201] For cases where the entire system changes (such as type II changes), this invention selects the model architecture source as the source range of the reference pool, that is, it selects to switch to other models to better adapt to the current scenario, thereby improving the accuracy of model calculation.
[0202] Furthermore, by combining actual needs to recommend different model update strategies, this invention can avoid excessive model updates, thereby reducing the update pressure on the model.
[0203] Please see Figure 2 This invention provides a model update system, comprising:
[0204] A model pool construction module is used to construct a model pool for a group of wafer rigs, the group of wafer rigs including at least two wafer rigs, and the model pool including: (1) at least one primary model, and the primary model is created corresponding to at least one of the wafer rigs; (2) at least one secondary model, and the secondary model is created uniformly corresponding to at least two wafer rigs.
[0205] The scene change recognition module is used to identify the scene change type of the target machine. The scene change type includes at least one of the following: process parameter change, model output change, machine object change, component update change, and model performance change; and the target machine is one of the wafer machines in the machine group.
[0206] The reference pool selection module is used to select a reference pool for model updating according to the scene change type. The reference pool is used to define the source range of reference data for updating the target model. The target model is the first-level model or the second-level model corresponding to the target machine. The source range includes at least one of the following: (A) training data source; (B) model architecture source.
[0207] The training data sources include at least one of the following: (1) the first operating data of the target machine; (2) the second operating data of a similar machine; and the similar machine is the wafer machine that has the same working scenario as the target machine. The model architecture sources include: (1) the first-level model of the similar machine; (2) the second-level model corresponding to the target machine.
[0208] The update module is used to update the target model based on the reference pool.
[0209] In some embodiments, the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements the model update method as described in any embodiment of the present invention.
[0210] In some embodiments, this application also provides a schematic block diagram of the structure of a computer device, please see... Figure 3 Computer programs can be used in situations such as Figure 3 It runs on the computer device shown. Figure 3As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The memory may include non-volatile storage media and internal memory. The non-volatile storage media may store an operating system and computer programs. The computer programs include program instructions that, when executed, cause the processor to perform arbitrary methods. The processor provides computational and control capabilities to support the operation of the entire computer device. The internal memory provides an environment for the execution of the computer programs in the non-volatile storage media; when executed by the processor, these programs cause the processor to perform arbitrary methods. The network interface is used for network communication, such as sending assigned tasks. Those skilled in the art will understand that... Figure 3 The structures shown are merely block diagrams of a portion of the structure related to the present application and do not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than shown in the figures, or combine certain components, or have different component arrangements. It should be understood that the processor may be a Central Processing Unit (CPU), but it can also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.
[0211] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.
[0212] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a computer terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.
[0213] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A model update method, characterized in that, include: S201, construct a model pool for the machine group, the machine group includes: at least two wafer machines, and the model pool includes: (1) at least one primary model, and the primary model is created corresponding to at least one of the wafer machines; (2) at least one secondary model, and the secondary model is created uniformly corresponding to at least two wafer machines; S202, identify the scene change type of the target machine, wherein the scene change type includes at least one of the following: process parameter change, model output change, machine object change, component update change, model performance change; and the target machine is one of the wafer machines in the machine group; S203, Select a reference pool for model update according to the scenario change type, wherein the reference pool is used to define the source range of reference data for updating the target model, the target model is the first-level model or the second-level model corresponding to the target machine, and the source range includes at least one of the following: (A) training data source; (B) model architecture source; The training data sources include at least one of the following: (1) the first operating data of the target machine; (2) the second operating data of a similar machine; and the similar machine is the wafer machine that has the same working scenario as the target machine; the model architecture sources include: (1) the first-level model of the similar machine; (2) the second-level model corresponding to the target machine; S204, updating the target model according to the reference pool; including: A reference model is selected from the reference pool. When the reference model includes a first-level model of the similar machine, the model architecture of the target model is updated using the reference model, including the following steps: Identify the historical events experienced by the similar machines during a set time period, and the corresponding times of occurrence; When the historical event is of the same type as the scene change, the historical event is identified as a matching event; otherwise, the historical event is identified as a non-matching event. The reference weights of the reference model are determined based on the matching event and the occurrence time. The target model is updated based on the reference weights and the reference model; Before updating the target model based on the reference weights and the reference model, the method further includes the following steps: Identify non-matching events that occur after a match has occurred; Compare the correlation between the non-matching events and the matching events; If the correlation degree is greater than the preset correlation degree, the reference weight is maintained or reduced. If the correlation degree is less than or equal to the correlation degree, the corresponding reference model will be removed from the reference pool.
