Model update resource scheduling method, system and product

Abstract

The invention relates to the technical field of model training, in particular to a model updating resource scheduling method and system and a product, and the method comprises the steps: building a model pool for a machine group which comprises at least two wafer machines; obtaining a first model score of the primary model and a second model score of the secondary model; the difference between the first model score and the second model score is calculated, and the difference is defined by subtracting the second model score from the first model score; allocating updating resources for the model according to the difference, wherein the updating resources are used for defining updating cost invested by the model; the method comprises the following steps: when the difference is greater than a preset first difference value, allocating a first update resource to a primary model; when the difference is smaller than a preset second difference value, allocating a second update resource to the primary model; when the difference is larger than or equal to the second difference value and smaller than or equal to the first difference value, current update resources of the first-level model are maintained. According to the invention, the model training efficiency of the semiconductor machine can be greatly improved.

Description

Technical Field

[0001] This invention relates to the field of model training technology, specifically to a method, system, and product for scheduling resource updates for models. Background Technology

[0002] Semiconductor manufacturing involves multiple processes, each of which can be broken down into several steps and stages, completed by different machines. Each machine is equipped with multiple sensors, and to achieve high-precision monitoring, high-frequency, wide-range data acquisition is required, resulting in a single machine generating massive amounts of data. Traditional technologies have proposed several machine model training methods to address the characteristics of large data volumes and multiple acquisition dimensions in semiconductor manufacturing.

[0003] For example, patent application CN109765863A proposes a cloud platform-based edge computing method for device parameters, including the following steps: Step 1, collecting device information from multiple data sources; Step 2, caching, batching, and multi-consumer processing of the device information collected in Step 1; Step 3, preprocessing the data obtained in Step 2; Step 4, uploading the preprocessed data from Step 3 to a public cloud, and using the public cloud algorithm platform for model training and optimization; Step 5, moving the trained and optimized model to the local machine and compiling the corresponding project files; Step 6, deploying the model program to a public machine at the production edge, and using the model for prediction, alarm, control, and parameter suggestions.

[0004] For example, patent application CN115310390A proposes a model training method, process diagnosis method, apparatus, and storage medium. It trains a machine learning model to learn the mapping relationship between device electrical parameters and corresponding device characteristic parameters, obtaining a trained device parameter prediction model that can be used to diagnose semiconductor process problems. The device parameter prediction model obtains predicted device characteristic parameters based on the input device electrical parameters. The deviation between the predicted device characteristic parameters and standard device characteristic parameters is compared to determine if there is an anomaly, identifying the process problem corresponding to the abnormal predicted device characteristic parameters. This application innovatively utilizes a machine learning model to accurately predict device characteristic parameters and, based on its anomaly analysis of process problem judgments, quickly locates the problematic process step and corresponding equipment, greatly improving diagnostic efficiency and shortening the diagnostic time. Furthermore, it accurately obtains device characteristic parameters based on the trained device parameter prediction model with high accuracy.

[0005] For example, patent application CN118521855A proposes 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] However, due to the complexity of wafer fabrication processes, AI models still face challenges such as high cost and low accuracy when deployed in practical applications. Summary of the Invention

[0007] The purpose of this invention is to provide a method, system, and product for updating resource scheduling of models, which partially solves or alleviates the above-mentioned shortcomings in the prior art and can greatly improve the model training efficiency of semiconductor equipment.

[0008] To solve the aforementioned technical problems, the present invention specifically adopts the following technical solution: A first aspect of the present invention is to provide a method for updating resource scheduling of a model, comprising the steps of: S301, construct a model pool for a group of equipment, the group of equipment includes: at least two wafer rigs, and the model pool includes: (1) at least one primary model, and the primary model is created corresponding to at least one of the equipment; (2) at least one secondary model, and the secondary model is created uniformly for at least two equipment; wherein, the secondary model is trained or updated by at least two primary models according to the set model weights; S302, obtain the first model score of the first-level model and the second model score of the second-level model; S303, calculate the difference between the first model score and the second model score, wherein the difference is defined as: the first model score minus the second model score; S304, Allocate update resources to the model based on the difference, wherein the update resources are defined as the update cost invested by the model, and the update resources are defined by any one of update frequency, monitoring frequency, and monitoring scale; wherein, S304 includes the following steps: When the difference is greater than a preset first difference value, the first update resource is allocated to the first-level model; When the difference is less than a preset second difference value, a second update resource is allocated to the first-level model; When the difference is greater than or equal to the second difference value and less than or equal to the first difference value, the current update resources of the first-level model are maintained. Wherein, the first updated resource is smaller than the second updated resource.

