Parameter adjusting method of ion implantation machine

Abstract

The invention discloses a parameter adjustment method of an ion implantation machine. The method comprises the following steps: acquiring a plurality of first parameters of an ion beam; acquiring historical data and screening at least one group of historical data close to the plurality of first parameters, wherein each group of historical data at least comprises a plurality of first parameters; inputting the plurality of first parameters and the historical data into a hybrid deep neural network model, and obtaining a predicted output variable; and adjusting a third parameter of the ion implantation machine according to the predicted output variable. According to the parameter adjustment method of the ion implantation machine provided by the invention, the third parameter of the machine is adjusted by adopting the hybrid deep neural network model, so that automatic ion beam parameter adjustment can be realized, the manual participation rate is reduced, and meanwhile, the working efficiency and the working accuracy of the machine which works for a long time can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ion implantation machines, and in particular to a parameter adjustment method of an ion implantation machine. Background Art

[0002] As the process nodes of integrated circuit manufacturing are reduced, ion implantation technology has gradually replaced the diffusion process and become an indispensable technology in the integrated circuit manufacturing process because of its great advantages in controllability of doping concentration, doping size and doping angle. Ion implantation technology ionizes gas molecules into ions, then accelerates the ions to the target energy through an acceleration unit, and finally implants them into the surface of the silicon wafer. During the ion implantation process, the implantation depth can be controlled by controlling the ion energy, and the implantation time and dose can be controlled by controlling the beam size.

[0003] Existing ion implantation machines have the function of automatically adjusting the ion beam, but the accuracy is low, and the longer the machine works, the greater the chance of errors. Therefore, staff are needed. After the machine's working time reaches a preset time, the machine will automatically report an error. After receiving the error message, the staff will manually adjust the ion beam parameters to improve the accuracy of ion implantation.

[0004] like Figure 1 and Figure 2 As shown, in the existing ion implantation machine, the parameter adjustment method includes: step S101: the ion implantation machine completes the ion implantation of a wafer; step S102: judging whether the ion implantation time of a menu of the machine is less than or equal to the preset time, wherein the ion implantation time of a menu refers to the time for the machine to automatically adjust the ion beam; step S103: if it is less than or equal to the preset time, the ion implantation machine automatically adjusts the parameters of the ion implantation beam and then continues to implant the ion of the next wafer; step S104: if it is greater than the preset time, manually check and adjust the parameters of the ion implantation beam and then continue to implant the ion of the next wafer. Wherein, reference Figure 2 For ion implantation machines of different models and manufacturers, the automatic parameter adjustment time and manual parameter adjustment time are different. For example, the IBV model ion implantation machine takes 3.8 hours to complete the ion implantation of a wafer, so the automatic parameter adjustment time of the machine is 3 hours, and the time required for manual parameter adjustment is 0.8 hours; similarly, the manual parameter adjustment time of the IAN model ion implantation machine is 1.2 hours within a 4.4-hour process, and the manual parameter adjustment time of the IDA model ion implantation machine is 1.1 hours within a 5.8-hour process. Summary of the invention

[0005] In view of the above problems, the present application provides a method for adjusting parameters of an ion implantation machine tool, aiming to improve the accuracy of the ion implantation machine tool and further improve the efficiency at the same time.

[0006] According to a first aspect of the present invention, there is provided a method for adjusting parameters of an ion implantation machine tool, including:

[0007] Obtain a plurality of first parameters of an ion beam; obtain historical data and screen at least one set of historical data close to the plurality of first parameters, and each set of historical data includes at least the plurality of first parameters; input the plurality of first parameters and the historical data into a hybrid deep neural network model, and obtain a predicted output variable; adjust a third parameter of the ion implantation machine tool according to the predicted output variable.

[0008] Optionally, the training steps of the hybrid deep neural network model include: obtaining historical data of the ion implantation machine tool and empirical data of technicians; preprocessing the historical data and the empirical data, and dividing them into training set data and test set data; constructing a hybrid deep neural network model, the hybrid deep neural network model includes an input layer, a hidden layer and an output layer; input the training set data into the input layer of the hybrid deep neural network model, and obtain a predicted output variable at the output layer; determine whether the error between the predicted output variable and the actual output variable meets the precision error; if not, adjust the weights and bias values of the hidden layer, and input the training set data into the input layer of the hybrid deep neural network again, and obtain a predicted output variable at the output layer, and compare the predicted output variable with the actual output variable again; if it meets, use the test set data to verify the hybrid deep neural network model, and determine that the parameters of the hybrid deep neural network model have reached the optimum; and add an online learning model to the hybrid deep neural network model, collect online data in real time, and optimize and fine-tune the hybrid deep neural network according to the online data.

