Searching device, searching method, and semiconductor device manufacturing system

Abstract

To enable a user to utilize reference processing data, which is optimal for searching for a target processing condition, among a large amount of accumulated reference processing data without requiring the user to have specialized knowledge in machine learning, this search device for searching for a manufacturing condition corresponding to a desired processing result by semiconductor manufacturing equipment by predicting the manufacturing condition corresponding to the desired processing result using a trained model is configured to, if a trained model is generated by transfer learning using first data and second data and the generated trained model does not satisfy a predetermined determination criterion, regenerate the trained model by transfer learning using the first data and additional second data.

Description

[Technical field]

[0001] The present invention relates to a searching apparatus and a searching method for searching for manufacturing conditions that realize desired processing results, and a semiconductor device manufacturing system. [Background technology]

[0002] In semiconductor manufacturing, it is necessary to set appropriate processing conditions to obtain the desired processing results. As semiconductor devices continue to become finer and processing control parameters increase, it is expected that in the future, processing conditions to obtain the desired processing results (reduced machine differences or high accuracy) will be derived by machine learning. Here, the processing conditions consist of at least one or more control parameter items of the processing equipment.

[0003] In recent years, the introduction of new materials and the increasing complexity of device structures have expanded the range of control for processing equipment, resulting in the addition of many new items to the processing conditions. Optimizing the processing conditions is essential to fully utilize the performance of the processing equipment. For this reason, attention has been focused on a method that uses machine learning to derive processing conditions that will achieve the good processing results desired by process developers. Here, the processing result consists of at least one or more items that indicate the shape and properties of the sample on which the processing was performed. Hereinafter, this good processing result will be referred to as the "target processing result."

[0004] The target processing result will be explained using an example of an etching process for a material to be etched on a silicon (Si) wafer 11. Figure 1 shows cross-sectional views of two locations, near the center 12 and near the edge 13, of the surface of the Si wafer 11 after the entire wafer and after the etching process. The material to be etched 14 formed on the surface of the Si wafer 11 is removed by etching, and the amount of etching 15 at that location can be estimated by measuring the difference in height between the surface 16 before etching, shown by the dashed line.

[0005] The etching rate and the in-plane uniformity of the etching rate can be calculated from the in-plane distribution data of the etching amount15 and the time required for etching. If the etching rate is an item of the processing result, the target processing result is defined as a specified value or a range of specified values, such as "etching rate of 50 nm / min" or "etching amount of 20 nm with in-plane variation within 5%." The processing conditions that realize such target processing results are called "target processing conditions."

[0006] The method of deriving target processing conditions by machine learning is generally carried out in the following procedure. First, a target processing result is set. Meanwhile, multiple basic processing conditions are determined, processing based on the basic processing conditions is performed on the sample, processing data consisting of the basic processing conditions and their processing results is obtained, and an initial processing database is constructed. A model describing the correlation between the basic processing conditions and their processing results is estimated by machine learning based on the initial processing database. Hereinafter, such a model will be referred to as an input / output model, since it describes the input / output relationship y=f(x), assuming that the processing conditions are input x and the processing results are output y. Based on the estimated input / output model, processing conditions (referred to as "predicted processing conditions") that satisfy the target processing result are predicted.

[0007] Next, a verification experiment is carried out using the obtained predicted processing conditions. That is, processing based on the predicted processing conditions is executed, and it is determined whether the obtained processing result is the target processing result. If the target processing result is obtained, the predicted processing conditions are set as the target processing conditions, and the verification experiment is terminated. On the other hand, if the target processing result is not obtained, the processing data obtained in the verification experiment is added to the database to update the input / output model, and prediction of processing conditions and the verification experiment are repeated until the target processing result is obtained.

[0008] In such a method of deriving target processing conditions, the accuracy of the input / output model used to predict the target processing conditions is important. Figure 2 is a graph showing the correlation (input / output relationship) between processing conditions and processing results. Here, the dashed line 21 represents the true input / output relationship, while the solid line 22 and the dashed line 23 represent the input / output relationships represented by input / output model A and input / output model B, respectively. The accuracy of the input / output model can be evaluated as the similarity to the true input / output relationship shown by the dashed line. In this case, the input / output relationship of input / output model A (solid line 22) is similar to the true input / output relationship (dashed line 21), and the accuracy of input / output model A is high. On the other hand, the input / output relationship of input / output model B (dashed line 23) deviates from the true input / output relationship (dashed line 21), and the accuracy of input / output model B is low.

[0009] Processing results based on predicted processing conditions obtained based on low-precision input-output models are likely to deviate from the target processing results. Therefore, the number of verification experiments required to obtain the target processing conditions increases. This increases the process development period and process development costs, including experimental expenses and labor costs. To avoid such a situation, it is necessary to improve the accuracy of the input-output model.

[0010] In order to improve the accuracy of the input / output model, a method of constructing a large-scale initial process database in advance is considered. However, this method requires repeating the process many times to construct the initial process database, and is not a fundamental solution for reducing the process development period and cost.

[0011] As a method for improving the accuracy of the input-output model while suppressing the amount of processing data acquired for constructing the initial processing database, there is a method for utilizing processing data acquired in a process other than the process for which processing conditions are to be derived (called the "target process"). Specifically, based on a database (called the "reference processing database") of processing data (called the "reference processing data") acquired in a process (called the "reference process") other than the target process, an input-output model (called the "reference input-output model") that describes the input-output relationship in the reference process is estimated, and the estimated reference input-output model is referred to for prediction in the target process.

