A semiconductor PVD thin film thickness prediction and closed-loop correction method and system

Abstract

A semiconductor PVD film thickness prediction and closed-loop correction method and system. The method comprises: constructing a structured view and a text view; training a text view teacher model, and obtaining a lightweight student prediction model through double-view knowledge distillation, so that it outputs an initial film thickness prediction result based on only structured process data in the online deployment stage. A human-machine collaborative process knowledge graph is constructed, a multi-source evidence score is calculated based on range constraints, similar working condition neighbors, measurement point consistency, text semantic constraints and engineer feedback, and a candidate correction amount is generated under the constraint of the maximum correction step; whether to execute the candidate correction is determined through hard gating and / or preference gating, while recording the trigger source, evidence edge, correction amplitude, gating result and execution state to form a traceable evidence chain. The disclosure forms a closed loop of prediction, process constraint verification, candidate correction, gating decision, evidence tracing and feedback learning, and is suitable for PVD film thickness prediction and cross-working condition migration.

Description

Technical Field

[0001] This disclosure relates to the field of semiconductor manufacturing process quality prediction and intelligent control technology, specifically to a semiconductor PVD thin film thickness prediction and closed-loop correction method and system, applicable to film thickness prediction, process constraint correction, engineer feedback learning, and cross-condition migration prediction in semiconductor thin film deposition processes such as AlCu and WTi. Background Technology

[0002] In semiconductor manufacturing, Physical Vapor Deposition (PVD) is a crucial process for forming metal layers, barrier layers, and functional thin films. Film thickness directly affects subsequent etching, resistivity, electromigration reliability, and device yield. PVD film thickness is influenced by a multitude of factors, including process power, chamber pressure, gas flow rate, temperature, deposition time, target condition, equipment condition, material type, and batch drift, exhibiting nonlinearity, coupling, and time-varying characteristics.

[0003] Existing methods for controlling film thickness mainly rely on statistical process control, manual empirical rules, or ordinary machine learning regression models. Statistical process control struggles to handle complex nonlinear relationships; manual empirical rules depend on engineers' experience and have limited generalization ability; while ordinary machine learning models can predict film thickness based on historical data, they typically output a single predicted value, lacking consideration for process feasibility, interpretability of prediction results, and the ability to reuse human feedback.

[0004] In practical engineering applications, existing methods have the following main shortcomings: First, the prediction results lack process constraints, and the model output may exceed the reasonable range of history or be inconsistent with similar working condition samples and multiple film thickness measurement points; Second, the prediction process lacks audit records, making it difficult for engineers to determine whether the predicted value has been corrected, why it was corrected, and based on which historical samples or process rules; Third, manual review experience is difficult to reuse, and engineers' acceptance or rejection of candidate correction results is often only used for a single judgment and is not transformed into subsequent automation strategies; Fourth, the stability of cross-material, cross-chamber, or cross-working condition migration is insufficient. For example, when migrating from AlCu process to WTi process, the input feature distribution and film thickness target distribution may change, and directly using the original model can easily lead to an increase in prediction error. Summary of the Invention

[0005] To at least solve one of the above-mentioned technical problems, the present invention provides the following technical solution.

[0006] A method for predicting and correcting the thickness of semiconductor PVD thin films, comprising the following steps:

[0007] Step S1: Obtain semiconductor PVD process input data and film thickness measurement data to form a film thickness prediction sample set;

[0008] Step S2: Construct a structured view based on the structured process data in the PVD process input data, construct a text view based on the text process data in the PVD process input data, and establish a sample correspondence between the structured view, the text view, and the target film thickness dimension.

[0009] Step S3: Train the text view teacher model and generate dual-view distilled soft labels based on the prediction output of the text view teacher model;

[0010] Step S4: Train the student prediction model using the real film thickness label and the dual-view distillation soft label, so that the student prediction model outputs the initial film thickness prediction result based on the structured process data when deployed;

[0011] Step S5: Construct a human-machine collaborative process knowledge graph, which includes process feature nodes, film thickness target nodes, evidence record nodes, and engineer feedback nodes, as well as range constraint edges, similar working condition nearest neighbor edges, measurement point consistency edges, text semantic constraint edges, and engineer feedback edges.

[0012] Step S6: Based on the human-machine collaborative process knowledge graph, generate a multi-source evidence score for the initial film thickness prediction result, and generate a candidate correction amount constrained by the maximum correction step size according to the multi-source evidence score;

[0013] Step S7: Input the candidate correction amount into the hard-gated model and / or the preference-gated model, determine whether to perform candidate correction, and obtain the corrected film thickness prediction result;

[0014] Step S8: Record the trigger source, evidence edge, correction range, gating result and execution status corresponding to the candidate correction to form a traceable evidence chain;

[0015] Step S9: Update the evidence edge weights in the preference gating model and the human-machine collaborative process knowledge graph based on the engineer's acceptance or rejection feedback on the candidate modifications;

[0016] Step S10: Output the final film thickness prediction result, evidence record, and manual review prompt.

[0017] Furthermore, the semiconductor PVD process input data includes one or more of the following: process power, chamber pressure, gas flow rate, temperature, deposition time, target status, chamber status, and equipment status; and the film thickness measurement data includes one or more target dimensions of film thickness.

[0018] Furthermore, the text view teacher model employs a text encoder or language model embedding to extract process semantic constraints from process recipes, alarm logs, maintenance records, and engineer notes; the dual-view distillation soft label is generated based on the prediction results of the text view teacher model on the validation set after confidence calibration.

[0019] Furthermore, the training loss of the student prediction model includes a weighted sum of the following three loss components: the supervision loss between the prediction result of the student prediction model and the true film thickness label; the distillation loss between the prediction result of the student prediction model and the dual-view distillation soft label; and the consistency loss between the structured view representation and the text view representation of the student prediction model.

[0020] Furthermore, the range constraint edges are used to characterize the allowable range constraints for each film thickness target dimension. The allowable range includes a soft range determined based on the film thickness data in the training set and a hard range used to determine whether the initial prediction is abnormal. The similar working condition nearest neighbor edges are used to characterize the nearest neighbor relationship between the sample to be predicted and historical samples similar to the process input. The measurement point consistency edges are used to characterize the correlation constraints between different film thickness target dimensions. The text semantic constraint edges are used to characterize the semantic association between the text process data and the film thickness target. The engineer feedback edges are used to characterize the engineer's acceptance or rejection label for candidate corrections.

[0021] Furthermore, the multi-source evidence score includes the range evidence score, the neighboring evidence score for similar working conditions, the measurement point consistency evidence score, the text constraint evidence score, and the historical engineer feedback evidence score; the multi-source evidence score is the weighted sum of the five types of evidence scores.

[0022] Furthermore, the candidate correction amount includes a fusion of one or more of the following correction components: a range pullback correction component generated based on the range constraint edge, used to pull back the predicted value that exceeds the allowable range to the soft range; a nearest neighbor support correction component generated based on the similar working condition nearest neighbor edge, used to make the predicted value approach the average film thickness of historical samples of similar working conditions; and a measurement point consistency correction component generated based on the measurement point consistency edge, used to make the predicted value of each film thickness target dimension satisfy the correlation constraint between measurement points.