[0009] Optionally, both the historical data and the empirical data include 12 third parameters and 5 first parameters of the ion implantation machine tool, the first parameters are used as input variables of the hybrid deep neural network model, and the third parameters are used as actual output variables for comparison with the predicted output variables of the hybrid deep neural network model.

[0010] Optionally, the third parameters include: arc starting current, gas flow rate, front-back position of three axes, tilt position of three axes, left-right position of three axes, ion source magnetic field current, arc starting voltage, focusing voltage, ion screening magnetic field current, energy screening magnetic field current, parallelism correction magnetic field current, small beam current filter; the first parameters include: element mass, energy, beam current uniformity, beam current size, and beam current state.

[0011] Optionally, the steps of constructing the hybrid deep neural network model include: obtaining the linear relationship and / or non-linear relationship between variables in the historical data and empirical data; establishing an index system for the hybrid deep neural network model according to the relationship between the variables, and confirming the input variables and output variables in the index system; confirming the connection relationship, iteration relationship, activation function, and loss function between the input layer, hidden layer, and output layer in the hybrid deep neural network model according to the linear relationship and / or non-linear relationship.

[0012] Optionally, the activation function includes the linear activation function ReLU and the non-linear activation function Sigmoid, and the loss function is the mean square error.

[0013] Optionally, the hidden layer of the hybrid deep neural network model includes a linear part and a non-linear part; the linear part includes four fully connected linear layers for learning the linear relationship of input variables; the non-linear part includes a total of four fully connected non-linear layers and linear layers for learning the non-linear relationship of the input variables.

[0014] Optionally, the steps of preprocessing the historical data and the empirical data include: deleting abnormal data and / or null data in the historical data and the empirical data; performing normalization processing on the remaining historical data and empirical data.

[0015] Optionally, the training set data is 80%, and the test set data is 20%.

[0016] Optionally, the hybrid deep neural network model narrows the adjustment range of the predicted output variables according to the historical data.

[0017] The unexpected technical effect of this application is:

[0018] According to the parameter adjustment method of the ion implantation machine tool in the embodiment of this application, a hybrid deep neural network model is used to adjust the parameters of the machine tool. The error value of the hybrid deep neural network model after training is less than the error accuracy. Therefore, the accuracy of the hybrid deep neural network model in this application is very high. Thus, the ion implantation machine tool using this hybrid deep neural network model can not only achieve automatic adjustment of machine tool parameters, reduce the manual participation rate, but also improve the working efficiency and working accuracy of the machine tool for long-term working machine tools.

[0019] Furthermore, for the parameter adjustment method of the ion implantation machine provided in the present application, the basic structure of the hybrid deep neural network model is built by technicians based on empirical data of their work, and then historical data of the machine and empirical data of technicians are used for training to improve the accuracy of the model. In addition, an online learning function is added to the hybrid deep neural network model, which can further optimize the model in real time according to the usage data of the machine, so that the model is more matched with the machine, effectively overcoming the influence brought by the hardware changes of the machine over working time, and thus making the accuracy of the model higher and higher.

[0020] Furthermore, for the parameter adjustment method of the ion implantation machine provided in the present application, during the process of using the hybrid deep network model to adjust the parameters of the machine, historical data that is closest in time and space to the current parameters will also be detected as a reference, so that the hybrid deep neural network model can narrow the range of parameter adjustment and improve the efficiency and accuracy of parameter adjustment during this parameter adjustment process. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Through the following description of the embodiments of the present application with reference to the drawings, the above and other objects, features, and advantages of the present application will become clearer. In the drawings:

[0022] Figure 1 The flowchart of the parameter adjustment method of the ion implantation machine in the prior art is shown;

[0023] Figure 2 The automatic working time and manual working time of different machines in the prior art are shown;

[0024] Figure 3 The schematic structural diagram of the ion implantation machine according to an embodiment of the present invention is shown;

[0025] Figure 4 The flowchart of the parameter adjustment method of the ion implantation machine according to an embodiment of the present invention is shown;

[0026] Figure 5 The flowchart of the construction idea of the hybrid deep neural network model according to an embodiment of the present invention is shown;

[0027] Figure 6 The training flowchart of the hybrid deep neural network model according to an embodiment of the present invention is shown;

[0028] Figure 7 The flowchart of the parameter adjustment method by technicians according to an embodiment of the present invention is shown;

[0029] Figure 8 The structural diagram of the hybrid deep neural network model according to an embodiment of the present invention is shown;

[0030] Figure 9 Shows a schematic diagram of the binning discretization among 12 output variables of the hybrid deep neural network model according to an embodiment of the present invention. Detailed implementation manners

[0031] The present application will be described in more detail below with reference to the accompanying drawings. In each of the drawings, like elements are denoted by like reference numerals. For clarity, the various parts in the drawings are not drawn to scale. In addition, some well-known parts may not be shown. The present application can be presented in various forms, and some examples will be described below.