[0009] In some embodiments, when the difference is greater than a preset first difference value, the method further includes the following step: Calculate the fault data of the wafer machine corresponding to the primary model in the historical period, wherein the fault data includes: the number and / or type of faults; Determine the fault indicators based on the fault data; The update resources or model weights of the first-level model are adjusted according to the fault indicators, including: If the fault indicator is greater than or equal to a set first threshold and less than or equal to a set second threshold, then the current updated resource can be maintained or reduced. When the fault indicator is less than the second threshold, the model weight is reduced; When the fault indicator is greater than the first threshold, the model weight is reduced or the current update resources are increased.

[0010] In some embodiments, correcting the fault indicator includes the following steps: Identify the machine maintenance record of the wafer fabrication equipment, wherein the machine maintenance record includes: maintenance time point; The cycle of the primary model returning to stability after undergoing machine maintenance is identified based on the machine maintenance records. If the cycle is less than the set first process cycle, the current fault index is maintained; When the cycle is greater than or equal to the first process cycle and less than the set second process cycle, the fault index is increased. When the cycle is longer than the second process cycle, the update association between the first-level model and the second-level model will be removed.

[0011] In some embodiments, the step of setting the first process cycle includes: Identify the remaining service life of the machine; The first process cycle is determined based on the remaining service life, wherein the higher the remaining service life, the shorter the first process cycle.

[0012] In some embodiments, the steps further include: When the score of the second model is lower than the preset first score for more than the preset first time, the update resources of the corresponding first model are increased. And / or, when the score of the first model is greater than the preset second score, the influence weight of the first-level model on the second-level model is increased.

[0013] In some embodiments, the steps further include: If the score of the first model is lower than the preset second score for more than the preset second time, the corresponding first-level model will be marked as an eliminated model.

[0014] In some embodiments, it also includes: The discarded models are updated or replaced using the model pool.

[0015] In some embodiments, updating or replacing the obsolete model using the model pool includes: Select at least two replacement models from the model pool; The at least two replacement models are used for trial runs, and recommended replacement models are selected based on the trial run results to replace the eliminated models.

[0016] A second aspect of the present invention is to provide a model update resource scheduling system, comprising: 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 for at least two wafer rigs; wherein, the secondary model is trained or updated by at least two primary models according to the set model weights; The model score acquisition module is used to acquire the first model score of the first-level model and the second model score of the second-level model. The difference calculation module is used to calculate the difference between the first model score and the second model score, wherein the difference is defined as: the first model score minus the second model score; An update resource allocation module is used to allocate update resources to the model based on the difference, wherein the update resources are defined as the update cost invested by the model; the update resources are defined by any one of update frequency, monitoring frequency, and monitoring scale; wherein, the update resource allocation module is used for: When the difference is greater than a preset first difference value, a first update resource is allocated to the first-level model; when the difference is less than a preset second difference value, a second update resource is allocated to the first-level model; when the difference is greater than or equal to the second difference value and less than or equal to the first difference value, the current update resource of the first-level model is maintained; wherein, the first update resource is less than the second update resource.

[0017] 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 an update resource scheduling method for a model as described in any embodiment of the present invention.

[0018] Beneficial technical effects: This invention proposes a model difference-based resource update constraint scheduling method, which can adaptively allocate and schedule limited model update resources based on the actual operating conditions of the machine, and can better avoid the waste of computing power.

[0019] In other words, for scenarios where model updates are costly and update resources are limited, this invention chooses to adjust the model's iteration update frequency based on the model maturity (or adaptability to the machine) reflected by the difference between the first-level model and the second-level model, thus achieving a good balance between update resource investment and model accuracy.

[0020] It is important to emphasize that this invention does not aim to achieve an absolutely high score or absolute accuracy for a single model. Instead, it aims to ensure that all models meet the usability standard by allocating update resources in a restricted manner, thereby maintaining the effective operation of the model pool and ensuring the feasibility of the dual-path update mechanism of the model pool.

[0021] In this invention, particular care is taken in allocating update resources for the high-resolution model to avoid generating false high scores (such as insufficient fault experience leading to overfitting of the model to certain faults and a lack of generalization ability). Specifically, for the high-resolution first-level model, before reducing the execution of update resources, fault indicators are introduced to carefully verify the reliability of the first-level model from the perspective of the equipment, thereby reducing the risk of error amplification from false high scores on the wafer equipment.

[0022] Specifically, this invention selectively incorporates the distribution of fault types and fault numbers of wafer rigs in the historical process while reducing update resources. This allows for differentiated decision-making on the first-level model at the level of update resources and contribution to the second-level model. This enables careful reduction of update resources without excessive attention, avoiding being misled by falsely high scores from wafer rigs.

