Remote plasma source dissociation rate detection method and device

Abstract

The invention belongs to the technical field of plasma sources, and provides a remote plasma source dissociation rate detection method and device. The method comprises the following steps: sampling a near infrared spectrum of a plasma source to be detected; performing feature prediction on the near infrared spectrum by using a pre-constructed ionization identification model to obtain a spectral feature prediction value; optimizing the spectral feature predicted values through an optimization model, and determining an optimal solution in the spectral feature predicted values; and obtaining the concentration of the residual gas according to the optimal solution, and determining the dissociation rate of the remote plasma source. The spectrum prediction value is obtained through the ionization identification model, and then the optimal solution of the characteristic spectrum data is determined through the optimization model, so that an error value with smaller difference with the result can be obtained, the detection precision of the remote plasma dissociation rate is improved, and the digital representation of the dissociation rate is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of plasma sources, and particularly to a method and device for detecting the dissociation rate of a remote plasma source. Background Art

[0002] A remote plasma source (RPS) is an advanced plasma generation device, also known as a remote high-density plasma generator, and is commonly used for the cleaning and etching processes of process chambers in the integrated circuit manufacturing process. Its core function is to efficiently and stably ionize the gas used for cleaning (NF3) or the gas process gas (NH3, O2), thereby greatly enhancing the activity of the reaction gas and the consistency of the process. Compared with traditional plasma sources, there is a physical separation between the plasma generation region and the processing region in the remote plasma source. After the plasma is generated, it is transported to the processing region, and the active particles (such as free radicals, ions, and neutral particles) diffuse during the transportation process. This design enables the plasma to act on the surface of the material to be processed more uniformly and efficiently.

[0003] The dissociation rate is one of the important parameters characterizing the reaction efficiency of the remote plasma source. It refers to the proportion of the reaction gas dissociated into active species (such as free radicals, single-core ions, multi-core ions, and neutral atoms) in a strong electric field environment. A higher and more stable dissociation rate means that at the same input power, a higher plasma density can be obtained, thereby reducing the cleaning time of the process chamber, improving the processing efficiency and quality. Therefore, the measurement of the dissociation rate and its stability is crucial.

[0004] Currently, the following main defects exist in the detection of the remote plasma dissociation rate: (1) The existing detection methods have insufficient accuracy, so they cannot effectively evaluate the influence of external factors on the dissociation rate of the RPS; (2) The high cost and complexity of the traditional mass spectrometry method make real-time monitoring difficult; (3) The low sensitivity of the spectroscopic technique to single atoms or ions; the insufficient energy of the atmospheric pressure ionization technique leads to incomplete information; (4) The Langmuir probe method has the problem of contaminating the plasma, resulting in errors in the data.

[0005] The above defects limit its application in high-precision scenarios such as semiconductor manufacturing, affecting the application efficiency and reliability in the plasma etching process. Summary of the Invention

[0006] Aiming at the defects in the prior art, the present invention provides a method and device for detecting the dissociation rate of a remote plasma source to solve the problems of insufficient accuracy in the current dissociation rate detection and inapplicability to high-precision scenarios such as semiconductor manufacturing.

[0007] In a first aspect, a method for detecting the dissociation rate of a remote plasma source provided by the present invention includes: Sampling the near-infrared spectrum of the plasma source to be detected; Using a pre-constructed ionization discrimination model to predict the characteristics of the near-infrared spectrum to obtain a spectral characteristic prediction value; the construction method of the ionization discrimination model includes: Based on the support vector machine algorithm, determine the optimization problem corresponding to the ionization discrimination model. The objective function and constraint conditions in the optimization problem are respectively: Objective function , represents the variable coefficient vector, represents the bias value vector, represents the allowable fitting error; Constraint condition , is the mapping function, represents the regularization coefficient, represents the input spectral matrix, represents the infrared data, represents the physical and chemical value, represents the number of samples; The kernel function of the optimization problem is , represents the width coefficient of the kernel function; The predicted value of the ionization discrimination model , are Lagrange multipliers, where and are obtained by introducing the KT condition for solution; Optimize the spectral characteristic prediction value through the optimization model to determine the optimal solution in the spectral characteristic prediction value; Obtain the residual gas concentration according to the optimal solution to determine the dissociation rate of the remote plasma source.