[0012] Patent Document 1 describes that "a computer for determining control parameters of a process performed on a sample includes a memory unit for storing a first model showing a correlation between a first process output obtained by measuring a first sample used in manufacturing and processed, and a second process output obtained by measuring a second sample that is easier to measure than the first sample and processed, and a second model showing a correlation between a control parameter of a process performed on the second sample and the second process output, and an analysis unit for calculating a target control parameter of a process performed on the first sample based on a target process output that is the first process output, the first model, and the second model," thereby "calculating optimal control parameters while reducing costs for process development." Patent Document 1 also describes, as an example, the use of a "qualitative real sample-substitute sample relationship model in which the larger B is, the larger A is" as the first model when a variable of the process output of a substitute sample, which is the second sample, is A and a variable of the process output of a real sample, which is the first sample, is B.

[0013] Patent Document 2 describes a processing condition search device that searches for processing conditions for a target process, the processing condition search device including a target processing result setting unit that sets a target processing result in the target process, a processing database that stores target processing data that is a combination of processing conditions and processing results in the target process, and a reference processing database that stores reference processing data that is a combination of processing conditions and processing results in a reference process, and a learning database that uses the target processing data to estimate an input / output model of the target process that is an input / output model between the target explanatory variables and the target objective variables, with the processing conditions of the target processing data as target explanatory variables and the processing results as target objective variables. The document states that the system "has a supervised learning execution unit, a transfer learning execution unit that estimates an input / output model of the target process using a reference input / output model between the reference explanatory variables and the reference objective variable, and the target processing data, with the processing conditions of the reference processing data as reference explanatory variables and the processing results as reference objective variables, and the target processing data, a transfer feasibility determination unit that determines whether the input / output model of the target process will be estimated by the supervised learning execution unit or the transfer learning execution unit, and a processing condition prediction unit that predicts the processing conditions that will realize the target processing result using the input / output model of the target process," thereby "searching for target processing conditions while reducing process development period and process development costs."

[0014] Furthermore, Patent Document 2 describes an example in which a combination of simulation results and simulation conditions obtained by simulating the target process is used as the reference processing database, rather than data obtained by actual processing using a processing device as reference processing data. [Prior art documents] [Patent documents]

[0015] [Patent Document 1] JP 2019-47100 A [Patent Document 2] JP 2021-182182 A Summary of the Invention [Problem to be solved by the invention]

[0016] In the method for determining process control parameters described in Patent Document 1, the process data of the second sample is used as reference process data to estimate a reference input-output model. The process conditions of the first sample are determined by referring to the reference input-output model. For this method of predicting the process in the target process by referring to the reference input-output model to be effective, several conditions must be satisfied.

[0017] 3A is a graph showing the input-output relationship (solid line 30) of an estimated input-output model and the true input-output relationship (dashed line 20) of a target process based on processing data consisting of processing results obtained by setting multiple basic processing conditions for the target process. In this example, there are few basic processing conditions set (black dots represent processing data; the same applies to Figs. 3B and C below), and the accuracy of the input-output model is low.

[0018] 3B is a graph showing the input-output relationship of the reference input-output model (solid line 31) estimated based on the reference process data stored in the reference process database for the reference process and the true input-output relationship of the reference process (dashed line 21). In this example, the reference process database is large-scale, so the accuracy of the reference input-output model is high.

[0019] Fig. 3C is a graph showing the input / output relationship (solid line 32) of an input / output model estimated by transfer learning referring to the reference input / output model shown in Fig. 3B and the true input / output relationship (dashed line 20) of a target process. The processing data of the target process used in transfer learning is the same as that in Fig. 3A, but since the true input / output relationship (dashed line 20) of the target process is similar to the true input / output relationship (dashed line 21) of the reference process, the accuracy of the input / output model estimated by transfer learning is improved over the accuracy of the input / output model shown in Fig. 3A.

[0020] Here, the true input-output relationships f and g being similar includes not only the case where they are roughly the same, but also the case where the input-output relationships are roughly the same except for the difference in constants or coefficients. That is, when f ≒ g or f ≒ ag + b holds. For example, if the target process and the reference process are both etching processes for the same sample, and only the processing time differs, 10 seconds and 100 seconds, respectively, the basic function characteristics are the same even if the processing results differ by a factor of about 10. That is, f ≒ 10g holds for the true input-output relationship, and the application of transfer learning is expected to be effective.

[0021] In this way, the technique of utilizing reference processing data of a reference process (transfer learning) is effective when, for example, the true input-output relationship between the target process and the reference process is similar, or the reference input-output model is more accurate than an input-output model estimated from only the target processing data. However, it is not necessarily effective when these conditions are not met.

[0022] In semiconductor processing, there are a wide variety of samples, processing equipment, and processing processes, so there are generally many candidates for reference processing data. However, depending on the selection of reference processing data, the accuracy of the input / output model may not be improved as much as expected.

[0023] For example, even if the target process and the reference process are the same etching process and the processing result item for both processes is the etching amount, if the material of the etched film of the sample to be processed is different, the characteristics of the etching rate against the processing conditions will be significantly different. Therefore, there is a risk that the true input-output relationship will not be similar in the first place.

[0024] Furthermore, even if reference processing data with a similar true input-output relationship is selected in order to estimate a reference input-output model, if the reference processing data is extremely small and a reference input-output model with sufficient accuracy cannot be obtained, it may not be possible to obtain an improvement in accuracy by referring to the reference input-output model.

[0025] If such inappropriate reference processing data is used, it is not possible to improve the accuracy of the input / output model to be predicted, and this may lead to an increase in process development time and costs.

