Magnetron sputtering parameter adjusting method and device, electronic equipment and storage medium

Abstract

The invention discloses a magnetron sputtering parameter adjusting method, a magnetron sputtering parameter adjusting device, electronic equipment and a storage medium, time sequence characteristics of process parameters are extracted through a Bi-LSTM network, dynamic changes in a thin film preparation process are well captured, and the accuracy of target material life prediction is improved through effective extraction and analysis of the time sequence characteristics; meanwhile, sputtering effect analysis is conducted by introducing the current service life of the target, and the accuracy is higher than that of thin film performance analysis conducted through a full-connection network only through technological parameters. And the generation of the life prediction result and the sputtering effect evaluation index is unified in a one-time reasoning process of a network model, so that the parameter adjustment efficiency is improved. The genetic algorithm is introduced to optimize the Bi-LSTM network, and the genetic algorithm is deeply customized according to magnetron sputtering parameters, so that the efficiency, accuracy and generalization ability of the prediction model are further improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of thin film deposition, and particularly relates to a magnetron sputtering parameter adjustment method and device, an electronic device, and a storage medium. BACKGROUND

[0002] The magnetron sputtering technology is a new thin film preparation technology developed in the field of vacuum coating in the 1970s. It is a high-speed and low-temperature sputtering technology, and has the advantages of simple equipment, easy control, large coating area, and strong adhesion.

[0003] In the magnetron sputtering process, as the target material is used, the surface state and sputtering rate thereof will change, thereby affecting the performance and quality of the thin film. Therefore, it is necessary to dynamically adjust the process parameters.

[0004] Although the technical personnel in the field are aware that the process parameters need to be dynamically and real-time adjusted, the adjustment process relies on manual experience, and an accurate and effective dynamic adjustment method has not been established, further leading to inaccurate and low-efficiency process parameter adjustment in the magnetron sputtering process.

[0005] That is, in the existing magnetron sputtering process, there is the technical problem of inaccurate and low-efficiency process parameter adjustment. SUMMARY

[0006] The application focuses on the various problems in the prior art, and provides a magnetron sputtering parameter adjustment method, device, electronic device, and storage medium.

[0007] According to a first aspect of the application, a magnetron sputtering parameter adjustment method is provided, comprising the following steps:

[0008] First, the target material condition before sputtering is evaluated, and the process parameters are set. Then, based on the target material condition before sputtering and the process parameters, a prediction model is used to simultaneously predict the target material condition after sputtering and the sputtering effect, wherein the prediction model is optimized by a genetic algorithm. Finally, based on the target material condition after sputtering and the sputtering effect, the optimal process parameters are determined.

[0009] Preferably, based on the target material condition before sputtering and the process parameters, a prediction model is used to simultaneously predict the target material condition after sputtering and the sputtering effect, specifically comprising the following steps:

[0010] Step S1: constructing a prediction model based on Bi-LSTM; wherein the input parameters of the prediction model include a first input variable and a second input variable; the output parameters include a first output variable and a second output variable; the first input variable is the current target material life Q, and the first output variable is the predicted target material life Q p, the second input variable at least includes sputtering power P, sputtering pressure p and bias voltage Vb, and the second output variable is resistivity p, film thickness T and thickness uniformity difference U;

[0011] Step S2: based on the model of the magnetron sputtering device and the related data in the magnetron sputtering process, the input parameters and output parameters for training the prediction model are constructed, and the prediction model is trained;

[0012] Step S3: based on the genetic algorithm, a plurality of sets of second input variables for optimizing the prediction model are generated, and the input parameters for optimizing the prediction model are constructed;

[0013] Step S4: based on the plurality of sets of second input variables for optimizing the prediction model and the first input variable obtained by calculation, a plurality of sets of output parameters are obtained through the prediction model, and then a population with high fitness is determined through the genetic algorithm; the fitness is determined by resistivity p, film thickness T and thickness uniformity difference U;

[0014] Step S5: based on the population with high fitness, the optimal first output variable and second output variable are determined.

[0015] Preferably, in the above step S1, the prediction model includes an input layer, a first Bi-LSTM layer, a first attention layer, a second Bi-LSTM layer, a second attention layer, a full connection layer, and an output layer;

[0016] Among them, the second Bi-LSTM layer and the second attention layer are multiple, and the number of the second Bi-LSTM layer and the second attention layer is the same and staggered;

[0017] Each second Bi-LSTM layer receives the output of at least one channel from the first attention layer or the second attention layer of the previous layer.

[0018] Preferably, in the above step S2, based on the model of the magnetron sputtering device and the related data in the magnetron sputtering process, the input parameters and output parameters for training the prediction model are constructed, and the prediction model is trained, specifically:

[0019] S21: a prediction model is bound to a model of a magnetron sputtering device through a unique identification code;

[0020] S22: the thickness change △h of the target material before and after each sputtering is calculated by laser interference method, and the current thickness d of the target material is calculated based on the thickness change △h and the initial thickness d0 of the target material now , so as to obtain the target life change △Q before and after each sputtering:

[0021]

[0022] Wherein, k, b are constant terms, which are fitted by experiments, and k, b are different in different target life stages;

[0023] S23: Record the sputtering time stamp, sputtering duration, sputtering power P, sputtering pressure p, bias voltage Vb, target temperature and other parameters of each run of the magnetron sputtering equipment, and the resistivity p, film thickness T and thickness uniformity difference U and other parameters after each run of the magnetron sputtering equipment;

[0024] S24: Based on the data obtained in the above steps S22-S23, the prediction model is trained for multiple rounds.

