Learning model generation method, learning model generation device, and learning model generation program

Abstract

A forward-direction model for predicting a processing result for a substrate from processing conditions pertaining to the substrate is generated on the basis of a plurality of actual datasets. Using the forward-direction model, a pseudo processing result for the substrate is predicted from individual pseudo processing conditions pertaining to a plurality of substrates, whereby a plurality of pseudo datasets are generated. A reverse-direction model for predicting a processing condition pertaining to the substrate from a processing result for the substrate is generated on the basis of at least some of the plurality of pseudo datasets.

Description

Learning model generation method, learning model generation device, and learning model generation program

[0001] The present invention relates to a learning model generation method, a learning model generation device, and a learning model generation program for generating a learning model.

[0002] In a substrate processing apparatus, a processing solution is supplied to a substrate under set processing conditions, thereby performing processing such as etching or cleaning on the substrate. In recent years, a trained model generated by machine learning has been used to set the processing conditions for the substrate. For example, Patent Literature 1 describes a substrate system including a substrate processing apparatus and a setting system, which is provided together with a learning device.

[0003] The memory unit of the learning device stores a learning data set in which multiple learning process conditions are associated with multiple process results. The learning process conditions include a number of parameters, such as the concentration of the process liquid, the temperature of the process liquid, the amount of the process liquid supplied, or the rotation speed of the substrate. The process results include, for example, the uniformity of the amount of processing within the surface of the substrate. The learning device finds certain rules in the learning data set and generates a trained model that expresses those rules.

[0004] In the substrate processing apparatus, a plurality of processing conditions are input to the generated trained model, whereby a plurality of processing results are acquired, and an image based on the acquired plurality of processing results is displayed on the display unit. A user checks the image displayed on the display unit and selects one processing condition corresponding to one of the plurality of processing results. The selected processing condition is set in the setting system as the execution processing condition for processing the substrate.

[0005] JP 2023-120961 A

[0006] As disclosed in Patent Document 1, it is relatively easy to generate a trained model (hereinafter referred to as a forward model) that predicts processing results from processing conditions. However, because processing conditions include a large number of parameters as described above, it is more difficult to generate a trained model (hereinafter referred to as a backward model) that predicts processing conditions from processing results than to generate a forward model. Therefore, it is desirable to develop a trained model generation method, a trained model generation device, and a trained model generation program that can easily generate a backward model.

[0007] An object of the present invention is to provide a learning model generation method, a learning model generation device, and a learning model generation program that can easily generate a trained model that predicts processing conditions from processing results.

[0008] A learning model generation method according to one aspect of the present invention includes generating a forward model that predicts a substrate processing result from the substrate processing conditions based on a plurality of actual data sets that indicate the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing result; generating a plurality of pseudo data sets that indicate the relationship between the pseudo substrate processing conditions and the pseudo substrate processing result by using the forward model to predict a pseudo substrate processing result from each of the plurality of pseudo substrate processing conditions; and generating a reverse model that predicts the substrate processing conditions from the substrate processing result based on at least a portion of the plurality of pseudo data sets, and is executed by a computer.

[0009] A learning model generation device according to another aspect of the present invention includes a forward model generation unit that generates a forward model that predicts a substrate processing result from the substrate processing conditions based on a plurality of actual data sets that indicate the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing result; a pseudo data generation unit that generates a plurality of pseudo data sets that indicate the relationship between the substrate processing conditions and the pseudo substrate processing result by predicting a pseudo substrate processing result from each of the plurality of pseudo substrate processing conditions using the forward model generated by the forward model generation unit; and an inverse model generation unit that generates an inverse model that predicts a substrate processing condition from the substrate processing result based on at least a portion of the plurality of pseudo data sets generated by the pseudo data generation unit.

[0010] A learning model generation program according to yet another aspect of the present invention causes a computer to perform the following steps: generating a forward model that predicts a substrate processing result from the substrate processing conditions based on a plurality of actual data sets that indicate the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing result; generating a plurality of pseudo data sets that indicate the relationship between the substrate's pseudo processing conditions and the substrate's pseudo processing result by using the forward model to predict a pseudo substrate processing result from each of the plurality of pseudo substrate processing conditions; and generating a reverse model that predicts a substrate's processing conditions from the substrate processing result based on at least a portion of the plurality of pseudo data sets.

[0011] According to the present invention, it is possible to easily generate a trained model that predicts processing conditions from processing results.