[0026] In addition, in machine learning, models generally learn based on known input data and output data, so when reusing a model that has already completed learning through transfer learning, the explanatory variables of the input data input to the model must be equivalent to those input during learning, even if the values ​​are different. For example, if there is a trained model that predicts "etching amount" based on the three input conditions of "temperature," "pressure," and "processing time," it is not possible to input "power" into this model to predict the "etching amount." It is also not possible to provide data that is missing "temperature," so some value must be input.

[0027] In Patent Document 2, the transfer learning execution unit performs learning using a reference input / output model and target processing data. In this case, it is considered that the input of the reference input / output model and the explanatory variables of the target processing data basically correspond to each other in many cases, but in the transfer learning in the above-mentioned case where "a combination of a simulation result and a simulation condition obtained by simulating a target process is used as a reference processing database, instead of data obtained by actually processing with a processing device as the reference processing data", it is not always possible to align the input of the reference input / output model and the explanatory variables of the target processing data.

[0028] For example, in actual processing conditions, "temperature," "pressure," and "processing time" are input as experimental conditions, but when simulating them with a physical simulator, the temperature term may not be able to be handled due to the convenience of the simulation model. There are also many other cases that are not easy to handle with a simulator, such as when incorporating the physics of response to pulses with a period of several milliseconds or more into a simulation that handles time development on a time scale of microseconds or less, or when wanting to handle metadata such as processing date and time. Furthermore, conversely, it is also assumed that there may be parameters, such as the calculation conditions used in the simulation, that affect the reference processing data but are not included in the explanatory variables of the target processing data.

[0029] The above-mentioned differences in input data formats apply not only when the target processing data is the result of actual processing and the reference processing data is the result of simulation, as described above, but also when the target processing data is the result of simulation and the reference processing data is the result of actual processing. Furthermore, even if both are the results of actual processing, a parameter that can be handled by one device may not be able to be handled by the other device due to a slight change in the system state of the processing device.

[0030] When you want to perform transfer learning using two sets of data with different explanatory variables like this, you can either preprocess the data by deleting the explanatory variables that are not present in one set and substituting some constant value or predicted value, or you can change the network structure of the neural network model.

[0031] The former method of deleting explanatory variables or inputting constant or predicted values ​​requires data processing, and the model cannot take into account explanatory variables that have been deleted or input with constant values, resulting in reduced accuracy. The latter method of changing the network structure has a degree of freedom in itself, but at the same time, it is necessary to avoid problems such as overfitting and negative transfer, making it difficult for users who are unfamiliar with machine learning to do it on their own. In addition, in order to avoid problems such as overfitting and negative transfer, it is also difficult to select appropriate data to search for target processing conditions from a large reference processing database.

[0032] The present invention solves the problems of the conventional technology described above, and provides a search device and search method, as well as a semiconductor device manufacturing system, that continuously and automatically accumulates reference processing data while at the same time searching for manufacturing conditions for utilizing optimal reference processing data to search for target processing conditions from a large amount of accumulated reference processing data, without the user needing specialized knowledge of machine learning. [Means for solving the problem]

[0033] In order to solve the above-mentioned problems, the present invention provides a search device that searches for manufacturing conditions corresponding to desired processing results by predicting manufacturing conditions corresponding to desired processing results of a semiconductor manufacturing device using a learning model, in which a learning model is generated by transfer learning using first data and second data, and if a predetermined judgment criterion is not satisfied by the generated learning model, the learning model is regenerated by transfer learning using the first data and the added second data. The learning model includes a reference model, and the reference model is a model generated based on explanatory variables of the second data and objective variables of the second data. It was configured as follows.

[0034] In order to solve the above-mentioned problems, the present invention provides a searching method for searching for manufacturing conditions corresponding to a desired processing result of a semiconductor manufacturing device by predicting the manufacturing conditions corresponding to the desired processing result using a learning model, the searching method comprising the steps of: generating a learning model by transfer learning using first data and second data; and, if a predetermined judgment criterion is not satisfied by the generated learning model, regenerating the learning model by transfer learning using the first data and the added second data. The learning model includes a reference model, and the reference model is a model generated based on explanatory variables of the second data and objective variables of the second data. I did so.

[0035] In order to solve the above-mentioned problems, the present invention provides a semiconductor device manufacturing system including a platform in which semiconductor manufacturing equipment is connected via a network and an application for predicting manufacturing conditions corresponding to desired processing results of the semiconductor manufacturing equipment using a learning model is implemented, the application executes a step of generating a learning model by transfer learning using first data and second data, and a step of regenerating the learning model by transfer learning using the first data and the added second data when a predetermined judgment criterion is not satisfied by the generated learning model. The learning model includes a reference model, and the reference model is a model generated based on explanatory variables of the second data and objective variables of the second data. It was configured as follows. Effect of the Invention

[0036] According to the present invention, it is possible to search for target processing conditions while suppressing the process development period and cost. In addition, even during the period when the target process is not actually being processed, the automatic reference processing data acquisition execution unit can automatically and continuously improve the prediction accuracy. [Brief description of the drawings]