[0008] As can be seen from the above technical solutions, the detection method provided by the present invention obtains the spectral prediction value through the ionization discrimination model, and then determines the optimal solution of the characteristic spectral data through the optimization model, which can obtain an error value with a smaller difference from the result, improve the detection accuracy of the remote plasma dissociation rate, and realize the digital characterization of the dissociation rate.

[0009] Optionally, before the near-infrared spectrum is input into the ionization discrimination model, it also needs to be preprocessed, specifically including baseline correction, vector normalization, and smoothing and noise reduction.

[0010] Optionally, the construction method of the optimization model includes: Establish a multi-dimensional Taylor network; Determine the weights of each product term in the input layer of the multi-dimensional Taylor network through the sparrow search algorithm; Compare the expected output of the multi-dimensional Taylor network with the output of the output layer. If the error exceeds the threshold range, update the weights through backpropagation; after the maximum number of training times is satisfied, output the optimal weights.

[0011] Optionally, it further includes evaluating the optimal solution, including: Use the partial least squares regression method to compare the optimal solution with the pre-acquired reference characteristic spectrum to obtain the prediction deviation amount; Compare the prediction deviation amount with the preset deviation value 、 for comparison. When y > , it is determined that the optimal solution is not close to the reference characteristic spectrum. When y < , it is determined that the optimal solution is close to the reference characteristic spectrum. When ≤ y ≤ , it indicates that the determination is unstable or cannot be determined. < .

[0012] Optionally, the reference characteristic spectrum is obtained by using the centroid method after collecting the characteristic spectrum set multiple times .

[0013] In a second aspect, a remote plasma source dissociation rate detection device provided by the present invention includes: An acquisition module for sampling the near-infrared spectrum of the plasma source to be detected; An ionization discrimination module for using a pre-constructed ionization discrimination model to perform feature prediction on the near-infrared spectrum to obtain a spectral feature prediction value; the construction method of the ionization discrimination model includes: Based on the support vector machine algorithm, determine the optimization problem corresponding to the ionization discrimination model. The objective function and constraint conditions in the optimization problem are respectively: Objective function , represents the variable coefficient vector, represents the bias value vector, represents the allowable fitting error; Constraint condition , is the mapping function, represents the regularization coefficient, represents the input spectral matrix, represents the infrared data, represents the physical and chemical value, represents the number of samples; The kernel function of the optimization problem is , represents the width coefficient of the kernel function; The predicted value of the ionization discrimination model , is a Lagrange multiplier, where and are obtained by solving through the introduction of the KKT conditions; An optimization module for optimizing the predicted spectral feature values through an optimization model to determine the optimal solution among the predicted spectral feature values; A dissociation calculation module for obtaining the residual gas concentration based on the optimal solution and determining the dissociation rate of the remote plasma source.

[0014] Adopting the above technical solution, the present application has the following beneficial effects: The detection method provided by the present invention obtains the spectral prediction value through the ionization discrimination model, and then determines the optimal solution of the characteristic spectral data through the optimization model, which can obtain an error value with a smaller difference from the result, improves the detection accuracy of the remote plasma dissociation rate, and realizes the digital characterization of the dissociation rate. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the description of the specific embodiments or the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0016] Figure 1 Shows a structural block diagram of a remote plasma source dissociation rate detection device provided by an embodiment of the present invention; Figure 2 Shows a flowchart of a remote plasma source dissociation rate detection method provided by an embodiment of the present invention; Figure 3 Shows a schematic diagram of the process expansion of a remote plasma source dissociation rate detection method provided by an embodiment of the present invention; Figure 4 Shows a schematic diagram of the process of an optimization model provided by an embodiment of the present invention; Figure 5 Shows a schematic diagram of a multi-dimensional Taylor network in the optimization model provided by an embodiment of the present invention; Figure 6 Shows a linear schematic diagram of the residual concentration and the characteristic spectral peak provided by an embodiment of the present invention; Figure 7 Shows a schematic diagram of the process of a dissociation rate detection method for control data provided by an embodiment of the present invention; Figure 8 Shows a comparison diagram of the standard deviation effect of the characteristic spectral peak obtained by the detection method provided by an embodiment of the present invention; Figure 9Shows a comparison schematic diagram of the detection method provided by the embodiments of the present invention. Detailed implementation manners