[0025] Preferably, in the above step S3, a plurality of sets of second input variables for optimizing the prediction model are generated based on a genetic algorithm, and input parameters for optimizing the prediction model are constructed, specifically:

[0026] S31: Determine the sputtering power P, sputtering pressure p, and bias voltage Vb in the input parameters as the decision variables of the genetic algorithm;

[0027] S32: Take the resistivity p, film thickness T, and thickness uniformity difference U in the output parameters as the target function construction factor, and construct the target function in the form of multi-objective or single objective;

[0028] S33: Set a constraint condition for the decision variables in the above step S31;

[0029] S34: Based on the decision variables that meet the constraint condition, randomly generate N chromosomes [P n , p n , Vb n ], wherein n=1, 2...N, n is a natural number.

[0030] Preferably, in the above step S4, based on the plurality of sets of second input variables for optimizing the prediction model and the first input variables obtained by calculation, a plurality of sets of output parameters are obtained through the prediction model, and then a population with high fitness is determined through the genetic algorithm, specifically:

[0031] S41: Input the N chromosomes [P n , p n , Vb n ] as the plurality of sets of second input variables for optimizing the prediction model, and the current target life Q as the first input variable obtained by calculation, into the trained prediction model, and output the predicted target life Q pn , the predicted resistivity p n , the predicted film thickness T n , and the predicted thickness uniformity difference U n corresponding to each chromosome, wherein n=1, 2...N, n is a natural number;

[0032] S42: retaining chromosomes with a difference between the current target material life Q and the predicted target material life Q not greater than a set threshold value pn determining respective ranks, performing non-dominated sorting, and obtaining a plurality of non-dominated solution sets;

[0033] S43: for any non-dominated solution set, calculating a crowding distance of each chromosome on a target function construction factor;

[0034] S44: retaining chromosomes with high ranks and large crowding distances.

[0035] Preferably, after the above step S44, further comprising:

[0036] S45: dividing the entire target material life cycle into three flat periods of an initial period, a stable period, and a decay period, and a first transition period from the initial period to the stable period, and a second transition period from the stable period to the decay period;

[0037] S46: for each of the chromosomes with high ranks and large crowding distances retained in the above step S44, determining a target material life cycle in which the chromosome is located according to the predicted target material life Q p ; in the initial period, the stable period, or the decay period, performing gene mutation on each of the chromosomes with high ranks and large crowding distances retained; in the first transition period or the second transition period, performing gene crossover on each of the chromosomes with high ranks and large crowding distances retained;

[0038] S47: merging a new generation population composed of the chromosomes subjected to gene mutation or gene crossover with a last generation population to form a new population with a size of N chromosomes;

[0039] S48: repeating the above steps S41-S47 until a termination condition is met to determine a population with high fitness.

[0040] According to a second aspect of the present application, there is provided a device for adjusting magnetron sputtering parameters, which specifically comprises:

[0041] an evaluation setting module configured to evaluate a target material condition before sputtering and set process parameters;

[0042] a model construction module configured to construct a prediction model based on Bi-LSTM;

[0043] a training data construction module configured to train the prediction model;

[0044] a model optimization module configured to optimize the prediction model based on a genetic algorithm;

[0045] a result output module configured to output a target material condition after sputtering, a sputtering effect, and determine optimal process parameters.

[0046] In a third aspect, the present application provides a computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of any of the above methods.

[0047] In a fourth aspect, the present application provides a computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, which, when executed by the processor, implements the steps of any of the above methods.

[0048] Compared with the prior art, the present application has the following beneficial effects:

[0049] 1) The present application extracts the time sequence characteristics of the process parameters through the Bi-LSTM network, which can well capture the dynamic changes in the film preparation process. With the use of the target material, the process parameters will change accordingly due to factors such as wear and aging of the target material and the running state of the equipment. These dynamic changes can be effectively extracted and analyzed through time sequence characteristics, improving the accuracy of target material life prediction.

[0050] 2) The present application extracts time sequence characteristics through the use of Bi-LSTM network, and simultaneously introduces the current life of the target material for sputtering effect analysis, which is more accurate than using only process parameters to analyze film performance through a fully connected network.

[0051] 3) The present application unifies the generation of life prediction results and sputtering effect evaluation indexes in one inference process of a network model, improving the parameter adjustment efficiency.

