Data processing method and system for optimizing process parameters of vacuum coating

Abstract

The application provides a data processing method and system for optimizing vacuum coating process parameters, comprising: based on the constraint range and parameter dimension of the process parameters, obtaining the actual test process parameters through space filling design, and corresponding quality parameters and actual state data, constructing an initial training sample set and a preliminary proxy model; based on the target quality parameter of the process parameter optimization, driving the proxy model to iteratively explore the process parameter space for global optimization through a global optimization algorithm, obtaining candidate process parameters and performing physical testing, in the case that the difference between the actual quality parameter and the target quality parameter is greater than a preset threshold, re-updating the proxy model and performing global optimization again until the iteration termination condition is met, and taking the candidate process parameters as the optimization result. Thus, the problems of low efficiency, low accuracy and poor optimization effect in the related art of vacuum coating process parameter optimization are solved.

Description

Technical Field

[0001] This invention relates to the field of vacuum coating technology, and in particular to a data processing method and system for optimizing vacuum coating process parameters. Background Technology

[0002] Vacuum deposition is a key technology for preparing functional thin films. Specifically, it involves vaporizing or sputtering materials in a vacuum environment and depositing them onto a substrate surface using physical or chemical methods to form a thin film. It is widely used in high-end manufacturing fields such as semiconductors, optical components, display panels, and precision instruments. Product quality is highly dependent on precise process parameters. Traditional vacuum deposition process formulation development mainly relies on a trial-and-error approach. Engineers set multiple sets of process parameters based on experience, conduct physical experiments, test the quality parameters of the finished products, and then readjust the process parameters based on experience, repeating this cycle until the required process parameters are obtained. However, this method has a long development cycle and is very costly.

[0003] With technological advancements, several methods based on traditional experimental design have emerged, such as the Taguchi method, response surface methodology, single AI prediction models, and pure simulation-driven optimization methods. However, due to the large number and strong nonlinearity of process parameters in vacuum coating, these traditional experimental design methods often suffer from low fitting accuracy and struggle to accurately capture the complex interactions between various parameters. Single AI prediction models are prone to getting trapped in local optima and cannot efficiently explore high-dimensional parameter spaces. Pure simulation-driven optimization algorithms fail to accurately reflect real-world factors such as equipment aging and batch-to-batch material variations, leading to accumulated errors and ultimately low accuracy.

[0004] There is currently no effective solution to the problems of low efficiency, low accuracy, and poor optimization effect in the optimization of vacuum coating process parameters in related technologies. Summary of the Invention

[0005] The present invention provides a data processing method and system for optimizing vacuum coating process parameters, which at least solves the problems of low efficiency, low accuracy and poor optimization effect in the optimization of vacuum coating process parameters in related technologies.

[0006] This invention provides a data processing method for optimizing vacuum coating process parameters, comprising: based on the constraint range and parameter dimensions of the process parameters, obtaining no more than a preset number of actual test process parameters, as well as corresponding quality parameters and actual state data, through space filling design, and constructing an initial training sample set; constructing a preliminary surrogate model based on the initial training sample set, the surrogate model being used to predict the quality parameters corresponding to the process parameters; based on the target quality parameter of the process parameter optimization, driving the surrogate model to iteratively explore the process parameter space for global optimization through a global optimization algorithm to obtain candidate process parameters; conducting physical experiments based on the candidate process parameters to obtain actual quality parameters and actual state data; when the difference between the actual quality parameter and the target quality parameter is greater than a preset threshold, supplementing the candidate process parameters, actual state data, and actual quality parameters into the training sample set, updating the surrogate model, and performing global optimization again until the iteration termination condition is met; when the iteration termination condition is met, using the candidate process parameters as the optimization result.

[0007] This invention also provides a data processing system for optimizing vacuum coating process parameters, comprising: an intelligent model module, used to acquire no more than a preset number of actual test process parameters, corresponding quality parameters, and actual state data through space-filling design based on the constraint range and parameter dimensions of the process parameters, and construct an initial training sample set; constructing a preliminary surrogate model based on the initial training sample set, the surrogate model being used to predict the quality parameters corresponding to the process parameters; an optimization engine module, used to drive the surrogate model to iteratively explore the process parameter space for global optimization based on the target quality parameters optimized by the process parameters, and obtain candidate process parameters; an experiment execution module, used to conduct physical experiments according to the candidate process parameters, and obtain actual quality parameters and actual state data; when the difference between the actual quality parameters and the target quality parameters is greater than a preset threshold, supplementing the candidate process parameters, actual state data, and actual quality parameters to the training sample set, updating the surrogate model, and performing global optimization again until the iteration termination condition is met; when the iteration termination condition is met, the candidate process parameters are taken as the optimization result.

[0008] This invention also provides an electronic device, including: a processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of the preceding descriptions.

[0009] This invention also provides a non-transitory machine-readable medium storing computer instructions, wherein the computer instructions are used to cause the computer to perform any of the methods described above.

[0010] This invention also provides a computer program product, including a computer program / instructions, characterized in that, when the computer program / instructions are executed by a processor, they implement the method described in any one of the above-described embodiments.

[0011] The data processing method for optimizing vacuum coating process parameters provided in this invention, based on the constraints and dimensions of the process parameters, obtains no more than a preset number of actual tested process parameters, along with corresponding quality parameters and actual state data, through space-filling design to construct an initial training sample set. A preliminary surrogate model is then constructed based on this initial training sample set to predict the quality parameters corresponding to the process parameters. Based on the target quality parameters for process parameter optimization, a global optimization algorithm drives the surrogate model to iteratively explore the process parameter space for global optimization, obtaining candidate process parameters. This combines artificial intelligence and optimization algorithms to query and solve for process parameters, accurately capturing the complex relationships between multiple parameters and performing global optimization to ensure the accuracy of candidate process parameters.