[0017] The embodiments of the technical solution of the present invention will be described in detail below with reference to the accompanying drawings. The following embodiments are only used to illustrate the technical solution of the present invention more clearly, so they are only examples and cannot be used to limit the protection scope of the present invention. It should be noted that unless otherwise specified, the technical terms or scientific terms used in this application should have the ordinary meanings understood by those skilled in the art to which the present invention belongs.

[0018] As Figure 1 shown, a remote plasma source (RPS) is used to dissociate process gases, and processes such as thin film deposition cleaning are carried out under vacuum conditions; an etching test station is used to etch thin films on materials such as silicon wafers and wafers. At present, accurate measurement of the dissociation rate of the plasma source cannot be achieved. Therefore, in one embodiment, as Figure 2 shown, a method for detecting the dissociation rate of a remote plasma source is provided, including: S1. Sampling the near-infrared spectrum of the plasma source to be detected.

[0019] Using a BRUKER TENSORⅡ Fourier transform infrared spectrometer, the infrared spectrum of the remote plasma source at different dissociation rates is measured. The original spectral data is collected by Result software and transmitted to TQAnalyst6.2 software for data preprocessing and calculation; the infrared spectrum collection conditions of the remote plasma source are: spectral range 4000~10000cm -1 , the number of scans is 32~128 times, and the resolution is 4~16cm -1 .

[0020] The collected infrared spectral data is preprocessed in the spectral range of 4000~10000cm -1 to eliminate interference and standardize the spectrum; the preprocessing steps include the processes of baseline correction, vector normalization, smoothing and noise reduction, and data calibration. Data preprocessing can eliminate the influence of factors such as offset and baseline change, weaken the influence of non-important features on the infrared spectral curve, retain and highlight the effective information, and ensure a good correlation between the spectral data and the output dissociation rate.

[0021] 1. Baseline correction: Multiple scattering correction / Polynomial fitting correction to remove instrument noise, baseline drift or background scattering; Multivariate scattering correction uses the average of the spectral curves of all samples as the standard spectrum, and then performs a univariate linear regression analysis on the spectrum of each sample to calculate the regression coefficient and regression constant. Finally, the original spectrum of each sample is subtracted by the regression constant and divided by the regression coefficient to correct the relative tilt of the spectral baseline, thereby eliminating the interference information in the spectrum.

[0022] The spectral matrix of the test sample is , and its average spectrum is , and the x linear regression formula is: (1) In the formula, m is the multi-dimensional spectral wavenumber variable, is the number of multi-dimensional spectral samples.

[0023] Analyze the spectral data of the sample and calculate and values, and based on this, deduce the scattering equation after spectral preprocessing: (2) 2. Vector normalization: Perform unit vector normalization on the entire spectrum to eliminate the influence of sample concentration differences on spectral intensity; Calculate the average spectrum value and standard deviation, input the infrared spectral data , calculate the average spectrum value of the sample, and calculate the standard deviation s of the spectrum: (3) In the formula, n represents the number of samples, ; Perform normalization processing on the sample spectrum , and the formula is as follows: (4) Vector normalization reduces the interference signal of the spectrum and increases the stability of the model.

[0024] 3. Smoothing and noise reduction: Use the Savitzky-Golay smoothing algorithm to filter out high-frequency noise while retaining the peak shape; The convolutional smoothing Savitzky-Golay algorithm performs polynomial fitting through local data, uses the least squares method to solve the smoothing factor, and removes noise from the signal. The formula is as follows: (5) Among them, represents the smoothing factor; H represents the normalization factor; the smoothing factor fits the best signal trend according to the distribution of data points collected within the window.

[0025] 4. Feature extraction: Mining key information in the spectrum (such as peak position, peak height, peak area, peak width, peak shape).

