A remote plasma source dissociation rate detection method

Abstract

The present application belongs to the technical field of remote plasma source, and provides a remote plasma source dissociation rate detection method, comprising: collecting infrared spectrum data of the remote plasma source; inputting the infrared spectrum data collected multiple times into a pre-trained plasma source ionization identification model to obtain characteristic spectrum peak values respectively and to obtain the average value of the characteristic spectrum peak values; in the collection process, compensating environmental interference through a parameter control mechanism to make the characteristic spectrum peak values positively correlated with the residual gas concentration; and calculating the dissociation rate based on the average value of the characteristic spectrum peak values. The present application realizes real-time monitoring of the dissociation rate of the remote plasma source, and controls the environmental interference through the parameter control mechanism to make the characteristic spectrum peak values positively correlated with the residual gas concentration, so that the dissociation rate can be solved by replacing the residual gas concentration with the characteristic spectrum peak values, thereby ensuring the accuracy and being suitable for real-time closed-loop control of the detection of the dissociation rate of multi-element gas plasma in semiconductor processes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of semiconductor integrated circuit manufacturing, and in particular to a remote plasma source dissociation rate detection method. BACKGROUND

[0002] Remote plasma sources are widely used in the fields of semiconductors, photovoltaics, and chemical industry. Dissociation rate detection is a key technical link in semiconductor manufacturing and surface treatment processes. Remote plasma sources (RPS) can effectively avoid damage to sensitive devices by physically separating the plasma generation area from the processing area, while providing high-activity free radicals and neutral particles, and are widely used in chip etching, surface cleaning, thin film deposition, and other processes. Dissociation rate, as a core parameter for measuring plasma reaction efficiency, directly affects process effectiveness and product quality. For example, in fluorine-based plasma cleaning, the dissociation rate of NF3 and other reaction gases determines the concentration of active fluorine radicals, which in turn affects the removal efficiency and uniformity of silicon dust.

[0003] The existing technology has the following main defects in remote plasma dissociation rate detection: high cost and complexity of traditional mass spectrometry, difficulty in real-time monitoring; low sensitivity of spectroscopy to single atoms or ions; insufficient energy of atmospheric ionization technology leading to incomplete information; interference problem of probe method; and insufficient accuracy of rapid detection methods. These defects limit the application efficiency and reliability of the technology in high-precision scenarios such as semiconductor manufacturing. SUMMARY

[0004] In view of the defects in the prior art, the present application provides a remote plasma source dissociation rate detection method to solve the problems of real-time detection difficulty and large interference in the current remote plasma source dissociation rate detection.

[0005] In a first aspect, the present application provides a remote plasma source dissociation rate detection method, comprising:

[0006] Collecting infrared spectrum data of the remote plasma source;

[0007] Inputting the multiple collected infrared spectrum data into a pre-trained ionization discrimination model to obtain feature spectrum peak values respectively, and calculating the average value of the feature spectrum peak values; during the collection process, compensating for environmental interference through a parameter regulation mechanism to make the feature spectrum peak values positively correlated with the residual gas concentration; the parameter regulation mechanism makes the feature spectrum peak values positively correlated with the residual gas concentration, comprising:

[0008] Establishing a target function about the process parameter compensation feature spectrum peak value; the target function , is the state calibration value of the current moment feature spectrum peak value, a sampling value of a characteristic spectral peak value at a current moment; a sampling value of a gas pressure, a chamber temperature, and a gas flow at a current moment k, a sampling value of a gas pressure, a chamber temperature, and a gas flow at a previous moment k-1, respectively corresponding weight proportions;

[0009] adjusting the chamber temperature , the gas pressure , and the gas flow in real time according to the target function to minimize the target function

[0010] calculating the dissociation rate based on the average value of the characteristic spectral peak value.

[0011] According to the technical solution, the method for detecting the dissociation rate of the remote plasma source is provided, the dissociation rate of the remote plasma source is monitored in real time, the environmental interference is regulated by the parameter regulation mechanism, the molar absorption coefficient is compensated, the positive correlation between the characteristic spectral peak value and the residual gas concentration is formed, the characteristic spectral peak value is directly replaced by the residual gas concentration, the dissociation rate is solved, the precision is ensured, and the real-time closed-loop control of the dissociation rate of the multi-element gas plasma in the semiconductor process is applicable.