[0052] 4) The present application introduces a genetic algorithm to optimize the Bi-LSTM network, and deeply customizes the genetic algorithm according to the magnetron sputtering parameters. On the one hand, based on the predicted target material life, different types of gene mutations and different degrees of gene mutations are performed on multiple decision variables constituting the chromosome genes; on the other hand, different iteration steps can be selected according to the adjustment complexity of the decision variables during population iteration. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The present application provides a flowchart of a magnetron sputtering parameter adjustment method;

[0054] Figure 2 The present application provides a network architecture diagram based on a Bi-LSTM prediction model;

[0055] Figure 3 The present application provides a schematic diagram of optimizing the prediction model based on a deeply customized genetic algorithm. DETAILED DESCRIPTION

[0056] The technology described below can have various modifications and can be various embodiments, and is described in detail below with specific embodiments in conjunction with the accompanying drawings. However, this does not mean that the technology described below is limited to specific embodiments. It should be understood that the present application includes all similar modifications, equivalents and alternatives without departing from the spirit and technical scope of the technology described below.

[0057] Embodiment 1

[0058] The present application provides a method for adjusting magnetron sputtering parameters, comprising the following steps:

[0059] First, the target material condition before sputtering is evaluated, and the process parameters are set; then, based on the target material condition before sputtering and the process parameters, the target material condition after sputtering and the sputtering effect are simultaneously predicted by a prediction model, wherein the prediction model is optimized by a genetic algorithm; finally, based on the target material condition after sputtering and the sputtering effect, the optimal process parameters are determined.

[0060] The target material condition before sputtering is the current target material life, and the process parameters at least include sputtering power, sputtering pressure and bias voltage; the target material condition after sputtering is the predicted target material life, and the sputtering effect includes resistivity, film thickness and thickness uniformity difference. Considering the predicted target material life and the sputtering effect, the desired result is selected based on the genetic algorithm, and then the corresponding process parameters are determined as the optimal process parameters.

[0061] As shown in Figure 1 Based on the target material condition before sputtering and the process parameters, the target material condition after sputtering and the sputtering effect are simultaneously predicted by a prediction model, specifically comprising the following steps:

[0062] Step S1: Construct a prediction model based on Bi-LSTM; wherein the input parameters of the prediction model include first input variables and second input variables; the output parameters include first output variables and second output variables.

[0063] The first input variable and the first output variable are the same variable, and the second input variable and the second output variable are both multiple.

[0064] Specifically, the first input variable is the current target material life Q, and the first output variable is the predicted target material life Q p The second input variable at least includes sputtering power P, sputtering pressure p and bias voltage Vb, and the second output variable is resistivity p, film thickness T and thickness uniformity difference U.

[0065] Furthermore, the prediction model includes an input layer, a first Bi-LSTM layer, a first attention layer, a second Bi-LSTM layer, a second attention layer, a fully connected layer, and an output layer; there are multiple second Bi-LSTM layers and second attention layers, and the number of second Bi-LSTM layers and second attention layers are the same and they are interleaved; each second Bi-LSTM layer receives the output of at least one channel from the first attention layer or the second attention layer above.

[0066] Specifically, the network architecture design of the prediction model is as follows: Figure 2 As shown, it includes:

[0067] Input Layer: The number of neurons in the input layer equals the number of input features, including the current lifetime of the target material, the working timestamp, and various process parameters during magnetron sputtering, such as sputtering time, sputtering power, pressure, and bias voltage. These parameters are arranged in a time series and used as the input to the model.

[0068] Bidirectional Bi-LSTM Layer: This layer consists of two LSTM layers: a forward LSTM layer for processing forward time-series information and a backward LSTM layer for processing backward time-series information. The number of neurons in both LSTM layers can be the same, typically between 64 and 256. For example, each LSTM layer can be set to have 128 neurons. The outputs of the forward and backward LSTM layers are concatenated at each time step to form a vector containing both forward and backward information. This invention uses a four-channel Bi-LSTM network to enhance its performance in predicting target lifetime, resistivity, film thickness, and thickness uniformity differences.

[0069] Attention Layer: An attention layer is added after each Bi-LSTM layer in each channel to enable the Bi-LSTM network to automatically learn which feature parameters are more important, such as sputtering power potentially affecting resistivity more strongly than sputtering pressure or bias voltage. Specifically, the attention layer learns a weight vector [α_power, α_pressure, α_bias, ...], where a larger learned α_power indicates that sputtering power contributes more to the prediction.

[0070] In addition to the first Bi-LSTM layer, each Bi-LSTM layer receives not only the output of the previous attention layer for its current channel, but also the outputs of the attention layers for other channels, for feature fusion, thus better capturing global feature information. Figure 2 The network framework with two Bi-LSTM layers is shown.

[0071] In addition to the Bi-LSTM of the first layer, the Bi-LSTM layer of each layer can be the output of the attention layers of all other channels or the output of the attention layers of part of all other channels when inputting the output of the attention layers of other channels, depending on the number and importance of the important feature parameters in the weight vector learned by the attention layer of the first layer.

[0072] Fully connected layer: used for fusing and mapping high-dimensional features to further extract features. The fully connected layer can contain multiple neurons, the number of which gradually decreases to form a pyramid-like structure to reduce the complexity and parameter number of the model and prevent overfitting. For example, the first fully connected layer can have 256 neurons, the second fully connected layer can have 128 neurons, and the third fully connected layer can have 64 neurons.

[0073] Output layer: the output layer has four neurons for outputting the predicted target life value, the predicted film resistivity, the predicted film thickness, and the predicted film thickness uniformity difference.

