Multi-frequency modulation plasma surface treatment method and system

Abstract

The present invention discloses a multi-frequency modulation plasma surface treatment method and system, which relates to the field of semiconductor manufacturing technology; the system comprises an acquisition module, a processing module and a control module; the acquisition module is responsible for acquiring etching information and characteristic data of a target plasma, such as optical, electrical and gas composition data, and aligning these data to a unified time axis; the processing module analyzes these data through a preset machine learning model, generates process state evaluation information, and generates the first process parameters of the target plasma based on these information using a genetic algorithm; the control module multi-frequency modulates a plasma source according to these parameters to optimize the physical properties of the plasma; the present invention improves the semiconductor etching effect by optimizing the synchronization of multi-source data in the multi-frequency modulation plasma surface treatment process and the physical properties of the plasma.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and in particular to a multi-frequency modulation plasma surface treatment method and system. Background Art

[0002] In the field of semiconductor manufacturing, etching technology is a crucial step in the process of manufacturing integrated circuits. It involves transferring specific patterns onto thin films or other underlying materials on silicon wafers. Plasma etching technology is widely used because it can provide highly anisotropic etching and is suitable for miniaturization processes. Plasma etching technology usually relies on a single-frequency RF source to excite and maintain the plasma. Although this method works well in some applications, it faces some limitations when processing complex or large-area wafers. Especially in high-end applications that require both high etching rates and fine pattern fidelity, existing RF sources are usually fixed at a certain frequency, which limits the ability to adjust the physical properties of the plasma, such as ion density and energy distribution, which are critical for optimizing the etching process. In addition, the shortcomings of single-frequency RF sources in power distribution and regulation may lead to energy waste, especially when the plasma source needs to quickly adjust the output according to different process steps.

[0003] For example, the Chinese patent with the authorization announcement number CN117936425B discloses an etching device and method for precise etching and thin film removal, including: a plasma etching chamber for etching thin films on wafers to be removed under the action of plasma; an observation window arranged on one side of the plasma etching chamber; an information acquisition module arranged outside the observation window, and collecting the initial characteristic information of the plasma in the plasma etching chamber through the observation window; an information processing module for processing the characteristic information of the plasma to obtain the final characteristic information after information processing; the information feedback module obtains adjustment parameters according to the final characteristic information after information processing, and feeds the adjustment parameters back to the plasma control module; the plasma control module adjusts the reaction parameters of the plasma etching chamber according to the adjustment parameters. This scheme accurately captures the characteristics of the plasma in the etching chamber based on the information acquisition module and the information processing module, thereby improving the accuracy of etching and removing thin films.

[0004] The above methods have the problems raised by the background technology. The existing RF sources are usually fixed at a certain frequency, which limits the ability to adjust the physical properties of the plasma, such as ion density and energy distribution. In addition, the shortcomings of the single-frequency RF source in power allocation and regulation may lead to energy waste, especially when the plasma source needs to quickly adjust the output according to different process steps. Summary of the invention

[0005] In view of the deficiencies in the prior art, the present invention discloses a multi-frequency modulation plasma surface treatment method and system for optimizing the physical properties of plasma under multi-frequency modulation.

[0006] In a first aspect, an embodiment of the present disclosure provides a multi-frequency modulation plasma surface treatment system, including: a collection module, a processing module and a control module;

[0007] The acquisition module is used to obtain etching process information and obtain characteristic data of the target plasma under multi-frequency modulation through sensors, including optical data, electrical data and gas composition data; and align different types of characteristic data to a unified time axis;

[0008] The processing module is used to input the characteristic data of the target plasma aligned to the same time axis into a preset machine learning model to generate process state evaluation information; based on the process state evaluation information, optimize using a genetic algorithm to generate a first process parameter of the target plasma;

[0009] The control module is used to perform multi-frequency modulation on a plasma source that generates the target plasma based on the first process parameter to adjust the physical properties of the target plasma.

[0010] As an optional implementation, the acquisition module includes: an optical sensor, an electrical sensor and a mass spectrometer;

[0011] The optical sensor is installed at the window of the plasma etching chamber and is used to monitor the optical data of the target plasma in real time, including: emission spectrum and light intensity;

[0012] The electrical sensor is directly connected to the plasma source and is used to measure the electrical data of the plasma, including: voltage, current and power;

[0013] The mass spectrometer is connected to the plasma etching chamber through a gas sampling system, and is used to monitor the composition and ion types of the reaction gas, and to analyze the gas composition in the plasma in real time to generate gas composition data.

[0014] As an optional implementation, the acquisition module further includes a data processing unit;

[0015] The data processing unit is used to pre-process the optical data, electrical data and gas composition data of the target plasma, and align different types of characteristic data to the same time axis.

[0016] As an optional implementation manner, aligning different types of feature data to the same time axis includes:

[0017] For the optical data, for each data collection time point, a cubic spline interpolation method is used to calculate the optical data at the time point according to adjacent known optical data points;

[0018] For the electrical data, for each data collection time point, a linear interpolation method is used to calculate the electrical data at that time point based on adjacent known electrical data points.

[0019] For the gas composition data, at each data collection time point, a Kalman filter is used to correct and interpolate the gas composition data to calculate the gas composition data at that time point.

[0020] As an optional implementation, the Kalman filter includes:

[0021] an interaction noise covariance matrix, wherein the elements of the interaction noise covariance matrix are defined as the impact of changes in different gas components on the measurement of other gas components during the processing of the target plasma;

[0022] The Kalman filter is further configured with a state prediction mechanism and an error update mechanism;

[0023] The state prediction mechanism predicts the state of the current time step by applying the state transfer matrix to the state estimate of the previous time step;

[0024] The error updating mechanism updates the state estimation and error covariance through the observation matrix, the actual measurement value and the actual measurement noise covariance, thereby adjusting the state estimation value of the gas composition.

[0025] As an optional embodiment, before the processing module is used to input the characteristic data of the target plasma aligned to the same time axis into the preset machine learning model, it also includes:

[0026] The historical characteristic data of the target plasma under multi-frequency modulation is standardized and converted into dimensionless data;

[0027] Extracting spectral features, voltage features, current features, power features, and concentrations of different gas components and their changing trends from the dimensionless data to generate standardized feature data;

[0028] Constructing a data matrix based on the standardized characteristic data; wherein each row of the data matrix represents a data sample of characteristic data of a target plasma, and each column of the data matrix represents a feature;

[0029] Calculate the covariance matrix of the data matrix, and perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors; wherein the eigenvalues ​​represent the importance of the principal component, and the eigenvectors represent the direction of the principal component;

[0030] Based on the size of the eigenvalue, a principal component whose cumulative contribution rate of the eigenvalue reaches more than 90% is selected;

[0031] Projecting the data matrix onto the principal components to generate a first dimensionally reduced data set;

[0032] The preset machine learning model is constructed using the first dimension reduction data set.

