Search device, search program, and plasma processing device

Abstract

The parameter compression unit compresses the first input parameter value in a manner that can be restored by the parameter restoration unit, thereby generating a first compressed input parameter value with a reduced number of control parameters. The model learning unit learns a prediction model using learning data that is a group of the first compressed input parameter value and a first output parameter value, wherein the first output parameter value is a processing result obtained by giving the first input parameter value as a plurality of control parameters to a processing device. The processing condition search unit uses the prediction model to infer a second compressed input parameter value corresponding to a target output parameter value.

Description

Technical Field

[0001] The present invention relates to a search device for searching for an optimal solution, a search program, and a plasma processing device having a function of optimizing processing. Background Art

[0002] In recent years, new materials have been introduced into semiconductor devices to improve their performance, and the structure of semiconductor devices is becoming three-dimensional and complex. In addition, nanometer-level precision is required in the processing of current cutting-edge semiconductor devices. Therefore, semiconductor processing equipment needs to process a variety of materials into various shapes with extremely high precision, and it is inevitable that it will become a device with a large number of control parameters (input parameters).

[0003] Along with this, in order to fully exert the performance of semiconductor processing equipment, it is necessary to determine several or even dozens of control parameters. Therefore, as the performance of the device improves and the device becomes more complex, it becomes increasingly difficult to find the combination of control parameters that can obtain the desired processing results. This will cause the device development to be prolonged and become a cause of increased development costs. Therefore, a function or device that semi-automatically searches for the optimal control parameters and easily exerts the performance of the device is sought.

[0004] Patent Document 1 discloses a method and an apparatus for autonomously searching for processing conditions that can provide a desired processed shape through model learning in conjunction with an etching apparatus and a processed shape evaluation apparatus.

[0005] On the other hand, when the search for control parameters is understood as an optimal solution search problem, in order to efficiently search, it is necessary to reduce the number of search parameters and narrow the search range. Patent Documents 2, 3, and 4 disclose methods for reducing the number of model parameters. Patent Documents 5 and 6 disclose methods for narrowing the search range.

[0006] Prior art literature

[0007] Patent Literature

[0008] Patent Document 1: Japanese Patent Application Publication No. 2019-040984

[0009] Patent Document 2: JP-A-2019-046380

[0010] Patent Document 3: Japanese Patent Application Publication No. 2015-162113

[0011] Patent Document 4: Japanese Patent Application Publication No. 2019-079214

[0012] Patent Document 5: Japanese Patent Application Publication No. 2017-157112

[0013] Patent Document 6: Japanese Patent Application Publication No. 2017-102619 Summary of the invention

[0014] Problems to be solved by the invention

[0015] In the method of patent document 1, a learning data set including a small number of processing conditions and processing results is prepared. After the model is learned, the learned model is used to reversely calculate the processing conditions that give the desired processing results. If the number of processing condition parameters is about 10, the processing conditions can be obtained in a few minutes. A verification experiment is carried out on the searched processing conditions. If the desired results are not obtained, the experimental results are added to the learning data set, and the learning of the model is repeated. Since the model learning, processing condition estimation, and experimental verification are repeated several times, the time required for processing condition estimation needs to be the same as the time required for the experiment (the longest is about 1 hour).

[0016] On the other hand, the processing conditions (hereinafter referred to as recipes) of current etching devices have become multi-step types as the materials to be processed become more complex, and the processed shapes become more miniaturized and complicated. In the multi-step type, processing is performed while changing the processing conditions of the etching device for each step. Since the number of steps reaches 10 to 30, and there are dozens of control parameters for each step, the total number of control parameters that must be determined reaches hundreds. Although Patent Document 1 discloses the use of a random search method as a processing condition search method, it is impossible for current computer capabilities to obtain the optimal solution of a parameter space of hundreds of dimensions in a practical computing time.

[0017] In Patent Document 2, the model is set to two stages in order to simplify the model. The output of the previous stage model is compressed and used as the input parameter of the next stage, thereby simplifying the model of the next stage. However, the original number of input parameters is not reduced.

[0018] In Patent Document 3, duplication of model parameters is deleted to simplify the model, but the number of input parameters is not reduced.

[0019] Patent Document 4 uses a dimensionality reduction method such as principal component analysis, but this is to reduce the number of hyperparameters of the model, and does not achieve a reduction in input parameters.

[0020] In Patent Documents 5 and 6, a simulator for simulating a device is used to narrow the search range of control parameters, but this cannot be used if a simulator does not exist.

[0021] As described above, no prior art document discloses a method for efficiently searching for an optimal solution for a model having a large number of search parameters. An object of the present invention is to provide a search device and a search program that can search for an optimal solution in a practical time even for a model having a large number of search parameters.

[0022] Means for solving problems

[0023] A search device of one embodiment of the present invention searches for input parameter values ​​assigned to a plurality of control parameters set in a processing device so that a processing result of a given processing performed by the processing device satisfies a target output parameter value. The search device comprises: a processor; a memory; and a search program, which is stored in the memory and is executed by the processor to search for input parameter values ​​that satisfy the target output parameter value. The search program comprises a parameter compression unit, a model learning unit, a processing condition search unit, a parameter restoration unit, and a convergence determination unit. The parameter compression unit compresses a first input parameter value in a manner that can be restored by the parameter restoration unit to generate a first compressed input parameter value having a reduced number of control parameters. The model learning unit performs the first compression on the first input parameter value. A prediction model is learned using learning data of a group of a completed input parameter value and a first output parameter value, wherein the first output parameter value is a processing result obtained by giving the first input parameter value as a plurality of control parameters to a processing device, a processing condition search unit uses the prediction model to infer a second compressed input parameter value corresponding to a target output parameter value, a parameter restoration unit generates a second input parameter value from the second compressed input parameter value by adding the value of the control parameter deleted by the parameter compression unit, and a convergence determination unit determines whether the second output parameter value converges to a given range of the target output parameter value, wherein the second output parameter value is a processing result obtained by giving the second input parameter value as a plurality of control parameters to the processing device.