[0074] Step S2: based on the model of the magnetron sputtering device and the related data in the magnetron sputtering process, the input parameters and output parameters for training the prediction model are constructed, and the prediction model is trained. Specifically, the following steps are included:

[0075] S21: a prediction model is bound to a model of a magnetron sputtering device through a unique identification code.

[0076] The prediction model is responsible for predicting the target life, the resistivity, thickness, and uniformity difference of the output film of the target under a certain set of process parameters. A prediction model is bound to a magnetron sputtering device of a certain model corresponding to a certain target. Each prediction model has a separate identification code. In the case of using the prediction model to adjust the process parameters, the system will load the corresponding prediction model according to the model of the target device and make predictions of the life and sputtering effect.

[0077] S22: the thickness change amount Δh of the target before and after each sputtering is calculated by the laser interference method, and the current thickness d of the target is calculated based on the thickness change amount Δh and the initial thickness d0 of the target. now Thus, the target life change amount ΔQ before and after each sputtering is obtained:

[0078]

[0079] where k and b are constant terms, which are fitted through experiments and are different at different target life stages.

[0080] In the data preparation stage, the target material life needs to be calibrated first. Among them, the target material life is determined by laser interference to determine the percentage of the used life of the target material. It is mainly through the measurement of the thickness change of the target material to indirectly reflect the degree of loss, and then to calculate the percentage of the remaining life. Specifically, first, the laser interferometer is installed and debugged, and before each sputtering of the target material, the laser beam is accurately irradiated to the specified position on the surface of the target material. After the target material completes a sputtering, start the laser interferometer to measure and record the movement of the interference fringes. According to the relationship formula between the movement of the interference fringes and the thickness change, the thickness change amount of the target material at different time points is calculated:

[0081]

[0082] Among them, Δh represents the thickness change of the target material, λ represents the wavelength of the incident light in vacuum, ΔN represents the number of interference fringe movements, and n represents the refractive index of the medium.

[0083] Given the initial thickness of the target material and the thickness change amount Δh of the target material before and after each sputtering, the life change amount ΔQ of the target material before and after each sputtering can be obtained, so as to infer the target material life before sputtering and the target material life after sputtering. The target material life before each sputtering is taken as one of the input parameters of the prediction model in the training stage, and the target material life after sputtering is taken as one of the output parameters of the prediction model in the training stage.

[0084] S23: Record the sputtering time stamp, sputtering time, sputtering power P, sputtering pressure p, bias voltage Vb, target material temperature and other parameters of each running of the magnetron sputtering equipment as one of the input parameters of the prediction model in the training stage; Record the resistivity ρ, film thickness T and thickness uniformity difference U and other parameters after each running of the magnetron sputtering equipment as one of the output parameters of the prediction model in the training stage.

[0085] S24: Based on the data obtained in the above steps S22-S23, the prediction model is trained for multiple rounds.

[0086] After completing the data preparation through steps S22-S23, the data is subjected to missing value checking, outlier detection and processing, cleaning and unification and other operations.

[0087] In the missing value checking stage, if the power, pressure or other parameters are not recorded in some samples, the following methods can be used for processing:

[0088] (1) Delete samples with serious missing values, such as samples corresponding to a time stamp missing multiple data such as power, pressure, etc.

[0089] (2) Use mean, median or interpolation method to fill in;

[0090] (III) Categorical variables can be filled with mode or "missing" label.

[0091] In the outlier detection and processing phase, use box plot (IQR), Z-score or visualization tools to identify outliers; for outlier samples, you can choose to delete (if it is obvious that the input error) or limit within a reasonable range (such as using clip limiting); or set as a special category for model learning (such as extreme sample of air pressure).

[0092] In the cleaning and unification phase, first need to carry out unit standardization, such as power unification for watt (W), pressure for pa (Pa or mTorr); Then define the loss function and optimizer, use mean square error (MSE) or mean absolute error (MAE) as loss function. The optimizer selects Adam optimizer, which can adaptively adjust the learning rate, automatically adjusts the step size during the training process according to the update of the parameters, which helps to speed up the convergence of the model and avoid falling into local optimal solution.

[0093] Then model training, input training data into the model, according to the set of hyperparameters and optimizers for training. In each round of training, the model will calculate the error between the predicted value and the true value according to the loss function, and update the parameters of the model through the back propagation algorithm. For Bi-LSTM model, the back propagation process will be carried out in the forward and reverse LSTM layers to update the parameters in both directions.

[0094] During the training process, the performance of the model is evaluated regularly on the validation set, and the hyperparameters are adjusted according to the loss value of the validation set, such as adjusting the learning rate, increasing or decreasing the number of neurons in the LSTM layer, adjusting the structure of the fully connected layer, etc., to avoid overfitting or underfitting of the model. When the performance of the model on the validation set reaches the optimal, save the parameters of the model.

[0095] Finally, model evaluation, use test set to evaluate the trained model, calculate root mean square error (RMSE), mean absolute percentage error (MAPE) and other evaluation indexes, to measure the prediction accuracy and generalization ability of the model.

[0096] Step S3: generating a plurality of second input variables for optimizing the prediction model based on a genetic algorithm, and constructing input parameters for optimizing the prediction model.

[0097] According to the trained prediction model, the target material life and the sputtering effect are predicted, and the optimal process parameters are determined.

