Process parameter determination method and apparatus for metal compound thin film deposition

Abstract

This application relates to a method and apparatus for determining process parameters in metal compound thin film deposition. The method constructs a two-stage cascaded process prediction model. The first stage, a feature extraction model, performs nonlinear transformations and feature extraction on the process parameters, extracting intermediate feature information that comprehensively characterizes the sputtering process state. The second stage, a probabilistic prediction model, establishes a probabilistic mapping relationship between the intermediate feature information and the thin film response information, decoupling the complex nonlinear coupling relationships between multiple parameters in reactive sputtering. It outputs predicted values ​​and corresponding confidence information. Based on the predicted values ​​and confidence information, process parameters are screened, considering both target performance requirements and prediction reliability requirements. This improves the efficiency and reliability of process parameter screening, reduces the number of invalid experiments, and enables the prediction of key response quantities such as thin film deposition rate and chemical composition even with limited experimental data, shortening the process development cycle.

Description

Technical Field

[0001] This application relates to the field of integrated circuit manufacturing technology, and in particular to a method and apparatus for determining process parameters for metal compound thin film deposition. Background Technology

[0002] With the continuous evolution of integrated circuit manufacturing technology, especially after entering the nanoscale, advanced devices have placed unprecedentedly stringent demands on the performance consistency, controllability, and process repeatability of functional thin film materials. Reactive sputtering is a key process in integrated circuit manufacturing for preparing metal compound thin films. By allowing a metal target to react in situ with reactive gases (such as oxygen and nitrogen) in a plasma environment, highly controllable metal compound thin films with controllable composition and properties are deposited on the substrate. These films are widely used in many functional units and interconnect structures, such as diffusion barrier layers, memory devices, and thin-film transistors. The electrical properties of metal compound thin films are highly sensitive to their oxidation stoichiometry; therefore, precise control of oxygen content during deposition is crucial.

[0003] However, pulsed DC reactive sputtering suffers from severe multi-parameter coupling and nonlinearity issues. Process parameters such as sputtering power, operating pressure, and oxygen flow rate collectively determine the target surface state and film composition, and are prone to "target poisoning," leading to process instability. Current process parameter optimization primarily relies on orthogonal experiments or empirical tuning, which is not only time-consuming and material-intensive but also difficult to transfer across different equipment or material systems. Furthermore, conventional monitoring methods, such as emission spectroscopy, suffer from expensive equipment and response lag, failing to achieve in-situ, low-cost monitoring of the target surface state. Summary of the Invention

[0004] Therefore, it is necessary to provide a method and apparatus for determining process parameters for metal compound thin film deposition to address the aforementioned technical problems.

[0005] In a first aspect, this application provides a method for determining process parameters for metal compound thin film deposition, the method comprising:

[0006] The process parameters, process status information, and corresponding film response information of the reactive sputtering preparation of metal compound thin films are obtained to construct a modeling dataset; the process status information includes at least the target voltage.

[0007] A process prediction model is constructed, which includes a first-stage feature extraction model and a second-stage probabilistic prediction model. The first-stage feature extraction model takes the combination of process parameters as input and extracts intermediate feature information that characterizes the state of the sputtering process. The second-stage probabilistic prediction model takes the intermediate feature information extracted in the first stage as input and outputs the predicted value and the corresponding confidence information.

[0008] Using the target metal compound thin film as the optimization target, based on the predicted values ​​and corresponding confidence information output by the process prediction model, process parameter combinations that meet the optimization target and have a confidence level higher than a preset threshold are selected.

[0009] In one embodiment, the construction of the process prediction model includes:

[0010] The training data in the modeling dataset is input into the neural network for training to obtain the first-stage feature extraction model after training.

[0011] The intermediate feature information output by the first-stage feature extraction model is used as input to train the Gaussian process regression model, establishing a probability mapping relationship between the intermediate feature information and the thin film response information, and obtaining the trained second-stage probability prediction model.

[0012] In one embodiment, the second-stage probabilistic prediction model employs a Gaussian process regression model, wherein the covariance function of the Gaussian process regression model adopts a composite kernel function structure; the composite kernel function includes a first kernel function that measures the correlation between intermediate feature information and a second kernel function that characterizes the influence of observation noise.

[0013] The prediction results output by the second-stage probability prediction model include the prediction mean and the prediction variance. The prediction mean is used as the predicted value of the thin film response information, and the prediction variance is used as the corresponding prediction confidence information.

[0014] In one embodiment, the process status information further includes the compound coverage on the target surface; obtaining the compound coverage on the target surface includes:

[0015] Based on the differences in electrical properties between metal targets and their corresponding compounds, a correlation mapping model is established between relevant parameters of the reactant gas, the working voltage of the target, and the coverage.

[0016] The collected target working voltage is input into the correlation mapping model to obtain the corresponding target surface compound coverage.

[0017] In one embodiment, it further includes:

[0018] Based on the relationship between the target voltage and the flow rate of the reactant gas, the hysteresis characteristics of the oxidation state of the target surface are identified, and the target state is divided into a metallic range, an unsaturated metal oxide range, and a saturated oxide range according to the hysteresis characteristics.

[0019] In one embodiment, the construction of the process prediction model further includes:

[0020] During the training process of the first-stage feature extraction model, a weighted loss function is used to assign loss weights to the training data, and the loss weight of the training data in the unsaturated metal oxide range is greater than the loss weight of the training data in the metallic range and the saturated oxide range.

[0021] In one embodiment, it further includes:

[0022] Based on the process parameter combinations with confidence levels below a preset threshold within the unsaturated metal oxide range, new process parameter information, process state information, and corresponding thin film response information are obtained and added to the modeling dataset.

[0023] The process prediction model is trained using the updated modeling dataset.

[0024] In one embodiment, during retraining, the weighted loss function assigns a loss weight to the newly added training data, and the loss weight for the newly added training data is higher than the loss weight for the original training data in the modeling dataset.

[0025] In one embodiment, the metal compound film is a niobium oxide film; the target voltage range corresponding to the unsaturated metal oxide range is 345V~380V; the loss weight of the training data in the unsaturated metal oxide range is 2 to 10 times that of the loss weight of the training data in the metallic range and / or the saturated oxide range.