[0024] Effects of the Invention

[0025] Even processing conditions with a large number of parameters can be optimized in a short time. Other issues, structures and effects than those described above will be clarified through the following description of the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is a diagram showing an example of a system configuration of a semiconductor manufacturing system.

[0027] Figure 2 This is a block diagram showing the hardware structure of the search device.

[0028] Figure 3 It is a flowchart for determining the optimal processing conditions based on the target processing results.

[0029] Figure 4A It is a diagram for explaining a method of compressing input parameters.

[0030] Figure 4B This is a diagram for explaining the method of restoring input parameters.

[0031] Figure 5 It is a diagram for explaining a method of performing a partial parameter search.

[0032] Figure 6 It is a graph showing the relationship between the number of partial search parameters and the total number of searches.

[0033] Figure 7 It is a diagram showing the meaning of output parameters.

[0034] Figure 8 is the target output value.

[0035] Fig. 9 This is a graph showing the maximum and minimum values ​​allowed for input parameters.

[0036] Fig.10 : is a diagram showing an example of initial input parameters (1 sample response amount).

[0037] Fig.11 It is a diagram showing output parameter values ​​(processing results) for initial input parameters.

[0038] Fig.12 This is a diagram showing the change in the number of parameters when parameter compression is performed.

[0039] Fig.13 This is a diagram showing input parameters after parameter compression (compressed input parameters).

[0040] Fig.14 This is a graph showing the values ​​of the searched recipe candidates.

[0041] Fig.15 This is a graph showing output prediction values ​​for searched recipe candidates.

[0042] Fig.16 This is a schematic diagram of a plasma etching device with a function for searching for optimal processing conditions.

[0043] Fig.17 It is a diagram showing an example of the system configuration of a processing and forming system.

[0044] Fig.18 : is a diagram showing an example of initial input parameters (1 sample response amount).

[0045] Fig.19 This is a diagram showing input parameters after parameter compression (compressed input parameters). DETAILED DESCRIPTION

[0046] The inventors conducted a study using a multi-step recipe for an etching device as an example and found that there are unused or fixed parameters in all samples. There is no need to search for these parameters. In addition, the periodicity of the alternating control parameter values ​​can often be seen in the odd-numbered steps and the even-numbered steps. Such a feature can be seen, for example, in a process where the desired shape is formed while the etching step and the sidewall protection step are alternately performed. Therefore, although the number of all parameters is huge, not all parameters necessarily fluctuate independently, so the number of control parameters can be reduced to a certain extent based on these features.

[0047] However, if this is the case alone, there will still be situations where there are many parameters and it is difficult to search. In this case, it is preferred to perform a high-speed search through an approximate search method. Various methods have been proposed for optimization problems as mathematical analysis, and there are methods that can also be used in topics with huge parameters. If the topic is such that the solution is completed after one parameter search, it is acceptable even if the search takes several days. However, in topics such as recipe search, since the model is built and the recipe is predicted based on a small amount of learning data as in the autonomous search method of patent document 1, it is necessary to repeat the addition of data and the updating of the model several times. Furthermore, since the model is made based on a small amount of learning data, the prediction accuracy of the model is low for parameter values ​​that are far away from the learning data points, and it is meaningless to search such areas in detail. Therefore, compared to spending time searching for the optimal solution for a low-precision model, a method that shortens the search time even for an approximate solution is preferred.

[0048] According to the above, by reducing the data as a pre-processing for preparing the learning data and then using an approximate search method in searching for the optimal solution, the optimal solution can be searched in a practical time even for a recipe with a large number of parameters.

[0049] Example 1

[0050] Figure 1 1 is a diagram showing an example of a system configuration of a semiconductor manufacturing system. The semiconductor manufacturing system 10 includes a search device 100 , an input device 103 , an output device 114 , a processing device 111 , and an evaluation device 112 .

[0051] The processing device 111 is a device for processing a semiconductor or a semiconductor device including a semiconductor. The content of the processing of the processing device 111 is not particularly limited. For example, it includes a photolithography device, a film forming device, a pattern processing device, an ion implantation device, and a cleaning device. The photolithography device includes, for example, an exposure device, an electron beam drawing device, and an X-ray drawing device. The film forming device includes, for example, CVD (Chemical Vapor Deposition), PVD (Physical Vapor Deposition), an evaporation device, a sputtering device, and a thermal oxidation device. The pattern processing device includes, for example, a wet etching device, a dry etching device, an electron beam processing device, and a laser processing device. The ion implantation device includes, for example, a plasma doping device and an ion beam doping device. The cleaning device includes, for example, a liquid cleaning device and an ultrasonic cleaning device.