[0012] Physical experiments are conducted based on candidate process parameters to obtain actual quality parameters and actual state data. If the difference between the actual quality parameters and the target quality parameters exceeds a preset threshold, the candidate process parameters, actual state data, and actual quality parameters are added to the training sample set. The surrogate model is then updated, and global optimization is performed again until the iteration termination condition is met. The training sample set is dynamically updated to allow the artificial intelligence model to adapt to long-term iterations with changing data. When the iteration termination condition is met, the candidate process parameters are used as the optimization result. This yields process parameters that have been experimentally verified to meet the target quality parameters, ensuring the accuracy of the process parameters.

[0013] By combining data updates, artificial intelligence models, optimization algorithms, and closed-loop experimental verification, not only is the accuracy of the final process parameters guaranteed, but also the optimization efficiency is high, solving the problems of low efficiency, low accuracy, and poor optimization effect in vacuum coating process parameter optimization. Attached Figure Description

[0014] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely some embodiments of the present invention, and those skilled in the art can obtain other embodiments based on these drawings without creative effort.

[0015] Figure 1 This is a flowchart of a data processing method for optimizing vacuum coating process parameters according to an embodiment of the present invention.

[0016] Figure 2 This is a schematic diagram of a data processing system for optimizing vacuum coating process parameters, according to an embodiment of the present invention.

[0017] Figure 3 This is a schematic diagram of the structure of the electronic device created by this invention. Detailed Implementation

[0018] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0019] The core logic of vacuum coating is to transform the target material into gaseous particles through physical / chemical means in a vacuum environment, and then deposit, nucleate, and grow them on the surface of the substrate to be coated, ultimately forming a dense and uniform functional thin film. According to the technology type, it can be divided into "sputtering coating," "evaporation coating," and "ion plating," among which sputtering coating is the most widely used in high-end fields such as semiconductors, optical components, and display panels.

[0020] Taking magnetron sputtering as an example, the vacuum coating process includes: first, cleaning the substrate and the environment through pretreatment operations such as cleaning the electrode and evacuating the vacuum. Then, an inert gas or reactive gas is introduced into the closed cavity where the electrode is located, and the working pressure of the cavity is stabilized.

[0021] Then, power is applied to the target material through radio frequency or DC power supply, which ionizes the inert gas to generate plasma. The plasma generated by the plasma bombards the target surface at high speed under the action of electric field, knocking out the target atoms / molecules. The sputtered target particles move towards the substrate in a vacuum environment, adsorb and nucleate on the substrate surface. By adjusting the substrate temperature and deposition time, the particles are grown in an orderly manner into a dense thin film.

[0022] Further processing, such as annealing and cooling / depressurization, can be performed as needed. Finally, quality inspection is conducted to obtain the quality parameters of the finished film layer.

[0023] The entire vacuum coating process described above involves more than ten parameters, such as target power, working pressure, gas flow rate, and deposition time. These parameters also exhibit strong nonlinear interactions. Traditional design methods, such as the Taguchi method / response surface methodology, are only applicable to low-dimensional linear / weakly nonlinear scenarios and cannot effectively handle them.

[0024] In high-end manufacturing sectors such as semiconductors, optical components, and display panels, the precision requirements for thin film quality parameters, such as refractive index, hardness, and adhesion, are ±0.1% to ±1%. Simultaneously, these sectors face fierce market competition and necessitate shortening R&D cycles and reducing production costs. However, the trial-and-error approach is generally inefficient, time-consuming, and costly, and it cannot balance precision and efficiency, thus failing to meet industry demands.

[0025] Moreover, the development of traditional vacuum coating processes relies on the experience of engineers. Senior engineers need 5 to 10 years of experience to accurately adjust parameters, which makes it difficult to pass on the process and limits the efficiency of research and development due to personnel limitations.

[0026] Therefore, embodiments of the present invention provide a data processing method for optimizing vacuum coating process parameters, such as... Figure 1 As shown, it includes:

[0027] Step S101: Based on the constraint range and parameter dimension of the process parameters, obtain no more than a preset number of actual test process parameters, as well as the corresponding quality parameters and actual state data through space filling design, and construct an initial training sample set;

[0028] Step S102: Construct a preliminary surrogate model based on the initial training sample set. The surrogate model is used to predict the quality parameters corresponding to the process parameters.

[0029] Step S103: Based on the target quality parameters of the process parameter optimization, the proxy model is driven by a global optimization algorithm to iteratively explore the process parameter space for global optimization and obtain candidate process parameters.

[0030] Step S104: Conduct physical experiments based on candidate process parameters to obtain actual quality parameters and actual state data;

[0031] Step S105: If the difference between the actual quality parameter and the target quality parameter is greater than the preset threshold, the candidate process parameters, actual state data and actual quality parameters are added to the training sample set, the surrogate model is updated again and global optimization is performed again until the iteration termination condition is met.

[0032] Step S106: If the iteration termination condition is met, the candidate process parameters are taken as the optimization result.

[0033] The above steps can be executed by a processor or server to optimize the process parameters of vacuum coating as needed. The data processing method for optimizing vacuum coating process parameters combines artificial intelligence models and optimization algorithms to query and solve for the optimal process parameters. This not only accurately captures the complex relationships between multiple parameters but also performs global optimization to ensure the accuracy of candidate process parameters. Furthermore, the training sample set can be dynamically updated to allow the artificial intelligence model to adapt to long-term iterations of data changes. Through actual experimental verification, process parameters that meet the target quality parameters are obtained, ensuring the accuracy of the process parameters. By combining data updates, artificial intelligence models, optimization algorithms, and closed-loop experimental verification, not only is the accuracy of the final process parameters guaranteed but also high optimization efficiency is achieved, solving the problems of low efficiency, low accuracy, and poor optimization results in vacuum coating process parameter optimization.