[0026] Secondly, this application provides a process parameter determination apparatus for metal compound thin film deposition, comprising:

[0027] The acquisition module is used to acquire process parameter information, process state information, and corresponding film response information during the reactive sputtering preparation of metal compound thin films, and to construct a modeling dataset; the process state information includes at least the target voltage;

[0028] The storage module stores a process prediction model, which includes a first-stage feature extraction model and a second-stage probabilistic prediction model. The first-stage feature extraction model takes a combination of process parameters as input and extracts intermediate feature information that characterizes the sputtering process state. The second-stage probabilistic prediction model takes the intermediate feature information extracted in the first stage as input and outputs the predicted value and the corresponding confidence information.

[0029] The screening module takes the target metal compound thin film as the optimization target and, based on the predicted value output by the process prediction model and the corresponding confidence information, screens out the combination of process parameters that meets the optimization target and has a confidence level higher than a preset threshold.

[0030] The aforementioned method and apparatus for determining process parameters for metal compound thin film deposition constructs a two-stage cascaded process prediction model. The first-stage feature extraction model performs nonlinear transformation and feature extraction on the process parameters, extracting intermediate feature information that can comprehensively characterize the sputtering process state. The second-stage probabilistic prediction model establishes a probabilistic mapping relationship between the intermediate feature information and the thin film response information, decoupling the complex nonlinear coupling relationship between multiple parameters in the reactive sputtering process, and outputs predicted values ​​and corresponding confidence information. Based on the predicted values ​​and confidence information, process parameters are screened, taking into account both target performance requirements and prediction reliability requirements. This improves the efficiency and reliability of process parameter screening, reduces the number of invalid experiments, and enables the prediction of key response quantities such as thin film deposition rate and chemical composition under limited experimental data conditions, thus shortening the process development cycle. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 A flowchart of a method for determining process parameters for metal compound thin film deposition provided as an exemplary embodiment of this application;

[0033] Figure 2 A schematic diagram of the overall flow of a method for determining process parameters for metal compound thin film deposition provided as an exemplary embodiment of this application;

[0034] Figure 3 A schematic diagram illustrating the process of data processing by the process prediction model provided in an exemplary embodiment of this application;

[0035] Figure 4 A framework diagram of the process prediction model provided for an exemplary embodiment of this application;

[0036] Figure 5 A structural block diagram of a process parameter determination apparatus for metal compound thin film deposition provided as an exemplary embodiment of this application;

[0037] Figure 6 An internal structural diagram of a computer device provided for an exemplary embodiment of this application. Detailed Implementation

[0038] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0039] In novel non-volatile memories, selectors, and new devices for brain-like and neuromorphic computing, multivalent transition metal compound thin films have become a research frontier due to their highly tunable electrical properties (such as resistance and switching thresholds). For example, metal oxide thin films such as TaOx and HfOx have shown great potential in threshold switching, memristor, and resistive random access memory; while oxide semiconductor materials such as ZnO and IGZO have made significant progress in the application of thin-film transistors. Similarly, metal nitride thin films also play an important role in interconnects, electrodes, and diffusion barriers. The modulation and integration of metal compound thin films are one of the key driving forces for the development of integrated circuits in the post-Moore's Law era.

[0040] However, the key physical and electrical properties of multivalent metal compound thin films are highly sensitive to their stoichiometry. For example, different oxygen contents in metal oxides correspond to different proportions of metal valence states, and these valence state variations directly affect the film's conductivity mechanism, defect distribution, band structure, and interface properties. Multivalent metal oxides exhibit different phases and electrical properties within different oxygen content ranges, with the substoichiometric region generally considered a crucial component for achieving the desired device performance. Therefore, achieving precise control over oxygen or nitrogen content during thin film deposition is a prerequisite for realizing the target device performance.

[0041] Currently, when using DC reactive sputtering to prepare metal compound thin films, the target surface remains predominantly metallic at low oxygen flow rates, making it difficult to obtain stable oxide films. As the oxygen flow rate increases, the target surface gradually becomes covered by insulating oxides or nitrides, leading to a decrease in sputtering rate, unstable discharge, and even arc extinction – a problem commonly referred to as "target poisoning." The coverage of the target compound directly determines the poisoning state, affecting process efficiency and product quality. In pulsed DC reactive sputtering, multiple process parameters are coupled and collectively determine the chemical state of the target surface and plasma characteristics, making the control of thin film composition a complex and highly nonlinear problem.

[0042] The method for determining process parameters for metal compound thin film deposition provided in this application can be implemented by a terminal. The terminal can be, but is not limited to, various personal computers, laptops, smartphones, tablets, etc.

[0043] In one exemplary embodiment, such as Figure 1As shown, a method for determining process parameters for metal compound thin film deposition is provided, including the following steps S101 to S103. Wherein:

[0044] Step S101: Obtain process parameter information, process state information and corresponding film response information during the reactive sputtering preparation of metal compound thin films, and construct a modeling dataset; the process state information includes at least the target voltage.

[0045] In this embodiment, refer to Figure 2 The pulsed DC reactive sputtering deposition process of multivalent metal oxide thin films (NbOx) is used as an example for illustration. Due to the multivalent nature of niobium, NbOx thin films exhibit a wide range of electrical behaviors, from metallic conductivity to semiconductor to insulator, under different oxygen content conditions. Controlling its oxidation stoichiometry is crucial for obtaining the performance of the target device. In this embodiment, a pulsed DC magnetron reactive sputtering process is used to prepare NbOx thin films. The sputtering system includes a metal Nb target, a pulsed DC power supply, a vacuum reaction chamber, a reaction gas supply device, and a process monitoring device. The reaction gases include argon and oxygen, which are independently regulated by mass flow controllers.

[0046] Before thin film deposition, the reaction chamber is evacuated to achieve the preset initial vacuum conditions. Argon gas is then introduced for ignition, and oxygen is introduced after the sputtering process stabilizes, forming a pulsed DC reactive sputtering process. By adjusting the output power of the pulsed DC power supply, the working pressure of the chamber, and the flow ratio of argon and oxygen, NbOx thin film deposition under different sputtering conditions can be achieved.