[0037] [Figure 1] 1 is a perspective view of a wafer and an enlarged cross-section of the surface near the center and edge of the wafer. [Diagram 2] FIG. 1 is a diagram for explaining the background of the present invention, and is a graph showing the correlation (input / output relationship) between processing conditions and processing results. [Figure 3A] This is a graph showing the relationship between processing conditions (input) and processing results (output) to explain the problem of the present invention, and shows the input / output relationship of an estimated input / output model when few basic processing conditions are set and the accuracy of the input / output model is low, and the true input / output relationship of the target process. [Figure 3B] This is a graph showing the relationship between processing conditions (input) and processing results (output) to explain the problem of the present invention, and shows the input / output relationship of a reference input / output model estimated based on reference processing data and the true input / output relationship of the reference process. [Figure 3C]This is a graph showing the relationship between processing conditions (input) and processing results (output) to explain the problem of the present invention, and shows the input / output relationship of an input / output model estimated by transfer learning that references a reference input model and the true input / output relationship of a target process. [Figure 4] 1 is a block diagram showing a schematic configuration of a processing condition search system according to a first embodiment of the present invention. [Diagram 5] FIG. 1 is a block diagram showing the concept of a transfer learning model using a neural network according to a first embodiment of the present invention. [Figure 6] FIG. 2 is a front view of a screen showing an example of a GUI (ROI data selection manager) provided to a user by a model explanation unit according to the first embodiment of the present invention. [Figure 7] FIG. 11 is a front view of a screen showing an example of a GUI (model optimization termination criterion setting) provided to a user by a transfer learning model evaluation unit 45 according to the first embodiment of the present invention. [Figure 8] 1 is a flowchart showing a process from the start of an operation to prediction of a target processing condition according to the first embodiment of the present invention. [Figure 9] 13 is a flowchart showing a procedure in which a computer automatically expands a reference process database during a period in which no operation for searching for a processing condition is performed according to the second embodiment of the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0038] The present invention is a search system that searches for desired manufacturing conditions for semiconductor manufacturing equipment through machine learning, which predicts the desired manufacturing conditions for the semiconductor manufacturing equipment using a model constructed through transfer learning of data from a physical simulator.

[0039] Generally, physical simulations are unable to take into account all parameters under actual processing conditions, and conventional machine learning using neural networks is unable to learn task data with different features and labels using a single model; however, the present invention aims to solve these problems by using a network structure that employs transfer learning.

[0040] That is, in the present invention, the characteristics of the model are set in advance by a "model explanation unit" to prevent negative transfer, and the model obtained as a result of transfer learning is evaluated by a "transfer learning model evaluation unit." If the evaluation value does not exceed a threshold as a result of the model evaluation, simulation data of the conditions necessary to improve the accuracy of the transfer learning model is automatically generated by an attached computer ("automatic execution unit for acquiring reference process data"), and transfer learning is performed again.

[0041] As a result, an optimal transfer learning model for predicting the target processing results set by the user is constantly constructed and updated automatically, making it possible to shorten and reduce the time required for recipe optimization to reduce machine / component differences by utilizing a large amount of training data generated by simulations that are less expensive than actual processing.

[0042] Hereinafter, an embodiment of the present invention will be described with reference to the drawings. However, the present invention is not to be interpreted as being limited to the description of the embodiment shown below. It is easily understood by those skilled in the art that the specific configuration can be changed without departing from the idea or intent of the present invention. In addition, the position, size, shape, etc. of each configuration shown in the drawings in this specification may not represent the actual position, size, shape, etc. in order to facilitate understanding of the invention. Therefore, the present invention is not limited to the position, size, shape, etc. disclosed in the drawings, etc. EXAMPLES

[0043] In this embodiment, in order to be able to search for target processing conditions while suppressing the process development period and cost, a processing condition search device for searching for processing conditions of a target process is provided, which includes a target processing result setting unit that sets a target processing result in the target process, a learning database including a target processing database that stores target processing data which is a combination of processing conditions and processing results in the target process, and a reference processing database that stores reference processing data which is a combination of processing conditions and processing results in a reference process, and a reference input / output model between the reference explanatory variables and the reference objective variables, using the reference processing data, with processing conditions as reference explanatory variables and processing results as reference objective variables. We will explain an example that is configured with a model explanation unit that explains the features, a transfer learning execution unit that uses target processing data, with the processing conditions of the target processing data as target explanatory variables and the processing results as target objective variables, and estimates an input / output model of the target process using the target explanatory variables, the target objective variable, and a reference input / output model, a transfer learning model evaluation unit that evaluates the transfer learning model, which is a model of the target process input / output estimated by the transfer learning execution unit, a reference processing data acquisition automatic execution unit that adds new reference processing data to the reference processing database based on the evaluation by the transfer learning model evaluation unit, and a processing condition prediction unit that predicts the processing conditions that will achieve a target processing result using the transfer learning model.

[0044] FIG. 4 is a block diagram showing an example of the configuration of a processing condition search system 40 according to the first embodiment. The processing condition search system 40 includes a database unit 410 that stores data of a target process and data of a reference process, a transfer learning execution and evaluation unit 420 that performs transfer learning using the data stored in the database unit 410 and evaluates a learning model created, an automatic reference process data acquisition execution unit 46 that acquires reference process data when the transfer learning model evaluated by the transfer learning execution and evaluation unit 420 does not clear the target, a processing condition prediction unit 47, a target processing result setting unit 48, and an output unit 49.

[0045] The database unit 410 is configured to include a target process database 41 and a reference process database 42, and the transfer learning execution / evaluation unit 420 is configured to include a model explanation unit 43, a transfer learning execution unit 44, and a transfer learning model evaluation unit 45. Each of the components is connected to each other directly or via a network.

[0046] The target process database 41 stores target process result data, which is a combination of past process conditions Xp and process results Yp in the target processing device. The type and content of the process performed by the processing device here are not limited. The processing device includes, for example, a lithography device, a film forming device, a pattern processing device, an ion implantation device, a heating device, a cleaning device, and the like.