[0034] The aforementioned process parameters may include core process parameters, auxiliary process parameters, environmental and equipment status parameters, and substrate / target pretreatment parameters. Core process parameters for vacuum coating may include at least one of the following: target power, working gas pressure, gas flow rate, deposition time, and substrate temperature.

[0035] Auxiliary process parameters may include at least one of the following: substrate bias, target-substrate spacing, target current, magnetic field strength, gas partial pressure ratio, deposition rate, etc. Environmental and equipment status parameters include at least one of the following: cavity vacuum level, vacuum pump speed, cavity wall temperature, plasma density, gas purity, target purity and density, etc. Substrate / target pretreatment parameters include substrate cleaning parameters such as ultrasonic cleaning power and cleaning time; target pretreatment parameters such as target pre-sputtering time and target baking temperature; and substrate preheating temperature, etc.

[0036] Since core process parameters have a more significant impact on quality, this embodiment uses core process parameters as an example for optimization. In actual implementation, other process parameters only need to be added according to requirements. Target power, which is the electrical power applied to the target, typically ranges from 100 to 500 W. Working gas pressure, which is the gas pressure inside the coating chamber (unit: Pa), is determined by both the vacuum pump power and the working gas supply power, and ranges from 10⁻⁻⁴. 5 ~10⁻¹ Pa.

[0037] Gas flow rate, also known as the volume of gas entering the cavity per unit time (unit: sccm, standard cubic centimeters per minute), includes the flow rate of inert gas and the flow rate ratio of reactant gas. The inert gas is generally argon (Ar), with a flow rate ranging from 20 to 100 sccm. The reactant gases include oxygen and nitrogen, with an O2 / N2 flow rate ranging from 5 to 20 sccm.

[0038] Deposition time, also known as the total time from the start of sputtering on the target to its stop (unit: min), is a direct parameter controlling the film thickness, with a value ranging from 5 to 60 min. Substrate temperature, also known as the temperature of the substrate during the coating process (unit: °C), ranges from 80 to 300 °C.

[0039] The above quality parameters include core optimization indicators and basic constraint indicators: the core optimization indicator is a single objective, selected from at least one of the following: film thickness uniformity, transmittance, hardness, adhesion, refractive index; the basic constraint indicators are other quality requirements that need to be met, selected from at least one of the following: film thickness deviation ≤ ±2nm, adhesion ≥ 4B grade, refractive index fluctuation ≤ ±0.02, hardness ≥ 5H.

[0040] If the difference between the actual quality parameters and the target quality parameters is greater than a preset threshold, the candidate process parameters, actual state data, and actual quality parameters are added to the training sample set, the training sample set is updated, and the next iteration cycle begins. Based on the updated training sample set, steps S101-S102 are executed again to update the surrogate model, and steps S103-S104 are executed again to perform global optimization and obtain the corresponding candidate process parameters and actual quality parameters.

[0041] By selecting different values ​​within the range, different combinations of process parameters can be formed. Initially, when there is no historical data, multiple combinations of process parameters can be selected, and corresponding quality parameters can be obtained through multiple physical experiments. This forms multiple sets of effective samples as initial training data, providing real training data that covers the core parameter space for the training of the artificial intelligence model. This ensures that the artificial intelligence model can learn the real mapping relationship between process and quality, avoiding the lack of accuracy caused by pure simulation data.

[0042] As an optional approach, step S101 above, based on the constraint range and parameter dimensions of the process parameters, obtains no more than a preset number of actual test process parameters, as well as corresponding quality parameters and actual state data, through space-filling design, and constructs an initial training sample set, including: selecting Latin hypercube design to generate multiple sets of process parameters according to the constraint range and parameter dimensions of the process parameters; determining the corresponding quality parameters for each set of process parameters through multiple actual tests; standardizing the data format of the process parameters and corresponding quality parameters, and removing outliers exceeding the normal value range based on the 3σ principle; and using each set of process parameters, actual test state data, and corresponding quality parameters as a set of training samples to form a training sample set.

[0043] The space-filling design described above can be a Latin hypercube design. Process parameters are actively controllable decision variables. Based on their constraint range and parameter dimensions, 30-50 sets of process parameters are generated through Latin hypercube design. The constraint range of these process parameters is also the value range of each parameter, and the parameter dimension is also the parameter type; different types of process parameters represent corresponding parameter dimensions.

[0044] By employing Latin hypercube design (LHD), multiple sets of process parameter combinations that uniformly cover the parameter space are generated within the constraints of process parameters. This ensures that the selected process parameter combinations can cover the key areas of the parameter space, avoids underfitting of the model caused by blind sampling, and improves the model's generalization ability to different parameter combinations.

[0045] The corresponding quality parameters were determined through multiple actual tests. Specifically, multiple repeated experiments were conducted for each combination of process parameters, and the average value of the quality parameters was taken as the quality parameter of that combination of process parameters. This reduced the impact of random errors in a single experiment on the sample labels and improved the reliability of the training data.

[0046] The aforementioned data format standardization allows for the normalization of process parameters and quality parameters. By employing the 3σ principle, outlier samples with substandard quality parameters caused by equipment malfunctions or other non-core process parameters are eliminated. This removes interference from parameter scale differences in AI model training, prevents outliers from misleading the AI ​​model's learning, and improves the training efficiency and fitting accuracy of the AI ​​model.