[0023] Even in order to maintain the generality of the secondary model, we avoid focusing on all high-scoring models and instead appropriately exclude models that may have falsely high scores. In other words, this embodiment will make appropriate sacrifices during the model borrowing process to avoid learning overfitting information. Attached Figure Description

[0024] 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.

[0025] Figure 1 A flowchart illustrating a model update resource scheduling method provided by the present invention; Figure 2 A schematic diagram of the structure of a model update resource scheduling system provided by the present invention; Figure 3 This is a schematic block diagram of the structure of a computer device provided by the present invention. Detailed Implementation

[0026] 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.

[0027] 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.

[0028] 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.

[0029] 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.

[0030] In this document, "and / or" includes any and all combinations of one or more of the listed related items.

[0031] In this article, "multiple" means two or more, that is, it includes two, three, four, five, etc.

[0032] 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.

[0033] 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.

[0034] Definition of noun: A wafer can refer to a silicon chip used to manufacture silicon semiconductor circuits.

[0035] A wafer fabrication unit (or simply unit) can refer to any one or more processing modules or devices on a wafer fabrication line.

[0036] 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 (or operational data, which is used as training data) is collected and recorded for real-time monitoring, fault diagnosis, yield analysis, and predictive maintenance.

[0037] 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.

[0038] In this paper, a primary model can be a dedicated model created specifically for a single wafer rig within a group of rigs, adapted only to the rig's own operating characteristics (such as process type, operating status, hardware attributes, etc.).

[0039] The 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.

[0040] A group of equipment can refer to a collection that contains at least two wafer fabs.

[0041] For example, multiple machines in a production line can be grouped into a machine group.

[0042] For example, a machine group refers to a collection of machines that have the same or similar attributes (such as machine type, hardware configuration or process formula, etc.).

[0043] For example, multiple etching machines used to perform the same process can be grouped into a single machine group.

[0044] 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.

[0045] 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.

[0046] 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.

[0047] 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.

[0048] In other words, AI models face significant technical challenges in both the application deployment and maintenance phases.

[0049] Furthermore, since wafer fabrication involves ultra-precise processing techniques, even with similar equipment or processes, there may be differences in the adaptability of AI models, and these differences may have a significant impact on production line progress.

[0050] Therefore, even a mature AI model system still faces significant adaptation challenges when deployed on a new production line, which places extremely high cost requirements on the deployment and application of AI models.

[0051] Please see Figure 1 This invention provides a method for updating resource scheduling in a model, comprising the following steps: S301, construct a model pool for a group of equipment, the group of equipment includes: at least two wafer equipment, and the model pool includes: (1) at least one primary model, and the primary model is created corresponding to at least one of the equipment; (2) at least one secondary model, and the secondary model is created uniformly for at least two equipment. S302, obtain the first model score of the first-level model and the second model score of the second-level model; S303, calculate the difference between the first model score and the second model score, wherein the difference is defined as: the first model score minus the second model score; S304, Allocate update resources to the model based on the difference, wherein the update resources are defined as the update cost invested by the model; wherein, S304 includes the following steps: When the difference is greater than a preset first difference value, the first update resource is allocated to the first-level model; Alternatively, when the difference is less than a preset second difference value, a second update resource is allocated to the first-level model; Alternatively, when the difference is greater than or equal to the second difference value and less than or equal to the first difference value, the current update resources of the first-level model are maintained; Wherein, the first updated resource is smaller than the second updated resource.

[0052] In some embodiments, the update resource is defined by any one of update frequency, monitoring frequency, and monitoring scale.

[0053] The monitoring scale refers to the level of detail in the model monitoring. For example, the larger the scale, the more indicators can be detected, and the higher the corresponding cost.

[0054] In some embodiments, the unified creation of the secondary model corresponding to at least two machines can be achieved by creating a secondary model using group machine data from at least two machines (which can be considered as a group of machines), wherein the specific model architecture / model type can be selected by the user. Preferably, the model architecture / model type is the same as that of the primary model.

[0055] Furthermore, in some embodiments, a fusion algorithm can be used to fuse two or more first-level models to form a second-level model.

[0056] For example, in some embodiments, engineers may be allowed to select different fusion methods from recommended fusion algorithms, such as weighted fusion, teaching / meta-learning, random forest, etc.

[0057] This invention provides a dual-path model interaction update scheme. This dual-path interaction update can meet the current AI model adaptation requirements of the production line while incorporating the key commonalities of the machines, so as to simultaneously train a set of alternative schemes that can be quickly called upon.

[0058] The dual-route model interaction update scheme includes: Set up a personalized model for each machine (equivalent to a first-level model) and a general model for a group of machines (equivalent to a second-level model).

[0059] In actual operation, both personalized and general models can be updated at a certain frequency.