[0033] As an optional implementation, the processing module performs the following steps when optimizing using a genetic algorithm based on the process state evaluation information to generate the first process parameter of the target plasma:

[0034] Generate an initial population where each individual represents a set of possible power output and frequency combinations;

[0035] designing a fitness function based on the process state evaluation information and the target plasma etching process information;

[0036] Performing selection, crossover and mutation operations to explore the solution space, and dynamically adjusting the fitness function based on the process state evaluation information;

[0037] The mutation operation is a non-uniform mutation, including: performing Gaussian mutation on the power output and performing exchange mutation on the frequency combination;

[0038] Repeat the above steps to evaluate the fitness values ​​of the newly generated individuals, and select individuals whose fitness meets the first target threshold to enter the next generation until the fitness value converges or reaches the preset number of iterations;

[0039] In response to the fitness value converging or reaching a preset number of iterations, a first process parameter of the target plasma is generated based on the current individual.

[0040] As an optional implementation, the control module includes: a power amplifier, a radio frequency source, and a power distribution network;

[0041] The power amplifier is used to amplify the power of the radio frequency signal generated by the radio frequency source, and each target frequency corresponds to an independent power amplifier;

[0042] The RF source generates a RF signal of target frequency and power based on the first process parameter and inputs the signal into the power amplifier; wherein the RF source includes a plurality of frequency generating units, each of which is used to generate a target frequency;

[0043] The power distribution network is used to distribute the power of the amplified radio frequency signal to the plasma source.

[0044] As an optional implementation manner, when the control module performs multi-frequency modulation on the target plasma based on the first process parameter to adjust the physical properties of the target plasma, the control module performs the following steps:

[0045] The RF source sets the frequency of the RF signal generated by each of the frequency generating units based on the target frequency, and sends the frequency to the power amplifier corresponding to each of the target frequencies;

[0046] The power amplifier amplifies the output power of the radio frequency signal at each frequency based on the output power of the target frequency;

[0047] The power distribution network receives radio frequency signals of various frequencies after amplifying the output power and distributes them to different electrodes or regions of the plasma source to adjust the physical properties of the target plasma.

[0048] In a second aspect, the present invention also provides a multi-frequency modulation plasma surface treatment method, comprising:

[0049] Obtain etching process information and obtain characteristic data of target plasma under multi-frequency modulation through sensors, including optical data, electrical data, and gas composition data; align different types of characteristic data to the same time axis;

[0050] Inputting the characteristic data of the target plasma aligned to the same time axis into a preset machine learning model to generate process state evaluation information; optimizing using a genetic algorithm based on the process state evaluation information to generate a first process parameter of the target plasma;

[0051] Based on the first process parameter, a plasma source of the target plasma is multi-frequency modulated to adjust the physical properties of the target plasma.

[0052] Compared with the prior art, the present invention has the following beneficial effects: the present invention collects optical data, electrical data and gas composition data of plasma under multi-frequency modulation, and makes full use of these characteristic data to evaluate the current semiconductor etching situation. The integration of such multi-dimensional data provides more comprehensive monitoring and control capabilities for the etching process.

[0053] The established machine learning model is used to evaluate the etching process in real time. Based on the evaluation results, the improved genetic algorithm is used to optimize the output power and frequency combination. This method can dynamically adjust the multi-frequency modulation parameters according to the target etching effect, actually change the physical properties of the plasma, and optimize the etching quality and consistency.

[0054] For the data used to build the machine learning model, the present invention adopts multiple methods to synchronize the time to ensure the consistency and accuracy of the data. In addition, by applying the principal component analysis method to reduce the dimension of the data, the present invention improves the processing efficiency and prediction accuracy of the machine learning model, thereby optimizing the control and prediction effect of the entire etching process. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A schematic diagram of a multi-frequency modulation plasma surface treatment system provided by an embodiment of the present invention;

[0056] Figure 2 A flow chart of a multi-frequency modulation plasma surface treatment method provided by an embodiment of the present invention;

[0057] Figure 3 A schematic diagram of the computer device structure provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] In order to make the purpose, technical scheme and advantages of the embodiments of the present disclosure clearer, the technical scheme in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the drawings in the embodiments of the present disclosure. Obviously, the described embodiments are only part of the embodiments of the present disclosure, rather than all of the embodiments. The components of the embodiments of the present disclosure generally described and shown here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present disclosure is not intended to limit the scope of the present disclosure claimed for protection, but merely represents the selected embodiments of the present disclosure. Based on the embodiments of the present disclosure, all other embodiments obtained by those skilled in the art without making creative work belong to the scope of protection of the present disclosure.

[0059] The multi-frequency modulation plasma surface treatment method and system provided by the embodiments of the present disclosure are described below.

[0060] See also Figure 1 , which is a schematic diagram of a multi-frequency modulation plasma surface treatment system provided by an embodiment of the present invention, comprising: a collection module 110, a processing module 120 and a control module 130;

[0061] The acquisition module 110 is used to obtain etching process information and obtain characteristic data of the target plasma under multi-frequency modulation through sensors, including optical data, electrical data, and gas composition data; and align different types of characteristic data to the same time axis;

[0062] The processing module 120 is used to input the characteristic data of the target plasma aligned to the same time axis into a preset machine learning model to generate process state evaluation information; based on the process state evaluation information, a genetic algorithm is used for optimization to generate a first process parameter of the target plasma;

[0063] The control module 130 is used to perform multi-frequency modulation on the plasma source of the target plasma based on the first process parameter to adjust the physical properties of the target plasma.

[0064] The etching process information includes: etching depth information, etching profile and sidewall quality, etching target material and etching uniformity information, etc. The above information is used to describe the etching target for the semiconductor device.

[0065] In a specific implementation, the acquisition module 110 acquires various parameter data in plasma etching in real time and sends them to the processing module 120 and the control module 130 for use. The acquisition module 110 is equipped with a variety of high-precision sensors to cover the monitoring requirements of key parameters.

[0066] The characteristic data is used to reflect the characteristics of the plasma, which are the physical properties of the plasma. By analyzing these physical properties, we can further obtain the influence of different physical properties of the plasma on the etching effect.

[0067] As an optional implementation, the acquisition module 110 includes: an optical sensor, an electrical sensor, and a mass spectrometer;

[0068] The optical sensor is used to monitor the optical data of the plasma, including emission spectrum and light intensity, which can reflect the ion density and energy distribution of the plasma. The optical sensor is installed at the window of the plasma etching chamber and can capture the changes in the plasma state in real time.