[0012] Optionally, the state calibration value of the characteristic spectral peak value at the current moment is determined according to , and the method comprises the following steps.

[0013]

[0014] wherein, is a cumulative sampling time at the current moment k, is a cumulative sampling time at the previous moment k-1, is an average value of the characteristic spectral peak value output by the ionization identification model when the infrared spectral data at the previous moment k-1 are input multiple times; A is a time-varying frequency factor, and B is an influence factor of the switching frequency change rate.

[0015] Optionally, the chamber temperature , the gas pressure , and the gas flow are adjusted in real time according to the target function to minimize the target function

[0016] ​​​​​​​When the chamber temperature is detected to be greater than a temperature threshold value , adjust the cooling water flow rate of the remote plasma source to stabilize the chamber temperature at a preset temperature value;

[0017] According to the relationship between the state value of the characteristic spectral peak at the current time and the sampling value of the characteristic spectral peak at the current time , adjust the gas pressure and the gas flow rate ; comprising:

[0018] When , no adjustment is made;

[0019] When , slowly reduce the gas pressure P and slowly increase the gas flow rate ;

[0020] When , slowly increase the gas pressure P and slowly reduce the gas flow rate .

[0021] Optionally, the weight proportion , , is determined according to the importance degree of the temperature T, the gas pressure P, and the gas flow rate Q to the characteristic spectral peak determined by the random forest algorithm.

[0022] Optionally, the training method of the ionization identification model comprises:

[0023] Based on the support vector machine algorithm, an optimization problem corresponding to the ionization identification model is determined, and the objective function and the constraint condition in the optimization problem are respectively:

[0024] The objective function , represents a variable coefficient vector, represents a bias value vector, represents an allowed fitting error, represents a regularization coefficient;

[0025] The constraint condition , is a mapping function, represents an input infrared spectrum matrix, represents infrared spectrum data, R represents a real number set, and d represents the dimension of the infrared spectrum data , represents the characteristic spectral peak corresponding to the output of the ionization identification model input, represents the number of samples corresponding to the scanning times; ​

[0026] The kernel function for the optimization problem is , For the sample data to be classified, This represents the sample data that has been classified. Represents the width coefficient of the kernel function;

[0027] The characteristic spectral peak , For Lagrange multipliers, among which and The solution is obtained by introducing the KT condition.

[0028] Alternatively, the dissociation rate can be calculated using the following formula:

[0029] dissociation rate , The characteristic spectral peaks corresponding to the concentration of the gas to be measured are input in the initial state. The value is the average of the characteristic spectral peaks output by the ionization identification model, which is formed by inputting infrared spectral data collected multiple times at the current time k.

[0030] By adopting the above technical solution, this application has the following beneficial effects:

[0031] This invention provides a remote plasma source dissociation rate detection method. By real-time monitoring of the dissociation rate of a remote plasma source and adjusting environmental interference through a parameter control mechanism to compensate for the molar absorptivity, a positive correlation is formed between the characteristic spectral peak and the residual gas concentration. This allows for direct replacement of the residual gas concentration with the characteristic spectral peak to solve for the dissociation rate, ensuring accuracy. This method is suitable for real-time closed-loop control of multi-element gas plasma dissociation rate detection in semiconductor processes. Attached Figure Description

[0032] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. 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.

[0033] Figure 1 A schematic diagram of a remote plasma source dissociation rate detection device provided in an embodiment of the present invention is shown;

[0034] Figure 2 A flowchart of a traditional dissociation rate acquisition method is shown;

[0035] Figure 3 A flowchart of a remote plasma source dissociation rate detection method provided by an embodiment of the present invention is shown;

[0036] Figure 4 This invention provides another flowchart of a remote plasma source dissociation rate detection method according to an embodiment of the present invention;

[0037] Figure 5 A flowchart illustrating the relationship between the characteristic spectral peaks and the importance of process parameters provided by this invention is shown.

[0038] Figure 6 A flowchart illustrating the decision tree construction process for the relationship between the characteristic spectral peaks and the importance of process parameters provided by this invention is shown.

[0039] Figure 7 This diagram illustrates the score ranking of the importance of characteristic spectral peaks and process parameters provided by the present invention.