[0060] For example, when the personalized model triggers the update mechanism, it can trigger the general model to update. At this time, the general model will be updated by merging the updated personalized model.

[0061] For example, when the general model has a high model score, but a certain personalized model has a low model score, the general model can be used to merge and update the personalized model.

[0062] Therefore, this embodiment is equivalent to providing a dual-path model interaction update method, in which the personalized model and the general model can have their own update trigger conditions (such as time period) and will affect each other during the update process.

[0063] In particular, it should be emphasized that this application adopts a scheme for coordinating and allocating update resources for multiple models. That is, this application does not aim to pursue the absolute accuracy of a single model, but rather to improve the overall availability of the model by limiting update and maintenance resources.

[0064] Specifically, in this embodiment, model update resources are allocated based on model scores, and considering the complexity of the wafer fabrication equipment, a restrictive resource update mechanism to prevent false high scores is adopted as follows: When the difference is greater than a preset first difference value, the method further includes: Step 1) Calculate the fault data of the wafer machine corresponding to the primary model in the historical period. The fault data includes: the number and / or type of faults. Step 2) Determine the fault indicators based on the fault data; For example, in some embodiments, the more types of faults there are, the greater the fault index.

[0065] For example, in some embodiments, the greater the number of faults, the greater the fault index.

[0066] For example, in some embodiments, the fault index can also be calculated by weighting the fault type and the number of faults.

[0067] Step 3) Adjust the update resources or model weights of the first-level model according to the fault indicators, including: (1) When the fault index is greater than or equal to the set first threshold and less than or equal to the set second threshold, it is permissible to maintain or reduce the current update resources; In this embodiment, the update resource can be either the update frequency or the monitoring attention level (which defines the human or computing power cost invested in monitoring). For example, the larger the update resource, the greater the monitoring attention, which may result in a higher frequency or more detailed performance monitoring of the model.

[0068] In this embodiment, when the fault index is in the middle range (such as when the fault types are relatively limited), it can be preliminarily regarded as the machine being in a relatively reasonable operating condition (such as the machine error being relatively controllable and no additional maintenance or replacement is required in the short term). At this time, when the model score of the first-level model is high, it is allowed to maintain a low update frequency.

[0069] Preferably, in this embodiment, it is also permissible to increase the model weights of the first-level model.

[0070] (2) When the fault index is less than the second threshold, the model weight is reduced; In this embodiment, when the fault index is low, it may indicate that the machine tool's operating scenario is relatively simple, and it has not experienced any faults or the fault types are relatively simple. In this case, it is permissible to use the machine tool to improve its processing quality by leveraging its compatibility with the machine tool. However, it is necessary to limit its influence on the secondary model, i.e., reduce the model weight.

[0071] (3) When the fault index is greater than the first threshold, the model weight is reduced or the current update resources are increased.

[0072] In this embodiment, when the fault index is large, it indicates that the machine's operating scenario is relatively complex, such as the machine's abnormality rate or failure rate being high. In this case, it is necessary to increase the update resources to monitor the machine in a relatively timely manner.

[0073] Meanwhile, in this embodiment, in order to maintain the generality of the second-level model, its absorption and learning of some high-scoring first-level models will be appropriately abandoned. Therefore, it is necessary to reduce the model weights of the first-level model.

[0074] Among them, model weights are used to define the influence weight of a first-level model in the fusion process when multiple first-level models are fused to obtain a first-level model.

[0075] Clearly, this invention does not focus on pursuing absolute high scores / absolute accuracy for a single model. Instead, it aims to ensure that all models can reach a usable standard by restricting the allocation of update resources, thereby maintaining the effective operation of the model pool and ensuring the feasibility of the dual-path mechanism of the model pool.

[0076] In this embodiment, special care is taken in the allocation of update resources for high-scoring models to avoid false high scores (such as insufficient fault experience, overfitting of the model to some faults, and lack of generalization ability).

[0077] Even in order to maintain the generality of the secondary model, we avoid focusing on all high-scoring models and instead appropriately exclude models that may have falsely high scores. In other words, this embodiment will make appropriate sacrifices during the model borrowing process to avoid overfitting information.

[0078] Among them, false high scores refer to personalized models that perform well on evaluation indicators and have scores significantly higher than general models. However, their high scores do not stem from effective learning of normal machine operating conditions and real fault modes. Instead, they are due to false steady states formed by overfitting to historical noise, single abnormal patterns, or a long period without new samples. In this case, the model's actual generalization ability and fault detection ability are seriously insufficient. If the model update frequency is reduced accordingly, it will be unable to respond to new faults, thereby masking the real machine risks.