[0012] FIG. 1 is a diagram showing an example of the configuration of a substrate processing system including a learning model generation device according to a first embodiment of the present invention. FIG. 2 is a schematic cross-sectional view showing the configuration of the substrate processing device of FIG. 1. FIG. 3 is a diagram for explaining the operation of the substrate processing device. FIG. 4 is a waveform showing fluctuations in nozzle position in a preliminary experiment. FIG. 5 is a block diagram showing the functional configuration of the learning model generation device. FIG. 6 is a flowchart showing an example of the flow of a learning model generation process. FIG. 7 is a flowchart showing an example of the flow of a learning model generation process in a second embodiment. FIG. 8 is a block diagram showing the functional configuration of a learning model generation device according to a third embodiment. FIG. 9 is a flowchart showing an example of the flow of a learning model generation process in the third embodiment.

[0013] 1. First Embodiment (1) Substrate Processing System A learning model generation method, a learning model generation device, and a learning model generation program according to embodiments of the present invention will be described below with reference to the drawings. In the following description, a substrate refers to a semiconductor substrate (wafer), a substrate for an FPD (Flat Panel Display) such as a liquid crystal display device or an organic EL (Electro Luminescence) display device, a substrate for an optical disk, a substrate for a magnetic disk, a substrate for a magneto-optical disk, a substrate for a photomask, a ceramic substrate, or a substrate for a solar cell, etc.

[0014] Fig. 1 is a diagram showing an example of the configuration of a substrate processing system including a learning model generation device according to a first embodiment of the present invention. As shown in Fig. 1, a substrate processing system 300 includes a learning model generation device 100 and a substrate processing device 200. The learning model generation device 100 generates a trained model such as an inverse model that predicts substrate processing conditions from the substrate processing results in the substrate processing device 200. Fig. 1 mainly shows the hardware configuration of the learning model generation device 100.

[0015] Specifically, the learning model generation device 100 includes a CPU (Central Processing Unit) 110, a RAM (Random Access Memory) 120, a ROM (Read Only Memory) 130, a storage device 140, an operation unit 150, a display device 160, an input / output I / F (Interface) 170, and a bus 180. The learning model generation device 100 may be configured, for example, by a computer. The CPU 110, the RAM 120, the ROM 130, the storage device 140, the operation unit 150, the display device 160, and the input / output I / F 170 are connected to the bus 180.

[0016] The RAM 120 is used as a working area for the CPU 110. The ROM 130 stores a system program. The storage device 140 includes a storage medium such as a hard disk or a semiconductor memory, and stores a learning model generation program for executing a learning model generation process that generates the above-described trained model. The learning model generation process will be described in detail later.

[0017] The learning model generation program may be stored in ROM 130. Alternatively, the learning model generation program may be provided in a form stored in a computer-readable storage medium 190 such as a CD (compact disc)-ROM or an SD memory card, and may be installed in storage device 140 or ROM 130 by being read out via input / output I / F 170 or the like. Furthermore, when input / output I / F 170 is connected to a communication network, the learning model generation program distributed from a server connected to the communication network may be installed in storage device 140 or ROM 130.

[0018] The operation unit 150 is an input device such as a keyboard, a mouse, or a touch panel. By operating the operation unit 150, a user can store an actual data set, etc., described later, in the storage device 140. The display device 160 is a display device such as a liquid crystal display device, and displays a GUI (Graphical User Interface) or the like for accepting operations by the user. The input / output I / F 170 is connected to the substrate processing apparatus 200. This makes it possible to obtain information indicating substrate processing conditions and the like from the substrate processing apparatus 200 and store the information in the storage device 140 or the like when a preliminary experiment, described later, is performed in the substrate processing apparatus 200.

[0019] (2) Substrate Processing Apparatus The substrate processing apparatus 200 performs substrate processing by supplying a processing liquid to a substrate under set processing conditions. In this example, the substrate processing is etching, and the processing liquid is an etching liquid. Figure 2 is a schematic cross-sectional view showing the configuration of the substrate processing apparatus 200 of Figure 1. As shown in Figure 2, the substrate processing apparatus 200 includes a spin motor 210, a spin chuck 220, a processing cup 230, a processing liquid supply unit 240, and a control device 250.