[0047] Lithography equipment includes exposure equipment, electron beam lithography equipment, and X-ray lithography equipment. Film formation equipment includes CVD, PVD, deposition equipment, sputtering equipment, and thermal oxidation equipment. Pattern processing equipment includes wet etching equipment, dry etching equipment, electron beam processing equipment, and laser processing equipment. Ion implantation equipment includes plasma doping equipment and ion beam doping equipment. Heating equipment includes resistance heating equipment, lamp heating equipment, and laser heating equipment. Cleaning equipment includes liquid cleaning equipment and ultrasonic cleaning equipment.

[0048] In the first embodiment, the processing apparatus is assumed to be a "dry etching apparatus," the processing conditions are "temperature," "pressure," "flow rate of gas A," "flow rate of gas B," "power," and "processing time," and corresponding values ​​actually implemented, and the processing result is "etching amount." The processing condition Xp items "temperature," "pressure," "flow rate of gas A," "flow rate of gas B," "input power," and "processing time" are called explanatory variables, and the processing result Yp item "etching amount" is called a target variable.

[0049] The reference process database 42 stores reference process result data, which is a combination of simulation conditions Xs and simulation results Ys in a simulation that simulates a target process. The type and content of the simulation are not limited here. In the first embodiment, the simulation content is assumed to be "electromagnetic field calculation in plasma using the finite element method," the simulation conditions are "pressure," "flow rate of gas A," "flow rate of gas B," and "power," and corresponding values ​​that are actually implemented, and the simulation results are "amount of A ions" and "amount of B ions," but the reference process database contains many more explanatory variables and objective variables.

[0050] In this way, the explanatory variables and their numbers do not need to match between the processing conditions Xp in the target process database 41 and the simulation conditions Xs in the reference process database 42, and the objective variables and their numbers do not need to match between the processing results Yp and the simulation results Ys. In the first embodiment, the explanatory variable items of Xs are a subset of the explanatory variables of Xp. A typical transfer learning model 50 using a neural network in such a case is shown in FIG.

[0051] In the example shown in FIG. 5, a reference model 51, as enclosed by a dashed line, is included in the transfer learning model 50, and in training the transfer learning model 50, the weights of this reference model 51 can be fixed or retrained (fine-tuned) as initial values.

[0052] In FIG. 5, the output of the reference model 51 is the amount of A ions (A + )511 and the amount of B ions (B + ) 512, but the type and number can be freely changed based on the knowledge of the user who operates the processing equipment to match the objective variable (here, “etching amount”) 52 of the target process for which the user ultimately wants to improve the prediction accuracy.

[0053] For example, in this target process, the user assumes the phenomenon that "A ions and B ions are generated from gas A and gas B using electricity, and these ions etch the wafer," and therefore it is believed that the "etching amount" can be predicted with high accuracy by setting the "amount of A ions" and "amount of B ions" as the output.

[0054] In the first embodiment, since the reference process data is based on a simulation, it is possible to relatively freely assign values ​​to explanatory variables without worrying about safety device constraints and interlocks, cost conditions, etc. (For example, high voltage conditions that exceed the withstand voltage performance of the device, low temperature conditions using a cooling function that ignores costs, etc.) For this reason, the reference process database 42 can include a large amount of data in which various parameters are comprehensively assigned.

[0055] Although it is possible to construct a transfer learning model using all the reference process data stored in the reference process database 42, here we consider using a more accurate model that is more specialized for the target process desired by the user. By selecting the data group used for transfer learning from the reference process data group stored in the reference process database 42 based on appropriate judgment, a transfer learning model with higher prediction accuracy can be constructed.

[0056] FIG. 6 is an example of a GUI (ROI data selection manager) 430 that the model explanation unit 43 provides to the user. This GUI 430 is displayed on the screen of the output unit 49. The model explanation unit 43 can display the features of the model on the GUI 430 by the XAI (Explainable AI) method selected and set by the XAI setting button 437 for the reference model created from the reference process data stored in the reference process database 42. There are various methods for XAI, but here, the PFI value of the reference model is calculated by the PFI (Permutation Feature Importance) method, and the values ​​are displayed in the GUI 430 in a ranking manner using bar graphs 433 and 434. In the first embodiment, the simulation condition Xs 4330 of the reference process database 42 has four parameters, namely, "pressure" 4331, "flow rate of gas A" 4332, "flow rate of gas B" 4333, and "power" 4334, so that the PFI ranking of the four elements is displayed.

[0057] The PFI value is expressed as the ratio of how much each explanatory variable contributes to the prediction accuracy of the model. This PFI value is greatly affected by the network structure of the model, and especially the data set used for learning.

[0058] While looking at the positions and distribution of data points 4321 in the data space in the graph 432 on the left side of the “ROI Data Selection Manager Model Description” window 431 in Figure 6, the user can select a data set to be used for training a reference model to be used in transfer learning in any manner, such as by creating new reference data by clicking the “Create new reference data” button 435 or setting detailed conditions for model selection by clicking the “Model detailed settings” button 436.

[0059] Graph 432 in Fig. 6 shows that 121 learning data sets for reference model 4322 are selected by ROI rectangular selection in the two-dimensional data distribution for "electricity" 4324 and "pressure" 4323. The PFI value output here takes some time to calculate depending on the amount of data, but the user can continue working by selecting the second ROI while waiting for the calculation.