[0098] The genetic algorithm is used to solve the process parameter combination under different life conditions, specifically:

[0099] S31: Determine the sputtering power P, sputtering pressure p, and bias voltage Vb in the input parameters as the decision variables of the genetic algorithm.

[0100] S32: Construct factors with the resistivity p, film thickness T, and thickness uniformity difference U in the output parameters as the target function to construct a target function in the form of multiple targets or a single target.

[0101] f1(·): Resistivity p, the lower the better;

[0102] f2(·): Film thickness T, the closer to the target thickness the better;

[0103] f3(·): Thickness uniformity difference U, the smaller the better.

[0104] If the target function in the form of multiple targets is constructed, the three indicators are processed separately, and the Pareto frontier is used.

[0105] If the target function in the form of a single target is constructed, a weighted method is adopted:

[0106] F = w1·p + w2·|T-T0| + w3·U

[0107] where T0 is the target film thickness, and w1-w3 are weights.

[0108] The assignment method for w1-w3 weights is as follows:

[0109] In the new target stage, the resistivity and uniformity of the formed film are usually good, at this time more attention should be paid to the stability of the film thickness change, that is, a higher weight value should be given to w2; After the target material enters the stable period, the target at this time is adjusted to focus on preventing the resistivity from deviating or exceeding the specification, at this time a higher weight value should be given to w1; After the target material enters the recession period, the film thickness and resistivity tend to be stable, but the thickness uniformity should be suppressed as much as possible, which is also the most challenging challenge at the end of the target material life, at this time, a higher weight value should be given to w3.

[0110] S33: Set constraints for the decision variables in the above step S31.

[0111] On the one hand, set upper and lower limit constraints for the sputtering power P and the bias voltage Vb. For example, sputtering power P ∈ [200, 500] (W), and bias voltage Vb ∈ [-100, -300] (V).

[0112] On the other hand, process window limits are imposed on the sputtering pressure p, such as sputtering pressure p not exceeding 5 mTorr.

[0113] S34: Based on the decision variables that satisfy the constraints, randomly generate N chromosomes [P n , p n , Vbn , wherein n=1, 2...N, n is a natural number, P n , p n , Vb n constitute a chromosome.

[0114] Step S4: based on the plurality of sets of second input variables for optimizing the prediction model and the first input variable obtained by calculation, a plurality of sets of output parameters are obtained by the prediction model, and then a population with high fitness is determined by the genetic algorithm.

[0115] The fitness is determined by the resistivity p, the film thickness T and the thickness uniformity difference U. The principle is that the lower the resistivity p, the smaller the film thickness error and the smaller the thickness uniformity difference U, the better. The fitness can be determined by constructing a multi-objective form of the objective function or a single-objective form of the objective function according to the value of the objective function.

[0116] Specifically includes the following steps:

[0117] S41: input N chromosomes [P n , p n , Vb n ] as a plurality of sets of second input variables for optimizing the prediction model, and input the current target material life Q as the first input variable obtained by calculation into the trained prediction model, and output the predicted target material life Q pn , the predicted resistivity p n , the predicted film thickness T n and the predicted thickness uniformity difference U n , wherein n=1, 2...N, n is a natural number.

[0118] S42: retain the chromosomes whose difference between the current target material life Q and the predicted target material life Q pn is not greater than a set threshold value, determine the respective grades, perform non-dominated sorting, and obtain a plurality of non-dominated solution sets.

[0119] The setting of the threshold value is used to exclude the prediction results that deviate obviously from the actual results, so as to perform the first screening on the population constituted by the chromosomes. In addition, the prediction model provided by the present application can simultaneously predict the target material life and the sputtering effect such as the resistivity p, the film thickness T and the thickness uniformity difference U, so as to simplify the first screening process of the chromosome population.

[0120] When the chromosomes are graded, the following principles are followed:

[0121] The first grade is the chromosome that is not dominated by any other individual;

[0122] The second grade is the chromosome that is only dominated by the first grade;

[0123] The third rank is the chromosome that is dominated by the first and second ranks.

[0124] Suppose the current situation of chromosome A and chromosome B is as follows:

[0125]

[0126] Where film thickness error ΔT = T - T0, T is the film thickness, and T3 is the target film thickness.

[0127] Since chromosome A is better than chromosome B in three targets (resistivity p, film thickness error ΔT, and thickness uniformity difference U), chromosome A dominates chromosome B.

[0128] If one target in chromosome A is better than chromosome B, and another target in chromosome B is better than chromosome A, then chromosome A and B do not dominate each other.

[0129] S43: For any non-dominated solution set, calculate the crowding distance of each chromosome on the objective function construction factor.