[0047] Reference Figure 2 During each thin film deposition process, the corresponding sputtering power, cavity working pressure, and oxygen flow rate are recorded as process parameter information for the preparation of metal compound thin films. During sputtering deposition, real-time process monitoring data, such as target voltage and target current, are collected as process status information to characterize changes in the target surface state and discharge characteristics during sputtering. After each sputtering deposition, parameters such as the film thickness, stoichiometry, and resistivity of the prepared compound thin film are collected as corresponding film response information. The oxygen content or stoichiometry of the compound thin film directly reflects the multivalent state distribution of niobium (Nb) and is an indicator for evaluating whether the metal compound thin film meets the target requirements.

[0048] In this embodiment, sputtering power, cavity working pressure, and oxygen flow rate can be used as process control parameters, and these process parameters can be combined and set according to preset parameter levels. For example, an orthogonal experimental design method can be used to configure the combination of process parameters to cover different parameter combination conditions within a limited number of experiments, while ensuring the balance between the levels of each factor.

[0049] After completing all sputtering experiments, the obtained process parameter information, process status information and corresponding thin film response information are organized and normalized or standardized for different types of data. Each set of process parameter combinations and its corresponding process monitoring data and thin film response information are used as a sample to construct a modeling dataset.

[0050] Step S102: Construct a process prediction model. The process prediction model includes a first-stage feature extraction model and a second-stage probabilistic prediction model. The first-stage feature extraction model takes the combination of process parameters as input and extracts intermediate feature information that characterizes the state of the sputtering process. The second-stage probabilistic prediction model takes the intermediate feature information extracted in the first stage as input and outputs the predicted value and the corresponding confidence information.

[0051] In this embodiment, refer to Figure 3 , Figure 4 A two-stage cascaded process prediction model was constructed. The first-stage feature extraction model adopted a deep neural network structure, consisting of multiple layers of linear mapping units connected in series. Nonlinear activation functions, batch normalization structures, and random deactivation regularization structures were introduced between each network layer to enhance the modeling ability of complex nonlinear process relationships. The standardized process parameter vector was input into the deep neural network, and nonlinear transformation and feature abstraction were performed through multiple cascaded hidden layers. The intermediate feature vector representing the sputtering process state was obtained from the output layer. The intermediate feature vector is a high-dimensional feature representation output by the deep neural network after performing nonlinear transformation on the input standardized process parameters.

[0052] The second-stage probabilistic prediction model can employ a Gaussian process regression model, using the intermediate feature vector extracted in the first stage as input to establish a probabilistic mapping relationship between the intermediate feature information and the thin film response information. The second-stage probabilistic prediction model outputs predicted values ​​and corresponding confidence information, where the predicted mean serves as the predicted value of the thin film response information, and the predicted variance serves as the corresponding confidence information, reflecting the model's confidence level in the prediction results under the current process conditions.

[0053] Step S103: Taking the target metal compound thin film as the optimization target, based on the predicted value output by the process prediction model and the corresponding confidence information, select the combination of process parameters that meets the optimization target and has a confidence level higher than the preset threshold.

[0054] In this embodiment, the stoichiometry of the target NbOx thin film can be used as the optimization objective. Candidate process parameter combinations are input into a trained process prediction model to obtain the corresponding predicted values ​​and prediction variances. When the prediction variance is lower than a preset threshold, it indicates that the process state corresponding to the process parameter combination has high coverage in the modeling dataset, and the model has high reliability in predicting its results. Process parameter combinations that simultaneously meet the stoichiometry requirements of the target metal compound thin film and have prediction variances lower than the preset threshold are selected as deposition process conditions to guide the implementation of subsequent thin film preparation processes.

[0055] The above-described method for determining process parameters for metal compound thin film deposition involves acquiring process parameter information, process state information, and corresponding thin film response information during reactive sputtering of metal compound thin films, and constructing a modeling dataset. The process state information includes at least the target voltage. A process prediction model is constructed, comprising a first-stage feature extraction model and a second-stage probabilistic prediction model. The first-stage feature extraction model takes the combination of process parameters as input and extracts intermediate feature information characterizing the sputtering process state. The second-stage probabilistic prediction model takes the intermediate feature information extracted in the first stage as input and outputs predicted values ​​and corresponding confidence information. Using the target metal compound thin film as the optimization objective, based on the predicted values ​​and corresponding confidence information output by the process prediction model, films that meet the optimization objective and have a confidence level higher than [a certain value] are selected. The process parameters are combined with preset thresholds. A two-stage cascaded process prediction model is constructed. The first stage feature extraction model performs nonlinear transformation and feature extraction on the process parameters to extract intermediate feature information that can comprehensively characterize the sputtering process state. The second stage probabilistic prediction model establishes a probabilistic mapping relationship between intermediate feature information and thin film response information, decoupling the complex nonlinear coupling relationship between multiple parameters in reactive sputtering, and outputs predicted values ​​and corresponding confidence information. Based on the predicted values ​​and confidence information, process parameters are screened, taking into account both target performance requirements and prediction reliability requirements. This improves the efficiency and reliability of process parameter screening, reduces the number of invalid experiments, and enables the prediction of key response quantities such as thin film deposition rate and chemical composition under limited experimental data conditions, thus shortening the process development cycle.

[0056] In some embodiments, refer to Figure 3 , Figure 4The process prediction model comprises a data preprocessing unit, a first-stage feature extraction head, and a second-stage Gaussian process regression module, connected sequentially. The data preprocessing unit uses a normalization scaler to standardize the input sputtering power, oxygen flow rate, chamber working pressure, and target voltage. The feature extraction head employs a deep neural network structure, consisting of multiple cascaded neural network layers. The feature extraction head includes a linear layer, a batch normalization layer, a linear unit with leakage correction, and a random deactivation layer, used to map the standardized process parameters into intermediate feature information characterizing the sputtering process state. The Gaussian process regression module uses a composite kernel function, including the white kernel function and the Marton kernel function, to construct a covariance matrix for probabilistic regression prediction based on the intermediate feature information. It also includes a data output layer that outputs predicted values ​​and confidence information.

[0057] In this embodiment, before training the model, the collected process parameter information and process status information are first uniformly organized and scaled to form a standardized input data format suitable for subsequent model training and prediction.