[0047] The training sample data includes input process parameters, such as preset parameter combinations including target power, gas flow rate, working pressure, substrate temperature, deposition time, etc., and actual state data, such as actual operating parameters collected during the experiment, including actual vacuum degree, actual gas flow rate, actual substrate temperature profile, etc., as well as output quality parameters of the finished film, such as the thickness, uniformity, and adhesion of the finished film.

[0048] By inputting process parameters and actual state data, artificial intelligence models can simultaneously learn the dual mapping relationship between ideal parameter design and industrial reality deviation, solving the problem of inaccurate predictions and poor robustness of models trained solely on process parameters when implemented in industry.

[0049] The inputs and outputs of training samples can be stored in a one-to-one correspondence to form a structured training sample set, which provides structured and easily readable data for training artificial intelligence models, improves model training efficiency, and facilitates subsequent sample set updates and management.

[0050] An artificial intelligence model is trained based on a training sample set, serving as a surrogate model for the global optimization algorithm. Specifically, the AI ​​model can be a machine learning regression model, including one or more combinations of Gaussian process regression, deep neural network models, gradient boosting tree models, random forest or support vector regression models, large-scale structured data models, large-scale tabular data models, and large-scale time series models. Training is performed using the input and output of the training samples to obtain the surrogate model for global optimization.

[0051] Replacing most of the physical experiments in the optimization of vacuum coating process parameters with artificial intelligence models significantly reduces the trial-and-error costs and time costs in the optimization process, while accurately capturing the nonlinear mapping of process parameters.

[0052] As an optional approach, the surrogate model is a tabular prior data fitting network (TabPFN). A preliminary surrogate model is constructed based on the initial training sample set, including: performing dimensionality adaptation on the initial training sample set; if the feature dimensions of the initial training sample set exceed a preset dimensionality threshold, using mutual information analysis to filter and crop core features to within the preset dimensionality threshold; if the feature dimensions of the initial training sample set are less than the preset dimensionality threshold, then zero-padding and scaling are performed to within the preset dimensionality threshold; the pre-trained tabular prior data fitting network is directly loaded, and the initial training sample set is input into the model for context learning to obtain the surrogate model.

[0053] TabPFN's core advantage lies in in-context learning and fast inference. The encapsulation must retain this feature while adapting to Optuna's objective function interface, taking a trial object as input and outputting a quantized target value. Seamless adaptation is achieved through trial runs. The objective function is directly compatible with the Optuna interface, supporting parallel computation (n_jobs=-1) without incurring additional optimization overhead.

[0054] The initial training sample set used to train TabPFN has a small number of training samples, usually no more than 50, and can be 10-30, that is, the range of the above-mentioned preset number does not exceed 50.

[0055] Preferred sampler: TPE sampler (Tree-Structured Parzen Estimator) – adaptable to mixed-type parameters, supports Bayesian optimization logic, and has strong exploration-utilization balancing capabilities.

[0056] Pruning strategy selection: MedianPruner (median pruning) – avoids invalid iterations and improves search efficiency; configuration: n_warmup_steps=3 (no pruning in the first 3 iterations to accumulate enough data), interval_steps=1 (pruning judgment is performed in each iteration).

[0057] Pruning strategies can reduce more than 40% of invalid iterations (such as parameters that are significantly worse than the historical best), and shorten the number of iterations in the optimization cycle.

[0058] The Bayesian logic of the TPE sampler, combined with the accurate prediction of TabPFN, greatly improves the global optimal solution hit rate.

[0059] In addition, TabPFN inputs must be purely numerical, with no missing values, a small sample size (≤1000 groups), and ≤100 features; the input data must be strictly aligned to these criteria. Preprocessing of the input data may include at least one of the following: unifying data types, standardization, dimensional and format adaptation, etc.

[0060] Data types are standardized. Process parameters: Discrete parameters (such as target type: silicon target / titanium target) use one-hot encoding (silicon target = [1,0], titanium target = [0,1]); actual state data (such as temperature curve fluctuation values) are averaged and converted into single values. Quality parameters: Discrete grades (adhesion 3B~5B) are quantified as 3~5, and continuous indicators (film thickness, uniformity) retain their original values ​​and will be standardized subsequently.

[0061] This ensures that the training data can meet the TabPFN input constraints, avoiding prediction errors caused by data format issues, and improving prediction accuracy by 10% to 15% after standardization.

[0062] For handling missing and outlier values: missing values ​​are filled with 0 and labeled after filling. Such data will be added first when updating samples in the future. Outliers: the above 3σ principle is reused to remove them and then fill them to avoid contaminating the training data.

[0063] The above standardization process standardizes both the input process parameters and the output quality parameters. The process parameters (input) are Z-score standardized (mean = 0, standard deviation = 1), using the formula: \(x_{norm}=\frac{x-\mu}{\sigma}\), eliminating the scale difference between power (100~500W) and air pressure (1e-3~1e-1Pa).

[0064] The quality parameters (output) are normalized to [0,1] through single-objective optimization, and then normalized after weighting according to the weights during multi-objective optimization to ensure that the objective function values ​​are comparable.

[0065] This improves data quality. After outlier removal, the prediction bias of TabPFN decreased from ±5% to within ±3%, providing a reliable basis for calculating the objective function value.

[0066] The above-mentioned dimension and format adaptation includes: dimension pruning: if the number of process parameters + actual state data dimensions > 100, core parameters are filtered through mutual information analysis (such as retaining the top 100 highly relevant features such as target power and air pressure); format conversion: saved as a NumPy array or PandasDataFrame, arranged in the column order of "process parameters + actual state data → quality parameters" to adapt to the set-valued input requirements of TabPFN.

[0067] After dimensional clipping, it is compatible with the 100-dimensional feature limit of TabPFN, while retaining core influencing factors and not losing key information.

[0068] Define the search space and numerical constraints for Optuna process parameters, delineate the legal search boundaries, and avoid invalid exploration. The aforementioned numerical constraints are also the basis for the above-mentioned legality verification.