[0026] S2. Use a pre-built ionization discrimination model to predict the features of the near-infrared spectrum and obtain the spectral feature prediction values.

[0027] The ionization discrimination model in this step is implemented based on the support vector machine model. When performing function fitting, the support vector machine model transforms the spectral data from a low-dimensional space to a high-dimensional space and uses equality constraints to replace the original inequality constraints in the SVM, reducing the computational complexity and improving the computational speed. The SVM has strong learning ability and provides an effective solution for complex problems such as limited sample data, high sparsity, high dimensionality, and non-linearity. The specific calculation process of the SVM is as follows: S201: Spectrum matrix Input, where represents the infrared data, represents the physical and chemical values, represents the number of samples; S202: Use a high-dimensional linear function to fit the data and set the prediction model as: (6) where represents the variable coefficient vector, represents the bias value vector, and use the mapping function to map the input data to a high-dimensional space.

[0028] S203: According to the support vector machine principle, and the optimal solutions of are obtained by minimizing a specific function, and the model optimization problem is transformed into the following form: (7) The constraint condition is: (8) where represents the regularization coefficient, represents the allowable fitting error.

[0029] S204: Introduce Lagrange multipliers and construct the Lagrangian function: (9) is the Lagrange multiplier. Take the partial derivatives of, , , , and introduce the KT conditions: (10) S205: Cancel variables , , to obtain a matrix equation: (11) where ; ; ; S206: According to previous research experience, different kernel functions perform differently when dealing with training samples of different scales and dimensions. The radial basis kernel function is selected as the kernel function for the optimization problem, and the expression is as follows: (12) where represents the width coefficient of the kernel function.

[0030] S207: Calculate and according to S205, transform the non - linear equation into a linear equation, input the test set sample x, and calculate the predicted value y: (13) S208: Randomly divide the sample data into a 70% training set and a 30% test set. Among them, the data in the training set is input into the optimization model for optimization.

[0031] S3. Optimize the predicted values of spectral features through the optimization model to determine the optimal solution among the predicted values of spectral features.

[0032] Among them, the construction method of the optimization model includes: Establish a multi - dimensional Taylor network; Determine the weights of each product term in the input layer of the multi - dimensional Taylor network through the sparrow search algorithm; Compare the expected output and the output of the output layer of the multi - dimensional Taylor network. If the error exceeds the threshold range, update the weights through backpropagation; after meeting the maximum number of training times, output the optimal weights.

[0033] Specifically, as Figure 4 shown, step S3 combines the local optimal solution of the sparrow search and the multi - dimensional Taylor network strategy through the self - adaptive sparrow multi - dimensional Taylor network (SSA - MTN) to avoid local optima and is suitable for multi - peak optimization problems.

[0034] The principle of applying the multi-dimensional Taylor network to the detection of gas dissociation rate by infrared spectroscopy is based on the learning and classification ability of neural networks. The local solutions output by the sparrow algorithm are optimized and used as the input layer variables. After passing through the hidden layer, the global optimal solution is output from the output layer. During the training process, the backpropagation algorithm is used to continuously adjust the network parameters to minimize the error between the network output and the actual characteristic spectrum, so as to accurately classify and identify the characteristic spectrum. Suppose the input layer of the multi-dimensional Taylor network has n nodes, the data processing layer has N(n,m) nodes, and the data processing layer realizes the weighted summation of the product terms of each power of the input variables. The multi-dimensional Taylor network uses polynomials composed of addition and multiplication to approximate non-linear functions and can be used to fit multi-variable functions. The basic principle of the multi-dimensional Taylor network is as follows: S301: Standardize all the eigenvalue of the sample training data set: The normalization process is to uniformly transform the eigenvalue sequence into the interval [0,1], mainly to prevent the problem that too large or too small sample data may cause non-convergence or slow convergence of data convergence. The calculation formula is as follows: (14) Among them, in an input eigenvalue sequence, M i is the input eigenvalue after normalization, N i is the input eigenvalue before normalization, N max is the maximum input eigenvalue, N min is the minimum input eigenvalue.