[0069] Electrical sensors are used to measure the electrical characteristics of the plasma, such as voltage, current, and power. These sensors are directly connected to the plasma source, i.e. the generator and power supply system, and can provide precise electrical data. This data helps to analyze the energy input and reactivity of the plasma.

[0070] Mass spectrometer, used to monitor the composition and ion species of the reaction gas. The mass spectrometer can analyze the gas composition in the plasma in real time and provide detailed chemical reaction information. The mass spectrometer is connected to the etching chamber through a gas sampling system, which can continuously collect and analyze reaction gas samples.

[0071] The plasma etching chamber is a closed environment specially designed for plasma processing, such as etching or deposition. It is the main place where plasma is generated and acts. The interior of the etching chamber is usually equipped with an optical window, allowing optical sensors to be installed here to monitor the state changes of the plasma in real time, such as emission spectrum and light intensity.

[0072] A plasma source is a device that generates plasma, usually including a radio frequency (RF) source and other auxiliary equipment, such as a gas flow control system. The plasma source is responsible for exciting the input gas (such as argon, hydrogen fluoride, etc.) to generate a plasma with high-energy particles, which are used to etch materials on silicon wafers. The plasma source is directly connected to electrical sensors, and the voltage, current, and power data measured by these sensors provide precise information about the plasma energy input and reactivity.

[0073] The gas sampling system is a device connected to the mass spectrometer, which is used to extract gas samples from the etching chamber and transmit them to the mass spectrometer for analysis. The system can monitor and analyze the changes in gas composition in the plasma, including the composition of the reactive gas and the ion species.

[0074] In a multi-frequency modulated plasma surface treatment system, these components interact with each other to optimize the etching process of semiconductors. For example, the plasma generated by the plasma source reacts with the silicon material in the etching chamber, and the progress of the chemical reaction is monitored in real time through a gas sampling system and a mass spectrometer to adjust the frequency and power of the RF source. The data provided by optical and electrical sensors is used to further accurately control and monitor the physical and chemical state of the plasma to ensure the efficiency and repeatability of the etching process.

[0075] In a specific implementation, a plasma source generates a target plasma, and its operating parameters such as RF power and frequency directly affect the physical and chemical properties of the plasma. The target plasma interacts with the semiconductor material in the etching chamber to achieve etching or deposition. The gas sampling system is used in conjunction with a mass spectrometer to analyze the gas composition in the target plasma and provide data support for adjusting the parameters of the plasma source.

[0076] As an optional implementation, the acquisition module 110 also includes a data processing unit.

[0077] The data processing unit is used to pre-process the optical data, electrical data and gas composition data of the target plasma, and align different types of characteristic data to the same time axis.

[0078] In a specific implementation, the data acquisition unit is equipped with a data processing chip for quickly processing a large amount of real-time data and transmitting the processed data to the processing module 120 .

[0079] In addition, the data acquisition unit integrates data synchronization and time stamp functions to ensure that all sensor data are collected and recorded at the same time point. This can avoid analysis errors caused by asynchronous data acquisition time and provide a data basis for subsequent data processing and control decisions.

[0080] In a specific implementation, the data acquisition unit receives data packets from the optical sensor, the electrical sensor, and the mass spectrometer via a high-speed data bus or a network interface, each of which contains a measurement value and a timestamp.

[0081] The received data is first denoised and filtered to remove sensor noise and environmental interference. This step ensures the clarity and accuracy of the data.

[0082] In order to ensure that different types of characteristic data can be analyzed at the same time point, the data acquisition unit needs to synchronize and align the data.

[0083] The data acquisition unit aligns all data based on the timestamp. Data synchronization is ensured by mapping data from different sources onto a unified timeline. For example, if the timestamps of optical data and electrical data are slightly different, the system will interpolate or weighted average to ensure that the various types of data at the same time can match.

[0084] As an optional implementation, different types of characteristic data are aligned to the same time axis, including:

[0085] For the optical data, for each data collection time point, the cubic spline interpolation method is used to calculate the optical data at that time point based on the adjacent known optical data points.

[0086] For electrical data, for each data collection time point, a linear interpolation method is used to calculate the electrical data at that time point based on adjacent known electrical data points.

[0087] For the gas composition data, for each data collection time point, a Kalman filter is used to correct and interpolate the gas composition data to calculate the gas composition data at that time point.

[0088] In a specific implementation, the acquisition timestamps of various sensors may not be the same, and these data need to be aligned to a unified time axis.

[0089] There are many reasons why the acquisition timestamps of each sensor are different. For example, each sensor, such as optical sensors, electrical sensors, and mass spectrometers, has its own specific response time and data processing speed. Electrical sensors can provide data almost in real time, while mass spectrometers can only output data after collecting and analyzing gas samples. These differences lead to inconsistent data acquisition and processing time. For example, improper equipment maintenance or environmental interference may cause some sensors to respond more slowly, which in turn affects the data acquisition time. For another example, different types of sensors may be set with different sampling rates, that is, the number of measurements per unit time. Some sensors may require more frequent data updates, such as electrical sensors, while other sensors may update more slowly, such as mass spectrometers. This difference in sampling rate directly affects the time point of data recording.

[0090] In a specific implementation, there are three different types of data sources: optical data, electrical data, and gas composition data.

[0091] Among them, optical data usually has high temporal and spatial resolution, but may be affected by environmental noise. In order to ensure the accuracy of the data, the cubic spline interpolation method is used for alignment. The specific steps are as follows: For each time point, the cubic spline interpolation method is used to calculate the optical data value at that time point based on the adjacent known optical data points. This interpolation method generates high-precision interpolation results and reduces the impact of noise by considering the smooth transition between data points.

[0092] Electrical data usually changes slowly and has high reliability and accuracy. In order to align, the linear interpolation method is used. The specific steps are as follows: For each time point, the electrical data value at that time point is calculated based on the adjacent known electrical data points. The linear interpolation method provides fast and reliable interpolation results through simple linear calculations, which is suitable for the slowly changing characteristics of electrical data.

[0093] Gas composition data reflects the chemical reaction process in the plasma during the multi-frequency modulated plasma etching process. However, since the data may have time delays and noise, the data needs to be corrected and interpolated to ensure that it is synchronized with other types of data. Kalman filter is an effective way to achieve this goal.

[0094] As an optional implementation, in order to cope with complex gas dynamics and enhance the performance of the Kalman filter, the interaction noise covariance is introduced. This parameter is designed to better handle the mutual influence between gas components, especially in a multi-component gas environment, where the changes in different gas components may be interdependent or interfere with each other.

[0095] The purpose of introducing the interaction noise covariance is to consider the interaction between gas components in the measurement update process of the Kalman filter. When measuring gases, changes in a single gas component may affect the measurement accuracy of other components due to the complexity of chemical reactions and physical interactions. The decomposition products may affect The detection readings of the gas and vice versa are also true. The traditional Kalman filter may not accurately estimate the gas concentration without considering this interaction, thus affecting the control accuracy of the entire etching process.