[0069] The pre-trained tabular prior data fitting network (TabPFN) is directly loaded, and the initial training sample set is input into the model for in-context learning. The training samples are cached for inference, and no additional iterative training is required to complete the initial proxy model construction.

[0070] As an optional approach, step S103 involves using a global optimization algorithm to drive a proxy model to iteratively explore the process parameter space for global optimization based on the target quality parameters optimized by the process parameters. This process includes: using the sampler of the global optimization algorithm to search the process parameter space and generate candidate process parameters; verifying the validity of the candidate process parameters; if the verification fails, entering a new round of iteration to regenerate candidate process parameters and accumulating the number of iterations; if the verification passes, inputting the candidate process parameters and corresponding actual state data into the proxy model to predict the core quality parameters; calculating the objective function value based on the optimization objective of the global optimization algorithm and the predicted core quality parameters; adjusting the parameter search strategy based on the objective function value, entering a new round of iteration, and accumulating the number of iterations; and, if the termination condition is met, selecting the candidate process parameters from the valid candidate process parameters that have the optimal objective function value and the highest prediction stability.

[0071] When the optimization objective is clear, the Optuna optimization framework defaults to "minimizing". Therefore, the desired objective, maximizing the quality parameters, is transformed into maximizing the negative of the objective function.

[0072] The TabPFN network, which fits tabular prior data, is used as the sole objective function. The objective function value is calculated solely using the prediction quality parameters output by TabPFN, without introducing other prediction models.

[0073] Optuna's sampler selection can be adapted based on the type of process parameters. For continuous parameters (such as power and gas pressure), the TPE sampler (Tree-Structured Parzen Estimator) is preferred. For high-dimensional parameters (>8 dimensions), the CMA-ES sampler is used, which has stronger global optimization capabilities.

[0074] Based on the constraints of the vacuum coating equipment and the feasibility of the process, the range, type, and step size of each parameter are precisely defined. For the initial parameters of the sampler, the "exploration priority" mode is activated in the first 30% of the iteration, that is, the gamma of the TPE sampler is 0.1, covering the core area of ​​the parameter space; in the latter 70%, the "utilization priority" mode is switched, that is, the gamma of the TPE sampler is 0.5, focusing on fine-tuning the optimal area.

[0075] The above-mentioned legality verification can be performed using a pre-created verification rule library. The sources of these rules can include equipment manuals, process manuals, safety specifications, expert experience, etc. Specifically, they can include necessary constraint verifications, such as equipment limits: target power ≤ 500W (power supply rated), working gas pressure ≥ 1e-3Pa (vacuum pump protection), substrate temperature ≤ 350℃ (substrate tolerance); safety specifications: target temperature ≤ 500℃ (to prevent melting), cavity pressure fluctuation ≤ ±10% (to avoid uneven film formation), etc. They can also include non-essential constraint verifications, such as a gas flow ratio Ar / O2 ≥ 1:1 (minimum inert gas percentage in reactive sputtering), deposition time ≥ 5min (film formation threshold); parameter coordination constraints: at high power (>400W), the substrate temperature must be ≤ 250℃ (to avoid excessive film stress), etc.

[0076] It can filter out illegal parameters to avoid equipment damage and experimental waste, thereby improving experimental efficiency. In addition, necessary constraint verification can avoid safety risks such as equipment overload and target melting, and comply with industrial production standards.

[0077] Once the validation is successful, candidate process parameters that meet the TabPFN input requirements will be used to predict quality parameters. It should be noted that TabPFN does not require retraining. Utilizing in-context learning, the prediction time for a single set of candidate parameters is <0.1s (GPU), representing a 3600-fold efficiency improvement compared to the Transformer proxy model (which requires hours for training and prediction).

[0078] Moreover, stable predictions are possible with only a few historical samples, typically ≤1000 samples, eliminating the need for extensive experimental data accumulation and lowering the barrier to industrial implementation. Integrated prediction reduces the prediction error of quality parameters to ≤3%, providing a reliable basis for calculating the objective function value.

[0079] In single-objective optimization scenarios, the objective function value is determined by the quality parameters predicted by TabPFN and the optimization objective. The closer the objective function value is to the optimization objective, the higher the objective function value will be.

[0080] By calculating the objective function value, TabPFN and Optuna are connected to construct a collaborative optimization mechanism.

[0081] The search strategy adjustment is based on Optuna's adaptive mechanism. The TPE sampler can update the probability distribution of optimal parameter regions based on historical objective function values. Parameter regions with higher objective function values ​​have higher subsequent sampling probabilities; regions that are not fully explored are kept under-explored through sampling diversity.

[0082] Optuna's pruning strategy triggers MedianPruner pruning if the objective function value of the current iteration is much lower than the historical best value, thus terminating the evaluation of that parameter early.

[0083] Iteration count accumulation: The iteration counts of valid parameters and invalid iteration counts are accumulated together to form a continuous iteration sequence. The accumulated count is synchronized to the Optuna study object in real time, and the termination condition is supported for real time judgment.

[0084] The adaptive mechanism allows for more focused and regional targeting, dynamically adjusting the search strategy to gradually focus Optuna towards regions with high objective function values. Compared to a fixed-strategy search, this significantly improves convergence speed. Furthermore, it reduces unnecessary computation, with pruning strategies eliminating poor parameters early, avoiding meaningless TabPFN predictions and experimental verifications, and saving computational resources.

[0085] The cumulative number of iterations is the maximum number of iterations. Termination conditions such as multiple consecutive periods without improvement provide data support to ensure that the optimization process is controllable.