[0032] The present application can be presented in various forms, and some examples will be described below.

[0033] Figure 3 Shows a schematic structural diagram of an ion implantation machine according to an embodiment of the present invention. Figure 4 Shows a flowchart of a method for adjusting ion beam parameters of an ion implantation machine according to an embodiment of the present invention. Figure 5 Shows a flowchart of the construction idea of a hybrid deep neural network model according to an embodiment of the present invention. Figure 6 Shows a training flowchart of a hybrid deep neural network model according to an embodiment of the present invention. Figure 7 Shows a flowchart of a method for adjusting ion beam parameters by a person skilled in the art according to an embodiment of the present invention. Figure 8 Shows a structural diagram of a hybrid deep neural network model according to an embodiment of the present invention. Figure 9 Shows a schematic diagram of the binning discretization among 12 output variables of the hybrid deep neural network model according to an embodiment of the present invention.

[0034] Reference Figure 3, the ion implantation machine tool 100 of the present application includes a control unit 101, an ion beam generation unit 102, a Faraday unit 103, an acceleration unit 104, and an ion implantation chamber 105. Among them, the control unit 101 is used to control the generation and shutdown of the ion beam, and adjust multiple third parameters of the machine tool; the ion beam generation unit 102 is connected to the control unit 101, and is used to generate an initial ion beam, and feedback the second parameters of some hardware in the ion beam generation unit 102 to the control unit 101; the Faraday unit 103 is connected to the ion beam generation unit 102 and the control unit 101, and is used to detect multiple first parameters such as the size, shape, position, and energy of the initial ion beam generated by the ion beam generation unit 102, and feedback the detected multiple first parameters to the control unit 101. The control unit 101 adjusts the third parameters inside the ion beam generation unit 102 according to the feedback first parameters and second parameters to obtain an initial ion beam that meets the recipe requirements; the acceleration unit 104 is connected to the ion beam generation unit 102 and the Faraday unit 103, and is used to accelerate the initial ion beam generated by the ion beam generation unit 102; the ion implantation chamber 105 is connected to the acceleration unit 104, and is used to perform ion implantation on the wafer according to the accelerated ion beam.

[0035] In this embodiment, a hybrid deep neural network model is set in the control unit 101. After comparing the first parameters fed back by the Faraday unit 103 with the recipe requirements, if they do not meet the recipe requirements, the predicted output variables can be obtained through the hybrid deep neural network model, and multiple third parameters of the ion implantation machine tool can be adjusted according to the predicted output variables. Since the third parameters affect the first parameters, the first parameters can be made to meet the recipe requirements by adjusting the third parameters, thereby improving the accuracy and working efficiency of the ion implantation machine tool.

[0036] The following will be combined with Figures 4 - 9 to describe in detail the parameter adjustment method of the ion implantation machine tool in the present application. The parameter adjustment method includes:

[0037] Step S201: Obtain multiple first parameters of the ion implantation machine tool.

[0038] In this step, the Faraday unit 103 detects various aspects such as the size, range, shape, and energy of the initial ion beam generated by the ion beam generation unit 102, thereby obtaining multiple first parameters of the ion implantation machine tool and feeding them back to the control unit 101. In addition, the ion beam generation unit 102 can also feedback the second parameters of each hardware that affects the first parameters of the ion implantation machine tool to the control unit 101, so that the control unit 102 can simultaneously obtain multiple first parameters characterizing the size, range, shape, energy, etc. of the ion beam and multiple second parameters characterizing the influence of hardware parameters on ion beam parameters.

[0039] In this embodiment, multiple first parameters characterizing aspects such as the size, range, shape, and energy of the ion beam are, for example, the output parameters of the ion implantation machine 100, including 5, namely Element Mass (Mass NO), Energy, Beam Current Uniformity, Beam Current, and Beam Size.

[0040] The second parameters of each hardware that affect the first parameters of the ion beam are the hardware wear parameters of the ion implantation machine 100, such as the wear parameters of the magnetic poles in the ion beam generation unit 102, etc.

[0041] Step S202: Obtain historical data and screen at least one set of historical data close to the multiple first parameters.