2. The method according to claim 1, characterized in that, S203 includes the following steps: When the scenario change type is a type of change, the source range of the reference pool includes: the training data source; and the type of change includes: process parameter changes and model performance changes; When the scenario change type is a type II change, the source range of the reference pool includes: the model architecture source; and the type II change includes: model output change, machine object change, and component update change.
3. The method according to claim 2, characterized in that, When the source range of the reference pool includes: model architecture source, correspondingly, S204 includes the following steps: Obtain the model score of the model, wherein the model includes: a first-level model of the similar machine and / or a second-level model corresponding to the target machine; The model whose score is greater than the set score is used as the reference model; The model architecture of the target model is updated using the reference model.
4. The method according to claim 1, characterized in that, The later the occurrence time, the higher the reference weight.
5. The method according to claim 1, characterized in that, S204 includes: Updated training data is obtained from the reference pool; The first-level model is updated using the updated training data.
6. The method according to claim 3, characterized in that, The input of the primary model is process parameters, and the output of the primary model is business rules. The business rules include at least one of the following categories: yield indicators, stability indicators, capacity indicators, and cost indicators. And / or, the update objective in S204 is to require the output of the first-level model to meet the business rules; And / or, Before applying the updated primary model to the target machine, the following steps are also included: Obtain process parameters based on the updated primary model; Test production was carried out using sample wafers under the stated process parameters; The updated primary model is verified based on the test production results, and after successful verification, the updated primary model is allowed to be applied to the target machine.
7. A model update system, characterized in that, include: A model pool construction module is used to construct a model pool for a group of wafer rigs, the group of wafer rigs including at least two wafer rigs, and the model pool including: (1) at least one primary model, and the primary model is created corresponding to at least one of the wafer rigs; (2) at least one secondary model, and the secondary model is created uniformly corresponding to at least two wafer rigs. The scene change recognition module is used to identify the scene change type of the target machine. The scene change type includes at least one of the following: process parameter change, model output change, machine object change, component update change, and model performance change; and the target machine is one of the wafer machines in the machine group. The reference pool selection module is used to select a reference pool for model updating according to the scene change type. The reference pool is used to define the source range of reference data for updating the target model. The target model is the first-level model or the second-level model corresponding to the target machine. The source range includes at least one of the following: (A) training data source; (B) model architecture source. The training data sources include at least one of the following: (1) the first operating data of the target machine; (2) the second operating data of a similar machine; and the similar machine is the wafer machine that has the same working scenario as the target machine. The model architecture sources include: (1) the first-level model of the similar machine; (2) the second-level model corresponding to the target machine. The update module is used to update the target model according to the reference pool; the update module is also used to perform: A reference model is selected from the reference pool. When the reference model includes a first-level model of the similar machine, the model architecture of the target model is updated using the reference model. The following steps are performed: Identify the historical events experienced by the similar machines during a set time period, and the corresponding times of occurrence; When the historical event is of the same type as the scene change, the historical event is identified as a matching event; otherwise, the historical event is identified as a non-matching event. The reference weights of the reference model are determined based on the matching event and the occurrence time. The target model is updated based on the reference weights and the reference model; Before updating the target model based on the reference weights and the reference model, the following steps are also performed: Identify non-matching events that occur after a match has occurred; Compare the correlation between the non-matching events and the matching events; If the correlation degree is greater than the preset correlation degree, the reference weight is maintained or reduced. If the correlation degree is less than or equal to the correlation degree, the corresponding reference model will be removed from the reference pool.
8. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the model update method as described in any one of claims 1 to 6.