[0023] Furthermore, the candidate correction amount is also subject to the truncation constraint of the maximum correction step size, which is used to limit the upper limit of the absolute value of a single candidate correction amount.

[0024] Furthermore, the hard-gated model determines whether to perform candidate correction based on the candidate correction magnitude, anomaly degree, and number of evidence sources; the preference-gated model takes the correction magnitude, range deviation, nearest neighbor deviation, measurement point consistency deviation, text constraint deviation, number of evidence sources, historical acceptance rate, target dimension number, and working condition identifier corresponding to the candidate correction amount as input features, and outputs the probability that the candidate correction is accepted. If the probability is higher than a preset threshold, the candidate correction is automatically performed; otherwise, it enters manual review.

[0025] Furthermore, the evidence record includes one or more of the following: sample number, film thickness target number, predicted value before correction, predicted value after correction, correction magnitude, trigger source, supporting evidence, gating method, acceptance probability, and execution status.

[0026] Furthermore, the method also includes a cross-operating condition adaptation step: extracting common input features and common film thickness target dimensions of the source operating condition and the target operating condition; calculating the distribution offset of the source operating condition and the target operating condition on the common input features, retaining features with offsets less than a preset threshold, and reducing the weight of features with offsets greater than or equal to the preset threshold; and performing covariance alignment on the features of the source operating condition so that the source operating condition training samples are statistically close to the target operating condition samples.

[0027] A semiconductor PVD thin film thickness prediction and closed-loop correction system includes:

[0028] The data acquisition module is configured to acquire semiconductor PVD process input data and film thickness measurement data;

[0029] The data preprocessing module is configured to perform verification, standardization, sample division and feature alignment on the data, construct a structured view based on the structured process data in the PVD process input data, construct a text view based on the text process data in the PVD process input data, and establish a sample correspondence between the structured view, the text view and the target film thickness dimension.

[0030] The initial prediction module is configured to output an initial film thickness prediction result based on the structured process data using a student prediction model. The student prediction model is trained by a combination of real film thickness labels and dual-view distillation soft labels generated by a text-view teacher model.

[0031] The process knowledge construction module is configured to construct a human-machine collaborative process knowledge graph, which includes process feature nodes, film thickness target nodes, evidence record nodes and engineer feedback nodes, as well as range constraint edges, similar working condition nearest neighbor edges, measurement point consistency edges, text semantic constraint edges and engineer feedback edges.

[0032] The candidate correction module is configured to generate a multi-source evidence score based on the human-machine collaborative process knowledge graph of the initial film thickness prediction result, and generate a candidate correction amount constrained by the maximum correction step size based on the multi-source evidence score.

[0033] The gating execution module is configured to input the candidate correction amount into the hard gating model and / or the preference gating model, determine whether to perform the candidate correction, and obtain the corrected film thickness prediction result.

[0034] The evidence recording module is configured to record the trigger source, evidence edge, correction range, gating result and execution status corresponding to the candidate correction, forming a traceable evidence chain;

[0035] The engineer feedback module is configured to receive feedback from engineers on whether they accept or reject candidate modifications, and update the evidence edge weights in the preference gating model and the human-machine collaborative process knowledge graph.

[0036] The cross-condition adaptation module is configured to perform common feature alignment, feature offset evaluation, feature filtering, feature weight reduction, and distribution alignment in cross-material or cross-condition migration scenarios.

[0037] The results output module is configured to output the final film thickness prediction results, evidence records, and manual review prompts.

[0038] Furthermore, the result output module is also used to store the initial film thickness prediction result, the corrected film thickness prediction result, the candidate correction record, the executed correction record, and the engineer feedback result into the historical database.

[0039] A semiconductor PVD thin film thickness prediction and closed-loop correction system includes: a processor; a memory coupled to the processor and storing computer program instructions; wherein, when the processor executes the instructions in the memory, it implements the method described in any of the preceding claims.

[0040] A computer-readable storage medium having stored thereon computer program instructions, characterized in that the computer program instructions, when executed by a processor, implement the method described in any of the preceding claims.

[0041] The beneficial effects of this disclosure are as follows:

[0042] (1) Transforming the pure data-driven film thickness prediction results into correctable, auditable, and iterative closed-loop prediction results improves the engineering credibility of the prediction results.

[0043] (2) Introducing film thickness range constraints, similar working condition nearest neighbor constraints and measurement point consistency constraints can reduce prediction results that exceed the reasonable range of the process or are inconsistent with similar samples.

[0044] (3) Generate evidence records for each correction to facilitate engineers to trace the formation process of the prediction results and meet the auditability requirements of the semiconductor manufacturing process.

[0045] (4) Use engineer feedback to train the preference gating model so that the experience of manual review can be continuously reused by the system and gradually increase the proportion of automatic execution.

[0046] (5) By using a cross-condition adaptation mechanism, the model’s ability to migrate between different PVD conditions such as AlCu and WTi has been improved.

[0047] (6) The teacher-student prediction structure is adopted, which takes into account both the utilization of text view information and the low latency requirements of online deployment. Attached Figure Description

[0048] Figure 1 This is a flowchart of the overall process for predicting and correcting the thickness of semiconductor PVD thin films disclosed herein.

[0049] Figure 2 This is a structural diagram of the semiconductor PVD thin film thickness prediction and closed-loop correction system disclosed herein;

[0050] Figure 3 This is a flowchart of the training process for the teacher prediction model and the student prediction model.

[0051] Figure 4 This document provides a flowchart for generating the process knowledge model and candidate corrections for the disclosed process.

[0052] Figure 5 This is a flowchart for engineer feedback and learning gating.

[0053] Figure 6 This is a schematic diagram of the system deployment and software interface of this disclosure. Detailed Implementation

[0054] The following is in conjunction with the appendix Figures 1 to 6 Examples and illustrations are provided to further illustrate this disclosure in detail. It is understood that the specific examples described herein are for illustrative purposes only and are not intended to limit the scope of this disclosure. It should also be noted that, for ease of description, only the parts relevant to this disclosure are shown in the accompanying drawings.

[0055] It should be noted that, where there is no conflict, the embodiments and features described in this disclosure can be combined with each other. The technical solutions of this disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0056] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this disclosure, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0057] Unless otherwise stated, the exemplary implementations / embodiments shown are to be understood as providing exemplary features of various details that provide ways in which the technical concepts of this disclosure can be implemented in practice. Therefore, unless otherwise stated, the features of various implementations / embodiments may be additionally combined, separated, interchanged and / or rearranged without departing from the technical concepts of this disclosure.

[0058] It should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a physical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this disclosure based on the specific circumstances.