[0086] When screening candidate process parameters, the top-N candidate process parameters with the best objective function values ​​are selected from all valid iterations. For the top-N parameters, TabPFN is used to perform integrated prediction again, and the group with the most stable prediction quality parameters is selected as the final candidate process parameters to avoid accidental good values.

[0087] This ensures global optimality and, except for accidental good values, guarantees the stability of the quality parameters of the candidate process parameters.

[0088] As an optional approach, before generating candidate process parameters by searching the process parameter space using the sampler of the global optimization algorithm according to the optimization objective, the method further includes: defining the optimization objective of the global optimization algorithm: maximizing the predicted value of a single core quality parameter, or minimizing the deviation between the single core quality parameter and the target value; performing data preprocessing on the initial training sample set, wherein the processed samples must meet the input requirements of the surrogate model: pure numerical values, no missing values, and consistent distribution after standardization; defining the process parameter search space and numerical constraints of the global optimization algorithm, and clarifying the type, range, and search step size of each parameter.

[0089] Define the optimization objective of the global optimization algorithm as follows: maximize the predicted value of a single core quality parameter, or minimize the deviation between the single core quality parameter and the target value. The basic constraint indicators must meet the preset requirements, such as adhesion ≥ 4B grade and refractive index fluctuation ≤ ±0.02.

[0090] The initial training sample set is preprocessed to ensure that the processed samples meet the input requirements of the surrogate model: pure numerical values, no missing values, and consistent distribution after standardization. The search space and numerical constraints for the process parameters of the global optimization algorithm are defined, specifying the type, range, and step size of each parameter (the step size should match the equipment control precision). An example is shown below.

[0091] Target power: continuous type, 100~500W, step size 10W;

[0092] Working air pressure: continuous type, 1e-3~1e-1Pa, logarithmic scale;

[0093] Ar / O2 flow ratio: Discrete type, selectable "1:1, 3:1, 5:1, 10:1";

[0094] Deposition time: Integer type, 5~60 min, step size 5 min;

[0095] Encapsulate the objective function: Embed the proxy model into the objective function. The objective function receives candidate process parameters, calls the proxy model to predict core quality parameters, calculates and returns the scalar objective function value according to the optimization objective, and adapts to the calling interface of the global optimization algorithm.

[0096] Choose a pruning strategy for the global optimization algorithm: Use the median pruning strategy, set n_warmup_steps=3 (no pruning in the first 3 iterations) and interval_steps=1 (pruning is performed in each iteration). If the objective function value of the current iteration is lower than the historical median, the evaluation of this parameter is terminated.

[0097] As an optional approach, the iteration termination condition includes at least one of the following: the number of iterations reaches the maximum number of iterations; the improvement of the optimal target value in multiple consecutive iterations is less than a preset threshold; the predicted quality parameter reaches the target quality parameter.

[0098] The maximum number of iterations mentioned above can be 20 to 50 in the scenario of optimizing vacuum coating process parameters. For large models with Transformer architecture, where prediction costs are low, the number of iterations can be appropriately increased to 30 to 70.

[0099] If the improvement in the optimal target value after multiple iterations is less than a preset threshold close to zero (e.g., if the optimal target value after 5 iterations is ≤0.1%), it indicates that the iteration is converging, and further iteration is less meaningful. If the predicted quality parameter reaches the target quality parameter, it means the optimization objective has been achieved, and further iteration should be discontinued.

[0100] Optionally, if the variance of the objective function value of the candidate parameters is ≤0.01 for 10 consecutive iterations, it indicates that no better combination of process parameters has been generated, and the iteration can be stopped.

[0101] When using the above iterative termination and submission methods simultaneously, priority sorting needs to be set to avoid confusion in triggering conditions. Possible priorities are: reaching the target quality parameter > continuous improvement value less than a preset threshold > reaching the maximum number of iterations. Clearly defining the termination boundary of Bayesian optimization avoids resource waste caused by infinite iteration, while ensuring that the optimization result meets quality requirements or converges to a stable state, balancing optimization efficiency and effectiveness.

[0102] The above-mentioned multiple iteration termination conditions can have different priorities. In an optional embodiment, it can first determine whether the target quality parameter has been reached, then determine whether the improvement of the optimal target value in multiple consecutive iterations is less than a preset threshold, which means that there is no significant quality improvement in multiple iterations, and finally determine whether the maximum number of iterations has been reached.

[0103] As an optional approach, when the difference between the actual quality parameters and the target quality parameters exceeds a preset threshold, before supplementing the training sample set with candidate process parameters, actual state data, and actual quality parameters, the method further includes: after obtaining the actual experimental results, updating the iteration count and determining whether the iteration count has reached a preset threshold; if the iteration count reaches the preset threshold and the experimental results meet the target quality parameters, then the candidate process parameters are used as the final optimization result; if the iteration count reaches the preset threshold and the experimental results do not fully meet the target quality parameters, then the candidate process parameters are used as the final optimization result, and the deviation from the target quality parameters is calculated.

[0104] In the process of optimizing process parameters, an iteration number threshold constraint is introduced to avoid infinite loops caused by long-term failure of experimental results to meet the target. At the same time, it ensures that the target quality parameters are approached as close as possible within a reasonable iteration range, taking into account both the practicality and economy of process optimization.

[0105] The preset threshold for the number of iterations can be set to 10. The number of iterations is updated after each experimental verification to avoid infinite iterations caused by inherent defects in the equipment or excessively high quality targets, thereby controlling R&D costs.

[0106] When the preset threshold is reached, the current candidate process parameters are taken as the final result. If, after the 10th iteration, the experimental results are still not fully satisfactory but are close to the target, the current optimal parameters are taken as the final result, and the deviation is noted in the report. A balance is struck between quality requirements and R&D feasibility to avoid R&D stagnation caused by excessive pursuit of perfection, ensuring the feasibility of the solution.