[0020] The spin motor 210 has a rotation shaft 211 and is disposed at the bottom of the substrate processing apparatus 200 with the rotation shaft 211 facing upward. The spin chuck 220 is attached to the upper end of the rotation shaft 211 of the spin motor 210 and holds the substrate W in a horizontal position. The spin chuck 220 rotates the substrate W around a vertical axis as the rotation shaft 211 of the spin motor 210 is driven to rotate.

[0021] The processing cup 230 is disposed to surround the periphery of the substrate W held by the spin chuck 220, and its inner wall receives the processing liquid splashed from the substrate W. An opening 231 is formed in the upper part of the processing cup 230. The upper surface of the substrate W is exposed upward through the opening 231. This allows processing to be performed on the upper surface of the substrate W from above. A waste liquid port 232 and an exhaust port 233 are formed in the lower part of the processing cup 230. The waste liquid port 232 guides the processing liquid received by the inner wall of the processing cup 230 to a waste liquid facility (not shown). The exhaust port 233 is connected to an exhaust facility in a factory and exhausts the atmosphere inside the processing cup 230 to the exhaust facility.

[0022] The processing liquid supply unit 240 includes a rotary support shaft 241, an arm 242, a nozzle 243, and a drive unit 244. The rotary support shaft 241 extends vertically outside the processing cup 230 and is rotatably supported by the drive unit 244. The arm 242 is provided to extend horizontally from the upper end of the rotary support shaft 241. The nozzle 243 has an outlet that can discharge the processing liquid. The nozzle 243 is attached to the tip of the arm 242 so that the outlet faces downward. The drive unit 244 includes a pulse motor, an air cylinder, or the like, and rotates the rotary support shaft 241.

[0023] The control device 250 includes, for example, a CPU, and controls the operation of the spin motor 210 and the drive unit 244 in accordance with set processing conditions. Fig. 3 is a diagram for explaining the operation of the substrate processing apparatus 200. Specifically, the control device 250 controls the operation of the spin motor 210 as indicated by arrow A in Fig. 3 to rotate the substrate W held by the spin chuck 220. Furthermore, the control device 250 controls the operation of the drive unit 244 as indicated by arrow B in Fig. 3 to reciprocate the nozzle 243 in an arc within a horizontal plane.

[0024] 3, the nozzle 243 moves from the initial position Pa to the turning position Pb and then returns to the initial position Pa. This reciprocating motion of the nozzle 243 is repeated. Here, the distance between the initial position Pa and the turning position P2 is greater than the radius of the substrate W, and the initial position Pa substantially overlaps with the outer peripheral edge of the substrate W in a planar view. Therefore, in a planar view, the nozzle 243 reciprocates across a portion of the substrate W with a length equal to at least the radius of the substrate W. In this case, as the substrate W rotates, the treatment liquid is supplied to the entire upper surface of the substrate W. As a result, the upper surface of the substrate W is treated.

[0025] Furthermore, in the substrate processing apparatus 200, a preliminary experiment is performed to generate a plurality of actual data sets. Each actual data set indicates the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing results, and is used to generate a forward model. The forward model is a trained model that predicts the processing results of the substrate W from the processing conditions of the substrate W. In the preliminary experiment, the above substrate processing is performed under various processing conditions, and information indicating the processing conditions is generated. In this example, the information indicating the processing conditions includes a waveform indicating fluctuations in the position of the nozzle 243.

[0026] Fig. 4 is an example of a waveform showing fluctuations in the position of the nozzle 243 in a preliminary experiment. As shown in Fig. 4, during substrate processing, the nozzle 243 reciprocates between an initial position Pa and a turning position Pb in a cycle of 2T. The user specifies the position of the nozzle 243 at multiple locations. For example, as shown in Fig. 4, the user specifies positions P2 to P11, which divide the area between positions P1 and P11 into 10 equal parts. In this example, the position of the nozzle 243 at each time (e.g., positions P2 to P11) becomes a feature quantity indicating the processing conditions for the substrate W.

[0027] The user also evaluates the processing results of the substrate W under each processing condition. In this example, the processing result is the etching amount of the substrate W, and is evaluated, for example, by image analysis of the substrate W. The user generates an actual data set by associating the identified processing conditions of the substrate W with the processing results of the substrate W under those processing conditions. A plurality of actual data sets corresponding to the plurality of processing conditions are generated and stored in the storage device 140 of the learning model generation device 100. In this example, approximately 100 actual data sets are stored in the storage device 140. The plurality of actual data sets may be stored in a storage device different from the storage device 140.