[0052] The processing device 111 processes the semiconductor or semiconductor device based on the processing conditions (input parameter values) input from the search device 100, and transfers it to the evaluation device 112. The evaluation device 112 measures the semiconductor or semiconductor device processed by the processing device 111 and obtains the processing result (output parameter value). The evaluation device 112 includes an optical monitor and a processing dimension measuring device using an electron microscope. It is also possible to take out a part of the semiconductor or semiconductor device processed by the processing device 111 as a fragment, and transport the fragment to the evaluation device 112 for measurement.

[0053] The search device 100 includes a central processing unit 104, a database 105, a parameter compression unit 106, a model learning unit 107, a processing condition search unit 108, a parameter restoration unit 109, a device control unit 110, and a convergence determination unit 113. The contents of each block will be described later using a flowchart.

[0054] The input device 103 includes an input interface such as a GUI and a storage medium reading device such as a card reader, and inputs data to the search device 100. In addition, the input device 103 receives not only data from the user but also measured values ​​from the evaluation device 112 and inputs them to the search device 100. The input device 103 includes, for example, a keyboard, a mouse, a touch panel, a storage medium reading device, and the like.

[0055] The output device 114 displays the processing conditions handed over from the search device 100 to the user as the optimal processing conditions 102. The display means include displaying on a display or writing to a file. The output device 114 includes, for example, a display, a printer, a storage medium writing device, and the like.

[0056] Figure 2A block diagram showing the hardware structure of the search device 100 is shown. The search device 100 searches for input parameter values ​​(processing conditions) of the processing device 111 that achieve the target output parameter value (target processing result) input from the input device 103. The search device 100 has a processor 116, a communication interface 115, a ROM 117, and a RAM 118. The communication interface 115 connects the processor 116 to the external input device 103, the output device 114, and the processing device 111. The communication interface 115, the ROM 117, and the RAM 118 are connected to the processor 116. The ROM 117 stores a table showing the settable range of the input parameters for the processing device 111, the constraint conditions for the parameters, the convergence judgment conditions, and the processing program executed in the processor 116. The RAM 118 stores the learning data, the learning model, etc. generated during the search process.

[0057] In addition, in Figure 1 In the correspondence, the database 105 is installed as the ROM 117 and the RAM 118, and each block in the search device is installed as a program (search program) stored in the ROM 117.

[0058] Figure 3 This is a flowchart executed by the search device 100 in the semiconductor manufacturing system 10 to determine the optimal processing conditions of the processing device according to the target processing result (target output parameter value).

[0059] First, regarding the processing performed by the processing device 111, the target processing result (target output parameter value) set as the target, the initial processing condition (initial input parameter value) selected as the parameter for controlling the processing device, and the target parameter number and the partial search parameter number 101 are delivered from the input device 103 to the central processing unit 104 (step S100). The target parameter number and the partial search parameter number will be described later.

[0060] Next, the central processing unit 104 stores the received target processing result and the selected initial processing condition in the database 105 , and transfers the selected initial processing condition to the device control unit 110 (step S101 ).

[0061] The device control unit 110 transfers the initial processing conditions to the processing unit 111. Alternatively, the user may input the initial processing conditions output by the device control unit 110 to the processing unit 111. The processing unit 111 performs processing according to the input initial processing conditions, performs evaluation in the evaluation unit 112, and inputs the obtained processing results (initial processing results, initial output parameter values) to the input device 103. The central processing unit 104 receives the initial processing results from the input device 103 (step S102). The central processing unit 104 transfers the initial processing conditions and the initial processing results to the convergence determination unit 113.

[0062] The convergence determination unit 113 compares the initial processing result with the target processing result to determine whether it has converged to the target processing result within a given accuracy (step S103). If converged, the initial processing condition that has converged to the target processing result is transferred to the output device 114, and the output device 114 outputs it as the optimal processing condition 102 (step S111).

[0063] In determining the convergence of the output parameter value (processing result), the sum of squares of the errors between the output parameter values ​​for all the output parameters used and the target output parameter value given by (Equation 1) can be used.

[0064] [Mathematical formula 1]

[0065]

[0066] Here, NP is the total number of output parameters used, y*i is the i-th target output parameter value, and yi is the i-th output parameter value (actual value).

[0067] On the other hand, if convergence has not yet occurred, a command to continue processing is sent from the convergence determination unit 113 to the central processing unit 104. The central processing unit 104 sends the initial processing conditions to the parameter compression unit 106, and the parameter compression unit 106 compresses the initial processing conditions (step S104). The parameter compression method will be described later using a specific example. Thereafter, the central processing unit 104 creates learning data including the compressed initial processing conditions (compressed initial input parameter values) and the initial processing results in the database 105 (step S105).

[0068] Next, the central processing unit 104 reads the learning data from the database 105 and sends it to the model learning unit 107. The model learning unit 107 learns a prediction model that establishes a relationship between the compression completion processing conditions (compression completion input parameter values) and the processing results (output parameter values) (step S106). As the prediction model, a neural network, a support vector machine, a kernel method, etc. can be used. The learned prediction model is transferred to the processing condition search unit 108.

[0069] Next, the processing condition search unit 108 uses the prediction model transferred from the model learning unit 107 and the constraints on the input parameters read from the database 105 to search for the processing conditions for the target processing result read from the database 105 (step S107). In the prediction model, since the processing conditions become input and the processing results become output, in order to reversely obtain the processing conditions based on the processing results, a partial space search based on the random search method is performed. The partial space search method will be described later using a specific example. The processing condition search unit 108 transfers the searched processing conditions (compressed additional input parameter values) to the parameter restoration unit 109.