[0035] S302: Initialize the MTN network and determine the structure of the MTN network, adopting a structure including an input layer, a hidden layer, and an output layer; where i represents the input layer nodes, j represents the hidden layer nodes, l represents the output layer nodes, and determine the number of input nodes m and the number of hidden layer nodes q, and the number of output layer nodes n; the input sample , the output is , the expected output is , and initialize the initial values of the weighted coefficients from the input layer to the hidden layer and from the hidden layer to the output layer and the bias ; The calculation formula for each layer is: (15) Among them, T(x) represents the activation function, represents the weight coefficient connecting the i-th neuron in the previous layer and the j-th neuron in the current layer, represents the bias of the j-th neuron in the current layer; For a multi-layer network, the calculation is performed using the feed-forward propagation method, that is, each layer is calculated according to Equation (15) until the last output layer; for the i-th neuron in the input layer, its output Xi is the i-th eigenvalue of the input data.

[0036] The Sparrow Search Algorithm (SSA) is based on two behaviors of sparrows in nature: foraging and anti-foraging, simulating the group dynamics and information dissemination mechanism of sparrows when searching for food; sparrows are divided into explorers and followers. Sparrows that can find better food are explorers, and their fitness function is higher; the remaining sparrows are followers and are affected by the direction of the explorers. The process of the Adaptive Sparrow Multi-dimensional Taylor Network Algorithm is as follows: S303: Initialization of the sparrow group position; Let the dimension of the sparrow population be , is the number of all sparrows, and H represents the dimension of the variable to be optimized. The position of the i-th sparrow can be described as , , . represents the position of the i-th sparrow with j-dimensional variables. The fitness of the sparrow population can be expressed as (16) where each row in represents the fitness value of a sparrow individual.

[0037] (17) where, , represent the upper and lower position boundaries of the population respectively.

[0038] S304: Update of the explorer position; (18) where t represents the current iteration number, represents the value of the j-th dimension of the i-th sparrow at the t-th iteration number, is the preset maximum iteration number, is a uniformly random number in (0, 1], Q is a random number from a standard normal distribution, is a 1×H matrix, and all elements in it are 1. , are the warning value and the safety threshold respectively. When , the sparrows conduct global search for food; when , the sparrows perform random walks according to a normal distribution. The explorers update their positions through global search to find potential optimal solutions; S305: Update of the follower position: (19) Among them, is the sparrow with the optimal fitness in the (t + 1)-th generation population, is the sparrow with the worst fitness in the t-th generation population; when holds, the sparrows with lower fitness do not obtain food, so they need to obtain more food; in other cases, it means that the sparrows randomly search for a position near the current optimal position. The followers adjust their positions through local search to optimize the current solution; S306: Reconnaissance and early warning, simulating the behavior of sparrows when facing the threat of predators, and enhancing the ability of the algorithm to jump out of the local optimal solution; (20) Where is a random number conforming to the normal distribution, and K is a random number between [-1, 1], which follows the standard normal distribution and can control the search step size; is a relatively small number to prevent the denominator from being zero. is the sparrow with the optimal fitness in the t-th generation population, is the worst fitness value of the current population, is the global optimal fitness value. When holds, the sparrows are at the edge of the population and are vulnerable to attacks by foragers, and at this time they fly towards the optimal sparrows; when holds, at this time the sparrows are in a dangerous position and are far from the position of the worst sparrows.