[0042] In this step, the control unit 101 receives the multiple first parameters, and then screens the historical data stored in the machine to obtain at least one set of historical data that is close to the multiple first parameters received this time in both time and space as a reference for subsequent parameter adjustment.

[0043] In this embodiment, each ion implantation of the ion implantation machine will store the corresponding first parameter, second parameter, and third parameter as the online data of the online learning model of the hybrid deep neural network model, and can also be used as the reference data during a certain parameter adjustment process. Therefore, each set of historical data includes multiple first parameters. By comparing the first parameters of the obtained ion beam with the first parameters in the historical data, a set of historical data that is close to the multiple first parameters received this time in both time and space is confirmed.

[0044] In addition, the control unit 101 also receives the second parameter. Since the second parameter will also indirectly affect the first parameter of the ion beam, the second parameter can also be used as a screening item when screening a set of historical data close to the first parameter in this step, so that the finally obtained historical data is not only close to the first parameter but also close to the second parameter.

[0045] Step S203: Input the first parameter, second parameter, and historical data into the hybrid deep neural network model and obtain the predicted output variable.

[0046] In this step, the hybrid neural network model of the control unit 101 receives the first parameter and the historical data, uses the historical data as a reference, and then makes a prediction based on the first parameter to output a predicted output variable. The output variable is specifically the value of multiple third parameters that can affect the first parameter, or the parameter adjustment range of multiple third parameters.

[0047] Specifically, in this embodiment, the third parameter is, for example, the input parameter of the ion implantation machine 100, including 12, namely: arc current, gas flow, three-axis front-back position (three-axis Gap), three-axis tilt position (three-axis Tilt), three-axis left-right position (three-axis Traverse), ion source magnetic field current (SMG), arc voltage (Arc Voltage), focus voltage (Focus), ion screening magnetic field current (SAM Current), energy screening magnetic field current (FEM Current), parallelism correction magnetic field current (Col Current), and small beam current filter (low beam aperture). That is, in this application, by adjusting these 12 third parameters, the 5 first parameters of the ion beam can be changed.

[0048] In addition, the hybrid deep neural network model uses the closest set of historical data as a reference, which can narrow the adjustment range of multiple third parameters in the output variables predicted by the hybrid deep neural network model, making the predicted output variables output by the hybrid deep neural network model more accurate. In addition, there is only one output result of the hybrid deep neural network model, and this one result includes the numerical values of multiple third parameters or the tuning ranges of multiple third parameters, that is, multiple data are packaged so that there is only one output result.

[0049] Furthermore, in this application, the hybrid deep neural network model is constructed by technicians based on the historical data and their experience data of the machine. In this application, referring to Figure 5 , the general process of constructing the hybrid neural network model includes five stages, namely: dataset collection, data processing, model selection and training, model optimization, and generating a prediction model. The following will combine Figure 5 and Figure 6 to describe the steps of constructing the hybrid neural network model of this application in detail.

[0050] Step S301: Obtain the historical data of the ion implantation machine and the experience data of the technician.

[0051] This step belongs to the dataset collection stage. Among them, the dataset for training the model comes from multiple groups of historical data stored in the ion implantation machine and multiple groups of experience data provided by technicians.

[0052] Step S302: Preprocess the historical data and experience data and divide them into training set data and test set data.

[0053] This step belongs to the data processing stage. In this embodiment, each set of historical data and empirical data includes multiple parameters, which include the aforementioned multiple first parameters, multiple second parameters, multiple third parameters, etc. Since the ion implantation of the machine tool is not all successful, in this step, not all of the multiple sets of historical data stored in the ion implantation machine tool and the multiple sets of empirical data provided by technicians are available data, and each set of data needs to be preprocessed to remove unusable data.

[0054] Among them, the steps of preprocessing historical data and empirical data include outlier and missing value filtering to delete unusable abnormal data and / or null data, etc., and using the data normalization (Min-Max method) method to normalize the remaining historical data and empirical data, so as to linearly scale the data to the specified range [0, 1] to improve the stability and efficiency when using this data to train the hybrid deep neural network model later. Among them, the formula for normalizing the data is:

[0055] (1)

[0056] (2)

[0057] In the above formulas (1) and (2), X and Y are the real data in the historical data and empirical data, X min is the minimum value of the data in the X sample, X max is the maximum value of the data in the X sample, Y min is the minimum value of the data in the Y sample, Y max is the maximum value of the data in the Y sample. Among them, the X sample refers to, for example, one of the historical data or empirical data, and the Y sample refers to the other of the historical data and empirical data.

[0058] Furthermore, it also includes dividing the historical data and empirical data into training set data and test set data. The training set data accounts for, for example, 80% of all the data, and the test set data accounts for, for example, 20% of all the data.