[0070] Next, the parameter restoration unit 109 restores the value of the control parameter deleted by the parameter compression unit 106 and adds it to the transferred processing condition (step S108), transfers the restored processing condition (added input parameter value) to the device control unit 110, and stores it in the database 105 via the central processing unit 104.

[0071] The device control unit 110 transfers the handed-over processing conditions (additional input parameter values) to the processing unit 111. Alternatively, the processing conditions output by the device control unit 110 may be input to the processing unit 111 by the user. The processing unit 111 performs processing according to the input processing conditions, performs evaluation in the evaluation unit 112, and inputs the obtained processing results (additional output parameter values) to the input device 103. The central processing unit 104 receives the processing results from the input device 103 (step S109). The central processing unit 104 hands over the processing conditions (additional input parameter values) and the processing results (additional output parameter values) to the convergence determination unit 113.

[0072] The convergence determination unit 113 compares the processing result (additional output parameter value (actual value)) with the target processing result (target output parameter value) to determine whether it has converged to the target processing result within a given accuracy (step S110). If it has converged, the processing condition that has converged to the target processing result is transferred to the output device 114, and the output device 114 outputs it as the optimal processing condition 102 (step S111).

[0073] On the other hand, if convergence has not yet occurred, a command to continue processing is sent from the convergence determination unit 113 to the central processing unit 104, and the central processing unit 104 adds and stores a set of processing conditions (added input parameter values) and processing results (added output parameter values ​​(actual values)) in the learning data set of the database 105, and transfers it to the parameter compression unit 106, where the processing conditions (input parameter values) of the learning data set are compressed (step S104). Thereafter, the central processing unit 104 creates learning data including the compressed processing conditions (compressed input parameter values) and processing results (output parameter values) in the database 105, thereby updating the learning data set (step S105).

[0074] Next, the estimation process from parameter compression (S104) to convergence determination (S110) is repeated until the processing result (actual performance value) converges to the target processing result. In this way, the optimal processing conditions for achieving the target processing result are autonomously searched.

[0075] Next, use Figure 4A The method of compressing the input parameters in the parameter compression unit 106 is described below. Figure 4A The first paragraph (uppermost paragraph) of illustrates processing conditions 400. In this example, since the input parameters are 7, S to Y, and the number of steps is 2, steps 1 to 2, the total number of input parameters is 14. In addition, the number of samples is 3, #1 to #3. As data compression methods, for example, the following methods A to D are considered.

[0076] Method A is a method that removes unused or fixed-value parameters in all samples. Figure 4A The second stage of FIG. 4 shows a compressed processing condition 401 in which method A is applied to the processing condition 400. In this case, since the input parameter T is not used and the input parameter U is a fixed value in all samples, the input parameters T and U of steps 1 and 2 are deleted, and the number of input parameters of the compressed processing condition 401 becomes 10. The deleted control parameter values ​​are saved in advance so that the parameter restoration unit 109 can restore them.

[0077] Method B is a method in which a small number (typically 1) of the same parameter values ​​in multiple steps of all sampled parameter values ​​are retained as representatives and the repeated parts are deleted. Figure 4A The third stage of FIG. 4 shows a compressed processing condition 402 in which method B is applied to the processing condition 400. In this case, since the input parameters S, V, and X are the same in steps 1 and 2 in all samples, the input parameters S, V, and X of the repeated step 2 are deleted, and the number of input parameters of the compressed processing condition 401 becomes 11. The proportional coefficient a and the intercept b are stored in advance so that the values ​​of the deleted parameters can be restored by the parameter restoration unit 109.

[0078] The restoration method in this embodiment is as follows Figure 4B As shown, regarding the input parameter value v of the input parameter α and the input parameter value w of the input parameter β, the input parameter values ​​(w1, w2, w3) for each sample of the input parameter β are restored from the input parameter values ​​(v1, v2, v3) for each sample of the input parameter α (when the number of samples is 3) by pre-saving the relationship w=av+b. Figure 4B The restoration method can be expressed in general terms as follows: method B is a method in which the input parameter α is limited to the parameter of step 1 and the input parameter β is limited to the parameter of step 2, the proportionality coefficient a=1, and the intercept b=0.

[0079] Method C is a method of retaining a small number (typically one) of the proportional input parameter groups as representatives and deleting the others. Figure 4A The fourth row of FIG. 4 shows a compressed processing condition 403 in which method C is applied to the processing condition 400. In this case, the input parameters V (10, 20, 20), X (20, 40, 40), Y (15, 30, 30) in step 1 and the input parameters S (10, 20, 20), V (10, 20, 20), X (20, 40, 40) in step 2 are in a proportional relationship (intercept 0) with respect to the input parameter S (10, 20, 20) in step 1, and Y (20, 30, 30) in step 2 is in a proportional relationship (intercept 10) with respect to the input parameter S (10, 20, 20) in step 1. In addition, the input parameter U (150, 150, 150) in step 2 is in a proportional relationship with respect to the input parameter U (100, 100, 100) in step 1. As a result, the input parameters V, X, Y in steps 1 and 2 and the input parameters S and U in step 2 are deleted, and the number of input parameters of the compressed processing condition 403 becomes 6. The proportional coefficient a and the intercept b are stored in advance so that the parameter restoration unit 109 can restore the values ​​of the deleted parameters.