[0040] Figure 8 The figure shows the experimental results comparing the remote plasma source dissociation rate detection method provided by the present invention with the traditional calibration method. Detailed Implementation

[0041] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are only used to more clearly illustrate the technical solution of the present invention and are therefore merely examples, and should not be construed as limiting the scope of protection of the present invention.

[0042] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0043] Figure 1 A schematic diagram of a remote plasma source dissociation rate detection module based on parameter coordinated control is shown; it consists of a remote plasma source, a gas ionization detection device, a Fourier transform infrared spectrometer, an RPS control platform, an etching platform, and a host computer system.

[0044] The remote plasma source is used to dissociate process gases and perform thin film deposition and cleaning processes under vacuum conditions; the etching stage is used to etch thin films on materials such as silicon wafers and crystal wafers.

[0045] A gas ionization detection device includes a maintenance gas input gas dissociation circuit control device connected to the dissociation chamber for controlling the amount of maintenance gas input to the dissociation chamber; the gas ionization detection device further includes an output detection unit connected to the dissociation chamber for detecting the amount of maintenance gas remaining in the gas in the dissociation chamber; the gas ionization detection device further includes an output calculation unit connected to the maintenance gas input gas dissociation circuit control device and the output detection unit for calculating the ionization rate based on the amount of maintenance gas input to the dissociation chamber and the amount of maintenance gas remaining in the gas output to the dissociation chamber.

[0046] The Fourier transform infrared spectrometer is used to collect infrared spectral data of the remaining sustaining gas; the RPS control station is used to control the process gas flow rate, vacuum pressure, water cooling system, etc. of the remote plasma source; the host computer system is used to perform preprocessing, feature extraction, pattern recognition and other operations on the infrared spectrum.

[0047] Based on the dissociation rate detection module described above, a dissociation rate acquisition method is currently provided, such as... Figure 2 As shown, it includes:

[0048] S110. First, the infrared spectra of the remote plasma source with different dissociation rates are collected by the Fourier transform infrared spectrometer; that is, the gas ionization detection device detects the amount M of the remaining gas in the gas of the dissociation chamber and the amount N of the input maintenance gas respectively, and the dissociation rate is 1-M / N.

[0049] S120. Establish a remote plasma source ionization identification model using infrared spectral data with different dissociation rates, that is, establish a dataset that corresponds one-to-one with the characteristic spectral peaks output by the ionization identification model for different residual gas concentrations, or use the Lambert-Beer law combined with the dynamic correction of the absorption coefficient to represent the relationship between the characteristic spectral peaks and the residual gas concentration.

[0050] The following uses NF3 as an example to illustrate the steps for calculating the real-time dissociation rate based on the Lambert-Beer law:

[0051] Step P1. Select the characteristic absorption peak

[0052] Identify the specific strong absorption peak of NF3 molecules in the infrared region, satisfying the linear relationship between absorbance and NF3 concentration;

[0053] Step P2. Calculate absorbance (A)

[0054] Extract the absorbance of the target characteristic peak from the measured spectrum:

[0055] (1)

[0056] in, The intensity of transmitted light in the presence of plasma. The background reference spectrum is the spectrum of pure NF3 gas or the spectrum of a cavity chamber without plasma.

[0057] Step P3. Apply Lambert-Beer Law to invert concentration.

[0058] (2)

[0059] The molar concentration of the remaining NF3 (mol·L) -1 ), The molar absorptivity of NF3 in the target wavelength band is given. The effective optical path length for an infrared beam to pass through plasma;

[0060] However, when the plasma source is operating, the actual ambient temperature is greater than 1000K, which reduces the molar absorptivity. If the value deviates from the normal temperature value, a high-temperature spectral database (such as HITRAN or custom experimental calibration) is required to correct the absorption coefficient. This step requires database calibration and also increases the complexity of the process.

[0061] S130. Using the same infrared spectral acquisition conditions, sample the infrared spectrum of the plasma source to be detected. Use a discrimination model to determine the infrared spectrum of the plasma source to be detected, and compare the detected characteristic spectral peaks with the database to obtain the current residual gas concentration. Then calculate the dissociation rate;

[0062] The dissociation rate F is calculated in real time as follows:

[0063] (3)

[0064] In equation (3), Input the initial concentration of NF3 gas;

[0065] When the initial concentration of the introduced NF3 gas is 100%, equation (3) can be simplified to:

[0066] (4)

[0067] However, using the Lambert-Beer law formula requires the prior establishment of a database and the comparison of the infrared spectrum to be sampled with the database spectrum, which makes it somewhat dependent on the database.