[0107] This invention also provides a data processing system for optimizing vacuum coating process parameters, such as... Figure 2 As shown, it includes:

[0108] The intelligent model module 21 is used to obtain no more than a preset number of actual test process parameters, as well as corresponding quality parameters and actual state data, based on the constraint range and parameter dimensions of process parameters through space filling design, and to construct an initial training sample set; based on the initial training sample set, a preliminary surrogate model is constructed, which is used to predict the quality parameters corresponding to the process parameters; the optimization engine module 22 is used to optimize the target quality parameters based on the process parameters, and to drive the surrogate model to iteratively explore the process parameter space for global optimization through a global optimization algorithm to obtain candidate process parameters; the test execution module 23 is used to conduct physical tests based on the candidate process parameters to obtain actual quality parameters and actual state data; if the difference between the actual quality parameters and the target quality parameters is greater than a preset threshold, the candidate process parameters, actual state data, and actual quality parameters are added to the training sample set, the surrogate model is updated again, and global optimization is executed again until the iteration termination condition is met; if the iteration termination condition is met, the candidate process parameters are used as the optimization result.

[0109] As an optional solution, it also includes: a data management module 24, which is used to store historical data and filter training samples based on historical data; and to receive and store candidate process parameters and experimental results, and update the training sample set.

[0110] The data processing system for optimizing vacuum coating process parameters described above, when executing any of the data processing methods for optimizing vacuum coating process parameters, achieves the same results as the methods described above, and will not be elaborated further here.

[0111] Embodiments of the present invention also provide a non-transitory machine-readable medium storing a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform a method according to an embodiment of the present invention.

[0112] Embodiments of the present invention also provide a computer program product, including a computer program, wherein the computer program, when executed by a computer's processor, is used to cause the computer to perform the method of an embodiment of the present invention.

[0113] An embodiment of the present invention also provides an electronic device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, causes the electronic device to perform the method of the embodiment of the present invention.

[0114] refer to Figure 3 The present invention will now describe a structural block diagram of an electronic device that can serve as an embodiment of the present invention, serving as an example of a hardware device applicable to various aspects of the present invention. The electronic device is intended to represent various forms of digital electronic computer devices, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0115] like Figure 3 As shown, the electronic device includes a computing unit 301, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 302 or a computer program loaded from a storage unit 308 into a random access memory (RAM) 303. The RAM 303 may also store various programs and data required for the operation of the electronic device. The computing unit 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0116] Multiple components in the electronic device are connected to I / O interface 305, including: input unit 306, output unit 307, storage unit 308, and communication unit 309. Input unit 306 can be any type of device capable of inputting information into the electronic device. Input unit 306 can receive input digital or character information and generate key signal inputs related to user settings and / or function control of the electronic device. Output unit 307 can be any type of device capable of presenting information and may include, but is not limited to, a display, speaker, video / audio output terminal, vibrator, and / or printer. Storage unit 308 may include, but is not limited to, disks and optical discs. Communication unit 309 allows the electronic device to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks, and may include, but is not limited to, modems, network cards, infrared communication devices, and / or wireless communication transceivers, such as Bluetooth devices, WiFi devices, WiMax devices, cellular communication devices, and / or the like.

[0117] The computing unit 301 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 301 include, but are not limited to, CPUs, graphics processing units (GPUs), various special-purpose artificial intelligence (AI) computing units, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any suitable processor, controller, microcontroller, etc. The computing unit 301 performs the various methods and processes described above. For example, in some embodiments, the method embodiments of the present invention can be implemented as computer programs tangibly contained in a machine-readable medium, such as storage unit 308. In some embodiments, part or all of the computer program can be loaded and / or installed on an electronic device via ROM 302 and / or communication unit 309. In some embodiments, the computing unit 301 can be configured to perform the methods described above by any other suitable means (e.g., by means of firmware).

[0118] Computer programs for implementing the methods of embodiments of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0119] In the context of embodiments of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable signal medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, or infrared systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0120] It should be noted that the term "comprising" and its variations used in the embodiments of this invention are open-ended, meaning "including but not limited to". The term "based on" means "at least partially based on". The term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments". The modifications of "one" and "a plurality" mentioned in the embodiments of this invention are illustrative and not restrictive, and those skilled in the art should understand that unless explicitly indicated otherwise in the context, they should be understood as "one or more".

[0121] 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, stored data, displayed data, etc.) involved in the embodiments of this invention are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0122] The steps described in the method embodiments provided by the present invention can be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of protection of the present invention is not limited in this respect.

[0123] The term "embodiment" in this specification refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily imply the same embodiment, nor does it imply independence or alternativeity from other embodiments. The various embodiments in this specification are described in a related manner, with reference to each other for similar or identical parts. In particular, for apparatus, device, and system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, and relevant details are referred to in the description of the method embodiments.

[0124] The above-described embodiments are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A data processing method for optimizing vacuum coating process parameters, characterized in that, include: Based on the constraints and dimensions of the process parameters, the space-filling design is used to obtain no more than a preset number of actual test process parameters, as well as the corresponding quality parameters and actual state data, to construct an initial training sample set. A preliminary proxy model is constructed based on the initial training sample set. The proxy model is used to predict the quality parameters corresponding to the process parameters. Based on the target quality parameters optimized by the process parameters, the proxy model is driven by a global optimization algorithm to iteratively explore the process parameter space for global optimization and obtain candidate process parameters. Physical experiments were conducted based on the candidate process parameters to obtain actual quality parameters and actual state data. If the difference between the actual quality parameter and the target quality parameter is greater than a preset threshold, the candidate process parameters, actual state data and actual quality parameters are added to the training sample set, the surrogate model is updated again and global optimization is performed again until the iteration termination condition is met. If the iteration termination condition is met, the candidate process parameters are taken as the optimization result.