[0059] The use of crosshairs and / or shading in the accompanying drawings is generally used to clarify the boundaries between adjacent components. Thus, unless otherwise stated, the presence or absence of crosshairs or shading does not convey or indicate any preference or requirement for the specific material, material properties, dimensions, proportions, commonalities between the illustrated components, or any other characteristics, properties, etc., of the components. Furthermore, in the accompanying drawings, the dimensions and relative dimensions of components may be exaggerated for clarity and / or descriptive purposes. When exemplary embodiments can be implemented differently, a specific process / method sequence may be performed in a different order than that described. For example, two consecutively described processes / methods may be performed substantially simultaneously or in the reverse order of their description.

[0060] When a component is referred to as being on or above another component, connected to or combined with another component, the component may be directly on, directly connected to or directly combined with the other component, or there may be intermediate components. However, when a component is referred to as being directly on, directly connected to or directly combined with another component, there are no intermediate components.

[0061] For descriptive purposes, this disclosure may use spatially relative terms such as below, under, below, above, above, higher, and sideways (e.g., in a sidewall) to describe the relationship between one component and another component as shown in the accompanying drawings. In addition to the orientations depicted in the drawings, the spatially relative terms are also intended to encompass different orientations of the device during use, operation, and / or manufacture. For example, if the device in the drawings is flipped, a component described as being below or under other components or features would subsequently be positioned above said other components or features. Thus, the exemplary term "below" can encompass both "above" and "below." Furthermore, the device may be otherwise positioned (e.g., rotated 90 degrees or in other orientations), thus interpreting the spatially relative descriptive terms used herein accordingly.

[0062] The terminology used herein is for the purpose of describing particular embodiments and is not intended to be limiting. As used herein, unless the context clearly indicates otherwise, the singular forms "a" and "the" are intended to include the plural forms as well. Furthermore, when the terms "contains" and / or "includes" and variations thereof are used in this specification, it indicates the presence of the stated features, integrals, steps, operations, parts, components, and / or groups thereof, but does not preclude the presence or addition of one or more other features, integrals, steps, operations, parts, components, and / or groups thereof. It should also be noted that, as used herein, the terms "substantially," "approximately," and other similar terms are used as approximate terms rather than as terms of degree, so that they are used to account for inherent deviations in measured, calculated, and / or provided values ​​that will be recognized by those skilled in the art.

[0063] In one embodiment, such as Figure 1 As shown, the semiconductor PVD thin film thickness prediction and closed-loop correction method disclosed herein includes the following steps:

[0064] Step (1): Acquisition of semiconductor thin film deposition process data, including: acquisition of PVD process data and acquisition of film thickness measurement data. Semiconductor thin film deposition process data is acquired from PVD equipment logs, process databases, or measurement databases. The semiconductor thin film deposition process data includes PVD process data X and film thickness measurement data Y. PVD process data X may include sensor signals, process power, pressure, temperature, gas flow rate, deposition time, chamber status, target status, etc. Film thickness measurement data Y may include multiple film thickness measurement points or multiple film thickness target dimensions. In one embodiment, the data includes AlCu thin film deposition data and WTi thin film deposition data, wherein the AlCu data includes 97 input features, the WTi data includes 108 input features, and the outputs of both include 17 film thickness target dimensions.

[0065] Step (2): Data preprocessing. A structured view is constructed based on the structured process data, and a text view is constructed based on the text process data. A sample correspondence is established between the structured view, the text view, and the target membrane thickness dimension. Subsequently, the input features are standardized; the target dimension for the membrane thickness output is selected; and the data is divided into training, validation, and test sets. For cross-condition tasks, the common input features and common target membrane thickness dimensions of the source and target conditions are extracted.

[0066] Step (3): Establish an initial prediction model for initial film thickness prediction. A teacher prediction model and a student prediction model are used. The teacher prediction model is used for offline training and can employ a more complex neural network structure; the student prediction model is used for online deployment and can employ a lighter network structure. The teacher prediction model learns the mapping relationship between PVD process input and film thickness output based on training samples and outputs the teacher prediction results on the validation set. The student prediction model receives the standardized process input and outputs the initial film thickness prediction results.

[0067] Step (4): Knowledge Distillation Training. A text-view teacher model is constructed during the offline training phase. The text-view teacher model can employ a text encoder, language model embedding, text feature regressor, or text semantic constraint extraction model to extract semantic information related to membrane thickness targets from the process text. Specifically, the prediction results of the teacher prediction model on the validation set are calibrated and used as soft labels; the system determines the distillation weights based on the teacher prediction errors for each membrane thickness target dimension, with higher weights for target dimensions with smaller errors and lower weights for target dimensions with larger errors. During student prediction model training, both real structured membrane thickness labels and teacher soft labels are used simultaneously to obtain an initial membrane thickness prediction model suitable for deployment.

[0068] Step (5): Constructing the process knowledge model. The process knowledge model includes three types of knowledge: First, film thickness range knowledge. Calculate the soft and hard ranges for each film thickness target dimension based on the film thickness data in the training set. The soft range is used to provide recommended correction targets, and the hard range is used to determine whether the predicted value is significantly abnormal. Second, neighbor knowledge for similar operating conditions. Establish a neighbor model based on the process input features of the training set. For the sample to be predicted, the system searches for historical samples with similar process conditions and calculates the mean film thickness of these historical samples as a local process reference value. Third, measurement point consistency knowledge. Establish a consistency model between measurement points based on the correlation between each film thickness target dimension in the training set. If the predicted value of a certain target dimension deviates too much from the reference value inferred by other related measurement points, it is considered that there is a consistency anomaly.

[0069] Step (6): Generate candidate corrections. After the student prediction model outputs the initial film thickness prediction result, the system uses the process knowledge model to determine whether correction is needed. When the initial predicted value exceeds the hard range of film thickness, a range pullback correction is generated to pull the predicted value back to the soft range of film thickness. When the initial predicted value deviates too much from the mean film thickness of neighboring films under similar operating conditions, a neighbor support correction is generated. When the initial predicted value deviates too much from the reference value given by the measurement point consistency model, a measurement point consistency correction is generated. When multiple corrections are triggered simultaneously in the same film thickness target dimension, the system comprehensively obtains candidate correction values ​​based on the severity and strength of evidence of each correction source.

[0070] Step (7): Gating Execution. Candidate corrections are not executed directly, but are first subject to gating. The gating includes hard gating and learning gating. Hard gating judges the candidate correction based on its deviation, correction magnitude, and evidence strength. If the candidate correction meets the preset conditions, it is executed automatically; otherwise, the candidate correction is only recorded without changing the prediction result. Learning gating is trained based on feedback from engineers. The system inputs the candidate correction features into the preference gating model and outputs the probability that the candidate correction is accepted by the engineer. If the probability is higher than the threshold, it is executed automatically; otherwise, it enters manual review.

[0071] Step (8): Evidence Recording. The system generates an evidence record for each candidate correction. The evidence record includes the sample number, film thickness target number, predicted value before correction, predicted value after correction, correction magnitude, whether range correction is triggered, whether nearest neighbor correction is triggered, whether consistency correction is triggered, nearest neighbor reference value, consistency reference value, gating method, acceptance probability, and execution status. This evidence record can be exported as a table or displayed in the software interface for engineers to review.