[0058] In this embodiment, the process parameter information and process status information collected during the operation of the pulsed DC reactive sputtering equipment are represented as the original input data vector:

[0059]

[0060] in, Represents the original process data input vector; Indicates the first The original values ​​of each process parameter or process state parameter; n represents the total dimension of the process parameters and process state parameters. The original input data should include at least sputtering power, chamber working pressure, oxygen flow rate, initial gas pressure, and process monitoring parameters related to the target voltage. Since the above-mentioned different process parameters differ significantly in physical meaning, units of measurement, and numerical distribution range, directly using them as model input can easily lead to bias in some parameters during model training and prediction.

[0061] To eliminate the influence of different physical dimensions and numerical scales on the modeling process, this embodiment performs standardization on the original input data vector. Standardization involves linearly transforming each dimension of the process parameters to represent them within a uniform numerical scale, as expressed in the following expression:

[0062]

[0063] in, This represents the vector of statistical means of each input parameter in the sample dataset. This represents the corresponding standard deviation vector. Standardization effectively suppresses the problem of a single parameter dominating the model training process due to its large numerical amplitude, and also improves the model's overall sensitivity and numerical stability to changes in different parameters. The standardized process parameter vector x, as a unified data input format, is input into the subsequent process prediction model for deep feature modeling and regression prediction.

[0064] In some embodiments, refer to Figure 3 , Figure 4 Step S102 involves constructing a process prediction model, including steps S1021-S1022.

[0065] Step S1021: Input the training data in the modeling dataset into the neural network for training to obtain the first-stage feature extraction model after training.

[0066] In this embodiment, process parameter information and process state information from the modeling dataset are used as training data, and the corresponding thin film response information is used as supervision labels to train the deep neural network. The training data (standardized process parameter information and process state information) is input into a feature extraction head composed of a multi-layer network structure. The deep neural network consists of multiple layers of linear mapping units connected in series, and nonlinear activation functions, batch normalization structures, and random deactivation regularization structures are introduced between each network layer to enhance the model's ability to model complex nonlinear process relationships. Through forward propagation calculations, intermediate feature information representing the sputtering process state is obtained from the output layer. During training, the network weight parameters are iteratively updated using a backpropagation algorithm, ensuring that the output intermediate feature information effectively retains key features related to the thin film response information while compressing the dimensionality and redundant information of the original input data. When the loss function on the validation set converges, training is stopped and the network parameters are fixed, resulting in the trained first-stage feature extraction model.

[0067] In this embodiment, the overall mapping relationship between the feature extraction head and the training data can be expressed as:

[0068]

[0069] in, Indicated by network parameters A defined depth nonlinear mapping function, This is intermediate feature information for the output process state.

[0070] In this embodiment, intermediate feature information is used to describe the overall operating state of the sputtering process under the training data. Through the nonlinear mapping of deep neural networks, the original process conditions (such as power, gas pressure, flow rate, and target voltage) that were originally in different parameter spaces and numerical scales are uniformly mapped to the same feature space for expression. In this way, the multidimensional and strongly coupled influence relationships of process parameters can be compressed and reflected in the finite-dimensional intermediate feature vector information, thereby reducing the input complexity of subsequent probabilistic regression modeling while retaining key process information.

[0071] Step S1022: Use the intermediate feature information output by the first-stage feature extraction model as input to train the Gaussian process regression model, establish the probability mapping relationship between the intermediate feature information and the thin film response information, and obtain the trained second-stage probability prediction model.

[0072] In this embodiment, the intermediate feature information output by the first-stage feature extraction model is used as input, and the corresponding thin film response information in the modeling dataset is used as the target output to train the Gaussian process regression model, establish the probability mapping relationship between the intermediate feature information and the thin film response information, and obtain the trained second-stage probability prediction model.

[0073] The intermediate feature information set extracted from the training samples is used as the input matrix of the Gaussian process regression model, and the corresponding thin film response information (such as the stoichiometry of NbOx thin films) is used as the observation vector to construct the training dataset. After training, for any input intermediate feature vector, the Gaussian process regression model can output the corresponding predicted mean (as the predicted value of the thin film response information) and predicted variance (as the confidence information of the prediction result), thereby realizing the probabilistic prediction of the thin film response information.

[0074] In this embodiment, through the construction and training of a two-stage cascaded process prediction model, the first-stage feature extraction model realizes the nonlinear mapping from the original process parameters to the intermediate feature information, while the second-stage probabilistic prediction model establishes the probabilistic mapping relationship between the intermediate feature information and the thin film response information, so that the overall model has both the strong feature expression capability of deep learning and the uncertainty quantification capability of Bayesian methods.

[0075] In some embodiments, the process state information obtained in step S101 during the reactive sputtering preparation of metal compound thin films includes at least the target voltage.

[0076] Reference Figure 2 In pulsed reactive sputtering, a periodic reverse electric field is applied to suppress charge accumulation on the target surface and maintain discharge stability. The sputtered target particles react with reactive gases (such as oxygen) in a plasma environment and deposit on the substrate surface to form a metal compound film.

[0077] This embodiment introduces in-situ dynamic hysteresis monitoring technology. By acquiring electrical parameters such as target voltage in real time, a mapping relationship is established between these parameters and the oxidation state of the target surface, enabling real-time monitoring of the target state. Utilizing the difference in secondary electron emission coefficients between metals and their oxides: as the oxide coverage of the target surface increases, the secondary electron emission coefficient increases, leading to a characteristic shift in the target voltage required to maintain glow discharge. Based on this mechanism, the target voltage can serve as an in-situ probe to reflect the degree of oxidation on the target surface in real time.

[0078] In some embodiments, the process status information further includes the target surface compound coverage; obtaining the target surface compound coverage in step S101 includes: based on the difference in electrical properties between the metal target and its corresponding compound, and according to the inverse relationship between the target working voltage and the target surface compound coverage (the higher the coverage, the lower the voltage), and the direct relationship between the reaction gas related parameters (such as oxygen flow rate, oxygen partial pressure, etc.) and the coverage (the higher the gas flow rate, the higher the coverage), establishing a correlation mapping model between the reaction gas related parameters, the target working voltage and the coverage; inputting the collected target working voltage into the correlation mapping model to obtain the corresponding target surface compound coverage; wherein, the target surface compound coverage, as a component of the process status information, is used together with the process parameter information as input to the process prediction model.