2. The method according to claim 1, characterized in that, Based on the constraints and dimensions of the process parameters, an initial training sample set is constructed by obtaining no more than a preset number of actual test process parameters, along with corresponding quality parameters and actual state data, through space-filling design. This sample set includes: Based on the constraint range and parameter dimensions of the process parameters, Latin hypercube design is selected to generate multiple sets of process parameters; For each set of process parameters, the corresponding quality parameters were determined through multiple actual tests. The process parameters and corresponding quality parameters are standardized in data format, and outliers that exceed the normal range are removed based on the 3σ principle. Each set of process parameters, actual test data, and corresponding quality parameters are used as a training sample to form a training sample set.

3. The method according to claim 2, characterized in that, The surrogate model is a tabular prior data fitting network. A preliminary surrogate model is constructed based on the initial training sample set, including: The initial training sample set is subjected to dimensionality adaptation processing; If the feature dimension of the initial training sample set exceeds a preset dimension threshold, core features are filtered and cropped to within the preset dimension threshold through mutual information analysis. If the feature dimension of the initial training sample set is less than the preset dimension threshold, then zero padding and scaling are performed to bring it within the preset dimension threshold. The pre-trained tabular prior data is directly loaded into the fitting network, and the initial training sample set is input into the model for context learning to obtain the surrogate model.

4. The method according to claim 1, characterized in that, Based on the target quality parameters optimized from the aforementioned process parameters, a global optimization algorithm drives the surrogate model to iteratively explore the process parameter space for global optimization, resulting in candidate process parameters, including: The sampler of the global optimization algorithm is used to search the process parameter space and generate candidate process parameters; The validity of the candidate process parameters is verified. If the verification fails, a new round of iterations is initiated to regenerate candidate process parameters, and the number of iterations is accumulated. If the verification passes, the candidate process parameters and the corresponding actual state data are input into the surrogate model to predict the core quality parameters; The objective function value is calculated based on the optimization objective of the global optimization algorithm and the predicted core quality parameters; Adjust the parameter search strategy according to the objective function value, enter a new round of iteration, and accumulate the number of iterations; If the termination condition is met, the candidate process parameter with the optimal objective function value and the highest prediction stability is selected from the legal candidate process parameters.

5. The method according to claim 4, characterized in that, Before generating candidate process parameters by searching the process parameter space using the sampler of the global optimization algorithm according to the optimization objective, the method further includes: The optimization objective of the global optimization algorithm is defined as: maximizing the predicted value of a single core quality parameter, or minimizing the deviation between the single core quality parameter and the target value; The initial training sample set is preprocessed, and the processed samples must meet the input requirements of the surrogate model: pure numerical values, no missing values, and consistent distribution after standardization. Define the process parameter search space and numerical constraints for the global optimization algorithm, and clarify the type, range and search step size of each parameter.

6. The method according to claim 4, characterized in that, The iteration termination condition includes at least one of the following: The maximum number of iterations has been reached. The improvement in the optimal target value after multiple consecutive iterations is less than a preset threshold; The predicted quality parameters meet the target quality parameters.

7. The method according to claim 1, characterized in that, If the difference between the actual quality parameter and the target quality parameter is greater than a preset threshold, before supplementing the candidate process parameters, actual state data, and actual quality parameters into the training sample set, the method further includes: After obtaining the actual test results, update the iteration count and determine whether the iteration count has reached a preset threshold. If the number of iterations reaches a preset threshold and the experimental results meet the target quality parameters, then the candidate process parameters are taken as the final optimization results. If the number of iterations reaches a preset threshold and the experimental results do not fully meet the target quality parameters, then the candidate process parameters are taken as the final optimization results, and the deviation between them and the target quality parameters is calculated.

8. The method according to any one of claims 1 to 7, characterized in that, The process parameters include core process parameters and actual status data; The core process parameters include at least one of the following: target power, working gas pressure, gas flow rate, deposition time, and substrate temperature; The actual state data includes at least one of the following: actual vacuum level, actual temperature curve, and actual gas flow rate fluctuation value; The quality parameters include core optimization indicators and basic constraint indicators. The core optimization indicators are single objectives, including at least one of the following: film thickness uniformity, transmittance, hardness, adhesion, and refractive index. The basic constraint indicators are other quality requirements that need to be met, including at least one of the following: film thickness deviation, adhesion, refractive index fluctuation, and hardness.

9. A data processing system for optimizing vacuum coating process parameters, characterized in that, include: The intelligent model module is used to obtain no more than a preset number of actual test process parameters, as well as corresponding quality parameters and actual state data, based on the constraint range and parameter dimensions of process parameters through space filling design, and to construct an initial training sample set. A preliminary proxy model is constructed based on the initial training sample set. The proxy model is used to predict the quality parameters corresponding to the process parameters. The optimization engine module is used to optimize the target quality parameters based on the process parameters. It drives the proxy model to iteratively explore the process parameter space for global optimization through a global optimization algorithm to obtain candidate process parameters. The test execution module is used to conduct physical tests based on the candidate process parameters to obtain actual quality parameters and actual state data. If the difference between the actual quality parameter and the target quality parameter is greater than a preset threshold, the candidate process parameter, actual state data, and actual quality parameter are added to the training sample set, the surrogate model is updated again, and global optimization is performed again until the iteration termination condition is met; if the iteration termination condition is met, the candidate process parameter is taken as the optimization result.

10. An electronic device, comprising: A processor and a memory storing a program, characterized in that the program includes instructions that, when executed by the processor, cause the processor to perform the method according to any one of claims 1 to 8.

Images