[0096] in, (Carbon tetrafluoride) is a widely used gas for etching silicon and its oxides, especially in the etching process on the surface of semiconductor devices; It's oxygen.

[0097] In a specific implementation, the interaction noise covariance S is a matrix whose structure is designed to describe the statistical dependencies between measurements of different gas components. and , S will be a 2x2 matrix where each element It indicates how the change of component i affects the measurement error of component j. The specific values ​​of this matrix can be estimated based on experimental data, chemical kinetic models, or through data-driven methods.

[0098] In a specific implementation, the system first initializes basic parameters including the state vector, error covariance matrix, state transfer matrix and observation matrix. The state vector includes the estimated concentrations of all key gas components. Optionally, the interaction noise covariance matrix described above can also be defined.

[0099] In each data collection cycle, the system first performs a prediction step to estimate the new state of each gas component using the state transfer matrix. Then, in the update step, the system collects actual gas concentration data and uses the observation matrix and the interaction noise covariance matrix to adjust and optimize the state estimate. This process includes calculating the Kalman gain and applying it to integrate the new measurement data, updating the state vector and the error covariance matrix.

[0100] Finally, the system aligns the corrected gas composition data with data collected from other sensors, such as optical and electrical sensors, ensuring that all data is based on a unified timeline. This is done through data interpolation or synchronization algorithms, ensuring that all data used in processing and control is accurate and consistent.

[0101] In this way, the processing module 120 can accurately synchronize and align the optical data, electrical data, and gas composition data, and fuse the data from different sources by weighted average to ensure the overall consistency and accuracy of the data, and provide reliable data support for feedback control and process optimization. In this way, the processing module 120 can effectively synchronize and align different types of data, and adopt appropriate methods according to their respective characteristics to improve the accuracy and reliability of the overall data fusion, and provide accurate data support for subsequent process control and optimization.

[0102] For example, in semiconductor manufacturing, there are two gas components that need to be precisely controlled and detected: and The two gases interact with each other during the plasma etching process and may affect each other's measurement accuracy.

[0103] The parameters are set as follows: state vector ,include and The concentration of , the state transfer matrix For example, if the gas concentration changes slowly, is the unit matrix; the observation matrix , which is also the unit matrix, indicating that the concentration of each gas is directly measured; the process noise covariance matrix , considering the stability of the system, it is set to ; Measurement noise covariance matrix , based on sensor performance, is set to ; Interaction noise covariance matrix , considering The decomposition may affect Measurement, set to ;

[0104] Data processing flow:

[0105] initialization: (initial and concentration estimation), (initial error covariance);

[0106] Real-time data collection based on the acquisition module 110: At time point k, the measured and The concentrations of are 102ppm and 48ppm respectively, that is, Among them, ppm (part per million) means one part per million. is the new measurement obtained from the mass spectrometer.

[0107] Implement Kalman filter;

[0108] State prediction formula:

[0109] ;

[0110] in; It represents the estimated value of the system state at time step k, given all previously observed information until time k-1 (i.e., excluding observations at time k). represents the state estimate at the previous time step k-1.

[0111] Using the above formula based on historical gas component concentrations (such as and ) to predict the concentration at the current time point. The state transition matrix here Describing how gas concentrations shift from one point in time to the next, it is usually assumed that these shifts are linear or nearly linear over short time scales, reflecting the natural processes of gas reactions and physical diffusion in the plasma.

[0112] Error covariance prediction formula:

[0113] ;

[0114] The above formula updates the uncertainty of gas concentration estimation. is the error covariance of the previous step, representing the uncertainty in the estimated state at the previous time step measurement.

[0115] Update steps:

[0116] ;

[0117] The above formula is used to calculate the Kalman gain, which determines the new measurement data (such as real-time measurement and The influence of concentration on state estimation. is the prediction error covariance matrix. It represents the uncertainty of the system state estimate at time k, given all the information at previous time k-1; is the observation matrix. It maps the state space to the measurement space and is used to convert the state variables into observable outputs; is the measurement matrix The transpose of is used in matrix multiplication for proper dimension matching and remapping the information in the observation space back to the state space; is the measurement noise covariance matrix, which describes the uncertainty or error in the measurement process, including sensor errors, etc. is the interaction noise covariance matrix.

[0118] The Kalman gain is the core of the filtering process, which determines how much "weight" should be given to new measurement data when updating the state estimate. This gain helps balance the weights of prediction and measurement, optimizing the accuracy of the state estimate.

[0119] ;

[0120] The above formula is used to update the estimate of gas concentration. The difference term in the formula Represents the error between the prediction and the actual observation, and the Kalman gain is used to adjust the contribution of this error to the final state estimation.

[0121] ;

[0122] The above formula is used to update the uncertainty of the state estimate. The introduction of the identity matrix I ensures that the updated covariance maintains the correct dimension and properties.

[0123] It can be understood that in the calculation example of the Kalman filter above, unless otherwise specified, all parameters are matrices or vectors. k|k-1 represents the prediction of the state at time k given all the information up to the deadline k-1. This is part of the prediction step of the Kalman filter and is used to represent the state prediction before the current observation. T represents the transpose of the matrix or vector. I represents the identity matrix, where the elements on the diagonal of the identity matrix are 1 and the elements in other positions are 0. -1 represents the inverse of the matrix, that is, the inverse matrix of a matrix is ​​multiplied by the original matrix to obtain the identity matrix.

[0124] Use the updated To correct and align gas composition data while ensuring synchronization with other sensor data (optical, electrical).

[0125] It is understandable that, since the specific parameters of each process, system, and equipment are not the same, the specific parameter settings of the above Kalman filter should be combined with actual needs. For example, the state transfer matrix can be determined based on the physical and chemical properties of the gas components. If the behavior of the gas components within the time interval is approximately linear, the state transfer matrix can be set to the unit matrix or the linear equation coefficient matrix. If the known gas concentration changes can be described by a first-order linear dynamic model, and these coefficients can be fitted by experimental data. For more complex dynamics, such as nonlinear behaviors affected by temperature and pressure changes, advanced mathematical models (such as physics-based models or data-driven machine learning models) can be used to estimate these coefficients.

[0126] The measurement matrix describes how the observables are obtained from the state variables. Usually, if each state variable is measured directly, the measurement matrix can be an identity matrix. If the observations are functions or combinations of the state variables, this relationship needs to be reflected in the measurement matrix. This can be determined from a physical understanding of the system or from experimental data.

[0127] The process noise covariance matrix reflects the uncertainty and external disturbances in the model predictions. This usually needs to be determined based on historical data. In some scenarios, the process noise covariance matrix can be dynamically adjusted by comparing the model predictions with the actual measured data.