[0039] S307: Update the optimal solution: Update the local best solution according to the fitness to guide the next search of the sparrow group; S308: Iterative training, repeat the above process, if the maximum number of iterations is reached, output the optimized weights; if not satisfied, return to step P2 to re-solve the local optimal solution by the sparrow algorithm; S309: Take the updated weights obtained in the sparrow search algorithm as input, feed them into the MTN network, and use the MTN to compare the predicted result with the actual result; if the error exceeds the threshold range, perform backpropagation to update the weights; During the backpropagation process, the activation function of the multi-dimensional Taylor network MTN adopts the Tansig function: (21) Perform Taylor expansion on the Tansig function activation function in formula (21), where x represents the sample data: (22) Give a general description of the Taylor expansion of the activation function T(x), where represents the Peano remainder after expansion: (23) After using the nth-order Taylor expansion of the activation function to replace the original activation function, the output of the jth hidden layer node in the data processing layer is : (24) Among them, represents the number of polynomial terms corresponding to the mth power expansion of the n-variable polynomial, , is the coefficient corresponding to each power product term, and ; Perform a linear combination of the outputs of all the hidden nodes in the network as the output of the MTN. is the combination coefficient, so the output of the output layer of the MTN is described in the following form: (25) Lemma 1: Any continuous function defined on a closed interval can be arbitrarily accurately approximated by a polynomial function; Lemma 2: For a continuous function defined on a closed interval, it can be approximated by . Among them is the total number of product terms in the approximation formula, is the power of the variable of the tth product term in the expansion; It can be seen from Lemma 1 and Lemma 2 that any continuous function defined on a closed interval can be approximated by a multi-dimensional Taylor network with arbitrary accuracy. Then, the expected output can be represented by Equation (26): (26) According to the principle of the backpropagation algorithm, replace the gradient descent method for model training, and calculate the minimization of the squared error for optimizing the output parameters: (27) When the E obtained from Equation (27) exceeds the preset threshold range, update the weights through backpropagation. Taking the output of the j node as an example for backpropagation, calculate the gradient of the relevant weights of the network in the way of the activation function, as shown in the following equation: (28) Among them, represents the accumulated value obtained by multiplying the product of each item in the input layer by the corresponding weight, and its value is as shown in the following equation: (29) Among them, m represents the number of times of the highest expansion item in the data processing layer of the multi-dimensional Taylor network, represents the input in the data processing layer of the multi-dimensional Taylor network The number of terms of the polynomial after m - power expansion respectively represent the weights corresponding to n nodes in the output layer ; Then it represents the power on in the j - th polynomial; Here, the variable is unknown, but its relative change can be measured, that is: , The relative change amount of (30) Calculate the average gradient of each weight from the gradient vector obtained in formula (30) 、 , and update the weights according to the following formula (31); (31) After the weight update in step S309, leave the best position and the best fitness; S310: Determine whether the maximum number of training times is satisfied. If it is satisfied, output the best weights according to the best position and the best fitness to obtain the optimal weight threshold of the multi - dimensional Taylor network, and then perform the characteristic infrared spectrum model determination; if it is not satisfied, return to step S303 to re - solve the local optimal solution by the sparrow algorithm.

[0040] In one embodiment, it further includes S5. Evaluate the optimal solution, including: Use the partial least - squares regression method to compare the optimal solution with the pre - obtained reference characteristic spectrum to obtain the prediction deviation amount; Compare the prediction deviation amount with the preset deviation value 、 . When y> , it is determined that the optimal solution is not close to the reference characteristic spectrum. When y< , it is determined that the optimal solution is close to the reference characteristic spectrum. When ≤y≤ , it indicates that the determination is unstable or cannot be determined, < .

[0041] Based on the optimal solution of the characteristic spectral data obtained after updating the weights by the SSA - MTN algorithm, use the partial least - squares regression method to conduct a comparative analysis with the sampling spectrum; the calculation process of the partial least - squares regression algorithm (PLSR) is as follows: S501: For the spectral matrix X and the physical - chemical value reference matrix Y , perform eigenvalue decomposition: (32) Among them, T 、U represents a score matrix, P , Q represents a loading matrix, E , F represents the model fitting error.

[0042] S502: Establish a multiple linear regression model for the score matrix T and U as follows: (33) where represents the error matrix, B represents the regression coefficient matrix, B The solution of (34) S503: The prediction formula for an unknown sample is as shown in Equation (35): (35) where x represents the spectral data of the unknown sample, which is an unknown quantity in the spectral matrix X, i.e., the optimal solution obtained in step S3, y represents the prediction deviation of the unknown sample.

[0043] In this embodiment, = 0.4, = 0.5.

[0044] When y > 0.5, it is determined that the sample does not belong to this class, indicating that the optimal solution obtained in step S3 is close to the reference quantity; when y < 0.4, it is determined that the sample belongs to this class, indicating that the optimal solution obtained in step S3 is not close to the reference quantity; when 0.4 ≤ y ≤ 0.5, it indicates that the determination is unstable or cannot be determined.