[0080] Method D is a variation of method C, which is a method that retains a small number (typically one) of input parameter groups that are not in a proportional relationship but have a high correlation (correlation coefficient greater than a certain threshold) as representatives and deletes the others. Figure 4A The fifth paragraph of FIG. 4 shows a compressed processing condition 404 in which method D is applied to the processing condition 400. In this case, in addition to the parameters deleted in method C, the input parameter W in step 1 is deleted, and the number of input parameters of the compressed processing condition 404 becomes 5. The proportional coefficient a and the intercept b that are fitted into a straight line by the least square method are stored in advance so that the parameter restoration unit 109 can restore the values ​​of the deleted parameters.

[0081] Thus, the original number of parameters, 14, can be reduced to 10 by method A, to 11 by method B, to 6 by method C, and to 5 by method D. Furthermore, Figure 4A As shown in the sixth paragraph (bottom) of the diagram, if method A and method D are combined, the number of parameters can be reduced to 2.

[0082] In this embodiment, if the number of input parameters is not less than the target number of parameters M specified by the user, a parameter number reduction process is performed. For example, if the input data is a single-step type, method A is applied, and if it is a multi-step type, method A and method B are applied together. If the number of input parameters is not less than M at this stage, method C is applied. If it is not reduced to less than the target number of parameters M after the application of method C, method D is applied. At this time, the threshold value of the correlation coefficient is automatically set so that the number of input parameters is less than the target number of parameters M when method D is applied.

[0083] If the compressed processing conditions are used to search for processing conditions, the values ​​of the deleted input parameters are sequentially restored according to the series of parameter compression methods used with respect to the searched compressed processing conditions. That is, the input parameters deleted by method A can be restored to the saved values, and the parameters deleted by methods B to D can be restored to the original representative values ​​using the saved proportional coefficient and intercept. In the case of method D, since approximation is added during compression, the restoration of the parameters is restoration of the approximate values.

[0084] Next, regarding the method of speeding up the search in the processing condition search unit 108, use Figure 5 Let M be the number of input parameters to be searched. Figure 5 In the example, since the input parameters are X1 to X5, M = 5). In this embodiment, in order to search at a high speed, instead of searching M input parameters at the same time, only a small number of input parameters (N is set as the number of partial search parameters N, N < M) are searched by random search, and the other input parameters ((MN)) are fixed to the best sampled parameter value in the learning data. Figure 5 In the example, N=2, and the values ​​of the best sampling parameters are set for the remaining 3 input parameters. Figure 5 In , the input parameters that use the best sampled values ​​are marked with * to show. Suppose that for N input parameters, all combinations of M input parameters are selected. Therefore, for M C N (=M! / (MN)!N!) combinations are searched. In this example, there are 5C2=10 combinations.

[0085] The values ​​of the N parameters to be searched are given as random numbers that follow a normal distribution (in addition, the average is the value of the parameter in the case of the best sampling, and the variance is the variance of all samples). If the number of search trials for the selected N parameters is S, the number of search points X for each parameter is XN = S, so X = S 1 / N , the total number of searches T is represented by T = S· M C N .

[0086] Figure 6 The relationship between the total search number T and the number of partial search parameters N is shown in the case where M=40 and X=3. When searching for all parameters in a round-robin manner with 3 criteria each, the total search number is But if N=4, the total number of searches is 7×10 6 The total number of searches can be reduced by 12 orders of magnitude. Although the larger the value of the partial search parameter number N, the higher the probability of obtaining a value close to the optimal solution, it is desirable to set N to about 4 to 5 from the perspective of calculation time. In addition, if Figure 6 As shown, when the total number of searches is N=17 to 39, the performance is worse than that of the round-robin method (N=40). This is because duplicate searches occur.

[0087] Since (MN) of the input parameters obtained by the search method have the same value as the best sampled parameter of the existing learning data, when M=40 and N=4, 1% of the parameter to be searched is set as a new value. In addition, in order to improve the efficiency of the autonomous search, multiple candidates (typically 5 to 10) are estimated for the candidate recipe (additional input parameter value) during the search.

[0088] The following description will be given by taking as an example a case where the processing device 111 is an etching device. Figure 7 is the output parameter in this example, and each output parameter represents the cross-sectional shape after processing. The cross-sectional shape after processing is obtained by reading using an electron microscope (evaluation device 112). In this example, the following five output parameters are used to describe the cross-sectional shape after processing, and the five output parameters are: the upper width average value Y1, the middle width minimum value Y2, the middle width maximum value Y3, the lower width average value Y4, and the channel depth Y5.

[0089] Figure 8 is an example of a target output parameter value input to the input device 103, giving a target output parameter value for Figure 7 The target values ​​(sizes) of the five output parameters Y1 to Y5 are shown. Figure 8 The target shape represented by the target output parameter value is equivalent to a vertical channel structure with a width of 20nm and a depth of 200nm.

[0090] Fig. 9 It is an extract of a table showing the maximum and minimum values ​​of the input parameters of the etching device 111 stored in advance in the database 105, which is determined by the prior setting of the engineer who uses it or the specifications of the device. In this example, 72 input parameters can be set. The parameters include the flow rate of various gases, gas pressure, microwave (MW) power, high frequency (RF) power, temperature, etching time, etc.