[0079] Furthermore, in some embodiments, the method further includes the step of correcting the fault indicator, which includes the steps of: 1. Identify the wafer rig's maintenance record, which includes: maintenance time point; Maintenance can refer to the engineer's correction or replacement of the components / parts in the machine tool.

[0080] 2. Identify the cycle in which the primary model recovers to stability after undergoing machine maintenance based on the maintenance records; During the wafer fabrication process, due to the precision requirements of wafer processing, even slight deviations after maintenance may cause fluctuations in the model's error. When the model's error fluctuation drops to a set threshold, or when the model fluctuation is in a stable range and the model recovers to the original set standard, the model is considered to have recovered to a stable state.

[0081] In this embodiment, the historical period of the model's recovery stability is introduced to indirectly verify the data feasibility of the equipment. Only when the equipment can quickly return to the steady-state range of the process within a preset number of wafers under a specified process formulation and layer is the equipment determined to have high reliability in terms of data quality.

[0082] 3. If the cycle is less than the set first process cycle, the current fault index is maintained; If the model recovers to stability relatively quickly, it indicates that the machine itself is probably in a relatively healthy and reliable state. Even if there is a fault, it is within the acceptable range. Therefore, it is permissible to maintain the fault index as a subsequent judgment criterion, or even allow the fault index to be appropriately reduced.

[0083] 4. When the cycle is greater than or equal to the first process cycle and less than the set second process cycle, the fault index is increased. When the model recovers to stability relatively slowly, but is still within an acceptable range, the failure index can be appropriately increased to allocate subsequent update resources more cautiously.

[0084] 5. When the cycle is greater than the second process cycle, the update association between the first-level model and the second-level model will be removed.

[0085] If the model recovers to stability too slowly, the update correlation between the first-level model and the second-level model will be directly severed. That is, the first-level model is allowed to be updated, but it will not contribute to the update of the second-level model for the time being.

[0086] In this embodiment, the connection between the primary model and the secondary model is severed in order to prevent the primary model from polluting the secondary model.

[0087] Furthermore, in some embodiments, the step of setting the first process cycle includes: Identify the remaining service life of the machine; The first process cycle is determined based on the remaining service life, wherein the higher the remaining service life, the shorter the first process cycle.

[0088] In this embodiment, the threshold selection for the first process cycle will be dynamically adjusted based on the expected lifespan of the machine itself.

[0089] For example, when the remaining service life is long, higher requirements will be placed on the operating conditions of the wafer rig, meaning that the model should also have a shorter recovery time.

[0090] In some embodiments, the models in the model pool can determine whether there are any abnormal conditions based on the machine's operating status and provide recommended adjustment parameters based on the machine's business rules.

[0091] For example, in some embodiments, the input to the model is process parameters, and the output of the primary model is business rules. The process parameters include process steps, process formulas, process cycles, etc.; the business rules include at least one of the following categories: yield indicators, stability indicators, capacity indicators, and cost indicators.

[0092] For example, in some embodiments, the model is set to meet specific business rules. When a change in a parameter of the model (such as ambient temperature) is detected, the model can recommend other recommended parameters (such as humidity, etching gas parameters, etc.) to adapt to the change in ambient temperature.

[0093] 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.

[0094] For example, yield can refer to the proportion of qualified products to total products when a machine is producing wafers.

[0095] For example, stability indicators can refer to the standard of stability of process parameters and hardware status of machine tools during production.

[0096] For example, a capacity indicator can refer to the standard of effective production capacity of a machine per unit time.

[0097] For example, cost indicators can refer to the control standards for resource consumption and loss during the machine production process.

[0098] It should be understood that the business rules output by the model are operating standards adapted to individual machines and tailored to actual equipment conditions. Typically, these rules can be manually set by engineers based on actual production tasks.

[0099] Furthermore, within the same group of machines, the machines typically perform similar or identical tasks and have similar process parameters. However, due to variations in performance details between different machines, such as differences in usage time and production frequency, minor adjustments to the process parameters are necessary. Therefore, different primary models can often have similar model architectures, but differ in the detailed parameters of the models.

[0100] For example, machine A may have a higher temperature threshold, while machine B with the same process parameters may have a relatively lower temperature threshold due to component aging.

[0101] In some embodiments, the secondary model can be trained from multiple primary models. For example, the secondary model can be obtained by fusing and training multiple primary models.

[0102] In some embodiments, the first model score can be obtained by calculating the error (e.g., root mean square error RMSE, mean absolute error MAE) between the predicted value and the actual production value of the first-level model on a specific machine. The smaller the error, the higher the model score.

[0103] For example, the predicted value can be the yield rate obtained from the forecast, while the actual production value can be the yield rate obtained from the actual inspection. If the actual production value is lower than the predicted value, it means that the model score is low.