[0045] Among them, the spectral matrix X in step S501 is an independent variable matrix, which is a matrix composed of all unknown spectral data; the physicochemical value reference matrix Y is obtained by using the centroid method after collecting the characteristic spectral set multiple times.

[0046] The measurement method of the spectral calibration value is to introduce M sccm of nitrogen trifluoride NF3 as the residual gas when the RPS is not working, N sccm of nitrogen N2 as the ionized gas, then the residual concentration is C = M / ( N + M ), and the dissociation rate is 1 - C ; Residual concentration of calibration value C It is necessary to collect the characteristic spectral sets multiple times After that, the centroid method is used to obtain it. The specific steps are as follows: The centroid method is used to calculate the centroid value of each element in the output quantity and its corresponding membership degree to obtain the output quantity C : (36) Among them, C i is the central value of the membership function interval corresponding to the output quantity, u i is C i the corresponding membership degree

[0047] S4. Obtain the residual gas concentration according to the optimal solution and determine the dissociation rate of the remote plasma source

[0048] From Figure 6 It can be seen that the characteristic spectral peak extracted by the method provided in this embodiment is positively correlated with different residual gas concentrations, and the error is small, which can basically reflect the current situation of the residual gas concentration of the gas. When the residual concentration C is determined, the current dissociation rate can be calculated

[0049] Under the conclusion that the characteristic spectral peak is positively correlated with the residual gas concentration and negatively correlated with the dissociation rate obtained in this embodiment, the effect of the detection method provided in this embodiment is further verified. Before that, the specific steps of the dissociation rate acquisition method for the control data are briefly described as follows, such as Figure 7 shown, including: S601. Collect the near-infrared spectra of different dissociation rates of the remote plasma source by a Fourier transform infrared spectrometer; that is, the gas ionization detection device respectively detects the amount M of the residual gas and the amount N of the input maintenance gas in the dissociation chamber, and the dissociation rate is 1 - M / N S602. Establish an ionization rate data set using the near-infrared spectral data of different dissociation rates. The data set includes the characteristic spectral peaks of the dissociation rate S603. Under the same near-infrared spectrum acquisition conditions, sample the near-infrared spectrum of the plasma source to be detected, and use the data set to determine the infrared spectrum of the plasma source to be detected

[0050] From Figure 8It can be seen that the standard deviation error between the characteristic spectral peak obtained by the method provided in this embodiment and the calibration value is less than 0.2, while the standard deviation error between the characteristic spectral peak collected in steps S601 - S603 and the calibration value is 0.3 - 0.7. It can be seen that after the update training by the method provided in this embodiment, for the optimal solution of the obtained characteristic spectral data, by using the partial least squares regression method to compare and analyze the optimal solution of the characteristic spectrum with the sampled spectrum, an error value with a smaller difference from the result can be obtained.

[0051] As Figure 9 It can be seen that the spectral data obtained by the method provided in this embodiment can well track the calibration value spectrum, and the concentration represented by the spectral peak has an error of 0.05 from the calibration value; while the error between the spectral peak collected in steps S601 - S603 and the calibration value is 0.35, indicating that the method provided in this embodiment has a higher acquisition accuracy for the characteristic spectral peak and can meet the requirement of improving the detection accuracy of the dissociation rate of the remote plasma source to a certain extent.

[0052] In one embodiment, as Figure 1 shown, there is also provided a device for detecting the dissociation rate of a remote plasma source, including: An acquisition module, configured to sample the near - infrared spectrum of the plasma source to be detected; An ionization discrimination module, configured to use a pre - constructed ionization discrimination model to perform feature prediction on the near - infrared spectrum to obtain a spectral feature prediction value; An optimization module, configured to optimize the spectral feature prediction value through an optimization model to determine the optimal solution in the spectral feature prediction value; A dissociation calculation module, configured to obtain the residual gas concentration according to the optimal solution and determine the dissociation rate of the remote plasma source.