[0059] Step S303: Construct a hybrid deep neural network model, including an input layer, a hidden layer, and an output layer.

[0060] This step belongs to the model selection and training stage. In this step, technicians use statistical knowledge to preliminarily explore the input variables and output variables in the historical data and empirical data to obtain the linear and non-linear relationships between the variables, as Figure 9 shown, which is convenient for better determining the algorithm used by the hybrid deep neural network model. Among them, Figure 9The abscissa represents time, and the ordinate represents the fitting degree. In the abscissa, 0 represents the current time, a negative number represents the time before the current time, and a positive number represents the time after the current time.

[0061] Specifically, in one embodiment, for example, taking the current time as the center, search for parameters of the same type three days before and after, then perform linear regression on two adjacent points, so as to fit the non-linear input-output characteristics and discover the rules among 12 third parameters; then establish an index system for the hybrid deep neural network model according to the linear and non-linear relationships among the variables, confirm the input variables and output variables in the index system; and confirm the connection relationship, iteration relationship, activation function, loss function, etc. between the input layer, hidden layer and output layer of the hybrid deep neural network model.

[0062] Among them, 5 first parameters (output parameters) of the ion implantation machine are used as input variables of the hybrid deep neural network model, and 12 third parameters (input parameters) of the ion implantation machine are used as output variables of the hybrid deep neural network model.

[0063] Further, referring to Figure 8 , the hybrid deep neural network model in this application includes an input layer, a hidden layer and an output layer. Specifically, the input layer is connected to multiple input variables y(k - 1)…y(k - nA)…y(k - nB); the hidden layer is connected to the input layer and includes multiple layers. The output of the previous hidden layer is used as the input of the next hidden layer. The multiple layers of the hidden layer are divided into a linear part and a non-linear part. The linear part includes, for example, 4 fully connected linear layers for learning the linear relationship of the input variables; the non-linear layer also includes, for example, a total of 4 fully connected non-linear layers and linear layers for learning the non-linear relationship of the input variables. Among them, W1 and W2 represent the information transmitted between the input layer and the hidden layer, and ν[φ(k)] represents the output result of the hybrid deep neural network model. The hybrid deep neural network model in this application can better adapt to complex data distributions by combining the linear part and the non-linear part.

[0064] In addition, the connection relationship, iteration relationship, activation function, loss function, etc. between the input layer, hidden layer and output layer of the hybrid deep neural network model can be confirmed according to the linear and / or non-linear relationships between the input variables and output variables in the historical data of the ion implantation machine and / or the empirical data of technicians. In one embodiment, for example, referring to the process of technicians manually adjusting parameters, as Figure 7 shown, the adjustment process includes:

[0065] Step S401: The machine alarms, and perform front-end fine-tuning on the machine parameters.

[0066] This front-end fine-tuning process includes:

[0067] Step S402: Fine-tune the three-axis parameters of the machine tool.

[0068] In this step, the three-axis parameters of the machine tool refer to the three-axis front-back position (three-axis Gap), three-axis tilt position (three-axis Tilt), and three-axis left-right position (three-axis Traverse) in the third parameter. During the adjustment of the third parameter of the machine tool, the purpose is to maximize the beam current in the first parameter. During the adjustment of the three-axis parameters, ensuring that the current beam current reaches the maximum means that the three-axis parameters have been adjusted. However, if the maximum beam current after adjusting the three-axis parameters still does not meet the recipe requirements, other parameters need to be further adjusted.

[0069] In this embodiment, the relationship between the three-axis parameters and the beam current is not clear, and it is necessary to detect the profile and uniformity of the ion beam to determine whether the three-axis parameters of the machine tool have been adjusted properly.

[0070] Therefore, according to the adjustment process, the connection relationship, iteration relationship, activation function, loss function, etc. between the input layer, hidden layer, and output layer of the hybrid depth neural network model can be roughly obtained, and more detailed confirmation needs to be adjusted through the training set and validation set data.

[0071] Step S403: Fine-tune the arc-starting current parameter of the machine tool.

[0072] In this step, the arc-starting current of the machine tool is fine-tuned to ensure that the current beam current reaches the maximum, which means that the arc-starting current parameter has been adjusted. However, if the maximum beam current after adjusting the arc-starting current parameter still does not meet the recipe requirements, other parameters need to be further adjusted.

[0073] In this embodiment, the magnitude of the arc-starting current is, for example, less than or equal to 3000 mA, and the beam current can be reduced by the arc-starting current. The arc-starting current is positively correlated with the beam current.