[0072] Step (9): Engineers provide feedback to update the process knowledge model. Engineers assign acceptance or rejection labels to candidate corrections. The system trains a preference gating model using candidate correction features and engineer labels as training data. Candidate correction features include correction magnitude, range bias, nearest neighbor bias, consistency bias, number of supporting sources, target dimension number, etc. After training, the preference gating model is used for automatic execution judgment of subsequent candidate corrections, in order to serve as the gating decision in the gating execution of step (7) mentioned above. This transforms the engineer's review experience into a reusable automated strategy.

[0073] Step (10): Cross-condition adaptation. For migration tasks between AlCu to WTi, WTi to AlCu, or between different chambers and different formulations, the system first extracts the common input features and common output targets of the source and target conditions. Then, based on the differences in the mean and standard deviation of each common input feature in the source and target conditions, the distribution offset is calculated, and the 85th percentile of the distribution offset of all common input features is used as the offset judgment threshold. Features with distribution offsets less than or equal to the threshold are retained, while features with distribution offsets greater than the threshold are downweighted. Further, the system can use covariance alignment to perform unsupervised adaptation of the source condition features, making the source condition training samples statistically close to the target condition samples. After adaptation, the prediction model and process knowledge model are trained using the source condition data and applied to the film thickness prediction of the target condition.

[0074] Step (11): Output Results: Output the final film thickness prediction result. The system outputs the final film thickness prediction result, initial prediction result, candidate correction record, executed correction record, manual review prompt, and historical stored results. When the actual film thickness measurement result is available, indicators such as prediction error, out-of-range ratio, and measurement point consistency violation rate can also be calculated for subsequent model updates and process trend analysis.

[0075] For the above embodiments, through the complete process from step (1) to step (11), an end-to-end closed-loop prediction system was established, which includes data acquisition, preprocessing, initial prediction, knowledge distillation, process knowledge modeling, candidate correction generation, gating execution, evidence recording, engineer feedback iteration, and cross-condition adaptation.

[0076] This embodiment integrates dual-view knowledge distillation, human-machine collaborative process knowledge graph (including five types of knowledge: range / nearest neighbor / consistency / text / feedback), bounded candidate correction, dual-gated decision-making, traceable evidence chain, and engineer feedback closed-loop iteration into a unified film thickness prediction and correction framework.

[0077] In one embodiment, such as Figure 2 As shown, this disclosure also provides a semiconductor PVD thin film thickness prediction and closed-loop correction system, including: a data acquisition module, a data preprocessing module, an initial prediction module, a process knowledge construction module, a candidate correction module, a gating execution module, an evidence recording module, an engineer feedback module, a cross-condition adaptation module, and a result output module, wherein...

[0078] The data acquisition module is used to acquire PVD process input data and film thickness measurement data;

[0079] The data preprocessing module is used to perform data verification, standardization, sample partitioning, and feature alignment.

[0080] The initial prediction module is used to train and call the teacher prediction model and the student prediction model, and output the initial film thickness prediction result;

[0081] The process knowledge construction module is used to construct film thickness range knowledge, nearest neighbor knowledge, and measurement point consistency knowledge.

[0082] The candidate correction module is used to generate candidate corrections based on the process knowledge model;

[0083] The gating execution module is used to determine whether to execute the candidate correction;

[0084] The evidence recording module is used to record the triggering source and execution status of candidate corrections;

[0085] The engineer feedback module is used to receive feedback from engineers on whether they accept or reject the feedback and to train a preference gating model.

[0086] The cross-condition adaptation module is used to perform common feature alignment, feature weight reduction, and distribution alignment.

[0087] The result output module is used to output the final film thickness prediction result, correction record, and manual review prompt.

[0088] In another embodiment, such as Figure 3 As shown, the training process of the teacher prediction model and student prediction model disclosed herein includes the following stages. The training data input includes: training process data (containing structured process parameters and unstructured text process data) and membrane thickness real labels (actual membrane thickness measurement results corresponding to the batch). The teacher prediction model adopts a high-complexity model architecture (such as deep neural networks, pre-trained language models with regression heads), extracting process semantic constraints from the text view and learning the nonlinear mapping between parameters and membrane thickness from the structured view. The teacher prediction model outputs the predicted values ​​and confidence scores for each membrane thickness dimension on the validation set. Confidence calibration is performed on the teacher prediction results, and distillation weights are assigned according to the prediction error of each dimension, with higher weights for dimensions where teacher accuracy is higher. The calibrated and weighted teacher prediction results serve as the teacher's soft labels. The student prediction model adopts a lightweight network architecture, using only structured process data as input. The total loss of the student model consists of three weighted components: hard label loss (deviation between student prediction and real label), distillation loss (deviation between student prediction and teacher soft label), and consistency loss (alignment loss between structured representation and text representation). The trained lightweight model can be deployed on industrial control computers, edge devices, or production line systems to achieve real-time initial film thickness prediction.

[0089] For example, such as Figure 4As shown, the process knowledge model and candidate correction generation process disclosed herein include: inputting the initial film thickness prediction result from the student model as the processing object of the correction step. The process knowledge model provides three types of knowledge and generates corresponding corrections. Film thickness range knowledge is obtained based on historical mass production data statistics and is divided into hard range (film thickness physical limit / process window boundary) and soft range (normal production statistical fluctuation range). When the initial prediction exceeds the hard range, the predicted value is pulled back to the soft range. Similar operating condition nearest neighbor knowledge is constructed based on historical samples to build a nearest neighbor retrieval library. Through process feature similarity matching, the historical batches that are closest to the operating conditions of the sample to be predicted are found. When the deviation between the initial prediction and the mean film thickness of similar operating condition samples is too large, the predicted value is adjusted with reference to the historical sample level. Film thickness measurement point consistency knowledge is learned based on historical data to learn the correlation rules of multiple measurement points within the chip. When the predicted value of a certain measurement point deviates from the reasonable range inferred by other measurement points, the predicted value of that measurement point is adjusted. The three types of correction components are weighted and fused, and then truncated by the maximum correction step size. The correction results are then truncated by the upper and lower limits of the process to ensure that the final candidate result always falls within the reasonable range of the process.

[0090] In another alternative, such as Figure 5 As shown, the engineer feedback and learning gating process disclosed herein includes:

[0091] Candidate corrections come from the candidate correction module. The system automatically generates a complete evidence file, which includes information such as the triggering reason, evidence scores in each dimension, correction range, and reference samples.

[0092] Engineers combine their domain experience with evidence and predictions to make a decision to accept or reject the application, and can provide reasons for rejection and suggestions for improvement.

[0093] The preference-gated model takes candidate correction feature vectors (correction magnitude, bias degree, amount of evidence, target dimension, working condition identifier, historical acceptance rate, etc.) as input and trains a binary classification model with engineer feedback labels as supervision, and outputs the probability that the correction is accepted by the engineer.