[0079] In some embodiments, step S101 further includes step S1011: based on the relationship between the target voltage and the flow rate of the reactant gas, identify the hysteresis characteristics of the oxidation state of the target surface, and divide the target state into a metallic range, an unsaturated metal oxide range and a saturated oxide range according to the hysteresis characteristics.

[0080] In this embodiment, the relationship between the target voltage and the flow rate of the reactant gas (oxygen in this embodiment) is obtained through in-situ dynamic hysteresis monitoring technology to identify the hysteresis characteristics of the target surface oxidation state. (Refer to...) Figure 2 The metallic region (θt=0) corresponds to a region with low oxygen flow rate. The target surface is mainly in a metallic state, resulting in a high sputtering rate but insufficient film oxidation. The unsaturated metal oxide region (0<θt<1) corresponds to a process window with moderate oxygen flow rate. The target surface is partially oxidized, and the film stoichiometry is adjustable within this region. It is a key fabrication region for the performance of the target device (such as the threshold switching characteristics of NbOx). The saturated oxide region (θt=1) corresponds to a region with high oxygen flow rate. The target surface is completely covered by oxide (target poisoning), resulting in a sharp decrease in sputtering rate and poor discharge stability.

[0081] Among them, the unsaturated metal oxide region corresponds to a narrow process window and a region where the target material state is extremely sensitive to changes in oxygen flow rate. In this region, the nonlinearity of the process response is high and the repeatability is low. By using in-situ dynamic hysteresis monitoring, the boundaries of the above three regions and the hysteresis characteristics of the target voltage can be accurately identified, providing key target material state characteristics for subsequent modeling.

[0082] In some embodiments, step S102, which constructs a process prediction model, further includes step S102-1: during the training process of the first-stage feature extraction model, a loss weight is assigned to the training data through a weighted loss function, and the loss weight of the training data in the unsaturated metal oxide range is greater than the loss weight of the training data in the metallic range and the saturated oxide range.

[0083] In this embodiment, in step S101, the relationship between the target voltage and the flow rate of the reaction gas (oxygen in this embodiment) is obtained by in-situ dynamic hysteresis monitoring technology. Three ranges of the oxidation state of the target surface have been identified in advance: metallic range, unsaturated metal oxide range and saturated oxide range.

[0084] In this embodiment, to make the first-stage feature extraction model focus more on the unsaturated metal oxide region, a weighted loss function is used during training to assign differentiated loss weights to training data falling into different regions. The weighted loss function can be pre-built or constructed during training. It assigns higher loss weights to training data within the unsaturated metal oxide region and lower loss weights to training data within the metallic and saturated oxide regions. This allows the model to focus more on learning training data within the unsaturated metal oxide region, emphasizing the mapping relationship between process parameter changes and target voltage fluctuations within this region. This weighted strategy makes the extracted intermediate feature information more sensitive to changes in the unsaturated metal oxide region, which is beneficial for improving the prediction accuracy of the subsequent second-stage probabilistic prediction model within this region.

[0085] In this embodiment, to further improve the prediction accuracy of the model, and addressing the problems of insufficient feature learning caused by the narrow process window, high nonlinearity of parameters, and low proportion of sample data in the unsaturated metal oxide range, this embodiment adds an auxiliary regression head to the network structure of the first-stage feature extraction model and constructs an interval-oriented weighted loss function to achieve reinforcement learning of the unsaturated metal oxide range.

[0086] The auxiliary regression head is connected to the output of the penultimate hidden layer of the feature extraction head. It is used to provide intermediate supervision signals during training. The intermediate supervision signals use the target voltage as the supervision label to guide the feature extraction head to prioritize the retention of key features that are strongly correlated with the target voltage change and the target surface state during feature learning, while weakening the interference of irrelevant and redundant features.

[0087] The weighted loss function consists of a basic loss term and an interval-weighted loss term, and its expression is:

[0088]

[0089] in, The basic loss term is the mean squared error loss, which is calculated based on the prediction error of the thin film response information. The interval-weighted loss term is calculated based on the target voltage prediction error output by the auxiliary regression head. These are weighting coefficients used to adjust the intensity of interval reinforcement.

[0090] For training data falling within the target voltage range corresponding to the unsaturated metal oxide range, the weighting coefficient λ is set to 2~10; for training data falling within the metallic and saturated oxide ranges, the weighting coefficient λ is set to 0.5~1. This weighting setting allows the model to assign a higher penalty weight to the prediction error of training data within the unsaturated metal oxide range during training, forcing the model to learn the nonlinear mapping relationship between process parameters and target state and thin film response within this range, thereby improving the prediction accuracy and reliability within the unsaturated metal oxide range.

[0091] In some embodiments, the method for determining process parameters for metal compound thin film deposition in this embodiment further performs step S104 after step S103: based on the process parameter combinations with confidence levels below a preset threshold within the unsaturated metal oxide range, new process parameter information, process state information, and corresponding thin film response information are obtained and added to the modeling dataset; the updated modeling dataset is used to train the process prediction model.

[0092] In step S103, if the predicted value does not meet the optimization objective and the confidence level is lower than the preset threshold, it is determined that these process parameter combinations have not been fully learned in the current model. Representative process parameter combinations are selected from these combinations for supplementary experiments to obtain new process parameter information, process state information and corresponding thin film response information. The new information is added to the modeling dataset, and step S102 is re-executed to train the process prediction model to update the process prediction model.

[0093] However, for process parameter combinations with confidence levels below a preset threshold within the unsaturated metal oxide range, even if the predicted values ​​meet the optimization objective, these parameter combinations are considered not yet fully learned in the current model. From these process parameter combinations with confidence levels below the preset threshold within the unsaturated metal oxide range, representative process parameter combinations (partial or complete) are selected for supplementary experiments to obtain new experimental process parameter information, process state information (including target voltage and the resulting target surface compound coverage), and corresponding thin film response information (such as stoichiometry). This new information is then added to the original modeling dataset to form an expanded modeling dataset. The expanded modeling dataset is used to train and update the process prediction model, enabling it to learn the process rules within the unsaturated metal oxide range, thereby improving the prediction accuracy for this range.