[0104] In some embodiments, the second model score can be obtained by calculating the average error or comprehensive error of the secondary model across all associated machines. The smaller the error, the higher the model score.

[0105] Understandably, the model score is primarily used to evaluate the model's accuracy and reliability. Specifically, the first and second model scores can also be obtained by the operators.

[0106] The applicant noted that in semiconductor manufacturing, model training and updates (especially those involving parameter tuning) consume massive amounts of computing resources and time. To address this, this invention proposes a model difference-based resource-constrained scheduling method for updates. This method adaptively allocates and schedules limited model update resources based on the actual operating conditions of the equipment, effectively avoiding wasted computing power.

[0107] In some embodiments, when the score difference is greater than a preset first difference value, i.e., the first model score is much greater than the second model score, it may indicate that the machine is performing well and the model has been well adapted to the physical characteristics of the machine (such as chamber matching degree and aging degree). In this case, the present invention chooses to allocate less first update resources to the first-level model, and imposes certain restrictions on the update behavior, thereby protecting the better adaptation state that the model has formed, avoiding invalid iterations, and achieving a better balance between update resource investment and model accuracy.

[0108] In some embodiments, when the difference is less than a preset second difference value, it may indicate that the current first-level model is significantly insufficient in adapting to the machine's physical characteristics, chamber state, aging degree, and other personalized features, or that the machine's performance is below the average level of similar machines, and there is significant room for improvement in model accuracy. In this case, the present invention adopts an update resource enhancement scheduling strategy: actively allocating higher-intensity second update resources to the first-level model, such as by increasing the iteration frequency, increasing training computing power, and expanding the sample size, to improve the first-level model's learning and fitting degree to the machine's personalized features, and to drive the model to quickly optimize to the optimal fit state.

[0109] In some embodiments, when the difference is greater than or equal to the second difference value and less than or equal to the first difference value, it may indicate that the current first-level model has a certain degree of adaptability. In this case, the present invention adopts a scheduling strategy that maintains the current update resources, which avoids excessive iteration leading to wasted update resources while ensuring that the model accuracy meets the requirements.

[0110] In other words, for scenarios where model updates are costly and update resources are limited, this invention chooses to adjust the model's iteration update frequency based on the model maturity (or adaptability to the machine) reflected by the difference between the first-level model and the second-level model, thus achieving a good balance between update resource investment and model accuracy.

[0111] In some embodiments, the steps further include: When the score of the second model is lower than the preset first score for more than the preset first time, the update resources of the corresponding first model are increased.

[0112] When the score of the second model corresponding to a certain type of machine consistently falls below a preset threshold (e.g., below 85 points for two consecutive process cycles), it indicates that the overall operating conditions, process characteristics, or environmental conditions of this type of machine have undergone a collective and trend-like shift. The original equilibrium model can no longer effectively represent the common patterns of this type of machine, and relying solely on independent machine updates cannot fundamentally solve the problem of declining overall adaptability. Alternatively, relying on independent machine updates may also lead to the consumption of a large amount of update resources.

[0113] In response, this invention proposes a group-based resource scheduling mechanism for updating resources. For example, by continuously monitoring the secondary model score, when it falls below a threshold for an extended period, a collective high-frequency update of all primary models under that type of machine can be triggered (i.e., the update resources of the primary models can be increased), and newly trained alternative balanced models can be imported from the model pool simultaneously.

[0114] This mechanism facilitates the transition from isolated updates and upgrades of a single machine to overall collaborative optimization of the machine group. It can quickly restore the accuracy of the overall model, improve the stability and consistency of machine group control, and further enhance the adaptability of the model pool in the face of changes in the machine group.

[0115] Simultaneously, when a persistent failure of a secondary model is detected, a collective high-frequency update of all primary models under that type of machine is triggered, and the system is simultaneously switched to a new, pre-trained alternative equilibrium model in the model cache. This unified replacement method integrates the previously scattered and multiple-execution updates of a single machine into a single global collaborative update, significantly reducing redundant computation and communication overhead, thereby saving update resource consumption.

[0116] In some embodiments, the step of improving the update resources of the corresponding first-level model includes: increasing the update frequency of the first-level model.

[0117] Alternatively, the steps to improve the update resources of the corresponding first-level model may include: increasing the training computing power of the first-level model and / or expanding the sample size.

[0118] In some embodiments, the larger the update resource, the higher the update frequency of the model.

[0119] In some embodiments, the greater the update resources, the greater the training computing power and / or sample size of the model.

[0120] In some embodiments, update resources can be used to define the investment cost or attention given to model training. Therefore, improving the update resources for the model is equivalent to increasing its attention, preferably by increasing the update frequency.