[0130] Suppose there are 5 chromosomes in the current non-dominated solution set (Front 1), and the three objective functions (multi-objective form objective functions) are:

[0131] Objective 1: Resistivity (p) → the smaller the better

[0132] Objective 2: Film thickness error (ΔT) → the smaller the better

[0133] Objective 3: Uniformity difference (U) → the smaller the better

[0134]

[0135]

[0136] The calculation method of the crowding distance on the objective 1, i.e. resistivity p, is as follows:

[0137] First, sort the chromosomes in ascending order according to the value of resistivity p:

[0138] Chromosome p E 2 D 4 C 6 B 8 A 10

[0139] Set the crowding distance of boundary solutions E and A to ∞

[0140] Then, the crowding distance of the intermediate solution is calculated as follows:

[0141] First, normalize the denominator:

[0142]

[0143] Then, the crowded distance of chromosome B, C, D under the target is calculated one by one:

[0144] C: ((B)-(D)) / 8 = 0.5

[0145] B: ((A)-(C)) / 8 = 0.5

[0146] D: ((C)-(E)) / 8 = 0.5

[0147] Similarly, the crowded distance of chromosome A to chromosome E in the film thickness error △T and the uniformity difference U target can be solved respectively, and the crowded distance of each chromosome in three targets is added to obtain the total crowded distance of each chromosome.

[0148] Similarly, when constructing a single-target form of the objective function, the crowded distance can be solved in a weighted sum manner.

[0149] S44: Reserve chromosomes with high ranking and large crowded distance.

[0150] Generally speaking, the chromosome with high non-dominated ranking and large crowded distance is the chromosome with high fitness; the population with high fitness is composed of chromosomes with high fitness.

[0151] S45: Divide the entire target material life cycle into three flat periods of initial period, stable period, and decay period, as well as the first transition period from the initial period to the stable period, and the second transition period from the stable period to the decay period.

[0152] First, according to the used life of the target material, the target material is divided into three flat period stages, i.e., the initial period is set to 0%-30% of the used life, the stable period is set to 30%-70% of the used life, and the decay period is set to 70%-100% of the used life.

[0153] In the initial period, the target surface is new and race track has not been formed, and the sputtering efficiency is high; in the stable period, the discharge is stable, the race track is formed, the process is most stable, and the film thickness and uniformity are best; in the decay period, the race track groove is deepened, and the target surface is aged, the deposition rate is decreased, the resistivity is increased, and the uniformity is decreased. See the following table for details:

[0154]

[0155] Among them, the used target material life is mostly the predicted target material life Q predicted by the prediction model pThe target material life prediction provided by the prediction model can be used not only for screening chromosomes, but also for determining the target material life after using a certain set of process parameters, thereby avoiding frequent use of the grating measurement method to calibrate the used life of the target material and reducing the operation complexity.

[0156] In the present solution, the grating measurement method can be used to calibrate the used life only when the training sample set is made, and otherwise, the predicted life of the model can be used as the used life of the model at the corresponding time point. Of course, in practice, the predicted life still needs to be corrected at a few time points, for example, the used life is measured by using the grating measurement method every 5 time points, and the predicted result of the model is used as the used life at the time points between the next grating measurement.

[0157] In addition, the stages within the ranges before and after the two nodes of 30% and 70% of the used target material life (within a positive or negative deviation of 5%) are defined as the first transition period and the second transition period, respectively. The target material life is in the flat period and the transition period, and thus the adjustment mode of the decision variables, i.e., the process parameters, contained in the chromosome should also be different.

[0158] S46: For each of the chromosomes with high ranking and large congestion distance reserved in the step S44, the predicted target material life Q p determining the target material life cycle in which the chromosomes are located; performing gene mutation on each of the chromosomes with high ranking and large congestion distance reserved in the initial period, the stable period or the decay period; and performing gene crossover on each of the chromosomes with high ranking and large congestion distance reserved in the first transition period or the second transition period.

[0159] Specifically, in the initial period stage, the surface of the new target material is relatively flat, and the sputtering efficiency is high, so a lower sputtering power can be appropriately used to avoid the thin film growth uneven or adversely affect the thin film structure caused by the too fast initial sputtering rate. With the increase of sputtering time, the power is gradually and steadily increased to the normal working range, so that the target material sputtering rate is stable, which is beneficial to the uniform growth of the thin film. In addition, a higher gas pressure can promote the uniformity of discharge, so that the ions can uniformly bombard the target material, which is beneficial to the formation of stable race track.

[0160] In the stable period stage, the sputtering state of the target material is relatively stable, and the stable process parameters can ensure the consistency of the sputtering rate, thereby ensuring the uniformity of the thin film. Therefore, the process parameters can not be adjusted in this stage or can be finely adjusted according to the thin film thickness monitoring result to accurately control the thin film growth rate.

[0161] In the recession stage, the target material will have non-uniform sputtering, serious surface erosion and other conditions, resulting in a decrease in sputtering efficiency. Too low sputtering rate will eventually have a negative impact on the resistivity, film thickness, and uniformity of the film. For example, when the sputtering rate is low, the atoms are deposited on the substrate surface at a slower speed, and impurities are more likely to be adsorbed on the surface of the film and incorporated into the film. The presence of impurities will change the electronic structure of the film, increase the electron scattering center, and thus increase the resistivity. Therefore, the power can be appropriately increased in this stage to maintain a certain sputtering rate, thereby ensuring the uniformity of the film thickness. However, the power increase should not be too large to avoid causing plasma instability or other process problems. The gas flow can be appropriately reduced and the bias voltage can be increased to compensate for the loss of sputtering rate. For the pressure, a higher gas pressure increase will result in an increase in the collision frequency of particles and gas molecules in the plasma, and the scattering effect on the particles during their journey to the substrate will be enhanced, resulting in a decrease in the number of particles reaching the substrate surface, thereby reducing the sputtering rate of the film. Therefore, to compensate for the loss of sputtering rate in this stage, the pressure can be appropriately reduced. For the bias voltage, increasing the bias voltage can enhance the ion bombardment effect on the target material, increase the sputtering yield of the target material atoms, and thus increase the sputtering rate to a certain extent.