[0128] The measurement noise covariance matrix reflects the statistical characteristics of the measurement errors and can be determined based on the technical specifications of the measurement equipment or through statistical analysis of actual measurement errors. For multi-sensor systems, if the accuracies of the sensors are known, the measurement noise covariance matrix can be directly set to the variance of these accuracies.

[0129] The interaction noise covariance matrix describes the mutual influence between different gas components. This can be determined through chemical kinetic simulations or controlled experiments by measuring the response of other components when one gas component is changed. In data-driven methods, the interaction noise covariance matrix can be estimated by statistically analyzing the interaction effects between multiple variables.

[0130] The initial value of the error covariance matrix is ​​usually large, indicating the uncertainty of the initial state estimate. Its subsequent values ​​are dynamically adjusted through the iterative update process of the Kalman filter. The update of this matrix reflects the accuracy of the system state estimate, which is gradually refined through the above prediction and update formulas.

[0131] In this way, the accuracy of gas component concentration estimation can be significantly improved by integrating the interaction noise covariance matrix into the update step of the Kalman filter. Specifically, this allows the filter to consider the effects caused by changes in other components when updating the state estimate, thereby more realistically reflecting the actual dynamics of the gas composition. This approach can improve the control accuracy of the etching process, because the accurate estimation of the gas component concentration directly affects the chemical and physical behavior of the etchant.

[0132] For the processing module 120:

[0133] In a specific implementation, the processing module 120 receives different types of characteristic data aligned to the same time axis, and inputs these characteristic data into a machine learning model trained based on historical data to generate process status assessment information.

[0134] As an optional implementation, principal component analysis (PCA) is used to process and analyze multiple feature data, simplify the data structure, identify key factors affecting process performance, and understand the relationship between various variables. Standardize data from different sources and convert them into dimensionless data. This can be achieved by subtracting the mean and dividing by the standard deviation. Standardization can eliminate the differences between data of different dimensions and make them comparable on the same scale. Extract spectral features, such as characteristic peaks, spectral intensity, spectral line width, etc. These features reflect information such as the ion density, energy distribution, and electron temperature of the plasma. Extract voltage, current, and power features, such as voltage fluctuation amplitude, power consumption rate, etc. These features reflect energy input and reaction activity. Extract the concentrations of different gas components and their changing trends, such as the instantaneous concentration of a specific gas, the average concentration change rate, etc. These features reflect the progress of the chemical reaction.

[0135] The standardized feature data is constructed into a data matrix, where each row represents a sample (for example, data from an etching process) and each column represents a feature (for example, spectral intensity, power consumption rate, etc.). The covariance matrix of the data matrix is ​​calculated to understand the correlation between different features. The covariance matrix is ​​subjected to eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors. The eigenvalue represents the importance of the principal component, and the eigenvector represents the direction of the principal component. The first few principal components are selected according to the size of the eigenvalue, which can explain most of the variance of the data. The principal components with a cumulative contribution rate of more than 90% of the eigenvalue are selected. The original data is projected onto the principal components to generate a reduced-dimensional data set. These data sets are easier to analyze and understand in the new space.

[0136] In a specific implementation, the data sent by the acquisition module 110 is first standardized. This step is necessary because the dimensions and magnitudes of the data sources may vary, and directly using these data may cause deviations or inaccuracies in the machine learning model. Standardization usually includes removing the mean and scaling the variance, that is, subtracting the mean from each feature and dividing it by its standard deviation, so that the mean of the resulting data is 0 and the standard deviation is 1. The data processed in this way is called dimensionless data, which is helpful for subsequent analysis and model training.

[0137] Next, key features are extracted from the standardized dimensionless data, including spectral features, voltage features, current features, power features, and the concentration of gas components and their changing trends. These features are important indicators for describing the state of the plasma, and they are used to capture key physical and chemical information during the process. Feature extraction not only involves the direct use of raw data, but may also include some derived calculations such as peak detection, average value, coefficient of variation, etc.

[0138] The extracted features are organized into a data matrix, where each row represents a data sample (i.e., an observation at a time point) and each column represents a feature. This data organization facilitates statistical analysis and training of machine learning models.

[0139] Calculate the covariance matrix of the data matrix, which is used to analyze the linear relationship between the features. Then perform eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors. The eigenvalue represents the variance of each principal component, that is, the importance of the component in the data; the eigenvector indicates the direction of each principal component.

[0140] Based on the size of the eigenvalues, principal components with a cumulative contribution rate of more than 90% are selected, which is a dimensionality reduction technique that can reduce the complexity of the model while retaining the most important information. These principal components constitute the new, reduced-dimensionality dataset.

[0141] The reduced-dimensional dataset is used to train a pre-set machine learning model, which will be used to predict and evaluate the status of the plasma processing process in real time, such as predicting future quality trends or potential processing anomalies.

[0142] As an optional implementation, the reduced-dimensional data is used to train a machine learning model to achieve classification and prediction of process status.

[0143] In a specific implementation, based on the reduced-dimensional data set, the processing module 120 divides the data into a training set and a test set to ensure that the model is trained and validated on different data to avoid overfitting. An appropriate machine learning model is selected, such as a support vector, random forest, or neural network, and the model is trained using the training set to learn the patterns and rules in the data. The performance of the model is evaluated using the test set.

[0144] In actual operation, the reduced-dimensional data collected in real time is input into the trained machine learning model, and the model can classify and predict the process status. For example, the model can predict that the current process status is "normal", "ion density is too high" or "power consumption is abnormal", etc. Based on the prediction results of the model, the processing module 120 generates process status evaluation information.

[0145] In a specific implementation, based on the process status evaluation information, the processing module 120 further generates a specific control strategy, namely, a first process parameter, to achieve process optimization and stability.

[0146] According to the specific surface treatment requirements, the energy distribution of ions in the plasma is adjusted. For example, in the deep etching process, higher energy ions are required to penetrate the material surface. The processing module 120 indicates to increase the high frequency power output and improve the average energy of the plasma according to the model prediction results.

[0147] During the surface cleaning process, ions with lower energy can meet the requirements. The processing module 120 instructs to adjust the low-frequency power to reduce the average energy of ions in the plasma to avoid excessive damage to the material surface.

[0148] In actual operation, the processing module 120 monitors the ion energy distribution in real time and dynamically adjusts the output and frequency combination of high-frequency and low-frequency power in combination with the model prediction results to achieve optimal energy distribution control.

[0149] In a specific implementation, the control strategy is optimized using a genetic algorithm to generate the first process parameters of the target plasma. The frequency combination of the multi-frequency power output is adjusted in real time according to different surface treatment requirements in the etching process information acquired by the acquisition module 110. The genetic algorithm calculates the optimal frequency combination based on real-time data and process requirements.