[0060] Using a GUI 430 such as that shown in FIG. 6, the user can check "what kind of model will be obtained by transfer learning when what kind of data is selected," while optimizing the reference model used in transfer learning in the transfer learning execution unit 44. However, it is not necessary to display the GUI 430 in this manner to have the user make a decision; a certain level of accuracy can be achieved even if transfer learning is performed automatically using all data accumulated in the reference process database 42, and so the GUI 430 is not essential.

[0061] Furthermore, if a judgment criterion is set in advance, for example, using a PFI value, it is possible for the model explanation unit 43 to automatically optimize the reference model used in transfer learning without user operation.

[0062] However, it should be noted that the PFI value explained by the model explanation unit 43 in the first embodiment is merely "how much each explanatory variable contributes to the prediction accuracy of the reference model predicting the amount of A ions and the amount of B ions", and is not essentially "how much each explanatory variable contributes to the determination of the amount of A ions and the amount of B ions". In addition, the user can arbitrarily set whether "the amount of A ions" and "the amount of B ions" of the reference model output are useful for predicting the "etching amount" (FIG. 5). In other words, it should be noted that it cannot be said with certainty that "if the prediction accuracy of the reference model is high, the prediction accuracy of the target model is high". However, if the model explanation unit 43 can be used while taking these points into consideration, it is possible to optimize the reference model used for transfer learning with high accuracy in a short time.

[0063] Finally, the user presses the “Perform Transfer” button 438 at the bottom right of FIG. 6, whereby the transfer learning execution unit 44 executes transfer learning.

[0064] The transfer learning model evaluation unit 45 evaluates the model created by the transfer learning execution unit 44, and if the evaluation result does not meet a certain standard, it determines that the cause is the network structure of the model and the reference process data, and commands the reference process data acquisition automatic execution unit 46 to automatically acquire and add the reference process data.

[0065] When the automatic reference process data acquisition execution unit 46 executes automatic acquisition and addition of reference process data and adds new reference process data to the reference process database 42, the data is again passed through the model explanation unit 43 and the transfer learning execution unit 44 and judged by the transfer learning model evaluation unit 45. Thereafter, this loop is repeated until the judgment criterion of the transfer learning model evaluation unit 45 is satisfied.

[0066] Basically, the more reference process data there is, the more improvement in prediction accuracy can be expected. Therefore, it is preferable that the reference process data acquisition automatic execution unit 46 continues to perform calculations under simulation conditions according to the design of experiments (DoE) and continue to accumulate data even when the transfer learning model evaluation unit 45 has not instructed it to automatically acquire data.

[0067] 7 is an example of a GUI (model optimization termination criterion setting) 450 provided to the user by the transfer learning model evaluation unit 45. First, the user sets the reference process data acquisition automatic execution in a reference process data acquisition automatic execution area 451 of the GUI 450. By selecting either an enable button 4511, a manual setting button 4512, or an disable button 4513, the user specifies whether or not to run a loop for improving the transfer learning model by adding reference process data using the reference process data acquisition automatic execution unit 46. At this time, rather than leaving it to automatic simulation conditions according to the design of experiments (DoE) proposed by the reference process data acquisition automatic execution unit 46, the user can manually specify the conditions themselves.

[0068] When automatic execution of reference process data acquisition is enabled by clicking the Enable button 4511, an end judgment criterion is set in the end judgment criterion setting area 452 of the GUI 450. If an end time is entered in the end time setting area 4531 and the end time is set by clicking the "End time set" button 4521, automatic execution of reference process data acquisition is repeated until the end time, even if the set criterion is not met, and the transfer learning model with the best verification results is sent to the processing condition prediction unit 47. If the set criterion is met, the transfer learning model is sent to the processing condition prediction unit 47 before the end time is reached.

[0069] The termination criterion set in the termination criterion setting area 452 in FIG. 7 will be described. (1) "Test data verification" is a verification method in which a model is evaluated using test data, which is a combination of processing conditions Xp and processing results Yp for several target processes prepared in advance by the user. This test data must not be data included in the target process database used for learning the model, and must be prepared separately, but it is the most appropriate model evaluation method. For example, in a model predicting "etching amount," the judgment condition is "relative error between actual etching amount and predicted etching amount verified with test data is <5%." The specified test data is selected by entering the name of the validation data set in the validation data set name input area 4532 and clicking the "Test validation data" button 4522.

[0070] (2) "XAI" is a verification method that uses the value obtained as a result of evaluating a model using the XAI method to make a judgment. For example, the PFI method is used for a transfer learning model, and a judgment is made based on whether the obtained PFI value satisfies a condition such as a certain value or above / below. For example, if a user has chemical and physical knowledge about the target process and thinks that "in this process, the influence of "power" should be greater than that of "pressure" in determining the "etching amount," the judgment condition is set to "PFI value of power > PFI value of pressure." By setting the verification condition (judgment condition) in the detailed setting area 4533 and clicking the "XAI" button 4523, the set verification condition (judgment condition) is applied and the evaluation result is judged.

[0071] (3) "Cross-validation" here refers to K-fold cross-validation. The entire training data used for learning is divided into K parts, one of which is taken as test data, and the rest is used as training data and evaluated in the same way as in (1). Similarly, a total of K evaluations are performed so that each of the K divided training data groups is used as test data once, and the average of the K evaluations is used to set a judgment criterion in the same format as in (1). The accuracy of the evaluation method is somewhat inferior to that of (1) due to the reduced training data, and the amount of calculations increases and the evaluation time is extended, but the user does not need to prepare test data in advance. The conditions for cross-validation are set by setting the conditions in the validation condition setting area 4534 and clicking the "Cross-validation" button 4524.