[0074] The adjustment of other parameters follows the same pattern, that is, after ensuring that the current beam current reaches the maximum, check whether it meets the recipe requirements. The fine-tuning of the parameters will not be described in detail hereafter.

[0075] Step S404: Fine-tune the ion source magnetic field current parameter of the machine tool.

[0076] In this embodiment, the ion source magnetic field current is positively and negatively correlated with the beam current.

[0077] Step S405: Fine-tune the gas flow parameter of the machine tool.

[0078] In this embodiment, when the beam current is greater than or equal to 800 mA, the beam current can be adjusted by adjusting the gas flow rate of the machine tool, and the gas flow rate is positively correlated with the beam current.

[0079] Step S406: Fine-tune the focusing voltage parameter of the machine tool.

[0080] In this embodiment, the focusing voltage is positively correlated with the beam current. During an adjustment process, for example, adjust the focusing voltage to the position where the beam current is the largest, then reduce the arc starting current to reduce the beam current, and then adjust the focusing voltage to the position where the beam current is the largest after reducing the arc starting current.

[0081] Step S407: Use a Faraday unit to detect the first parameter of the ion beam.

[0082] Step S408: Perform back-end fine-tuning on the machine tool parameters. The process of back-end fine-tuning includes:

[0083] Step S409: Perform sweep frequency adjustment on the machine tool.

[0084] In this step, sweep frequency adjustment (Sweep adjust) means that the machine tool sweeps the spot ion beam into a strip ion beam at a specific frequency to match the width of the wafer (wafer), where the width of the wafer is, for example, 300 mm.

[0085] Step S410: Fine-tune the arc starting current parameter of the machine tool again.

[0086] In this step, fine-tune the arc starting current of the machine tool again to match the first parameter of the ion beam. Among them, this step mainly matches the beam current (Beam current) in the first parameter of the ion beam.

[0087] Step S411: Detect the profile and uniformity of the ion beam.

[0088] In this step, if the first parameter of the ion beam still does not meet the recipe requirements, step S412 can be executed.

[0089] Step S412: Fine-tune the parallelism correction magnetic field current parameter of the machine tool.

[0090] In this embodiment, the correlation between the parallelism correction magnetic field current and the beam current is not clear, approximately -5% to +5%.

[0091] Step S413: End of parameter adjustment.

[0092] During the process of the above-mentioned technician fine-tuning the first parameter of the ion beam, the adjustment order of each parameter is not fixed and can be adjusted appropriately. However, during the process of adjusting the 12 third parameters of the machine to change the 5 first parameters, the 12 third parameters are divided into three groups with different priorities. Specifically, among the 12 third parameters, the parameters with high priority include arc current, three-axis front-back position (three-axis Gap), three-axis tilt position (three-axis Tilt), three-axis left-right position (three-axis Traverse), source magnetic field current (SMG), and focus voltage; the parameters with medium priority include gas flow; the parameters with low priority include ion screening magnetic field current (SAM Current), energy screening magnetic field current (FEM Current), parallelism correction magnetic field current (Col Current), low beam aperture, and arc voltage. However, generally, the arc voltage will not be adjusted.

[0093] Based on the process of the above-mentioned technician adjusting the third parameter to affect the first parameter, it can be confirmed that in the hybrid deep neural network model, the activation function of the linear layer adopts the rectified linear unit (ReLU). The rectified linear function has the advantages of simple calculation, reducing the problem of gradient disappearance, and accelerating the convergence speed. The activation function of the non-linear layer adopts the sigmoid function (Sigmoid Activation Function). The output range of the sigmoid function is between 0 and 1, and it has a smooth S-shaped curve feature. In the neural network, the sigmoid function is often used to map the output of the neuron to between 0 and 1. In the hybrid deep neural network model of this application, the adjustable range of the variable is limited by combining the engineer's parameter adjustment experience.

[0094] Furthermore, in this embodiment, the mean square error (MSE) is also used as the loss function of the hybrid deep neural network model to measure the difference between the predicted value and the true value of the model. Among them, the formula of the loss function (mean square error MSE) is:

[0095] (3)

[0096] In formula (3), n is the number of samples, y is the true value in the data, is the predicted value.

[0097] In addition, in this embodiment, the DataLoader of PyTorch is used to load data in batches, with each batch size being 64, so as to improve the training efficiency of the hybrid deep neural network model. During the training process of multiple epochs (representing the process of a complete dataset passing through the forward and backward propagations of the neural network), the model parameters are iteratively updated multiple times for the training data within each epoch until the model converges.