[0162] The adjustment scheme of each process parameter in different flat stages is shown in the following table:

[0163]

[0164] In the flat stage, the current optimal solution can be quickly found by the above-mentioned genetic mutation method; and in the transition stage, a new chromosome can be quickly and continuously constructed by the genetic exchange method.

[0165] Further, in the process of iterative calculation, one or more genes can be preferentially mutated or exchanged for actual needs until a target number of compliant chromosomes are screened or a preset maximum iteration threshold is reached. For example, most magnetron sputtering devices are usually equipped with a relatively precise power control system, which can achieve stable output and precise adjustment of power. The adjustment of gas pressure / intensity needs to be achieved by adjusting the gas flow, valve opening and other methods, which is relatively complex. To increase the adjustment efficiency, the pressure can be adjusted lastly, and different steps are taken to adjust the pressure to prevent the solution of the genetic algorithm from falling into local optimization. For example, assuming that the used life of the target material is 95%, the power and bias voltage are adjusted within the preset limit range every two genetic iterations, and the target number of compliant individuals is screened.

[0166] The iteration step of different process parameters is determined by the used life of the current target material.

[0167] S47: merging the new generation population composed of chromosomes with genetic mutation or genetic crossover with the last generation population to form a new population with N chromosomes.

[0168] The merged population is subjected to non-dominated sorting again, and the crowding distance is calculated to form a new population with N chromosomes, and the population size is kept constant.

[0169] S48: repeating the above steps S41-S47 until the end condition is met to determine the population with high fitness.

[0170] The so-called end condition generally refers to meeting the iteration number or convergence condition.

[0171] The output result after one round of genetic algorithm processing is as follows:

[0172]

[0173] Step S5: determining the optimal first output variable and second output variable based on the population with high fitness.

[0174] For example, at the initial time, the used life of the target material is 0%, and the used life of the target material at the next time point is predicted to be 5% after using a certain set of process parameters through the prediction model, and the set of process parameters is confirmed to be the optimal parameter through the above steps S41-S48, then it can be further confirmed that the used life of the target material at the next time point is 5%, and the predicted sputtering effect is obtained. By analogy, when the predicted used life exceeds 30%, the corresponding time point is confirmed as the time point when the target material enters the stable period.

[0175] Example 2

[0176] The application also provides a magnetic control sputtering parameter adjusting device, which specifically comprises:

[0177] An evaluation setting module is configured to evaluate the target material condition before sputtering and set the process parameters.

[0178] A model construction module is configured to construct a prediction model based on Bi-LSTM.

[0179] A training data construction module is configured to train the prediction model.

[0180] A model optimization module is configured to optimize the prediction model based on a genetic algorithm.

[0181] A result output module is configured to output the target material condition after sputtering, the sputtering effect, and determine the optimal process parameters.

[0182] Example 3

[0183] The application provides a computer readable storage medium, which stores a computer program, and the computer program realizes the steps of any of the foregoing methods when executed by a processor.

[0184] Embodiment 4

[0185] The application provides a computer device, which comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, and the computer program realizes the steps of any of the foregoing methods when executed by the processor.

[0186] Although the present application has been described in detail with general description and specific embodiments above, some modifications or improvements can be made on the basis of the present application. The above description is only the preferred embodiments of the present application, and does not limit the patent scope of the present application. Other changes and modifications made by those skilled in the art without departing from the spirit and protection scope of the present application are still included in the protection scope of the present application.

Claims

1. A method of adjusting magnetron sputtering parameters, characterized by, Firstly, the target condition before sputtering is evaluated, and process parameters are set; Then, based on the target condition before sputtering and the process parameters, a prediction model is used to simultaneously predict the target condition after sputtering and the sputtering effect, wherein the prediction model is optimized by a genetic algorithm; Finally, based on the target condition after sputtering and the sputtering effect, the optimal process parameters are determined.

2. The method of claim 1, wherein, The prediction model simultaneously predicts the target condition after sputtering and the sputtering effect based on the target condition before sputtering and the process parameters, and specifically includes the following steps: Step S1: constructing a prediction model based on Bi-LSTM; wherein, the input parameters of the prediction model include a first input variable and a second input variable; the output parameters include a first output variable and a second output variable; the first input variable is a current target material life Q, and the first output variable is a predicted target material life Q p , the second input variable at least includes sputtering power P, sputtering pressure p and bias voltage Vb, and the second output variable is resistivity p, film thickness T and thickness uniformity difference U; Step S2: Based on the model of the magnetron sputtering equipment and the related data in the magnetron sputtering process, the input parameters and the output parameters used to train the prediction model are constructed, and the prediction model is trained; Step S3: Based on the genetic algorithm, a plurality of second input variables for optimizing the prediction model are generated, and the input parameters for optimizing the prediction model are constructed; Step S4: Based on the plurality of second input variables for optimizing the prediction model and the first input variables obtained by calculation, a plurality of output parameters are obtained by the prediction model, and then a population with high fitness is determined by the genetic algorithm; the fitness is determined by the resistivity ρ, the film thickness T, and the thickness uniformity difference U; Step S5: Based on the population with high fitness, the optimal first output variable and the second output variable are determined.