[0072] (4) "Show details each time" is an option for users with more knowledge of transfer learning methods to check not only the XAI evaluation results and cross-validation results of the transfer learning model, but also the learning curve and parameter tuning results in detail and make their own judgment each time. Clicking button 4525 switches to a setting screen not shown in the figure, where the user can set the details.

[0073] When the "No end time setting (only once)" button 4526 is clicked, the end time is not set and the model optimization process is executed until the judgment criterion of the transfer learning model evaluation unit 45 is satisfied.

[0074] Finally, when the "Decid" button 454 is clicked, each condition set on the screen of the GUI 450 is sent to the processing condition search system 40 and set in the processing condition search system 40 as new conditions.

[0075] When using the processing condition search system 40 according to this embodiment, the user first inputs and specifies the type of processing result that is desired in the target process in the target processing result setting unit 48. For example, "etching amount" can be specified as "40 nm." There is no problem with operation even if there are multiple items, but higher accuracy can be expected with fewer items. The desired processing result can also be specified as a range, such as "30 nm to 50 nm."

[0076] The target processing result designated by the user is taken in by the processing condition prediction unit 47 after the transfer learning model that satisfies the criteria of the transfer learning model evaluation unit 45 is sent to the processing condition prediction unit 47. The processing condition prediction unit 47 optimizes the processing conditions that produce a predicted processing result closest to the target processing result set by the target processing result setting unit 48, using a root-finding algorithm such as Newton's method. The optimized processing conditions are provided to the user by means of GUI display on the screen of the output unit 49, or by means of saving in a CSV file.

[0077] FIG. 8 is a flowchart for explaining S1 to S11 from the start of a user's operation to the prediction of a target processing condition in the first embodiment.

[0078] S1: Set learning data already acquired for the target device for which the target processing conditions are to be predicted and stored in the target process database 41. If the transfer learning model evaluation unit 45 wants to perform "test data verification" as the termination criterion, set separate test data at this time.

[0079] S2: A target process result to be achieved by the target device is set in the target process result setting unit 48.

[0080] S3: The features of the latest reference model created by learning based on the reference process database are confirmed by the model explanation unit 43 using several XAI methods. When proceeding from S2 to S3, the model that can be confirmed is a reference model learned based on either (1) all reference process data, (2) reference process data selected in advance, or (3) reference process data selected at the previous time of use. When returning from S4 to S3, the learning reference process data used to learn the reference model can be selected on a screen for creating new reference data (not shown) by clicking the "Create new reference data" button 435 of the GUI 430 as shown in FIG. 6. Examples of XAI methods that can confirm the features of the model and learning data at this point include, but are not limited to, PFI (Permutation Feature Importance), SHAP (Shapley Additive exPlanation), PD (Partial Dependence), and ICE (Individual Conditional Expectation).

[0081] S4: Determine whether the PFI ranking obtained in S3 is appropriate for the value set in S2. If Yes, proceed to S5, and if No, return to S3.

[0082] S5: Perform transfer learning and output the transfer learning model.

[0083] S6: In the GUI shown in FIG. 7, in the model optimization termination criteria setting of the transfer learning model evaluation unit, check whether "End time set" is set.

[0084] S7: If "end time set" is set (Yes in S6), it is determined whether the end time has been reached.

[0085] S8: If the end time has been reached (Yes in S7), the processing condition prediction unit 47 outputs processing conditions that are expected to produce a predicted processing result closest to the target processing result. At this point, the series of user operations ends.

[0086] S9: If the end time has not been reached (No in S7), the transfer learning model evaluation unit 45 evaluates the accuracy of the model. In this embodiment, since "cross validation" 4525 is set in the end criterion setting area 452 of the transfer learning model evaluation unit 45 in the GUI 450 shown in Fig. 7, it is determined whether the cross validation result of the model exceeds the threshold set by the user in the target processing result setting unit 48. If the accuracy is equal to or exceeds the threshold set by the user (Yes in S9), proceed to S8; if not (No in S9), proceed to S10.

[0087] S10: In the reference process data acquisition automatic execution unit 46, new reference process data is calculated by DoE or user definition and added to the reference process database 42. Also, here, unlike the processing flow in FIG. 9 described later, by selecting "XAI" 4523 in the termination criteria setting area 452 of the transfer learning model evaluation unit in the GUI 450, a suggestion of the data space to be expanded by the XAI method can be obtained. For example, if the user has knowledge that "Gas A has a large effect on the etching amount," but the PFI value of Gas A calculated by the PFI method is small, it is useful to try to obtain reference process data in a data space with a focus on the parameters of Gas A.

[0088] S11: Using a new learning dataset with new reference processing data added, retrain the reference model. Proceed to S3 again based on the obtained model.

[0089] As described above, in this embodiment, the characteristics of the model are evaluated in advance by the "model explanation unit" to prevent negative transfer, and the model obtained as a result of transfer learning is evaluated by the transfer learning model evaluation unit. If the evaluation value does not exceed the threshold as a result of the model evaluation, simulation data of the conditions required to improve the accuracy of the transfer learning model is automatically generated by the reference process data acquisition automatic execution unit, and transfer learning is performed again.

[0090] This allows the optimal transfer learning model to be constantly and automatically constructed and updated to predict the target processing results set by the user, and makes it possible to shorten and reduce the time required for recipe optimization to reduce machine / component differences by utilizing large amounts of training data generated by simulations that are less expensive than actual processing.