[0098] The hybrid deep neural network model of this application uses the Adam optimizer to optimize the model parameters. At the same time, in order to prevent overfitting, an L2 regularization term and a dropout probability are also added to the hybrid deep neural network model.

[0099] Step S304: Input the training set data into the input layer of the hybrid deep neural network model, and obtain the predicted output variable at the output layer.

[0100] This step belongs to the model selection and training stage. In this step, the input variables in the training set data are input into the constructed hybrid deep neural network model, and after being processed by the network model, a predicted output variable is obtained at the output layer. Among them, the predicted output variable includes multiple third parameters, such as 12 third parameters of the ion implantation machine.

[0101] Step S305: Determine whether the error between the predicted output variable and the actual output variable meets the precision error.

[0102] This step belongs to the model selection and training stage. In this step, the third parameters in the predicted output variable output by the hybrid deep neural network model are compared with the third parameters in the actual output variable corresponding to the input variables in the corresponding training set in this training, to determine whether the error between the predicted output variable and the actual output variable meets the precision error. The precision error is, for example, the maximum acceptable error value. Meeting the precision error means that the error between the predicted output variable and the actual output variable is less than the precision error.

[0103] In this embodiment, if the error between the predicted output variable and the actual output variable meets the precision error, step S307 is executed; if the error between the predicted output variable and the actual output variable does not meet the precision error, step S306 is executed.

[0104] Step S306: Adjust the weights and bias values of the hidden layer, and again input the training set data into the input layer of the hybrid deep neural network model, and obtain the predicted output variable at the output layer.

[0105] This step belongs to the model selection and training stage. In this step, since the error value of the hybrid deep neural network model is relatively large, after adjusting the weights and bias values of each linear layer and / or non-linear layer in the hidden layer, the training set data is input into the input layer of the hybrid deep neural network model again, and the predicted output variable is obtained at the output layer. Then, it is determined again whether the error between the predicted output variable and the actual output variable meets the accuracy error. Steps S306 and S305 are repeated until the error between the predicted output variable and the actual output variable meets the accuracy error.

[0106] Step S307: Use the test set data to verify the hybrid deep neural network model to determine that the parameters of the pre-trained model have reached the optimal.

[0107] This step belongs to the model selection and training stage. In this step, if the predicted output variable obtained by using the training set data for the hybrid deep neural network model already meets the accuracy error, then the test set data is used to verify the hybrid deep neural network model again to determine that the parameters of the pre-trained model have reached the optimal.

[0108] In this embodiment, the training and test loss curves, accuracy curves can be plotted, and the training process and generalization performance of the model can be observed to determine whether the parameters of the model have reached the optimal, so as to finally generate the optimal pre-trained model.

[0109] Step S308: Add an online learning model to the hybrid deep neural network model, collect online data in real time, and optimize and fine-tune the hybrid deep neural network model according to the online data.

[0110] This step belongs to the model optimization stage. In this step, in order to further optimize the hybrid deep neural network model, an online learning model is also added to it. After adding the hybrid deep neural network model to the ion implantation machine tool, the online real data of the machine tool can be collected in real time. These real data are used to fine-tune the pre-trained model, and these data are stored in the database and the model is updated regularly. This process is repeated to improve the generalization performance, adaptability and robustness of the hybrid deep neural network model.

[0111] Step S204: Adjust the third parameter of the ion beam according to the predicted output variable.

[0112] In this step, the control unit 101 adjusts the third parameter of the ion implantation machine tool according to the predicted output variable output by the hybrid deep neural network model, so that the first parameter of the ion beam generated by the ion beam generation unit 102 meets the recipe requirements.

[0113] According to the ion beam parameter adjustment method of the ion implantation machine tool according to the embodiments of the present application, a hybrid deep neural network model is used to adjust the parameters of the machine tool. The error value of the hybrid deep neural network model after training is less than the preset error. Therefore, the accuracy of the hybrid deep neural network model in the present application is very high. Thus, the ion implantation machine tool using this hybrid deep neural network model can not only achieve automatic ion beam parameter adjustment, reducing the manual participation rate, but also improve the working efficiency and working accuracy of the machine tool for machine tools working for a long time.

[0114] Furthermore, for the ion beam parameter adjustment method of the ion implantation machine tool provided by the present application, the basic structure of the hybrid deep neural network model therein is built by technicians according to the working experience data, and then the historical data of the machine tool and the experience data of technicians are used for training to improve the accuracy of the model. In addition, the hybrid deep neural network model also adds an online learning function, which can further optimize the model in real time according to the usage data of the machine tool, so that the model can better match the machine tool, effectively overcoming the influence brought by the hardware changes of the machine tool over working time, and thus making the accuracy of the model higher and higher.