3. The method of claim 2, wherein, In the above step S1, the prediction model includes an input layer, a first Bi-LSTM layer, a first attention layer, a second Bi-LSTM layer, a second attention layer, a full connection layer, and an output layer; Wherein, the second Bi-LSTM layer and the second attention layer are multiple, and the number of the second Bi-LSTM layer and the second attention layer is the same and staggered; Each second Bi-LSTM layer receives the output of at least one channel from the first attention layer or the second attention layer of the previous layer.

4. The method of claim 3, wherein, In the above step S2, the input parameters and the output parameters used to train the prediction model are constructed based on the model of the magnetron sputtering equipment and the related data in the magnetron sputtering process, and the prediction model is trained, specifically as follows: S21: One prediction model is bound to one model of the magnetron sputtering equipment through a unique identification code; S22: The thickness change amount Ah of the target before and after each sputtering is calculated by laser interferometry, and the current thickness d of the target is calculated based on the thickness change amount Ah and the initial thickness do of the target now Thus, the target life change amount AQ before and after each sputtering is obtained: Wherein, k and b are constant terms, which are fitted through experiments, and different in different target life stages; S23: Record the sputtering time stamp, sputtering duration, sputtering power P, sputtering pressure p, bias voltage Vb, target temperature and other parameters of the magnetron sputtering equipment at each time of running, and the resistivity ρ, film thickness T and thickness uniformity difference U and other parameters of the magnetron sputtering equipment after each running; S24: Based on the data obtained in the above steps S22-S23, the prediction model is trained for multiple rounds.

5. The method of claim 4, wherein, In the above step S3, the plurality of second input variables for optimizing the prediction model are generated based on the genetic algorithm, and the input parameters for optimizing the prediction model are constructed, specifically as follows: S31: Determine the sputtering power P, sputtering pressure p, and bias voltage Vb in the input parameters as decision variables of the genetic algorithm; S32: Construct a target function with the resistivity p, film thickness T, and thickness uniformity difference U in the output parameters as the target function construction factors, and construct a multi-objective or single-objective form of the target function; S33: Set a constraint condition for the decision variables; S34: randomly generate N chromosomes [P n , p n , Vb n ] based on the decision variables satisfying the constraint conditions, wherein n = 1, 2...N, n is a natural number.

6. The method of claim 5, wherein, In the above step S4, the multiple sets of the second input variables for optimizing the prediction model are used to obtain multiple sets of the output parameters through the prediction model based on the first input variables obtained by calculation, and then the genetic algorithm is used to determine the population with high fitness, specifically: S41: inputting N chromosomes [P n , p n , Vb n ] as the second input variable groups for optimizing the prediction model, the current target material life Q as the first input variable obtained by calculation, into the trained prediction model, and outputting the predicted target material life Q pn , the predicted resistivity p n , the predicted film thickness T n , and the predicted thickness uniformity difference U n corresponding to each chromosome, wherein n = 1, 2...N, n is a natural number; S42: retaining the current target material life Q and the predicted target material life Q pn chromosomes whose difference between the current target material life Q and the predicted target material life Q is not greater than a set threshold value, determining respective ranks, performing non-dominated sorting, and obtaining a plurality of non-dominated solution sets; S43: For any non-dominated solution set, calculate the crowding distance of each chromosome on the target function construction factor; S44: Keep the chromosomes with high ranking and large crowding distance.

7. The method of claim 6, wherein, After the above step S44, further comprising: S45: Divide the entire target material life cycle into three flat periods of initial period, stable period, and attenuation period, as well as a first transition period from the initial period to the stable period and a second transition period from the stable period to the attenuation period; S46: For each of the chromosomes with high rank and large crowding distance reserved in step S44 above, according to the predicted target life Q p determining a target life cycle in which it is located; in the initial period, the stable period or the decay period, each of the chromosomes with high rank and large crowding distance reserved is subjected to gene mutation; in the first transition period or the second transition period, each of the chromosomes with high rank and large crowding distance reserved is subjected to gene crossover; S47: Merge the new generation population composed of chromosomes after gene mutation or gene exchange with the last generation population to form a new population with N chromosomes; S48: Repeat the above steps S41-S47 until the end condition is met to determine the population with high fitness.

8. A device for adjusting magnetron sputtering parameters for performing the steps of the method according to any one of claims 1 to 7, characterized in that The device comprises: An evaluation setting module for evaluating the target material condition before sputtering and setting process parameters; A model construction module for constructing a prediction model based on Bi-LSTM; A training data construction module for training the prediction model; A model optimization module for optimizing the prediction model based on a genetic algorithm; A result output module for outputting the target material condition after sputtering, sputtering effect, and determining the optimal process parameters.

9. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1-7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1-7.

Images