[0094] It is understandable that during the optimization of the process prediction model, the original training data and the newly added training data together constitute the expanded modeling dataset in order to maintain the model's predictive ability for the learned process space, while improving the prediction accuracy for the unsaturated metal oxide range.

[0095] Furthermore, during the retraining process, the weighted loss function assigns loss weights to the newly added training data, with the loss weights for the newly added training data being higher than the loss weights for the original training data in the modeling dataset.

[0096] In this embodiment, this weighted strategy enables the model to pay more attention to the newly supplemented training data during iterative optimization, thereby accelerating the improvement of the predictive certainty of the unsaturated metal oxide range while retaining the predictive ability for the learned process space. By repeating the above steps S102 to S104, the model's coverage of the process parameter space is gradually expanded, thereby improving the overall prediction accuracy and stability.

[0097] By repeatedly executing steps S103 to S104, a closed loop of process parameter screening and model iterative optimization based on prediction uncertainty is formed. Under limited experimental resources, the model's coverage of the process parameter space is gradually expanded, ultimately achieving efficient and stable preparation of the target metal compound thin film.

[0098] In some embodiments, process parameter information may include sputtering power, reactive gas flow rate (such as oxygen flow rate, nitrogen flow rate), working gas pressure, and parameters that affect plasma characteristics and thin film growth state during reactive sputtering.

[0099] Process status information can include parameters such as the intensity of characteristic peaks in plasma emission spectra and cavity wall temperature. The intensity of characteristic peaks in plasma emission spectra can be used to monitor the types and concentration changes of active particles in the plasma; the cavity wall temperature reflects the thermal environment during the deposition process, and reflects the crystallization state and defect distribution of the thin film.

[0100] Key response information for thin films can include parameters such as film thickness, chemical composition (e.g., oxidation stoichiometry, nitrogen stoichiometry), and resistivity. Chemical composition determines the valence state distribution and band structure of metal compound thin films and is an indicator for evaluating whether the film meets the expected targets; thickness reflects the yield and stability of the process; and resistivity reflects the conductivity of the film.

[0101] In some embodiments, the method for determining process parameters for metal compound thin film deposition in this embodiment is applicable to a variety of reactive sputtering methods, such as pulsed DC power supply sputtering (PDC), high-power pulsed magnetron sputtering (HiPIMS), or closed-loop feedback control systems.

[0102] In some embodiments, the process parameter determination method for metal compound thin film deposition in this embodiment can be used for the following metal compound thin films: tantalum oxide (TaOx), hafnium oxide (HfOx), zirconium oxide (ZrOx), aluminum oxide (AlOx), titanium oxide (TiOx), niobium oxide (NbOx), vanadium oxide (VOx), tungsten oxide (WOx), gallium oxide (GaOx), magnesium oxide (MgOx), ruthenium oxide (RuOx), zinc oxide (ZnO), cadmium oxide (CdO), hafnium zirconium oxide (HfZrO), tin oxide (SnOx), copper oxide (CuxO), nickel oxide (NiO), indium tin oxide (ITO), and oxygen. Metal oxide thin films such as indium zinc oxide (IZO), iron oxide (FeOx), strontium oxide (SrOx), indium gallium zinc oxide (IGZO), indium aluminum oxide (IAO), indium aluminum zinc oxide (IAZO), boron aluminum oxide (AlBO), strontium titanate (SrTiOx), zirconium titanate (ZrTiOx), and barium titanate (BaTiOx); and metal nitride thin films such as aluminum nitride (AlN), titanium nitride (TiN), tantalum nitride (TaN), gallium nitride (GaN), niobium nitride (NbN), vanadium nitride (VN), tungsten nitride (WN), copper nitride (CuxN), nickel nitride (NixN), and titanium aluminum nitride (TiAlN). The process parameter determination method for metal compound thin film deposition in this embodiment can also be applied to non-metallic nitride thin films such as hexagonal boron nitride (h-BN) and silicon nitride (SixNy). This application does not impose specific limitations in this regard.

[0103] It is understood that when the process parameter determination method of this embodiment is applied to nitride thin films, it can identify the characteristics of the nitridation state on the target surface based on the relationship between the target voltage and the reactant gas flow rate, and divide the target state into a metallic range, an unsaturated metal nitride range, and a saturated nitride range. The concept of other steps is the same as that of the parameter determination process for oxide thin films, and will not be repeated here.

[0104] In some embodiments, the metal oxide film is a niobium oxide (NbOx) film, and the process parameters for depositing the niobium oxide (NbOx) film using pulsed DC reactive sputtering are trained. The target voltage range corresponding to the unsaturated metal oxide range is 345V~380V. This range corresponds to the main preparation window of the substoichiometry of the niobium oxide (NbOx) film and is a key process range for achieving the target threshold switching characteristics. Based on the hysteresis characteristics of the target voltage changing with oxygen flow rate, the loss weight of the training data in the unsaturated metal oxide range is determined to be 2 to 10 times that of the loss weight of the training data in the metallic range and / or the saturated oxide range.

[0105] For example, in this training process, the weighting coefficient λ is set to 5 for training data in the unsaturated metal oxide range, and to 1 for samples in other ranges. This guides the feature extraction model to learn intermediate feature information that is highly sensitive to voltage fluctuations in the 345V~380V range, thus solving the problems of insufficient model learning and inadequate prediction accuracy caused by the low proportion of training data in the unsaturated metal oxide range and strong nonlinearity.

[0106] In this embodiment, during the model training phase, the first-stage feature extraction model and the second-stage probabilistic prediction model are jointly trained end-to-end. The intermediate feature vector output by the feature extraction head is used as the input to the Gaussian process regression. A loss function is constructed using the prediction error of the target process response (such as the stoichiometry of thin film oxidation), and the weight parameters of the deep neural network are iteratively updated using the backpropagation algorithm. Simultaneously, the Gaussian process regression model is optimized. The trained second-stage probabilistic prediction model can output the corresponding predicted value and confidence information for any input intermediate feature vector.

[0107] The second-stage probabilistic prediction model adopts a Gaussian process regression model, and the covariance function of the Gaussian process regression model adopts a composite kernel function structure. The composite kernel function includes a first kernel function that measures the correlation between intermediate feature information and a second kernel function that characterizes the influence of observation noise.