[0068] Therefore, in one embodiment, such as Figures 3-4 As shown, a method for detecting the dissociation rate of a remote plasma source is proposed, including:

[0069] S210. Acquire infrared spectral data from a remote plasma source.

[0070] Specifically, infrared spectral data of the remote plasma source was acquired using a BRUKER TENSOR II Fourier transform infrared spectrometer. The raw spectral data was acquired using Result software and then transferred to TQAnalyst software for data preprocessing and calculation. The infrared spectral acquisition conditions for the remote plasma source were: spectral range 400–4000 cm⁻¹. -1 The number of scans ranges from 32 to 128, with a resolution of 4 to 16 cm. -1 .

[0071] The collected infrared spectral data are in the range of 400~4000 cm⁻¹ -1 Preprocessing is performed within the spectral band to eliminate interference and standardize the spectrum. Data preprocessing can eliminate the influence of factors such as offset and baseline variation, reduce the impact of unimportant features on the infrared spectral curve, retain and highlight effective information, and ensure a good correlation between the spectral data and the output dissociation rate. The preprocessing steps include baseline correction, vector normalization, smoothing and noise reduction, and data calibration.

[0072] S211.1 Baseline Correction: Multiple scattering correction / polynomial fitting correction removes instrument noise, baseline drift, or background scattering. Baseline correction, specifically multiple scattering correction, uses the average of the spectral curves of all samples as the standard spectrum. Then, a univariate linear regression analysis is performed on the spectrum of each sample to calculate the regression coefficients and regression constants. Finally, the original spectrum of each sample is subtracted from the regression constant and divided by the regression coefficients to correct the relative tilt of the spectral baseline, thereby eliminating interference information in the spectrum.

[0073] The spectral matrix of the test sample is Its average spectrum is , The linear regression formula is:

[0074] (5)

[0075] In the formula, It is a multidimensional spectral wavenumber variable.

[0076] Analyze the spectral data of the sample and calculate and The value of is used to derive the scattering equation after spectral preprocessing:

[0077] (6)

[0078] S211.2 Vector Normalization: The entire spectrum is normalized using a unit vector to eliminate the influence of sample concentration differences on spectral intensity. The core idea of ​​the vector normalization algorithm is to center the mean of the spectral data and divide it by the standard deviation of the data to eliminate scale differences between data points, thereby improving the accuracy and efficiency of the model. The steps are as follows:

[0079] Calculate the average spectral value and standard deviation, and input the infrared spectral data. Calculate the average spectral value of the sample. Then calculate the standard deviation of the spectrum. :

[0080] (7)

[0081] In the formula, Indicates the number of samples. ;

[0082] For sample spectra The standardization process is as follows:

[0083] (8)

[0084] Vector normalization reduces spectral interference and increases model stability;

[0085] S211.3 Smoothing and Noise Reduction: Savitzky-Golay smoothing retains peak shape while filtering out high-frequency noise. The smoothing and noise reduction: The convolutional smoothing Savitzky-Golay algorithm performs multinomial fitting on local data and uses the least squares method to solve for the smoothing factor, removing noise from the signal. The formula is as follows:

[0086] (9)

[0087] in, Indicates the smoothing factor; The normalization factor represents the smoothing factor, which is based on the distribution of data points collected within the window to best fit the signal trend.

[0088] Feature extraction step: Mining key information (characteristic spectral peaks) in the spectrum.

[0089] S220. Input the infrared spectral data collected multiple times into the pre-trained ionization identification model, obtain the characteristic spectral peaks respectively, and calculate the average value of the characteristic spectral peaks; during the acquisition process, compensate for environmental interference through parameter adjustment mechanism so that the characteristic spectral peaks are positively correlated with the residual gas concentration.

[0090] After data preprocessing, a support vector machine regression algorithm is used to establish a remote plasma source ionization identification model using near-infrared spectral data with different dissociation rates, in order to output characteristic spectral peaks.