[0094] When the acceptance probability is higher than the preset threshold, the system will execute automatically without human intervention; when the acceptance probability is lower than the threshold, the initial prediction result will be retained and pushed to the manual review queue for final judgment by engineers.

[0095] Every piece of feedback from the engineers is added to the training set, continuously iterating and optimizing the preference gating model.

[0096] In one embodiment, a PVD film thickness prediction and closed-loop correction system is provided, comprising:

[0097] The data acquisition unit is configured to acquire PVD process input data and film thickness measurement data;

[0098] The dual-view construction unit is configured to construct a structured view based on the structured process data in the PVD process input data, and to construct a text view based on the text process data in the PVD process input data.

[0099] The distillation training unit is configured to generate dual-view distillation soft labels using a text-view teacher model, and to train a student prediction model based on the real film thickness labels and the dual-view distillation soft labels, so that the student prediction model outputs an initial film thickness prediction result.

[0100] The knowledge constraint unit is configured to generate multi-source evidence scores based on process parameter range constraints, historical sample nearest neighbor constraints, and film thickness measurement point consistency constraints.

[0101] The correction generation unit is configured to generate candidate correction amounts constrained by the maximum correction step size based on the multi-source evidence scores.

[0102] The gating decision unit is configured to determine whether to perform candidate modification through hard gating and / or preference gating.

[0103] The evidence management unit is configured to record the triggering source and execution status of candidate modifications to form a traceable chain of evidence;

[0104] The feedback learning unit is configured to update the preference gating model based on engineer feedback.

[0105] The data acquisition unit and the dual-view construction unit work together to transform unstructured textual process data into a structured representation that can participate in distillation training. The distillation training unit, through a teacher-student architecture, enables lightweight student models to inherit process semantic knowledge from the textual view during online deployment. The knowledge constraint unit introduces three types of process constraints—range, nearest neighbor, and consistency—to perform multi-dimensional verification of data-driven prediction results. The correction generation unit, through the maximum correction step size constraint, avoids prediction instability caused by excessively large single corrections. The gating decision unit, through a dual gating mechanism, aims to achieve a balance between automation and manual review. The evidence management unit and the feedback learning unit form a closed loop, and their technical role is to enable engineers' experience to be continuously reused and iterated.

[0106] In one embodiment, a method for predicting and correcting the thickness of a semiconductor PVD thin film is provided, comprising the following steps:

[0107] Step S1: Obtain semiconductor PVD process input data and film thickness measurement data to form a film thickness prediction sample set;

[0108] Step S2: Construct a structured view based on the structured process data in the PVD process input data, construct a text view based on the text process data in the PVD process input data, and establish a sample correspondence between the structured view, the text view, and the target film thickness dimension.

[0109] Step S3: Train the text view teacher model and generate dual-view distilled soft labels based on the prediction output of the text view teacher model;

[0110] Step S4: Train the student prediction model using the real film thickness label and the dual-view distillation soft label, so that the student prediction model outputs the initial film thickness prediction result based on the structured process data when deployed;

[0111] Step S5: Construct a human-machine collaborative process knowledge graph, which includes process feature nodes, film thickness target nodes, evidence record nodes, and engineer feedback nodes, as well as range constraint edges, similar working condition nearest neighbor edges, measurement point consistency edges, text semantic constraint edges, and engineer feedback edges.

[0112] Step S6: Based on the human-machine collaborative process knowledge graph, generate a multi-source evidence score for the initial film thickness prediction result, and generate a candidate correction amount constrained by the maximum correction step size according to the multi-source evidence score;

[0113] Step S7: Input the candidate correction amount into the hard-gated model and / or the preference-gated model, determine whether to perform candidate correction, and obtain the corrected film thickness prediction result;

[0114] Step S8: Record the trigger source, evidence edge, correction range, gating result and execution status corresponding to the candidate correction to form a traceable evidence chain;

[0115] Step S9: Update the evidence edge weights in the preference gating model and the human-machine collaborative process knowledge graph based on the engineer's acceptance or rejection feedback on the candidate modifications;

[0116] Step S10: Output the final film thickness prediction result, evidence record, and manual review prompt.

[0117] Steps S1 to S2, which acquire and construct dual-view data, provide a data foundation for subsequent cross-view knowledge distillation. Steps S3 to S4, through a teacher-student distillation architecture, enable the student model to inherit the process semantic knowledge from the text view while ensuring online inference efficiency. Step S5, which constructs a human-machine collaborative process knowledge graph containing five types of nodes and five types of edges, is used to uniformly represent process scope knowledge, nearest neighbor knowledge, consistency knowledge, text constraints, and engineer feedback as computable structured knowledge. Step S6, which generates multi-source evidence scores based on the knowledge graph and applies a maximum step size constraint, means that data-driven prediction results can be prevented from deviating from the reasonable range of the process. Step S7, which uses a dual gating mechanism to determine whether to perform corrections, aims to achieve a balance between automation efficiency and human reliability. Steps S8 to S9, which record the evidence chain and update the model based on feedback, enable the continuous reuse of human review experience and the closed-loop iteration of the system.

[0118] In one embodiment, a semiconductor PVD thin film thickness prediction and closed-loop correction system is provided, comprising:

[0119] The data acquisition module is configured to acquire PVD process input data and film thickness measurement data;

[0120] The initial prediction module is configured to output the initial film thickness prediction results based on the student model trained on dual-view distillation.

[0121] The process knowledge verification module is configured to generate a multi-source evidence score for the initial film thickness prediction result based on film thickness range constraints, nearest neighbor constraints, and measurement point consistency constraints.

[0122] The correction decision module is configured to generate candidate correction values ​​based on the multi-source evidence scores and determine whether to execute them through a gating mechanism.

[0123] The data acquisition module, initial prediction module, process knowledge verification module, and correction decision module constitute a complete prediction-verification-correction chain, which is to transform the purely data-driven prediction results into engineering-usable results that meet process constraints.

[0124] In another embodiment, a semiconductor PVD thin film thickness prediction and closed-loop correction system is provided, comprising:

[0125] An industrial communication interface is used to exchange process parameters and film thickness data with the PVD equipment control system and film thickness measurement equipment.

[0126] The processor is coupled to the industrial communication interface;

[0127] A memory, coupled to the processor, stores computer program instructions;

[0128] When the processor executes instructions in the memory, it is configured to implement the following units:

[0129] The data acquisition unit is configured to acquire PVD process input data and film thickness measurement data;

[0130] The initial prediction unit is configured to output the initial film thickness prediction result based on the student model trained on dual-view distillation.

[0131] The process knowledge verification unit is configured to generate multi-source evidence scores and candidate correction quantities based on the process knowledge graph.

[0132] The gating decision unit is configured to determine whether to perform candidate modification through hard gating and preference gating.