[0150] As an optional implementation, the processing module 120 performs the following steps when optimizing by using a genetic algorithm based on the process state evaluation information to generate the first process parameter of the target plasma:

[0151] Generate an initial population where each individual represents a set of possible power output and frequency combinations;

[0152] designing a fitness function based on the process state evaluation information and the target plasma etching process information;

[0153] Performing selection, crossover and mutation operations to explore the solution space, and dynamically adjusting the fitness function based on the process state evaluation information;

[0154] The mutation operation is a non-uniform mutation, including: performing Gaussian mutation on the power output and performing exchange mutation on the frequency combination;

[0155] Repeat the above steps to evaluate the fitness values ​​of the newly generated individuals, and select individuals whose fitness meets the first target threshold to enter the next generation until the fitness value converges or reaches the preset number of iterations;

[0156] In response to the fitness value converging or reaching a preset number of iterations, a first process parameter of the target plasma is generated based on the current individual.

[0157] In a specific implementation, a genetic algorithm is used to adjust the frequency combination of multi-frequency power output in real time according to different surface treatment requirements.

[0158] In the multi-frequency modulated plasma surface treatment process, the population initialization is used to cover the possible power allocation and frequency combinations, thus ensuring a wide search of the algorithm in the solution space.

[0159] Each individual represents a power allocation and frequency combination scheme. Considering the importance of high-frequency power, low-frequency power and frequency combination in this scenario, individuals can be represented in the following form:

[0160] ;

[0161] in, : High frequency power, : Low frequency power, : Frequency combination.

[0162] High frequency power and low frequency power The value range needs to be set according to the actual process requirements. For example, the high-frequency power range can be 20-100 watts, and the low-frequency power range can be 10-50 watts.

[0163] Frequency Combination The value of needs to be based on the process requirements. For example, common frequencies may include 2.45MHz, 13.56MHz, 27.12MHz, etc. Two or more frequencies can be selected for combination.

[0164] The size of the population determines how many individuals participate in the evolution in each generation. The population size is set to 50 to 100 individuals to ensure diversity and computational efficiency.

[0165] The individuals of the initial population are randomly generated within the set range. The high-frequency power, low-frequency power and frequency combination of each individual are randomly selected within the corresponding range.

[0166] In the setting of fitness function, the main optimization goals can include improving etching efficiency, optimizing material and energy utilization, and improving product quality. These goals reflect the key performance indicators pursued by the plasma processing system in operation, such as etching rate, energy consumption, material utilization efficiency, and surface quality of the final product.

[0167] To concretize these goals, the fitness function can include multiple quantitative indicators, such as: etching efficiency: evaluated by the material removal rate per unit time; material utilization: measure the effective use of materials and reduce waste; energy efficiency: calculated by the energy consumption of the system, higher efficiency means lower energy consumption; product quality: evaluated by detecting the surface quality of the product after etching.

[0168] These indicators can be combined into a comprehensive fitness function, for example:

[0169] ;

[0170] Among them, E, U, P, and G represent etching efficiency, material utilization, energy efficiency, and product quality, respectively. , , , are the corresponding weights, which are adjusted according to the priorities of different goals.

[0171] During implementation, the fitness function can be dynamically adjusted based on real-time process feedback or changes in the external environment. For example, if a significant change in energy costs is detected, the weight of energy efficiency can be appropriately adjusted to adapt to the new cost structure.

[0172] By repeatedly performing steps such as selection, crossover, and mutation, and evaluating the performance of each newly generated individual according to the fitness function, the system can gradually find the optimal solution. The whole process continues until the preset number of iterations is reached or the fitness converges, and finally determines the parameter configuration that can achieve the best process effect.

[0173] As an optional implementation, non-uniform mutation can be used in the process of genetic algorithm optimization. Non-uniform mutation is a mutation method to better adapt to the characteristics of different types of data. For the optimization problem in the multi-frequency modulation plasma surface treatment process, different mutation treatments can be performed on high-frequency power, low-frequency power and frequency combination.

[0174] The core idea of ​​non-uniform mutation is to use different mutation strategies for different parameters. For example, Gaussian mutation is used for power values ​​(mutation in a small range around the current value), and swap mutation is used for frequency combinations (swap two frequencies in a frequency combination).

[0175] In a specific implementation, the variation of high-frequency power and low-frequency power is varied through Gaussian distribution. Formula:

[0176] ;

[0177] in, is the new individual value, is the current individual value, It means the mean is 0 and the variance is Gaussian noise. Gaussian variation is performed on the high-frequency power and low-frequency power of each individual. Choose a suitable variance , in order to control the range of variation.

[0178] For example, the high-frequency power of an individual is 60 watts and the low-frequency power is 30 watts. After Gaussian mutation, the high-frequency power may become:

[0179] ;

[0180] The low frequency power may become:

[0181] ;

[0182] Among them, P high-new , P low-new They represent the high-frequency power and low-frequency power after Gaussian mutation, respectively.

[0183] In the specific implementation, the mutation of the frequency combination is achieved by exchanging two frequencies in the frequency list. This method ensures that the characteristics of the frequency combination remain unchanged, but the order is changed. Two frequencies in the frequency combination are randomly selected and their positions are exchanged. Ensure that the mutated frequency combination is within the allowed frequency range.

[0184] For example, suppose the frequency combination of an individual is [13.56MHz, 27.12MHz]. After the exchange mutation, the frequency combination may become:

[0185] ;

[0186] Among them, F comb-new Represents the frequency combination after exchange mutation.

[0187] In this way, through the above non-uniform variation method, the genetic algorithm can more effectively adapt to the different data characteristics in multi-frequency modulation plasma surface treatment and improve the optimization effect and efficiency.

[0188] For the control module 130:

[0189] In a specific implementation, the control module 130 controls the power output of different frequencies based on the first process parameter output by the processing module 120, and optimizes the ion energy distribution and density of the plasma, so as to achieve the best etching effect.

[0190] As an optional embodiment, the control module 130 includes: a power amplifier, a radio frequency source and a power distribution network; the power amplifier is used to increase the power of the radio frequency signal generated by the radio frequency source, and each target frequency corresponds to an independent power amplifier; the radio frequency source is used to generate a radio frequency signal of target frequency and power based on the first process parameters, and input it into the power amplifier; wherein the radio frequency source includes a plurality of frequency generating units, each frequency generating unit is used to generate a target frequency; the power distribution network is used to distribute the amplified radio frequency signal to the plasma source.