[0115] Furthermore, for the ion beam parameter adjustment method of the ion implantation machine tool provided by the present application, during the process of using the hybrid deep network model to adjust the parameters of the machine tool, historical data that is closest in time and space to the current parameters will also be detected as a reference, so that the hybrid deep neural network model can narrow the range of parameter adjustment in the current parameter adjustment process, improving the efficiency and accuracy of parameter adjustment.

[0116] As described above according to the embodiments of the present application, these embodiments do not elaborate on all details and do not limit the application to only the specific embodiments described. Obviously, many modifications and variations can be made according to the above description. The present specification selects and specifically describes these embodiments to better explain the principle and practical application of the present application, so that those skilled in the art can make good use of the present application and its modifications based on the present application. The present application is only limited by the claims and their full scope and equivalents.

Claims

1. A method for adjusting parameters of an ion implantation machine, comprising: Acquiring a plurality of first parameters of the ion implantation machine; Acquire historical data and filter at least one group of historical data close to the first parameters, each group of historical data including at least the first parameters; Inputting the plurality of the first parameters and the historical data into a hybrid deep neural network model and obtaining a predicted output variable; The third parameter of the ion implantation machine is adjusted according to the predicted output variable.

2. The parameter adjustment method according to claim 1, characterized in that: The training steps of the hybrid deep neural network model include: Obtain historical data of ion implantation equipment and experience data of technicians; Preprocessing the historical data and the empirical data, and dividing them into training set data and test set data; Constructing a hybrid deep neural network model, wherein the hybrid deep neural network model includes an input layer, a hidden layer, and an output layer; Inputting training set data into the input layer of the hybrid deep neural network model, and obtaining predicted output variables at the output layer; Determine whether the error between the predicted output variable and the actual output variable meets the precision error; If not, the weights and bias values ​​of the hidden layer are adjusted, and the training set data is input into the input layer of the hybrid deep neural network again, and the predicted output variables are obtained at the output layer, and the predicted output variables are compared with the actual output variables again; If yes, the hybrid deep neural network model is verified using the test set data to determine that the parameters of the hybrid deep neural network model have reached the optimum; and An online learning model is added to the hybrid deep neural network model, online data is collected in real time, and the hybrid deep neural network is optimized and fine-tuned according to the online data.

3. The parameter adjustment method according to claim 2, characterized in that: The historical data and the empirical data both include 12 third parameters and 5 first parameters of the ion implantation machine, The first parameter is used as an input variable of the hybrid deep neural network model, and the third parameter is used as an actual output variable to be compared with the predicted output variable of the hybrid deep neural network model.

4. The parameter adjustment method according to claim 3, characterized in that: The third parameters include: arc starting current, gas flow, three-axis front and rear position, three-axis tilt position, three-axis left and right position, ion source magnetic field current, arc starting voltage, focusing voltage, ion screening magnetic field current, energy screening magnetic field current, parallelism correction magnetic field current, and small beam filter; The first parameters include: element mass, energy, beam uniformity, beam size and beam state.

5. The parameter adjustment method according to claim 3, characterized in that: The steps of constructing the hybrid deep neural network model include: Obtaining linear and / or nonlinear relationships between variables in the historical data and empirical data; Establishing an index system of the hybrid deep neural network model according to the relationship between the variables, and confirming the input variables and output variables in the index system; The connection relationship, iteration relationship, activation function, and loss function among the input layer, hidden layer, and output layer in the hybrid deep neural network model are confirmed based on the linear relationship and / or nonlinear relationship.

6. The parameter adjustment method according to claim 5, characterized in that: The activation function includes a linear activation function ReLU and a nonlinear activation function Sigmoid, and the loss function is a mean square error.

7. The parameter adjustment method according to claim 5, characterized in that: The hidden layer of the hybrid deep neural network model includes a linear part and a nonlinear part; The linear part includes four fully connected linear layers for learning the linear relationship of input variables; the nonlinear part includes a total of four fully connected nonlinear layers and linear layers for learning the nonlinear relationship of the input variables.

8. The parameter adjustment method according to claim 2, characterized in that: The step of preprocessing the historical data and the empirical data includes: Deleting abnormal data and / or empty data in the historical data and the experience data; The remaining historical data and the experience data are normalized.

9. The parameter adjustment method according to claim 2, characterized in that: The training set data is 80%, and the test set data is 20%.

10. The parameter adjustment method according to claim 1, characterized in that: The hybrid deep neural network model narrows the adjustment range of the predicted output variable according to the historical data.

Images