[0133] The system architecture disclosed herein adopts a hybrid connection structure of serial plus feedback. For example... Figure 1 and Figure 2 As shown, the data preprocessing module receives data from the process data source and the film thickness measurement data source. After completing the construction of the structured view and the text view, the structured view is input into the initial prediction module (including the teacher prediction model and the student prediction model), and the text view is input into the teacher prediction model to generate distillation soft labels.

[0134] After the student prediction model outputs the initial film thickness prediction result, the result is simultaneously fed into the process knowledge model and the candidate correction module. The process knowledge model stores the three types of knowledge—film thickness range knowledge, similar working condition nearest neighbor knowledge, and measurement point consistency knowledge—based on the training set, in a queryable data structure for the candidate correction module to access.

[0135] After receiving the initial prediction results, the candidate correction module queries the evidence scores for each dimension of the process knowledge model. After weighted fusion and maximum step size truncation, it generates candidate correction values, which are then sent to the gate control execution module. The gate control execution module includes a hard gate control submodule and a preference gate control submodule. The output of the preference gate control submodule is continuously updated by the engineer feedback module.

[0136] The evidence recording module is connected to the gating execution module, the candidate correction module, and the engineer feedback module. It is used to record the trigger source, evidence edge, correction range, gating result, and engineer feedback label of the candidate correction.

[0137] When a migration task is required, the cross-condition adaptation module connects between the data preprocessing module and the initial prediction module to align and reduce the weight of the input features.

[0138] The results output module receives the final prediction results from the gating execution module, the evidence records from the evidence recording module, and the review prompts from the engineer feedback module, outputs them to the software display interface, and writes them to the historical database.

[0139] The forward propagation path of this architecture is: data source → preprocessing → initial prediction → candidate correction → gating execution → result output. The feedback path is: engineer feedback module → preference gating model update / process knowledge graph update, and actual measurement results → historical database → model iterative update. The two paths form a complete closed loop.

[0140] The hybrid connection structure with serial and feedback ensures efficient execution of the prediction-correction-output forward process, while the feedback path enables continuous optimization of the system based on engineer experience and real measurement results.

[0141] In one embodiment, the data acquisition module connects to the PLC controller of the PVD equipment via industrial communication protocols such as EtherCAT, Modbus TCP, or SECS / GEM to collect process parameters in real time, including process power, chamber pressure, gas flow rate, temperature, and deposition time. Film thickness measurement data is imported via a data interface from measurement results obtained from measuring equipment (such as an ellipsometry, profilometer, or XRF).

[0142] The initial prediction module and candidate correction module are implemented using an industrial-grade edge computing device, which includes a high-performance multi-core CPU and GPU accelerator card for performing neural network inference of the student prediction model and real-time querying of the process knowledge graph. Memory utilizes DDR4 / DDR5 RAM and NVMe SSDs to store model parameters, knowledge graph data, and evidence records.

[0143] The gating execution module is implemented by executing logical judgment programs via the CPU. The inference of the preference gating model can share GPU resources with the student prediction model. The evidence recording module is implemented through a database system, using SQLite or PostgreSQL for structured storage.

[0144] The engineer feedback module is implemented through a web interface or desktop application. Engineers can view evidence records and give acceptance / rejection labels through this interface. Feedback data is stored in the database and triggers incremental training of the preference gating model.

[0145] The cross-condition adaptation module can be implemented on an offline server, using the same hardware resources as the initial prediction module. When model migration is required, it performs common feature alignment, distribution offset evaluation, and covariance alignment calculation.

[0146] In another alternative approach, the system can be implemented using a cloud-edge collaborative architecture. The training phase (teacher model training, knowledge graph construction, and cross-condition adaptation) is executed on a GPU cluster in the cloud, while the deployment phase (student model inference, candidate correction generation, and gating execution) is executed on an industrial control computer at the production line edge. The cloud and edge synchronize model parameters and knowledge graph updates via MQTT or OPC UA protocols.

[0147] In another alternative, for production line environments with limited computing resources, the student prediction model can be quantized to INT8 precision and deployed on ARM-based embedded devices.

[0148] In one embodiment, the text-view teacher model includes a text embedding layer, a semantic encoding layer, a feature aggregation layer, and a regression output layer. The text embedding layer converts the process recipes, alarm logs, maintenance records, and engineer notes corresponding to the PVD process samples into text vectors. The semantic encoding layer employs a Transformer encoding structure and extracts contextual semantic relationships from the text process data through a multi-head self-attention mechanism. The feature aggregation layer aggregates the semantic encoding results to obtain a sample-level text-view representation. The regression output layer maps the text-view representation to the corresponding film thickness target dimension and outputs the teacher film thickness prediction result.

[0149] The text-view teacher model is trained under supervision using real film thickness measurement data corresponding to the text process samples. Text process data corresponding to the same process sample are associated according to sample identifiers and time information to form a sample-level text view, and a correspondence is established between the text view and the film thickness target vector. The supervised loss is calculated based on the deviation between the teacher's film thickness prediction results and the real film thickness measurements. Model parameters are updated through backpropagation, and model parameters are selected based on the prediction error on the validation set. After the teacher model is trained, teacher reliability weights are determined based on its prediction errors in each film thickness target dimension on the validation set, and the teacher's prediction results are calibrated to generate dual-view distillation soft labels for training the student prediction model.

[0150] In one embodiment, the student prediction model includes a structured process feature input layer, a structured feature encoding layer, a feature projection layer, and a film thickness prediction output layer. The structured feature encoding layer employs a lightweight multilayer perceptron structure, including at least one fully connected hidden layer and nonlinear activation units. The feature projection layer maps the intermediate representation of the structured view to a common representation space corresponding to the dimensions of the text view representation. The film thickness prediction output layer outputs the initial film thickness prediction results corresponding to each target film thickness dimension. The student prediction model is smaller in size than the text view teacher model.

[0151] The student prediction model is jointly trained using real membrane thickness labels, dual-view distillation soft labels, and text view representations. Its total training loss L is composed of the supervision loss λ1, the distillation loss λ2, and the cross-view loss λ3. Figure 1 The weighted sum of the consistency loss λ3, i.e.

[0152]

[0153] in, Used to measure the deviation between student predictions and actual film thickness labels; It is used to measure the deviation between student prediction results and dual-view distillation soft labels, and is weighted according to the teacher reliability weights corresponding to each film thickness target dimension; This is used to constrain the consistency between the structured view representation and the text view representation corresponding to the same process sample in a common representation space. After the model training is completed, the online deployment phase only inputs structured PVD process data into the student prediction model to output initial film thickness prediction results.

[0154] In one embodiment, the preference gating model employs a binary classification model. Its inputs include one or more of the following: candidate correction magnitude, range bias, similar working condition nearest neighbor bias, measurement point consistency bias, text constraint bias, number of evidence sources, historical acceptance rate, target dimension number, and working condition identifier. The output is the probability that a candidate correction is accepted. The preference gating model is trained using engineers' acceptance or rejection feedback on historical candidate corrections as supervisory labels, and determines an acceptance probability threshold based on the acceptance prediction precision and recall on the validation set. Candidate corrections whose acceptance probability reaches the preset threshold are automatically executed; otherwise, they undergo manual review.