[0091] When performing function fitting, the Support Vector Machine (SVM) model transforms spectral data from a low-dimensional space to a high-dimensional space and replaces the original inequality constraints in SVM with equality constraints, reducing computational complexity and improving computational speed. SVM has strong learning capabilities and provides an effective solution for complex problems such as limited sample data, high sparsity, high dimensionality, and nonlinearity.

[0092] The specific calculation process of SVM is as follows:

[0093] S221. Input spectral matrix ,in, Let R represent the set of real numbers and d represent the dimension of the infrared spectral data. express The output after inputting the ionization discrimination model Indicates the number of samples;

[0094] S222. Fit the data using a high-dimensional linear function, and set the prediction model as follows:

[0095] (10)

[0096] in, Represents the variable coefficient vector. Representing the bias value vector, using a mapping function This maps the input data to a higher-dimensional space.

[0097] S223. According to the principle of support vector machines. and The optimal solution is obtained by minimizing a specific function, and the model optimization problem is transformed into the following form:

[0098] (11)

[0099] The constraints are: (12)

[0100] in, Represents the regularization coefficient. This indicates the allowable fitting error.

[0101] S224. Introduce Lagrange multipliers to construct the Lagrange function:

[0102] (13)

[0103] right , , , Find the partial derivatives and introduce the KT conditions:

[0104] (14)

[0105] S225. Cancel out variables , The matrix equation is obtained as follows:

[0106] (15)

[0107] in, ; ; ;

[0108] S226. Based on previous research experience, different kernel functions perform differently when handling training samples of different sizes and dimensions. The radial basis function (RBF) kernel is chosen as the kernel function for the optimization problem, and its expression is as follows:

[0109] (16)

[0110] in, This represents the width coefficient of the kernel function.

[0111] S227. Calculated based on S225 and The nonlinear equation is transformed into a linear equation, and the input test set samples are used. Calculate the characteristic spectral peaks :

[0112] (17)

[0113] S228. Input the infrared spectral data collected from multiple scans into the ionization identification model to obtain the characteristic spectral peaks. Take the average value of the characteristic spectral peaks obtained from multiple scans for subsequent calculation of the dissociation rate.

[0114] The parameter control mechanism in step S220 compensates for the molar absorption coefficient by precisely controlling process parameters such as temperature, pressure, and gas flow rate of the cavity, so that the characteristic spectral peak and the residual gas concentration form a monotonic and stable proportional relationship, satisfying the characteristic peak intensity. In this case, peak intensity I can be directly used instead of concentration. The dissociation rate was calculated, ensuring accuracy.

[0115] The spectral peaks of the input gas are significantly affected by temperature, pressure, and gas flow rate. For example, as the temperature rises, the thermal motion of molecules intensifies, and the absorption lines broaden, leading to a decrease in the intensity of the spectral peaks. When the pressure rises, the frequency of molecular collisions increases, and the pressure broadening effect is significant. This broadening leads to energy dispersion, causing a decrease in peak height. When the gas flow rate increases, turbulence is enhanced, the gas distribution becomes more uniform, and the peak height tends to increase.

[0116] According to the Beer-Lambert law (2), when the process conditions are met, the molar absorptivity... With a fixed optical path length L, absorbance A is directly proportional to concentration c. ).although This holds true in most cases, but deviations may occur in the following situations: such as when When the concentration is too high (typically 0.1 mol / L), intermolecular interactions are enhanced, which may lead to peak broadening or saturation absorption, causing the linear relationship to fail. When the plasma source is operating, the actual ambient temperature for gas dissociation is greater than 1000 K, resulting in a higher molar absorptivity. Deviates from normal temperature value.