[0191] As an optional embodiment, the control module 130 performs the following steps when multi-frequency modulating the target plasma based on the first process parameter to adjust the physical properties of the target plasma: using the RF source to set the frequency of the RF signal generated by each frequency generating unit based on the target frequency included in the first power parameter, and sending it to the power amplifier corresponding to each target frequency; the power amplifier amplifies the output power of the RF signal of each frequency based on the output power of the target frequency; the power distribution network receives the RF signals of each frequency after the output power is amplified and distributes them to different electrodes or regions of the plasma source to adjust the physical properties of the target plasma.

[0192] A power amplifier is a device used to increase the power of an electrical signal, especially in radio frequency applications, so that the signal can effectively drive a load such as an antenna or plasma source.

[0193] An RF source is a device that generates RF electromagnetic waves and is used to generate signals of specific frequency and power to control various process equipment, such as plasma generators.

[0194] A power distribution network is a network system that receives an input power and distributes it to multiple outputs, ensuring that each part receives the appropriate amount of energy.

[0195] A plasma source is a device used to generate plasma, a gas containing free electrons and ions, which is commonly used for material processing and surface treatment.

[0196] In a specific implementation, based on the first process parameters, each frequency generating unit in the RF source generates a specific required frequency. These parameters include the required frequency and power level to meet specific physical processing requirements. The RF signal generated by the RF source is first input into the corresponding power amplifier. Each target frequency corresponds to an independent power amplifier to ensure that the signal reaches a sufficient power level before being transmitted to the plasma source. The power-amplified signal is sent to different electrodes or regions of the plasma source through a power distribution network. This step ensures that each part of the plasma source receives appropriate energy, thereby uniformly and effectively exciting the plasma.

[0197] In this way, by precisely controlling the power and frequency of each part of the plasma source, the generation and maintenance of the plasma can be optimized, thereby improving the processing efficiency and uniformity of the results. Multi-frequency modulation allows the system to be quickly and flexibly adjusted to different materials and processing requirements, enhancing the system's applicability and effectiveness. By optimizing the use of power output and frequency, the system can reduce unnecessary energy consumption and reduce operating costs.

[0198] In multi-frequency modulated plasma surface treatment systems, an effective power distribution network is key to ensure that RF power can be accurately and evenly distributed to each area of ​​the plasma source according to process requirements.

[0199] As an optional implementation, the power distribution network aims to achieve uniform distribution of RF signals and allow flexible adjustment of power output according to process variations.

[0200] In a specific implementation, a ring coupler or Wilkinson power splitter can be used. These passive power dividers can provide low insertion loss and good uniformity. The output of each splitter can be connected to an adjustable attenuator to achieve fine control of the output power.

[0201] The layout of the power distribution network should be designed according to the geometry of the plasma source. For example, if the plasma source has multiple independent reaction areas, a centralized distribution node can be designed to lead the main line from the RF source, and then distribute the power to each area through a splitter. Each branch should be designed with adequate shielding and proper impedance matching to reduce the reflection and attenuation of the RF signal.

[0202] To achieve dynamic control, the adjustable attenuator in each branch is managed by a control system, which can automatically adjust the attenuator settings in response to changes in the process based on sensor data (such as current, voltage, and spectral data) transmitted back from the plasma source.

[0203] It is understandable that managing each adjustable attenuator by a control system to achieve automatic adjustment is not the subject of the present invention and will not be elaborated here.

[0204] For example, in a plasma cleaning scenario for semiconductor wafers, the wafer diameter is 300 mm. The designed power distribution network uses a central splitter to divide the main RF signal into four areas. Each area is precisely controlled by an independent adjustable attenuator to meet the cleaning needs of different areas on the wafer. By real-time monitoring of the cleaning effect on the wafer surface and adjusting the power output of the corresponding area, the uniformity and efficiency of the cleaning process are guaranteed.

[0205] Based on the same inventive concept, a multi-frequency modulation plasma surface treatment method is also provided in an embodiment of the present invention. Since the principle of solving the problem by the method in the embodiment of the present invention is similar to the multi-frequency modulation plasma surface treatment system in the embodiment of the present invention, the implementation of the method can refer to the implementation of the system, and the repeated parts will not be repeated.

[0206] Reference Figure 2 FIG. 1 is a flow chart of a multi-frequency modulation plasma surface treatment method provided by an embodiment of the present disclosure, and the method includes steps S101 to S103, which are specifically as follows:

[0207] S101: Obtain etching process information, and obtain characteristic data of target plasma under multi-frequency modulation through sensors, including optical data, electrical data, and gas composition data; align different types of characteristic data to the same time axis;

[0208] S102: inputting the characteristic data of the target plasma aligned to the same time axis into a preset machine learning model to generate process state evaluation information; based on the process state evaluation information, optimizing using a genetic algorithm to generate a first process parameter of the target plasma;

[0209] S103: Based on the first process parameter, multi-frequency modulate the plasma source of the target plasma to adjust the physical properties of the target plasma.

[0210] The disclosed embodiment of the present invention also provides a computer device, such as Figure 3 FIG. 1 is a schematic diagram of a computer device structure provided by an embodiment of the present invention, including:

[0211] A processor 31 and a memory 32; the memory 32 stores machine-readable instructions executable by the processor 31, and the processor 31 is used to execute the machine-readable instructions stored in the memory 32. When the machine-readable instructions are executed by the processor 31, the processor 31 performs the following steps:

[0212] Obtain target plasma etching process information, and obtain characteristic data of target plasma under multi-frequency modulation through sensors, including optical data, electrical data, and gas composition data; align different types of characteristic data to the same time axis;

[0213] Inputting the characteristic data of the target plasma aligned to the same time axis into a preset machine learning model to generate process state evaluation information; optimizing using a genetic algorithm based on the process state evaluation information to generate a first process parameter of the target plasma;

[0214] Based on the first process parameter, a plasma source of the target plasma is multi-frequency modulated to adjust the physical properties of the target plasma.

[0215] The above-mentioned memory 32 includes internal memory 321 and external memory 322; the memory 321 here is also called internal memory, which is used to temporarily store the calculation data in the processor 31, as well as the data exchanged with the external memory 322 such as the hard disk. The processor 31 exchanges data with the external memory 322 through the internal memory 321.

[0216] The specific execution process of the above instructions can refer to the implementation content of a multi-frequency modulation plasma surface treatment system described in the embodiment of the present disclosure, which will not be repeated here.