[0121] In some embodiments, the steps further include: If the score of the first model is lower than the preset second score for more than the preset second time, the corresponding first-level model will be marked as an eliminated model.

[0122] In some embodiments, when the score of the first model is lower than a preset second score and continues for a preset second duration, it indicates that the first-level model has been unable to effectively adapt to the actual working conditions of the corresponding machine for a long period of time. At this time, even if continuous update resources are invested, it is difficult to restore it to the required accuracy level. If such a failed model is retained and maintained, it will continue to occupy update resources such as computing power, storage, and scheduling, resulting in invalid resource occupation and wasted update costs.

[0123] In some embodiments, it also includes: The discarded models are updated or replaced using the model pool.

[0124] In other words, this invention sets up an elimination mechanism for such primary models that are continuously failing: primary models that fail to meet the standards for a long time are marked as eliminated models, and update resources are stopped from being allocated to them, so as to avoid continuous cost consumption caused by invalid updates.

[0125] Furthermore, by promptly cleaning up these first-level models, limited update resources can be allocated to models that still have optimization potential, or in other words, resources can be tilted towards effective models. This can significantly reduce model update overhead.

[0126] In some embodiments, alternative models with the same process parameters (such as the same process type, the same process stage, etc.) can be selected from the model pool to replace the eliminated model.

[0127] In some embodiments, updating the obsolete model using the model pool may include: initializing parameters based on a second model of the same type of machine in the model pool, and combining the machine's historical operating data and real-time running data to perform rapid incremental updates and lightweight retraining on the updated first-level model.

[0128] It should be understood that this invention also proposes a hierarchical differentiated update strategy. Specifically, when the first-level model consistently scores low, a replacement mechanism is employed; when the second-level model consistently scores low, a global update mechanism is used. This differentiated approach—local updates for individual problems and global updates for overall problems—matches update resource investment with the scope of the problem's impact. This approach maximizes resource utilization and minimizes overall update overhead while ensuring model accuracy.

[0129] In some embodiments, updating or replacing the obsolete model using the model pool includes: Select at least two replacement models from the model pool; The at least two replacement models are used for trial runs, and recommended replacement models are selected based on the trial run results to replace the eliminated models.

[0130] For example, during a trial run using the at least two replacement models, model scores of the at least two replacement models can be obtained (e.g., based on prediction accuracy, fit, etc.), and the replacement model with the highest model score during the trial run can be used as the recommended replacement model.

[0131] In this embodiment, multiple replacement models can be selected for trial operation and the best one can be selected for replacement, so that the model pool can dynamically select the optimal solution according to the real-time operating conditions of different machines, thereby further improving the adaptability and reliability of the model under complex operating conditions.

[0132] In some embodiments, the steps further include: When the score of the first model is greater than the preset second score, the influence weight of the first-level model on the second-level model is increased.

[0133] In some embodiments, data can be extracted from high-scoring first-level models to retrain second-level models.

[0134] For example, when training a secondary model based on model a, the secondary model can be trained directly using the parameters of model a.

[0135] Alternatively, multiple first-level models can be referenced simultaneously, but the first-level model with a high score has a relatively larger influence weight on the second-level model.

[0136] For example, when training a secondary model based on models a and b, if the score of model a is higher than that of model b, the influence weight of the parameters of model a can be increased when training the secondary model.

[0137] In this embodiment, when the score of the first model is greater than a preset second score, it may indicate that the first-level model has a high degree of adaptability to the personalized characteristics, process characteristics, and physical state of the corresponding machine, and the personalized rules it has learned may have high reference value and representativeness. At this time, increasing the influence weight of the first-level model on the second-level model can apply the personalized experience with reference value to the group model, so that the group model can fully absorb and integrate the effective features and rules in the high-performing first-level model during the update and iteration process.

[0138] Please see Figure 2 In some embodiments, the present invention provides a model update resource scheduling system, comprising: A model pool construction module is used to construct a model pool for a group of equipment, the group of equipment 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 equipment; (2) at least one secondary model, and the secondary model is created uniformly for at least two equipments; The model score acquisition module is used to acquire the first model score of the first-level model and the second model score of the second-level model. The difference calculation module is used to calculate the difference between the first model score and the second model score, wherein the difference is defined as: the first model score minus the second model score; An update resource allocation module is used to allocate update resources to the model based on the difference, wherein the update resources are defined as the update cost invested by the model; wherein the update resource allocation module is used for: When the difference is greater than a preset first difference value, a first update resource is allocated to the first-level model; when the difference is less than a preset second difference value, a second update resource is allocated to the first-level model; when the difference is greater than or equal to the second difference value and less than or equal to the first difference value, the current update resource of the first-level model is maintained; wherein, the first update resource is less than the second update resource.