[0091] Fig.10 This is an example of the initial processing conditions (initial recipe) corresponding to 1 sampling. The initial processing conditions are set by engineers. In this example, the number of samplings is set to 16, among which 7 samplings are set to multi-step processing conditions including 4 steps, and 9 samplings are set to multi-step processing conditions including 7 steps. In this way, multi-step processing conditions with different numbers of steps can be mixed in the initial processing conditions. In this case, for the sampling of 4 steps, all parameters of steps 5 to 7 are set to 0, so that it is treated as 7 steps. Fig. 9 As explained in . Since the number of input parameters per step is 72, the total number of control parameters to be set in the end is 504.

[0092] Fig.11 It is aimed at Fig.10 The 16 samples obtained by the etching device 111 under the initial processing conditions shown were processed by the evaluation device 112, and the output parameter values ​​(actual values) of the processing results were measured. Figure 8 , but the best sample is sample 12.

[0093] Therefore, the parameter compression unit 106 compresses the initial processing conditions. Fig.12 The results of applying parameter compression methods A to D to the initial processing conditions are shown. 40 is set as the target number of parameters M in the parameter compression process. By setting the threshold value of the correlation coefficient in method D to 0.95, the initial 504 parameters can be reduced to 40, which is a reduction effect of about 92%. Fig.13 The 40 input parameters after reduction are shown. The initial processing conditions after compression are the initial input parameter values ​​after compression ( Fig.13 ) and the corresponding initial processing result ( Fig.11 ) groups (in this case, since there are 16 samples, there are 16 groups) constitute the initial learning data.

[0094] Fig.14 10 recipe candidates (compressed and additional input parameter values) found by the processing condition search unit 108 are shown. In the search, 4 is set as the number of partial search parameters N. Fig.14In the figure, the values ​​different from the optimal sample are marked with italic fonts, and all recipes have 4 parameters marked with italic fonts. The search time is about one and a half hours using a PC (Personal Computer), and the search can be performed in a practical time.

[0095] The parameter restoration unit 109 is Fig.14 After the compression is completed, the additional input parameter value is restored to restore the value of the deleted control parameter to restore the processing conditions (additional input parameter value) containing the number of parameters 504. The etching device 111 processes according to the input processing conditions, and the evaluation device 112 evaluates and obtains the processing result (additional output parameter value). For reference, Fig.15 Shows the Fig.14 The output prediction values ​​of 10 recipe candidates. Fig.11 Initial data show improvement but not yet Figure 8 If it is determined that the actual value, i.e., the processing result (additional output parameter value), has not converged to the target processing result, then the additional output parameter value is calculated. Figure 3 The learning data set is updated with 10 new sets of processing conditions (additional input parameter values) and processing results (additional output parameter values) obtained from the process, and the model learning is started again.

[0096] As a variation of the first embodiment, the function of the search device may be mounted in a control device included in the processing device. Fig.16 A schematic diagram of a plasma etching device having an optimal processing condition search function is shown as an example of a processing device. It is equipped with an antenna 56 for generating plasma, a high-frequency power supply 51 for applying a high-frequency voltage thereto, and a first high-frequency matcher 52. A first flow path 61 and a second flow path 62 are provided to introduce multiple gas species into a reactor 60. In addition, only two systems are illustrated here, but the number of flow paths is not particularly limited. By causing the high-frequency alternating electromagnetic field generated in the antenna 56 to act on the introduced mixed gas, an inductively coupled plasma 63 is generated from the reaction particles. In addition, the device is equipped with a substrate voltage generator 54 and a second high-frequency matcher 53 for performing processing based on the generated plasma 63. In addition, it is equipped with an end point determination device 55 that can monitor the changes in the plasma generated when processing the processing object, namely the substrate (sample) 59, and has a mechanism for feeding back the signal obtained by the end point determination device 55 to the first mass flow controller 57 and the second mass flow controller 58. In response to the signal from the endpoint determination device 55 , the gas flow rate in the first flow path 61 can be adjusted by the first mass flow controller 57 , and the gas flow rate in the second flow path 62 can be adjusted by the second mass flow controller 58 .

[0097] The control device 70 of the plasma etching device controls the high frequency power supply 51, substrate voltage generator 54, end point determination device 55 and other plasma generating devices of the device to perform etching processing on the substrate 59, and can install the ROM 117 ( Figure 2 ) is used to perform the search process described in the first embodiment. The input device 71 and the output device 72 of the plasma etching device respectively play the role of the search device ( Figure 1 ) in the input device 103 and the output device 114. In this way, the search device 100 can be installed as a function of the processing device 111 instead of being independently provided for the processing device 111, and the plasma generating device can be controlled by the searched input parameter value to perform etching processing.

[0098] Example 2

[0099] In the first embodiment, the present invention is described by taking a semiconductor manufacturing system including a processing device for processing a semiconductor or a semiconductor device including a semiconductor as an example, but the search device or search method of the present invention is not limited to the semiconductor manufacturing system. As the second embodiment, an example of applying the search device or search method to a processing molding system having an injection molding machine is described. In addition, for components having substantially the same functions as those in the first embodiment, the same reference numerals are marked and detailed descriptions are omitted, and the description is centered on the different parts.

[0100] Fig.17 The machining system 20 includes a search device 100 , an input device 103 , an output device 114 , a molding device 211 , and a measuring device 212 .