[0117] Based on the above conclusions, the parameter control mechanism provided in this embodiment includes:

[0118] S310. Establish an objective function for compensating characteristic spectral peaks of process parameters; objective function

[0119] (18)

[0120] in, The state calibration value is the characteristic spectral peak value at the current moment. This is the sampled value of the characteristic spectral peak at the current moment; , , These are the sampled values ​​of gas pressure, chamber temperature, and gas flow rate at the current moment. , , These are the sampled values ​​of gas pressure, chamber temperature, and gas flow rate at the previous time step k-1. , , These represent the corresponding weight proportions. Specifically, in this step, the current time k is the starting point, and the next time k+1 is the ending point. The characteristic spectral peak at the current time k needs to be collected from 32 to 128 infrared spectral data points. The characteristic spectral peaks of the multiple collected data points are obtained through the ionization discrimination model, and then the average value is taken to determine the value. If the number of scans is 32, and the difference between the current time k and the next time k+1 is 1 second, then with a sampling interval of 1 / 32 seconds, a total of 32 samples are taken starting from the current time k. The characteristic spectral peaks of the 32 collected infrared spectral data points are obtained respectively, and then the average value is taken as the value. .

[0121] S320. Adjust chamber temperature in real time according to objective function. Gas pressure and gas flow rate So that the objective function It is minimized.

[0122] Specifically, the state calibration value of the characteristic spectral peak at the current moment. The average value of the characteristic spectral peaks output by the ionization identification model can be obtained by inputting the infrared spectral data collected multiple times at the previous time k-1. To indicate:

[0123] (19)

[0124] in, The cumulative sampling time at the current time k. Let A be the cumulative sampling time of the previous time k-1; let A be the time-varying frequency factor; and let B be the factor affecting the rate of change of the switching frequency. To solve... The deviation from the normal temperature value leads to the problem of distortion of characteristic spectral peaks; this embodiment adjusts the chamber temperature in real time through the objective function (18). The characteristic spectral peaks of gas pressure P and gas flow rate Q are compensated under environmental interference.

[0125] Method for calculating the time-varying factor A of characteristic spectral peaks:

[0126] Record the peak increments at adjacent time points , ;

[0127] , ;

[0128] The rate of change R of the characteristic spectral peak can be expressed as: ,

[0129] It can be directly obtained from the increment formula at any given time. or .

[0130] The calculation method for the influence factor B of the rate of change of characteristic spectral peak switching frequency:

[0131] The rate of change influence factor B satisfies the following formula:

[0132] ;

[0133] B is obtained using an iterative method.

[0134] Optionally, step S320 includes:

[0135] S321. When the chamber temperature is detected. greater than the temperature threshold At that time, the flow rate of the remote plasma source cooling water was adjusted to stabilize the cavity temperature at the preset temperature value. The temperature was set to a value that was as low and stable as possible, i.e., the flow rate of the remote plasma source cooling water was changed to stabilize the cavity temperature at 600K. >600K, to minimize the impact of temperature The dynamic effects can be assessed by fine-tuning the set temperature based on the output of the state observer to find the optimal operating point. For example, by controlling the water flow rate to increase rapidly at 1 L / min per second, and using PID closed-loop control to change the cooling water flow rate of the remote plasma source, the chamber temperature can be stabilized at 600 K.

[0136] S322. State calibration value based on the current characteristic spectral peak value and the sampled value of the characteristic spectral peak at the current time The relationship between the gas pressure and the regulation of the gas pressure and gas flow rate Step S322 includes:

[0137] S322.1: When No adjustment is made at this time; this indicates the current sampled value. Able to track the state calibration value of the characteristic spectral peak at the current moment. If the temperature does not deviate from the normal temperature value, no process parameter compensation is required;

[0138] S322.2: When At that time, slowly decrease the gas pressure P and slowly increase the gas flow rate. At this time, the sampled value The value is too low; the gas pressure P needs to be gradually reduced. Reducing the pressure will decrease the Lorentz broadening, narrowing the spectral lines and lowering the peak value. Increasing the gas flow rate leads to a rebound in the characteristic spectral peak; a slow increase in gas flow rate enhances turbulence, homogenizes gas distribution, and causes the peak height to rise, thus improving the objective function. It is minimized.

[0139] S322.3: When At that time, the gas pressure P is slowly increased, and the gas flow rate is slowly decreased. At this time, the sampled value The pressure is too high, so the gas pressure P needs to be increased slowly while the gas flow rate Q is decreased, so that the objective function... It is minimized.

[0140] Specifically, in step S322.2, the gas pressure P is slowly reduced while the gas flow rate is slowly increased. The specific operation can be carried out by decreasing the gas pressure by 0.1 Torr per second and increasing the gas flow rate by 5 sccm per second; in step S322.3, the gas pressure P is slowly increased and the gas flow rate is slowly decreased. The specific operation can be carried out according to the gas pressure increasing by 0.1 Torr per second and the gas flow rate decreasing by 5 sccm per second.