[0139] In some embodiments, the present invention provides a computer program product comprising a computer program that, when executed by a processor, implements an update resource scheduling method for a model as described in any embodiment of the present invention.

[0140] 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 3 As 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 3The 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.

[0141] 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.

[0142] 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.

[0143] 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 method for updating resource scheduling in a model, characterized in that, Including the following steps: S301, construct a model pool for a group of equipment, the group of equipment includes: at least two wafer rigs, and the model pool includes: (1) at least one primary model, and the primary model is created corresponding to at least one of the equipment; (2) at least one secondary model, and the secondary model is created uniformly for at least two equipment; wherein, the secondary model is trained or updated by at least two primary models according to the set model weights; S302, obtain the first model score of the first-level model and the second model score of the second-level model; S303, calculate the difference between the first model score and the second model score, wherein the difference is defined as: the first model score minus the second model score; S304, Allocate update resources to the model based on the difference, wherein the update resources are defined as the update cost invested by the model, and the update resources are defined by any one of update frequency, monitoring frequency, and monitoring scale; wherein, S304 includes the following steps: When the difference is greater than a preset first difference value, the first update resource is allocated to the first-level model; When the difference is less than a preset second difference value, a second update resource is allocated to the first-level model; When the difference is greater than or equal to the second difference value and less than or equal to the first difference value, the current update resources of the first-level model are maintained. Wherein, the first updated resource is smaller than the second updated resource.

2. The method according to claim 1, characterized in that, When the difference is greater than a preset first difference value, the method further includes: Calculate the fault data of the wafer machine corresponding to the primary model in the historical period, wherein the fault data includes: the number and / or type of faults; Determine the fault indicators based on the fault data; The update resources or model weights of the first-level model are adjusted according to the fault indicators, including: If the fault indicator is greater than or equal to a set first threshold and less than or equal to a set second threshold, then the current updated resource can be maintained or reduced. When the fault indicator is less than the second threshold, the model weight is reduced; When the fault indicator is greater than the first threshold, the model weight is reduced or the current update resources are increased.

3. The method according to claim 2, characterized in that, Correcting the fault indicators includes the following steps: Identify the machine maintenance record of the wafer fabrication equipment, wherein the machine maintenance record includes: maintenance time point; The cycle of the primary model returning to stability after undergoing machine maintenance is identified based on the machine maintenance records. If the cycle is less than the set first process cycle, the current fault index is maintained; When the cycle is greater than or equal to the first process cycle and less than the set second process cycle, the fault index is increased. When the cycle is longer than the second process cycle, the update association between the first-level model and the second-level model will be removed.

4. The method according to claim 3, characterized in that, The steps for setting the first process cycle include: Identify the remaining service life of the machine; The first process cycle is determined based on the remaining service life, wherein the higher the remaining service life, the shorter the first process cycle.

5. The method according to claim 1, characterized in that, It also includes the following steps: When the score of the second model is lower than the preset first score for more than the preset first time, the update resources of the corresponding first model are increased. And / or, when the score of the first model is greater than the preset second score, the influence weight of the first-level model on the second-level model is increased.

6. The method according to claim 1, characterized in that, It also includes the following steps: If the score of the first model is lower than the preset second score for more than the preset second time, the corresponding first-level model will be marked as an eliminated model.

7. The method according to claim 6, characterized in that, Also includes: The discarded models are updated or replaced using the model pool.

8. The method according to claim 7, characterized in that, Updating or replacing the obsolete models using the model pool includes: Select at least two replacement models from the model pool; The at least two replacement models are used for trial runs, and recommended replacement models are selected based on the trial run results to replace the eliminated models.

9. A model update resource scheduling 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 for at least two wafer rigs; wherein, the secondary model is trained or updated by at least two primary models according to the set model weights; The model score acquisition module is used to acquire the first model score of the first-level model and the second model score of the second-level model. The difference calculation module is used to calculate the difference between the first model score and the second model score, wherein the difference is defined as: the first model score minus the second model score; An update resource allocation module is used to allocate update resources to the model based on the difference, wherein the update resources are defined as the update cost invested by the model; the update resources are defined by any one of update frequency, monitoring frequency, and monitoring scale; wherein, the update resource allocation module is used for: When the difference is greater than a preset first difference value, a first update resource is allocated to the first-level model; when the difference is less than a preset second difference value, a second update resource is allocated to the first-level model; when the difference is greater than or equal to the second difference value and less than or equal to the first difference value, the current update resource of the first-level model is maintained; wherein, the first update resource is less than the second update resource.

10. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the update resource scheduling method for the model as described in any one of claims 1 to 8.

Images