[0155] In one embodiment, the similar operating condition nearest neighbor model in the process knowledge graph performs Z-score normalization on the structured PVD process features before construction, with the normalization parameters determined based on the training set. The nearest neighbor model employs a K-nearest neighbor algorithm based on Euclidean distance to retrieve K historical samples similar to the sample to be predicted in the normalized process feature space, and determines a reference value for the film thickness of similar operating conditions based on the film thickness measurement data of the historical samples. Nearest neighbor models are constructed for each preset candidate K value, using the mean square error between the nearest neighbor film thickness reference value and the actual film thickness measurement value as an evaluation index, and selecting the K value with the smaller mean square error.

[0156] It should be noted that the above model structure and hyperparameters are one implementation method of this disclosure. The hidden layer size, learning rate, batch size, and loss weights of the text-view teacher model, student prediction model, and preference gating model are set as candidate ranges based on the training sample size and membrane thickness target dimension, and selected based on the validation set prediction error; the K value of the nearest neighbor model is determined based on the validation set nearest neighbor reference error; supervised loss, distillation loss, and cross-view loss... Figure 1The weights of consistency loss are initially set based on the numerical scale of each loss term and adjusted in conjunction with the film thickness prediction error of the validation set.

[0157] In one embodiment, the text view is constructed as follows: PVD process text data includes, but is not limited to: process formulation documents (containing descriptive text such as target film thickness, process gas ratio, and power setpoints), equipment alarm logs (containing alarm codes, alarm times, and alarm descriptions), maintenance records (containing component replacement times, maintenance types, and maintenance personnel notes), and engineer operation notes (containing reasons for manual intervention and operation descriptions). The text view construction steps include: First, aligning various types of text data according to sample timestamps, merging all text data corresponding to the same batch of processes into a text view for that batch of samples; Second, preprocessing the merged text (including removing special characters, unifying English capitalization, and Chinese word segmentation) to obtain the text view input sequence; Third, inputting the text view input sequence into the text view teacher model to extract process semantic constraint features related to the film thickness target. This construction method ensures a one-to-one correspondence between the text view and the structured view at the sample level.

[0158] In one embodiment, the specific determination method for the soft and hard ranges of film thickness is as follows: For the first... For each target thickness dimension, the actual film thickness measurement value corresponding to that target thickness dimension is obtained from the training set. Based on the quantile statistics of the actual film thickness measurement values, a soft range and a hard range are determined. The soft range represents the main distribution interval of the film thickness measurement values ​​under normal production conditions and is used to determine the reference target for range pullback correction. The hard range represents the anomaly judgment interval of the film thickness prediction results, and its range is wider than the soft range.

[0159] Specifically, let the first The training set film thickness measurement values ​​for each film thickness target dimension are , its first quantiles are denoted as Then the soft range and hard range They are represented as follows:

[0160]

[0161]

[0162] in, This is to ensure that the hard range encompasses the soft range;

[0163] The quantile (e.g., the 1st percentile) corresponding to the lower bound of the hard range is used to determine the hard lower limit of the film thickness prediction anomaly.

[0164] This represents the quantile (e.g., the 5th percentile) corresponding to the lower bound of the soft range, used to determine the soft lower limit of the normal distribution interval, serving as a reference target for pullback correction;

[0165] This represents the quantile (e.g., the 95th percentile) corresponding to the upper bound of the soft range, used to determine the soft upper limit of the normal distribution interval;

[0166] The quantile (e.g., the 99th percentile) corresponding to the upper limit of the hard range is used to determine the hard upper limit of the film thickness prediction anomaly.

[0167] The above variables satisfy That is, the hard range is wider than the soft range, and the soft range is contained within the hard range.

[0168] In one embodiment, the soft range is determined using the 5th to 95th percentiles of the film thickness measurements in the training set, and the hard range is determined using the 1st to 99th percentiles. For target film thickness dimensions with clear process windows or equipment-allowed film thickness boundaries, the hard range can also be determined by combining preset process upper and lower limits, ensuring that the soft range is within the hard range.

[0169] When the initial film thickness prediction value is within the hard range, no range pullback correction is generated based on the range constraint. When the initial film thickness prediction value is below the lower bound of the hard range or above the upper bound of the hard range, the target film thickness dimension is determined to trigger a range anomaly, and the corresponding boundary of the soft range is used as the range pullback reference value. For initial film thickness prediction values ​​below the lower bound of the hard range, candidate corrections are generated towards the lower bound of the soft range; for initial film thickness prediction values ​​above the upper bound of the hard range, candidate corrections are generated towards the upper bound of the soft range. The range pullback correction amount is further fused with one or more of the similar working condition nearest neighbor correction component and the measurement point consistency correction component, and is constrained by the maximum correction step size.

[0170] In summary, this disclosure systematically integrates dual-view knowledge distillation, human-machine collaborative process knowledge graph, multi-source evidence fusion correction, dual-gated decision-making, traceable evidence chain, engineer feedback iterative learning, and cross-condition adaptation mechanism into a unified closed-loop prediction framework. This breaks the technical inertia of the traditional PVD film thickness prediction's unidirectional open-loop approach, achieving a closed-loop prediction paradigm of prediction-verification-correction-audit-feedback-iteration. This framework upgrades film thickness prediction results from simple numerical outputs to reliable engineering results with process constraints, auditability, traceability, and continuous optimization, representing a significant technological advancement compared to existing purely data-driven models.

[0171] This disclosure also provides a computer program product. The methods of this disclosure can be implemented, in whole or in part, by software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer programs or instructions. When the computer program or instructions are loaded and executed, all or part of the processes or functions of this disclosure are performed.

[0172] Computer programs or instructions can be stored in a readable storage medium or transferred from one readable storage medium to another. For example, the computer program or instructions can be transferred from one website, computer, server, or data center to another website, computer, server, or data center via wired or wireless means. The readable storage medium can be any available medium capable of access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium, such as a floppy disk, hard disk, or magnetic tape; an optical medium, such as a digital video optical disc; or a semiconductor medium, such as a solid-state drive. The computer-readable storage medium can be a volatile or non-volatile storage medium, or it can include both volatile and non-volatile types of storage media.

[0173] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0174] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0175] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0176] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0177] Although specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, those skilled in the art should understand that the above examples are merely preferred embodiments of the present disclosure, and the scope of protection of the present disclosure is not limited thereto. Any equivalent modifications, substitutions, combinations, or variations made to the specific embodiments based on the technical concept of the present disclosure shall still fall within the scope of protection of the present disclosure. Technical details not described in detail in this disclosure can be implemented by referring to the prior art or common general knowledge.