[0141] In one embodiment, the weight proportions in the objective function , , The importance of temperature T, gas pressure P, and gas flow rate Q to the characteristic spectral peaks is determined based on the random forest algorithm. The random forest regression algorithm evaluates the relationship between process parameters such as chamber temperature T, gas pressure P, and gas flow rate Q and the characteristic spectral peaks. Figure 5 As shown, the steps are as follows:

[0142] Step P1: Bootstrap Sampling

[0143] A linear regression model can be used to fit an approximate linear relationship between process parameters such as chamber temperature T, gas pressure P, and gas flow rate Q and characteristic spectral peaks. T subsets are then drawn with replacement from the original dataset D. Each subset has a sample size of N. The probability that each sample is selected in a single sampling is... for:

[0144] (20)

[0145] The samples that were not selected constitute the out-of-bag (OOB) dataset, which is used for error estimation and feature importance analysis.

[0146] Step P2: Random selection of features

[0147] Let the total characteristic number be For each subset Construct a regression decision tree When a node splits, m features are randomly selected (m ≤ m). ).

[0148] The objective of node splitting in a regression tree is to minimize the mean squared error (MSE). For feature j and splitting thresholds, ... Select the optimal split pair :

[0149] (twenty one)

[0150] in, and These are the left and right child node sample sets after the split. and These are the predicted values ​​for the left and right child nodes, respectively. That is, the target mean of the subset samples:

[0151] (twenty two)

[0152] Step P3: Integration of Prediction Results

[0153] The final output is obtained by using the average prediction of all decision trees. Ensemble learning is used to reduce the uncertainty of predictions from individual trees, and the final output is as follows:

[0154] (twenty three)

[0155] in, This is the average result; This represents the prediction result of a single decision tree; T represents the number of decision trees. For example... Figure 6 As shown.

[0156] Step P4: Model Performance Evaluation Metrics

[0157] The root mean square error (RMSE) metric is used to quantify the accuracy of predicting characteristic spectral peaks.

[0158] (twenty four)

[0159] Determine whether the root mean square error of the characteristic spectral peaks meets the error accuracy requirement of 0.1; otherwise, return to step P1 for continued iteration until the maximum number of iterations is reached.

[0160] Step P5: Feature Importance Assessment Analysis

[0161] The importance of a feature is measured by the change in out-of-bag (OOB) error after eigenvalue permutation. After random permutation, the importance of the feature is measured by the change in out-of-bag error. A significant increase in error after permutation indicates that the feature has a significant impact on model prediction. The importance is calculated as follows:

[0162] (25)

[0163] in, Let be the OOB error of the t-th tree.

[0164] Step P6: Using the predicted mean of all decision trees as the final output, fit the characteristic spectral peaks formed by each combination of process parameters, and evaluate the relationship between the characteristic spectral peaks and the importance of each process parameter.

[0165] Step P7: Assign importance values ​​of each process parameter to the objective function of equation (18). Weighting of the middle chamber temperature T, gas pressure P, and gas flow rate Q , , .

[0166] Figure 7 The importance scores of the characteristic spectral peaks and process parameters were ranked. The importance relationships between the characteristic spectral peaks and gas pressure P, chamber temperature T, and gas flow rate Q were 0.68, 0.23, and 0.09, respectively, and were assigned to the objective function of equation (18). The weighting of gas pressure P, chamber temperature T, and gas flow rate Q , , ;Right now =0.68, =0.23, =0.09.

[0167] S230. Calculate the dissociation rate based on the average value of the characteristic spectral peaks. The dissociation rate is calculated using the following formula:

[0168] dissociation rate (26)

[0169] The characteristic spectral peaks corresponding to the concentration of the gas to be measured are input in the initial state. This represents the average value of the characteristic spectral peaks output by the k-ionization identification model at the current time.

[0170] The following uses NF3 as an example to verify the method provided in this embodiment. Table 1 shows the relationship between the residual gas concentration and the characteristic spectral peak. It can be seen that under optimized parameters, the linearity between the characteristic spectral peak and the concentration of NF3 is [not specified]. >0.95, achieving direct equivalent concentration without calibration.