[0053] The device for detecting the dissociation rate of a remote plasma source provided in the embodiments of the present application and the above - mentioned method for detecting the dissociation rate of a remote plasma source adopt the same inventive concept and can achieve the same beneficial effects, which will not be elaborated here.

[0054] The above - mentioned embodiments are only used to introduce the technical solutions of the present application in detail, but the descriptions of the above - mentioned embodiments are only used to help understand the method of the embodiments of the present invention and should not be construed as a limitation to the embodiments of the present invention. Any changes or substitutions that can be easily thought of by those skilled in the art should be covered within the protection scope of the embodiments of the present invention.

Claims

1. A remote plasma source dissociation rate detection method, characterized in that: include: Sampling the near infrared spectrum of the plasma source to be tested; Using a pre-built ionization identification model to perform feature prediction on the near infrared spectrum to obtain a spectral feature prediction value; The method for constructing the ionization identification model comprises: Based on the support vector machine algorithm, the optimization problem corresponding to the ionization identification model is determined, and the objective function and constraint conditions in the optimization problem are: Objective Function , represents the variable coefficient vector, represents the bias value vector, Indicates the allowable fitting error; Constraints , is the mapping function, represents the regularization coefficient, represents the input spectral matrix, Indicates infrared data, Indicates the physical and chemical value, Indicates the number of samples; The kernel function of the optimization problem is , Represents the width coefficient of the kernel function; Predictive value of the ionization discrimination model , is the Lagrange multiplier, where and Obtained by introducing KT condition solution; Optimizing the spectral feature prediction values ​​through an optimization model to determine the optimal solution among the spectral feature prediction values; The residual gas concentration is obtained according to the optimal solution, and the dissociation rate of the remote plasma source is determined.

2. The method according to claim 1, characterized in that Before the near infrared spectrum is input into the ionization identification model, it is also necessary to perform preprocessing, including baseline correction, vector standardization and smoothing noise reduction.

3. The method according to claim 1, characterized in that The method for constructing the optimization model comprises: Establishing a multidimensional Taylor network; The weights of each product term in the input layer of the multidimensional Taylor network are determined by the sparrow search algorithm; The expected output of the multidimensional Taylor network and the output layer output are compared. If the error exceeds a threshold range, the weights are updated by back propagation. When the maximum number of training times is met, the optimal weights are output.

4. The method according to claim 3, characterized in that Also included is an evaluation of the optimal solution, including: The optimal solution is compared with a pre-acquired reference characteristic spectrum using a partial least squares regression method to obtain a predicted deviation; The predicted deviation and the preset deviation value , For comparison, when y> When the optimal solution is determined to be not close to the reference characteristic spectrum, when y< When the optimal solution is close to the reference characteristic spectrum, ≤y≤ When the judgment is unstable or uncertain, < .

5. The method according to claim 4, characterized in that The reference characteristic spectrum is a set of characteristic spectra collected multiple times. Then the centroid method is used to obtain it.

6. A remote plasma source dissociation rate detection device, characterized in that: include: A collection module, used for sampling the near infrared spectrum of the plasma source to be detected; An ionization identification module, used to use a pre-built ionization identification model to perform feature prediction on the near-infrared spectrum to obtain a spectral feature prediction value; The method for constructing the ionization identification model comprises: Based on the support vector machine algorithm, the optimization problem corresponding to the ionization identification model is determined, and the objective function and constraint conditions in the optimization problem are: Objective Function , represents the variable coefficient vector, represents the bias value vector, Indicates the allowable fitting error; Constraints , is the mapping function, represents the regularization coefficient, represents the input spectral matrix, Indicates infrared data, Indicates the physical and chemical value, Indicates the number of samples; The kernel function of the optimization problem is , Represents the width coefficient of the kernel function; Predictive value of the ionization discrimination model , is the Lagrange multiplier, where and Obtained by introducing KT condition solution; An optimization module, used to optimize the spectral feature prediction value through an optimization model to determine the optimal solution among the spectral feature prediction values; The dissociation calculation module is used to obtain the residual gas concentration according to the optimal solution and determine the dissociation rate of the remote plasma source.

Images