[0091] According to the present invention, in a search system that uses machine learning to search for desired manufacturing conditions for semiconductor manufacturing equipment, generally, physical simulations are unable to take into account all of the parameters under actual processing conditions, and therefore conventional machine learning using neural networks was unable to learn task data with different features and labels using a single model. However, it has become possible to predict the desired manufacturing conditions for semiconductor manufacturing equipment using a model constructed with a network structure that uses transfer learning of data from a physical simulator. EXAMPLES

[0092] A second embodiment of the present invention will be described with reference to FIG. In this embodiment, in addition to the processing described in the first embodiment, during a period in which the user does not operate the device or method such as S1 to S3 described in the first embodiment using FIG. 8, the processing condition search system 40 automatically performs processing to expand the reference process database as shown in the flowchart in FIG. 9.

[0093] The procedure of the process of expanding the reference process database according to this embodiment will be described with reference to the flowchart shown in FIG.

[0094] S91: Since user operations always take priority, it is confirmed whether there is any user operation. That is, it is checked whether the "Execute Transfer" button 438 shown in Fig. 6 or the "Decid" button 454 shown in Fig. 7 has been pressed, and if Yes (if a user operation has been performed / is expected), it proceeds to the user operation processing described in the first embodiment with reference to Fig. 8, and executes steps S1 to S11. If No, it proceeds to S92.

[0095] S92: Calculate new reference process data according to DoE or user definition and add it to the reference process database.

[0096] S93: Every time reference processing data is added to the database, the reference model is trained using training data including the newly added reference processing data, that is, model training is performed using the entire reference processing data.

[0097] S94: The trained reference model is evaluated using various XAI methods (model interpretation calculation). The evaluation results and the trained model are saved in the system, and the user can load them at the time of the processing in S3 described in Figure 8.

[0098] According to this embodiment, in addition to the effects described in the first embodiment, the computer can automatically expand the reference process database during the period when the user does not operate the equipment and the method, so that the accuracy of the transfer learning model can be further improved, and it is possible to further shorten the recipe optimization period for reducing machine differences / component differences by utilizing a large amount of training data by simulation. The inventions according to the first and second embodiments can also be implemented as an application implemented on a platform. The platform is constructed on a cloud, and an application that executes processing on an OS and middleware runs on it. A user can access the platform from a terminal via a network and use the functions of the application constructed on the platform. The platform has a database, and stores data required for executing the application. Furthermore, semiconductor manufacturing equipment is also connected to the platform via a network to enable data exchange.

[0099] Although the invention made by the present inventor has been specifically described above based on the embodiments, it goes without saying that the present invention is not limited to the above embodiments and can be modified in various ways without departing from the gist of the invention. In other words, the present invention also includes those in which some of the configurations (steps) described in the above embodiments are replaced with steps or means having equivalent functions, or those in which some of the non-essential functions are omitted. [Explanation of symbols]

[0100] 40...processing condition search system, 41...target process database, 42...reference process database, 43...model explanation unit, 44...transfer learning execution unit, 45...transfer learning model evaluation unit, 46...reference process data acquisition automatic execution unit, 47...processing condition prediction unit, 48...target processing result setting unit, 49...output unit, 51...reference model, 430, 450...GUI, 451...reference process data acquisition automatic execution area, 452...termination judgment criterion setting area.

Claims

1. 1. A search device for searching for manufacturing conditions corresponding to a desired processing result of a semiconductor manufacturing device by predicting the manufacturing conditions corresponding to the desired processing result using a learning model, comprising: A learning model is generated by transfer learning using the first data and the second data; If the generated learning model does not satisfy a predetermined judgment criterion, a learning model is regenerated by transfer learning using the first data and the added second data; The learning model includes a reference model, A searching device characterized in that the reference model is a model generated based on explanatory variables of the second data and objective variables of the second data.

2. 2. The search device according to claim 1, the first data includes combination data of manufacturing conditions of the semiconductor manufacturing equipment and processing results under the manufacturing conditions of the semiconductor manufacturing equipment; The search device, wherein the second data includes data obtained by a simulation.

3. 2. The search device according to claim 1, A search device characterized in that, when transfer learning of the learning model is performed, the weights of the reference model are fixed, or the weights of the reference model are re-learned as initial values.

4. 2. The search device according to claim 1, A search device characterized in that the interpretation results of the reference model using a machine learning model interpretation method including PFI or SHAP are displayed on a user interface.

5. 2. The search device according to claim 1, A search device, characterized in that the first data and the second data are different in terms of the types or numbers of explanatory variables, or are in an inclusive relationship.

6. 2. The search device according to claim 1, A search device characterized in that the position and variance of the data space of the data group used for the transfer learning in the second data are displayed on a user interface.

7. A semiconductor device manufacturing system including a platform on which an application for predicting, by using a learning model, manufacturing conditions corresponding to desired processing results of the semiconductor manufacturing device is implemented, the system including: A step of generating a learning model by transfer learning using the first data and the second data; If the generated learning model does not satisfy a predetermined criterion, the application executes a step of regenerating a learning model by transfer learning using the first data and the added second data; The learning model includes a reference model, The semiconductor device manufacturing system according to claim 1, wherein the reference model is a model generated based on explanatory variables of the second data and objective variables of the second data.

8. 1. A method for searching for manufacturing conditions corresponding to a desired processing result of a semiconductor manufacturing apparatus by predicting the manufacturing conditions corresponding to the desired processing result using a learning model, comprising: generating a learning model by transfer learning using the first data and the second data; If the generated learning model does not satisfy a predetermined criterion, regenerating a learning model by transfer learning using the first data and the added second data; The learning model includes a reference model, A search method characterized in that the reference model is a model generated based on explanatory variables of the second data and target variables of the second data.

Images