Claims

1. A method for predicting and correcting the thickness of semiconductor PVD thin films using a closed-loop mechanism, characterized in that, Includes the following steps: Step S1: Obtain semiconductor PVD process input data and film thickness measurement data to form a film thickness prediction sample set; Step S2: Construct a structured view based on the structured process data in the PVD process input data, construct a text view based on the text process data in the PVD process input data, and establish a sample correspondence between the structured view, the text view, and the target film thickness dimension. Step S3: Train the text view teacher model and generate dual-view distilled soft labels based on the prediction output of the text view teacher model; Step S4: Train the student prediction model using the real film thickness label and the dual-view distillation soft label, so that the student prediction model outputs the initial film thickness prediction result based on the structured process data when deployed; Step S5: Construct a human-machine collaborative process knowledge graph, which includes process feature nodes, film thickness target nodes, evidence record nodes, and engineer feedback nodes, as well as range constraint edges, similar working condition nearest neighbor edges, measurement point consistency edges, text semantic constraint edges, and engineer feedback edges. Step S6: Based on the human-machine collaborative process knowledge graph, generate a multi-source evidence score for the initial film thickness prediction result, and generate a candidate correction amount constrained by the maximum correction step size according to the multi-source evidence score; Step S7: Input the candidate correction amount into the hard-gated model and / or the preference-gated model, determine whether to perform candidate correction, and obtain the corrected film thickness prediction result; Step S8: Record the trigger source, evidence edge, correction range, gating result and execution status corresponding to the candidate correction to form a traceable evidence chain; Step S9: Update the evidence edge weights in the preference gating model and the human-machine collaborative process knowledge graph based on the engineer's acceptance or rejection feedback on the candidate modifications; Step S10: Output the final film thickness prediction result, evidence record, and manual review prompt.

2. The method according to claim 1, characterized in that, In step S1, the semiconductor PVD process input data includes one or more of the following: process power, chamber pressure, gas flow rate, temperature, deposition time, target status, chamber status, and equipment status. The film thickness measurement data includes one or more target dimensions of film thickness.

3. The method according to claim 1, characterized in that, In step S3, the text view teacher model uses a text encoder or language model embedding to extract process semantic constraints from process recipes, alarm logs, maintenance records, and engineer notes; the dual-view distillation soft label is generated based on the prediction results of the text view teacher model on the validation set after confidence calibration.

4. The method according to claim 1, characterized in that, In step S4, the training loss of the student prediction model includes a weighted sum of the following three loss components: The supervised loss between the prediction results of the student prediction model and the actual film thickness label; The distillation loss between the prediction results of the student prediction model and the dual-view distillation soft label; The consistency loss between the structured view representation and the text view representation of the student prediction model.

5. The method according to claim 1, characterized in that, In step S5, the range constraint edge is used to characterize the allowable range constraint of each film thickness target dimension. The allowable range includes a soft range determined based on the film thickness data in the training set and a hard range used to determine whether the initial prediction is abnormal. The similar working condition nearest neighbor edge is used to characterize the nearest neighbor relationship between the sample to be predicted and historical samples similar to the process input. The measurement point consistency edge is used to characterize the correlation constraint between different film thickness target dimensions. The text semantic constraint edge is used to characterize the semantic association between the text process data and the film thickness target. The engineer feedback edge is used to characterize the engineer's acceptance or rejection label for the candidate correction.

6. The method according to claim 5, characterized in that, In step S6 The multi-source evidence score includes the range evidence score, the neighboring evidence score for similar working conditions, the measurement point consistency evidence score, the text constraint evidence score, and the historical engineer feedback evidence score; the multi-source evidence score is a weighted sum of the five types of evidence scores. The candidate correction amount includes a fusion of one or more of the following correction components: The range pullback correction component generated based on the range constraint edge is used to pull back the predicted value that exceeds the allowable range to the soft range; The nearest neighbor support correction component generated based on the nearest neighbor edge of the similar working condition is used to make the predicted value approach the average film thickness of historical samples of similar working conditions. The measurement point consistency correction component generated based on the measurement point consistency edge is used to ensure that the predicted values ​​of each film thickness target dimension meet the correlation constraints between measurement points.

7. The method according to claim 6, characterized in that, The candidate correction amount is also subject to the truncation constraint of the maximum correction step size, which is used to limit the upper limit of the absolute value of a single candidate correction amount.

8. The method according to claim 1, characterized in that, In step S7, the hard-gated model determines whether to perform candidate correction based on the candidate correction magnitude, anomaly degree, and number of evidence sources; the preference-gated model takes the correction magnitude, range deviation, nearest neighbor deviation, measurement point consistency deviation, text constraint deviation, number of evidence sources, historical acceptance rate, target dimension number, and working condition identifier corresponding to the candidate correction amount as input features, and outputs the probability that the candidate correction is accepted. If the probability is higher than a preset threshold, the candidate correction is automatically performed; otherwise, it enters manual review.

9. The method according to claim 1, characterized in that, It also includes a cross-operating condition adaptation step: extracting common input features and common film thickness target dimension of the source operating condition and the target operating condition; calculating the distribution offset of the source operating condition and the target operating condition on the common input features, retaining features with an offset less than a preset threshold, and reducing the weight of features with an offset greater than or equal to the preset threshold; and performing covariance alignment on the features of the source operating condition so that the training samples of the source operating condition are statistically close to the samples of the target operating condition.

10. A semiconductor PVD thin film thickness prediction and closed-loop correction system, characterized in that, include: The data acquisition module is configured to acquire semiconductor PVD process input data and film thickness measurement data; The data preprocessing module is configured to perform verification, standardization, sample division and feature alignment on the data, construct a structured view based on the structured process data in the PVD process input data, construct a text view based on the text process data in the PVD process input data, and establish a sample correspondence between the structured view, the text view and the target film thickness dimension. The initial prediction module is configured to output an initial film thickness prediction result based on the structured process data using a student prediction model. The student prediction model is trained by a combination of real film thickness labels and dual-view distillation soft labels generated by a text-view teacher model. The process knowledge construction module is configured to construct a human-machine collaborative process knowledge graph, which includes process feature nodes, film thickness target nodes, evidence record nodes and engineer feedback nodes, as well as range constraint edges, similar working condition nearest neighbor edges, measurement point consistency edges, text semantic constraint edges and engineer feedback edges. The candidate correction module is configured to generate a multi-source evidence score based on the human-machine collaborative process knowledge graph of the initial film thickness prediction result, and generate a candidate correction amount constrained by the maximum correction step size based on the multi-source evidence score. The gating execution module is configured to input the candidate correction amount into the hard gating model and / or the preference gating model, determine whether to perform the candidate correction, and obtain the corrected film thickness prediction result. The evidence recording module is configured to record the trigger source, evidence edge, correction range, gating result and execution status corresponding to the candidate correction, forming a traceable evidence chain; The engineer feedback module is configured to receive feedback from engineers on whether they accept or reject candidate modifications, and update the evidence edge weights in the preference gating model and the human-machine collaborative process knowledge graph. The cross-condition adaptation module is configured to perform common feature alignment, feature offset evaluation, feature filtering, feature weight reduction, and distribution alignment in cross-material or cross-condition migration scenarios. The results output module is configured to output the final film thickness prediction results, evidence records, and manual review prompts.

Images