[0108]

[0109] in, Represents the target process response function; Represent a Gaussian process; This represents the mean function, which is set to the zero mean function in this embodiment; The covariance function is represented by a composite kernel function structure to simultaneously model the similarity between process features and observation noise. The covariance function is as follows:

[0110]

[0111] in, and Let represent the eigenvectors of the i-th and j-th process states; C represents the amplitude coefficient of the kernel function; This represents the Marton kernel function (the first kernel function), used to characterize the smooth correlation between process features; The white noise kernel function (second kernel function) is used to describe the impact of observed noise. The covariance matrix is ​​constructed based on the training sample data.

[0112]

[0113] in, Represents the variance of observation noise. Let be the identity matrix. For any combination of process parameters to be predicted, its corresponding process state characteristics are: .

[0114] For any combination of process parameters to be predicted, the prediction results output by the Gaussian process regression model follow a Gaussian distribution, with the prediction mean and prediction variance as follows:

[0115]

[0116]

[0117] Among them, the predicted mean The prediction variance is used as the prediction result for the target process response. This is used to characterize the level of uncertainty in the prediction results under current process conditions. By simultaneously outputting the prediction mean and prediction variance, this embodiment can not only achieve high-precision prediction of the target process response, but also quantify the prediction reliability of the model in different process parameter ranges, thus providing a basis for subsequent uncertainty-based process parameter screening and experimental supplementation.

[0118] In one embodiment, the prediction results output by the second-stage probabilistic prediction model include the prediction mean and the prediction variance. The prediction mean serves as the predicted value of the thin film response information, and the prediction variance serves as the corresponding prediction confidence information.

[0119] During model training, Bayesian search is used to automatically optimize the hyperparameters of the kernel function of Gaussian process regression (such as the length scale of the Marton kernel, amplitude coefficient, noise variance, etc.) to maximize the marginal likelihood function and improve model performance.

[0120] An early termination condition is set during training: if the loss function on the validation set no longer decreases over multiple consecutive iterations, training is terminated early to prevent overfitting. When the early termination condition is met, the system saves the feature extraction header parameters, solidifying the trained first-stage model; then it saves the Gaussian process regression parameters, solidifying the optimized kernel function hyperparameters and the trained probability model; finally, it enters the data output stage, obtaining the trained process prediction model, which can be used for subsequent process parameter prediction and selection. If the early termination condition is not met, iterative training continues, and the weighted loss function and learning rate scheduling strategy are dynamically adjusted until the model converges. Through the above process, a trained two-stage cascaded process prediction model is obtained, combining the strong feature representation capability of deep learning with the uncertainty quantification capability of Bayesian methods.

[0121] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0122] Based on the same inventive concept, this application also provides a process parameter determination apparatus for metal compound thin film deposition, which implements the above-described method for determining process parameters for metal compound thin film deposition. The solution provided by this apparatus is similar to the solution described in the above-described method. Therefore, the specific limitations of one or more embodiments of the process parameter determination apparatus for metal compound thin film deposition provided below can be found in the limitations of the process parameter determination method for metal compound thin film deposition described above, and will not be repeated here.

[0123] In one exemplary embodiment, such as Figure 5As shown, a process parameter determination device for metal compound thin film deposition is provided, including an acquisition module 201, a storage module 202 and a screening module 203.

[0124] The acquisition module 201 is used to acquire process parameter information, process state information and corresponding film response information during the reactive sputtering preparation of metal compound thin films, and to construct a modeling dataset; the process state information includes at least the target voltage;

[0125] Storage module 202 stores a process prediction model, which includes a first-stage feature extraction model and a second-stage probabilistic prediction model. The first-stage feature extraction model takes the combination of process parameters as input and extracts intermediate feature information that characterizes the state of the sputtering process. The second-stage probabilistic prediction model takes the intermediate feature information extracted in the first stage as input and outputs the predicted value and the corresponding confidence information.

[0126] The screening module 203 takes the target metal compound thin film as the optimization target and, based on the predicted value and corresponding confidence information output by the process prediction model, screens out the combination of process parameters that meet the optimization target and have a confidence level higher than a preset threshold.

[0127] The modules in the aforementioned apparatus for determining process parameters for metal compound thin film deposition can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the operations corresponding to each module.

[0128] In some embodiments, the apparatus for determining process parameters for metal compound thin film deposition further includes a construction module for constructing a process prediction model, comprising: inputting data from a modeling dataset into a neural network for training to obtain a trained first-stage feature extraction model; using intermediate feature information output by the first-stage feature extraction model as input to train a Gaussian process regression model, establishing a probabilistic mapping relationship between the intermediate feature information and the thin film response information, and obtaining a trained second-stage probabilistic prediction model.

[0129] In some embodiments, the second-stage probabilistic prediction model adopts a Gaussian process regression model, and the covariance function of the Gaussian process regression model adopts a composite kernel function structure; the composite kernel function includes a first kernel function that measures the correlation between intermediate feature information and a second kernel function that characterizes the influence of observation noise; the prediction results output by the second-stage probabilistic prediction model include the prediction mean and the prediction variance, with the prediction mean serving as the predicted value of the thin film response information and the prediction variance serving as the corresponding prediction confidence information.

[0130] In some embodiments, the process status information further includes the compound coverage on the target surface. The acquisition module 201 includes a first acquisition unit, used to acquire the compound coverage on the target surface, and based on the difference in electrical properties between the metal target and its corresponding compound, establish a correlation mapping model between the relevant parameters of the reaction gas, the target working voltage, and the coverage; input the collected target working voltage into the correlation mapping model to obtain the corresponding compound coverage on the target surface.

[0131] In some embodiments, the acquisition module 201 further includes a division unit, which is used to identify the hysteresis characteristics of the oxidation state of the target surface based on the relationship between the target voltage and the flow rate of the reactant gas, and to divide the target state into a metallic range, an unsaturated metal oxide range and a saturated oxide range according to the hysteresis characteristics.

[0132] In some embodiments, the process parameter determination device for metal compound thin film deposition further includes an iteration module, which is used to obtain new process parameter information, process state information and corresponding thin film response information from process parameter combinations with confidence levels below a preset threshold within the unsaturated metal oxide range, add them to the modeling dataset, and iteratively update the process prediction model.