[0171] Table 1

[0172]

[0173] like Figure 8 The figure shows a comparison of experimental results between the remote plasma source detection dissociation rate detection method provided in this embodiment and the traditional calibration method; the traditional calibration method is as follows: steps S110-S130. From... Figure 8 It can be seen that the characteristic spectral peak based on the parameter control mechanism can track the calibration value spectrum very well, and the residual gas concentration represented by the characteristic spectral peak has an error of 0.05 with the calibration value. In this embodiment, without establishing a dataset for calibration, the acquisition accuracy of the characteristic spectral peak is high, which can meet the requirements of improving the detection accuracy of the dissociation rate of the remote plasma source to a certain extent.

[0174] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for detecting the dissociation rate of a remote plasma source, characterized in that, include: Acquire infrared spectral data from a remote plasma source; The infrared spectral data collected multiple times were input into a pre-trained ionization identification model to obtain the characteristic spectral peaks and calculate the average value of the characteristic spectral peaks. During the acquisition process, environmental interference was compensated through a parameter adjustment mechanism to make the characteristic spectral peaks positively correlated with the residual gas concentration. The method of compensating for environmental interference through parameter adjustment includes: Establish an objective function for compensating for characteristic spectral peaks related to process parameters; objective function , The state calibration value is the peak value of the characteristic spectrum of k at the current time. This represents the average value of the characteristic spectral peaks output by the k-ionization discrimination model at the current time. , , These are the sampled values ​​of gas pressure, chamber temperature, and gas flow rate at the current moment. , , These are the sampled values ​​of gas pressure, chamber temperature, and gas flow rate at the previous time step k-1. , , These are the corresponding weight proportions; Adjust the chamber temperature in real time according to the objective function. Gas pressure and gas flow rate So that the objective function Minimize; The dissociation rate is calculated based on the average value of the characteristic spectral peaks; State calibration value of the characteristic spectral peak at the current moment according to Determined, including: ， in, The cumulative sampling time at the current time k. The cumulative sampling time of the previous time k-1. The average value of the characteristic spectral peaks output by the ionization identification model is the infrared spectral data collected multiple times at the previous time k-1; A is the time-varying frequency factor, and B is the influence factor of the switching frequency change rate. The chamber temperature is adjusted in real time according to the objective function. Gas pressure and gas flow rate So that the objective function Minimize, including: When the chamber temperature is detected greater than the temperature threshold At the same time, the flow rate of the cooling water from the remote plasma source is adjusted to stabilize the cavity temperature at the preset temperature value; according to and The relationship between the gas pressure and the regulation of the gas pressure and gas flow rate ;include: when At that time, no adjustment is made; when At that time, slowly decrease the gas pressure P and slowly increase the gas flow rate. ; when At that time, the gas pressure P is slowly increased, and the gas flow rate is slowly decreased. .

2. The method according to claim 1, characterized in that, Weighting , , The importance of temperature T, gas pressure P, and gas flow rate Q to the characteristic spectral peaks is determined based on the random forest algorithm.

3. The method according to claim 1, characterized in that, The training method for the ionization discrimination model includes: Based on the support vector machine algorithm, the optimization problem corresponding to the ionization discrimination model is determined, and the objective function and constraints in the optimization problem are as follows: objective function , Represents the variable coefficient vector. Represents the bias value vector. This indicates the allowable fitting error. Represents the regularization coefficient; Constraints , For mapping functions, This represents the input infrared spectral matrix. Let R represent the set of real numbers, and d represent the infrared spectral data. Dimensions express Input the characteristic spectral peak value corresponding to the output of the ionization identification model. This indicates the number of samples corresponding to the number of scans. The kernel function for the optimization problem is , For the sample data to be classified, This represents the sample data that has been classified. Represents the width coefficient of the kernel function; The characteristic spectral peak , For Lagrange multipliers, among which and The solution is obtained by introducing the KT condition.

4. The method according to claim 1, characterized in that, The dissociation rate is calculated using the following formula: dissociation rate , The characteristic spectral peaks corresponding to the concentration of the gas to be measured are input in the initial state. The value is the average of the characteristic spectral peaks output by the ionization identification model, which is formed by inputting infrared spectral data collected multiple times at the current time k.

Images