[0217] It should be understood that in various embodiments of the present invention, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0218] In the description of this specification, the description with reference to the terms "one embodiment", "example", "specific example", etc. means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0219] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A multi-frequency modulation plasma surface treatment system, characterized in that: include: Acquisition module, processing module and control module; The acquisition module is used to obtain etching process information and obtain characteristic data of the target plasma under multi-frequency modulation through sensors, including optical data, electrical data and gas composition data; and align different types of characteristic data to the same time axis; The processing module is used to input the characteristic data of the target plasma aligned to the same time axis into a preset machine learning model to generate process state evaluation information; based on the process state evaluation information, a genetic algorithm is used for optimization to iteratively calculate the power output and frequency combination of the radio frequency signal to generate a first process parameter of the target plasma; The control module is used to perform multi-frequency modulation on a plasma source that generates the target plasma based on the first process parameter to adjust the physical properties of the target plasma; Wherein, the control module is configured with a plurality of frequency generating units, and the plurality of frequency generating units are used to output radio frequency signals of a plurality of frequencies and independently adjust the power of each of the radio frequency signals based on the first process parameter, so as to dynamically adjust the ion density and energy distribution of the target plasma; The aligning of different types of characteristic data to the same time axis includes: For the optical data, for each data collection time point, a cubic spline interpolation method is used to calculate the optical data at the time point according to adjacent known optical data points; For the electrical data, for each data collection time point, a linear interpolation method is used to calculate the electrical data at that time point based on adjacent known electrical data points; For the gas composition data, for each data collection time point, a Kalman filter is used, and an interactive noise covariance matrix is ​​introduced to correct and interpolate the gas composition data, and the gas composition data at that time point is calculated; wherein the interactive noise covariance matrix is ​​used to describe the statistical dependence between different gas component measurements, so as to introduce the interaction between the gas components in the measurement update process of the Kalman filter.

2. The multi-frequency modulation plasma surface treatment system according to claim 1, characterized in that: The acquisition module includes: an optical sensor, an electrical sensor and a mass spectrometer; The optical sensor is installed at the window of the plasma etching chamber and is used to monitor the optical data of the target plasma in real time, wherein the optical data includes: emission spectrum and light intensity; The electrical sensor is directly connected to the plasma source and is used to measure the electrical data of the plasma, wherein the electrical data includes: voltage, current and power; The mass spectrometer is connected to the plasma etching chamber through a gas sampling system, and is used to monitor the composition and ion types of the reaction gas, and to analyze the gas components in the plasma in real time to generate gas composition data.

3. The multi-frequency modulation plasma surface treatment system according to claim 2, characterized in that: The acquisition module also includes a data processing unit; The data processing unit is used to pre-process the optical data, electrical data and gas composition data of the target plasma, and align different types of characteristic data to the same time axis.

4. The multi-frequency modulation plasma surface treatment system according to claim 3, characterized in that: The Kalman filter comprises: an interaction noise covariance matrix, wherein the elements of the interaction noise covariance matrix are defined as the impact of changes in different gas components on the measurement of other gas components during the processing of the target plasma; The Kalman filter is further configured with a state prediction mechanism and an error update mechanism; The state prediction mechanism predicts the state of the current time step by applying the state transfer matrix to the state estimate of the previous time step; The error updating mechanism updates the state estimation and error covariance through the observation matrix, the actual measurement value and the actual measurement noise covariance, and adjusts the state estimation value of the gas composition.

5. The multi-frequency modulation plasma surface treatment system according to claim 4, characterized in that: Before the processing module is used to input the characteristic data of the target plasma aligned to the same time axis into a preset machine learning model, it also includes: The historical characteristic data of the target plasma under multi-frequency modulation is standardized and converted into dimensionless data; Extracting spectral features, voltage features, current features, power features, and concentrations of different gas components and their changing trends from the dimensionless data to generate standardized feature data; Constructing a data matrix based on the standardized characteristic data; wherein each row of the data matrix represents a data sample of characteristic data of a target plasma, and each column of the data matrix represents a feature; Calculate the covariance matrix of the data matrix, and perform eigenvalue decomposition on the covariance matrix to obtain eigenvalues ​​and eigenvectors; wherein the eigenvalues ​​represent the importance of the principal component, and the eigenvectors represent the direction of the principal component; Based on the size of the eigenvalue, a principal component whose cumulative contribution rate of the eigenvalue reaches more than 90% is selected; Projecting the data matrix onto the principal components to generate a first dimensionally reduced data set; The preset machine learning model is constructed using the first dimension reduction data set.

6. The multi-frequency modulation plasma surface treatment system according to claim 5, characterized in that: The method of optimizing based on the process state evaluation information using a genetic algorithm to generate the first process parameter of the target plasma comprises the following steps: Generate an initial population where each individual represents a set of power output and frequency combinations; designing a fitness function based on the process state evaluation information and the target plasma etching process information; Performing selection, crossover and mutation operations to explore the solution space, and dynamically adjusting the fitness function based on the process state evaluation information; The mutation operation is a non-uniform mutation, including: performing Gaussian mutation on the power output and performing exchange mutation on the frequency combination; Repeat the above steps to evaluate the fitness values ​​of the newly generated individuals, and select individuals whose fitness meets the first target threshold to enter the next generation until the fitness value converges or reaches the preset number of iterations; In response to the fitness value converging or reaching a preset number of iterations, a first process parameter of the target plasma is generated based on the current individual.

7. The multi-frequency modulation plasma surface treatment system according to claim 6, characterized in that: The control module includes: a power amplifier, a radio frequency source and a power distribution network; The RF source generates a RF signal of target frequency and power based on the first process parameter and inputs the signal into the power amplifier; wherein the RF source includes a plurality of frequency generating units, each of which is used to generate a target frequency; The power amplifier amplifies the power of the radio frequency signal generated by the radio frequency source based on the target frequency, and each target frequency corresponds to an independent power amplifier; The power distribution network is used to distribute the power of the amplified radio frequency signal to the plasma source.

8. The multi-frequency modulation plasma surface treatment system according to claim 7, characterized in that: The step of performing multi-frequency modulation on the target plasma based on the first process parameter to adjust the physical properties of the target plasma includes: The RF source sets the frequency of the RF signal generated by each of the frequency generating units based on the target frequency, and sends the frequency to the power amplifier corresponding to each of the target frequencies; The power amplifier amplifies the output power of the radio frequency signal at each frequency based on the output power of the target frequency; The power distribution network receives radio frequency signals of various frequencies after amplifying the output power and distributes them to different electrodes or regions of the plasma source to adjust the physical properties of the target plasma.

9. A multi-frequency modulation plasma surface treatment method, which is implemented based on the multi-frequency modulation plasma surface treatment system according to any one of claims 1 to 8, characterized in that: include: Obtain etching process information, and obtain characteristic data of target plasma under multi-frequency modulation through sensors, including optical data, electrical data, and gas composition data; Align different types of feature data to the same timeline; Inputting the characteristic data of the target plasma aligned to the same time axis into a preset machine learning model to generate process status assessment information; Based on the process state evaluation information, a genetic algorithm is used for optimization to generate a first process parameter of the target plasma; Based on the first process parameter, a plasma source of the target plasma is multi-frequency modulated to adjust the physical properties of the target plasma.

Images