[0101] The molding device 211 is a device for molding various materials such as plastic materials. The molding device 211 molds the material according to the processing conditions (input parameter values) input from the search device 100, and transfers the molded material to the measuring device 212. The measuring device 212 measures the dimensions of the material molded in the molding device 211 and obtains the processing shape data (output parameter values).

[0102] In Example 2, also according to Figure 3 The target processing result is given as a target size (target output parameter value) representing the shape of the molded product formed by the molding device 211.

[0103] The output parameter used in designating the target size input to the input device 103 is the same as Figure 8Similarly, this is performed by describing the shape of the molded product using a plurality of parameters.

[0104] The database 105 stores in advance a table showing the maximum and minimum values ​​of the input parameters of the forming device 211. Fig. 9 The same data structure determines the parameters and their maximum and minimum values ​​by the specifications of the molding device. For example, if the molding performed by the molding device 211 is performed by forming a cycle with four processes of plasticization, injection, pressure holding, and cooling and repeating it several times, it is necessary to set control parameters (input parameters) for each process. For example, there are control parameters such as cylinder temperature, motor speed, cylinder pressure in the plasticization process, screw speed in the injection process, resin pressure in the pressure holding process, and mold temperature in the cooling process. The maximum and minimum values ​​are defined for each parameter.

[0105] Fig.18 This is an example of the initial processing conditions (initial recipe) for 1 sampling amount. The initial processing conditions are set by the engineer. In this example, the number of cycles is 3, the number of parameters for 1 cycle is 6, and the total number of parameters is 18. The number of samples is set to 10, 10 types of initial processing conditions are set, the forming device 211 performs the processing, and the measuring device 212 obtains the output parameter value (actual value, initial processing result) representing the processing result.

[0106] As in Example 1, parameter compression methods A to D are applied to the initial processing conditions in sequence to compress the 18 input parameters to a value less than the target parameter number M. For example, if the target parameter number M = 10, we get Fig.19 The prediction model is learned using the compressed initial input parameter values ​​as shown and the set of the compressed initial input parameter values ​​and the corresponding initial processing results as initial learning data.

[0107] In this case, since the compressed initial input parameter is 10, although all parameters can be searched as objects, in order to achieve higher speed, it is desirable to set the partial search parameter number N to perform partial space search as in Example 1.

[0108] The present invention has been described above based on the embodiments. In addition, the present invention is not limited to the aforementioned embodiments, but includes various modifications and equivalent structures. For example, the aforementioned embodiments are described in detail in order to explain the present invention in an easy-to-understand manner, and the present invention is not necessarily limited to having all the structures described. In addition, a part of the structure of a certain embodiment can be replaced with the structure of other embodiments. In addition, the structure of other embodiments can be added to the structure of a certain embodiment. In addition, other structures can be added, deleted or replaced with a part of the structure of each embodiment.

[0109] In addition, the aforementioned structures, functions, processing units, processing means, etc. may be partially or entirely implemented in hardware by, for example, designing with integrated circuits, or may be implemented in software by interpreting and executing programs that implement the various functions by a processor. Information such as programs, tables, and files that implement the various functions may be stored in a storage device such as a memory, a hard disk, an SSD (Solid State Drive), or a recording medium such as an IC (Integrated Circuit) card, an SD card, or a DVD (Digital Versatile Disc).

[0110] In addition, the control lines and information lines are shown as those considered necessary for explanation, and are not limited to showing all the control lines and information lines necessary for installation. In fact, it can be considered that almost all the structures are connected to each other.

[0111] Description of Reference Numerals

[0112] 10: Semiconductor manufacturing systems,

[0113] 20: Processing and forming system,

[0114] 51: High frequency power supply,

[0115] 52: 1st high frequency matching device,

[0116] 53: Second high frequency matching device,

[0117] 54: Substrate voltage generator,

[0118] 55: End point determination device,

[0119] 56: Antenna,

[0120] 57: 1st mass flow controller,

[0121] 58: Second mass flow controller,

[0122] 59: Substrate,

[0123] 60: Reactor,

[0124] 61: 1st flow path,

[0125] 62: 2nd flow path,

[0126] 63: Plasma,

[0127] 70: Control device,

[0128] 71: Input device,

[0129] 72: output device,

[0130] 100: Search device,

[0131] 103: Input device,

[0132] 104: Central Processing Unit,

[0133] 105: Database,

[0134] 106: parameter compression unit,

[0135] 107: Model Learning Department,

[0136] 108: Processing condition search unit,

[0137] 109: Parameter recovery unit,

[0138] 110: Device control unit,

[0139] 111: Processing device,

[0140] 112: Evaluation device,

[0141] 113: Convergence determination unit,

[0142] 114: output device,

[0143] 115: Interface,

[0144] 116: Processor,

[0145] 117: ROM,

[0146] 118: RAM,

[0147] 211: Forming device,

[0148] 212: Measuring device.