[0133] In some embodiments, the building module includes a weight allocation unit, which assigns loss weights to the training data through a weighted loss function during the training process of the first-stage feature extraction model, wherein the loss weight of the training data in the unsaturated metal oxide range is greater than the loss weight of the training data in the metallic range and the saturated oxide range.

[0134] In some embodiments, the process parameter determination device for metal compound thin film deposition further includes an iteration module, which is used to obtain new process parameter information, process state information and corresponding thin film response information from process parameter combinations with confidence levels below a preset threshold within the unsaturated metal oxide range, and add them to the modeling dataset; and use the updated modeling dataset to train the process prediction model.

[0135] In some embodiments, the weight allocation unit is further configured to assign loss weights to the newly added training data during the retraining process, wherein the loss weights of the newly added training data are higher than the loss weights of the original training data in the modeling dataset.

[0136] In some embodiments, the metal compound film is a niobium oxide film; the target voltage range corresponding to the unsaturated metal oxide range is 345V~380V; the loss weight of the training data in the unsaturated metal oxide range is 2 to 10 times that of the loss weight of the training data in the metallic range and / or the saturated oxide range.

[0137] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When executed by the processor, the computer program implements a method for determining process parameters for metal compound thin film deposition.

[0138] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does 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 those shown in the figure, or combine certain components, or have different component arrangements.

[0139] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements a method for determining process parameters for metal compound thin film deposition.

[0140] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements a method for determining process parameters for metal compound thin film deposition.

[0141] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data that have been authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0142] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0143] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0144] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for determining process parameters for deposition of thin films of metal compounds, characterized in that, The method includes: The process parameters, process status information, and corresponding film response information of the reactive sputtering preparation of metal compound thin films are obtained to construct a modeling dataset; the process status information includes at least the target voltage. A process prediction model is constructed, which includes a first-stage feature extraction model and a second-stage probabilistic prediction model. The first-stage feature extraction model takes the combination of process parameters as input and extracts intermediate feature information that characterizes the state of the sputtering process. The second-stage probabilistic prediction model takes the intermediate feature information extracted in the first stage as input and outputs the predicted value and the corresponding confidence information. Using the target metal compound thin film as the optimization target, based on the predicted values ​​and corresponding confidence information output by the process prediction model, process parameter combinations that meet the optimization target and have a confidence level higher than a preset threshold are selected.

2. The method of claim 1, wherein: The construction process prediction model includes: The training data in the modeling dataset is input into the neural network for training to obtain the first-stage feature extraction model after training. The intermediate feature information output by the first-stage feature extraction model is used as input to train the Gaussian process regression model, establishing a probability mapping relationship between the intermediate feature information and the thin film response information, and obtaining the trained second-stage probability prediction model.

3. The method of claim 1, wherein: The second-stage probabilistic prediction model adopts a Gaussian process regression model, and the covariance function of the Gaussian process regression model adopts a composite kernel function structure; the composite kernel function includes a first kernel function that measures the correlation between intermediate feature information and a second kernel function that characterizes the influence of observation noise; The prediction results output by the second-stage probability prediction model include the prediction mean and the prediction variance. The prediction mean is used as the predicted value of the thin film response information, and the prediction variance is used as the corresponding prediction confidence information.

4. The method of claim 1, wherein: The process status information also includes the compound coverage on the target surface; Obtaining the compound coverage on the target surface includes: Based on the differences in electrical properties between metal targets and their corresponding compounds, a correlation mapping model is established between relevant parameters of the reactant gas, the working voltage of the target, and the coverage. The collected target working voltage is input into the correlation mapping model to obtain the corresponding target surface compound coverage.

5. The method for determining process parameters for metal compound thin film deposition according to claim 4, characterized in that, Also includes: Based on the relationship between the target voltage and the flow rate of the reactant gas, the hysteresis characteristics of the oxidation state of the target surface are identified, and the target state is divided into a metallic range, an unsaturated metal oxide range, and a saturated oxide range according to the hysteresis characteristics.

6. The method for determining process parameters for metal compound thin film deposition according to claim 5, characterized in that, The construction process prediction model also includes: During the training process of the first-stage feature extraction model, a weighted loss function is used to assign loss weights to the training data, and the loss weight of the training data in the unsaturated metal oxide range is greater than the loss weight of the training data in the metallic range and the saturated oxide range.

7. The method for determining process parameters for metal compound thin film deposition according to claim 6, characterized in that, Also includes: Based on the process parameter combinations with confidence levels below a preset threshold within the unsaturated metal oxide range, new process parameter information, process state information, and corresponding thin film response information are obtained and added to the modeling dataset. The process prediction model is trained using the updated modeling dataset.

8. The method for determining process parameters for metal compound thin film deposition according to claim 7, characterized in that, During retraining, the weighted loss function assigns loss weights to the newly added training data, and the loss weights for the newly added training data are higher than the loss weights for the original training data in the modeling dataset.

9. The method for determining process parameters for metal compound thin film deposition according to claim 6, characterized in that, The metal compound film is a niobium oxide film; the target voltage range corresponding to the unsaturated metal oxide range is 345V~380V; the loss weight of the training data in the unsaturated metal oxide range is 2 to 10 times that of the loss weight of the training data in the metallic range and / or the saturated oxide range.

10. A device for determining process parameters for metal compound thin film deposition, characterized in that, include: The acquisition module is used to acquire process parameter information, process state information, and corresponding film response information during the reactive sputtering preparation of metal compound thin films, and to construct a modeling dataset. The process status information includes at least the target voltage; The storage module stores a process prediction model, which includes a first-stage feature extraction model and a second-stage probabilistic prediction model. The first-stage feature extraction model takes a combination of process parameters as input and extracts intermediate feature information that characterizes the state of the sputtering process. The second-stage probabilistic prediction model takes the intermediate feature information extracted in the first stage as input and outputs the predicted value and the corresponding confidence information. The screening module takes the target metal compound thin film as the optimization target and, based on the predicted value output by the process prediction model and the corresponding confidence information, screens out the combination of process parameters that meets the optimization target and has a confidence level higher than a preset threshold.

Images