Claims

1. A search device for searching input parameter values ​​assigned to a plurality of control parameters set in a processing device so that a processing result of a given processing performed by the processing device satisfies a target output parameter value, wherein the search device is characterized in that: processor; Memory; and A search program is stored in the memory and is executed by the processor to search for the input parameter value that satisfies the target output parameter value. The search program includes a parameter compression unit, a model learning unit, a processing condition search unit, a parameter restoration unit, and a convergence determination unit. The parameter compression unit generates a first compressed input parameter value which is an input parameter value obtained by deleting values ​​of some control parameters of the first input parameter value so that the number of control parameters becomes less than the target number of parameters. The model learning unit learns the prediction model by using learning data that is a combination of the first compressed input parameter value and the first output parameter value, wherein the first output parameter value is a processing result obtained by giving the first input parameter value as the plurality of control parameters to the processing device, The processing condition search unit estimates a second compressed input parameter value corresponding to the target output parameter value using the prediction model. The parameter restoring unit generates a second input parameter value by restoring the value of the control parameter deleted by the parameter compressing unit from the second compressed input parameter value. The convergence determination unit determines whether a second output parameter value converges within a given range of the target output parameter value, wherein the second output parameter value is a processing result obtained by giving the second input parameter value as the plurality of control parameters to the processing device.

2. The search device according to claim 1, characterized in that: When the convergence determination unit determines that the second output parameter value has not converged to within the given range of the target output parameter value, the search program adds the group of the second input parameter value and the second output parameter value to the group of the first input parameter value and the first output parameter value to update the prediction model.

3. The search device according to claim 1, characterized in that: The given process includes a plurality of steps in which different values ​​are assigned to the plurality of control parameters. The search procedure searches for values ​​of the plurality of control parameters in the plurality of steps as the input parameter values.

4. The search device according to claim 1, characterized in that: The processing condition search unit estimates a plurality of second compressed input parameter values ​​using the prediction model.

5. The search device according to claim 1, characterized in that: The parameter compression unit deletes unused or fixed control parameter values ​​from among the first input parameter values, and stores the deleted control parameter values.

6. The search device according to claim 1, characterized in that: The parameter compression unit retains the value v of the first control parameter among the first input parameter values, deletes the value w of the second control parameter, and stores the values ​​of the coefficient a and the intercept b of w=av+b.

7. The search device according to claim 6, characterized in that: The value w of the second control parameter is proportional to the value v of the first control parameter, or the value v of the first control parameter and the value w of the second control parameter have a correlation coefficient greater than a given threshold value.

8. A search device for searching input parameter values ​​assigned to a plurality of control parameters set in a processing device so that a processing result of a given processing performed by the processing device satisfies a target output parameter value, the search device being characterized in that: processor; Memory; and A search program is stored in the memory and is executed by the processor to search for the input parameter value that satisfies the target output parameter value. The search program includes a model learning unit, a processing condition search unit, and a convergence determination unit. The model learning unit learns the prediction model by using learning data that is a combination of a first compressed input parameter value and a first output parameter value, wherein the first output parameter value is a processing result obtained by giving the first input parameter value as the plurality of control parameters to the processing device, The processing condition search unit estimates a second compressed input parameter value corresponding to the target output parameter value using the prediction model. The convergence determination unit determines whether the second output parameter value converges to a given range of the target output parameter value, wherein the second output parameter value is a processing result obtained by giving the second input parameter value as the plurality of control parameters to the processing device, The processing condition search unit changes the values ​​of some control parameters of the first compressed input parameter value and fixes the values ​​of other control parameters to the given control parameter values ​​of the learning data to search for an approximate solution. The first compressed input parameter value is an input parameter value obtained by deleting the values ​​of some control parameters of the first input parameter value so that the number of control parameters becomes less than the target parameter number, and the second input parameter value is an input parameter value obtained by restoring the values ​​of the control parameters deleted from the first compressed input parameter value from the second compressed input parameter value.

9. The search device according to claim 8, characterized in that: The processing condition search section fixes the value of the other control parameter to the value of the control parameter of the learning data that gives the best solution among the learning data.

10. A search program for searching for input parameter values ​​assigned to a plurality of control parameters set in a processing device so that a processing result of a given processing performed by the processing device satisfies a target output parameter value, the search program being characterized in that: The first step is to generate a first compressed input parameter value obtained by deleting values ​​of some control parameters of the first input parameter value so that the number of control parameters becomes less than the target parameter number; The second step is to learn a prediction model by using learning data that is a combination of the first compressed input parameter value and a first output parameter value, wherein the first output parameter value is a processing result obtained by giving the first input parameter value as the plurality of control parameters to the processing device; Step 3, using the prediction model to estimate a second compressed input parameter value corresponding to the target output parameter value; A fourth step of generating a second input parameter value by restoring the value of the control parameter deleted in the first step from the second compressed input parameter value; and In step 5, it is determined whether the second output parameter value converges within a given range of the target output parameter value, wherein the second output parameter value is a processing result obtained by giving the second input parameter value as the plurality of control parameters to the processing device.

11. The search program according to claim 10, characterized in that When it is determined in step 5 that the second output parameter value has not converged to within the given range of the target output parameter value, the group of the second input parameter value and the second output parameter value is added to the group of the first input parameter value and the first output parameter value to update the prediction model.

12. The search program according to claim 10, characterized in that In the third step, the values ​​of some control parameters of the first compressed input parameter value are changed, and the values ​​of other control parameters are fixed to the values ​​of control parameters of the learning data that gives the best solution among the learning data, so as to search for an approximate solution.

13. A plasma processing device for performing plasma processing on a sample using plasma, wherein the plasma processing device is characterized in that: Processing room; a plasma generating device that generates plasma within the processing chamber; and A control device that executes the search program described in claim 10, searches for input parameter values ​​assigned to the multiple control parameters set in the plasma processing device that satisfy the target output parameter value, and controls the plasma generating device so that plasma processing of the sample loaded in the processing chamber is performed by the searched input parameter values.

Images