[0028] (3) Learning Model Generation Device Fig. 5 is a block diagram showing the functional configuration of the learning model generation device 100. As shown in Fig. 5, the learning model generation device 100 includes, as functional units 10, a forward model generation unit 11, a pseudo data generation unit 12, and a backward model generation unit 13. The functional units 10 of the learning model generation device 100 are realized by the CPU 110 in Fig. 1 executing a learning model generation program. Some or all of the functional units 10 of the learning model generation device 100 may be realized by hardware such as electronic circuits.

[0029] The learning model generation device 100 is not limited to a configuration using only one CPU 110 as shown in FIG. 5 . A configuration using multiple processors, such as a configuration using multiple CPUs, can also be adopted. As the processor, in addition to a CPU, an MPU (Micro Processing Unit), a GPU (Graphics Processing Unit), a DSP (Digital Signal Processor), etc. can also be adopted. A combination of multiple types of processors can also be used. For example, with regard to the components shown in FIG. 5 , some components and the remaining components can be realized by different processors. Furthermore, a configuration including an FPGA (Field-Programmable Gate Array) or an ASIC (Application Specific Integrated Circuit) can also be adopted.

[0030] The forward model generator 11 acquires a plurality of actual data sets stored in the storage device 140 or the like. The forward model generator 11 also generates a forward model based on the acquired plurality of actual data sets. For example, to facilitate understanding, it is assumed that the processing conditions for the substrate W in each actual data set are expressed as X in a vector system. ０ = (x １ ０ , ..., x ｉ ０ ) where i is an integer indicating the number of parameters included in the processing conditions. In addition, the processing result of the substrate W in the actual data set is given by Y ０ In this case, if the forward model is T, then X ０ T=Y ０ Therefore, a forward model is generated by learning each real data set and searching for T so that the above equation holds.

[0031] The pseudo data generating unit 12 acquires pseudo processing conditions for the substrate W based on the processing conditions of the substrate W. In this example, the pseudo processing conditions for the substrate W are acquired by performing calculations on any of the parameters in the processing conditions for the substrate W. For example, X ０ = (x １ ０ , ..., x ｉ ０ ) by adding or subtracting a small shift amount to any of the components of the pseudo processing conditions X １ = (x １ １ , ..., x ｉ １ ) are generated. Depending on the magnitude of the shift amount or the component to be adjusted by the shift amount, pseudo processing conditions for a large number of substrates W can be generated. In this example, the number of pseudo processing conditions for substrates W generated is greater than the number of processing conditions for substrates W, for example, about 1000.

[0032] Furthermore, the pseudo data generation unit 12 uses the forward model generated by the forward model generation unit 11 to predict pseudo processing results for the substrate W from each of the pseudo processing conditions for the plurality of substrates W. In this way, the pseudo data generation unit 12 generates a plurality of pseudo data sets. Each pseudo data set indicates the relationship between the pseudo processing conditions for the substrate W and the pseudo processing results for the substrate W. As described above, the number of pseudo processing conditions for the substrate W is greater than the number of processing conditions for the substrate W. Therefore, the number of generated pseudo data sets is greater than the number of actual data sets obtained by actual measurement. For example, the number of pseudo data sets is approximately 1,000, and the number of actual data sets is approximately 100.

[0033] The inverse model generator 13 generates an inverse model based on a plurality of actual data sets generated by the forward model generator 11 and a plurality of pseudo data sets generated by the pseudo data generator 12. For example, if the pseudo processing result of the substrate W predicted by the pseudo data generator 12 is Y １ In this case, if the inverse model is R, then (Y ０ +Y １ ) R=X ０ +X １ Therefore, each real data set and each pseudo data set are learned, and R is searched for so that the above equation holds, thereby generating an inverse model.

[0034] Here, the relationship between the processing conditions for the substrate W and the processing results for the substrate W in each real data set is more likely to be accurate than the relationship between the pseudo processing conditions for the substrate W and the pseudo processing results for the substrate W in each pseudo data set. Therefore, in generating the inverse model, the inverse model generation unit 13 may train the inverse model with a weight that is greater for the real data set than for the pseudo data set.

[0035] (4) Learning Model Generation Process Fig. 6 is a flowchart showing an example of the flow of the learning model generation process. The learning model generation process is executed by the CPU 110 of the learning model generation device 100 as the CPU 110 executes a learning model generation program. The learning model generation process of Fig. 6 will be described below with reference to the learning model generation device 100 of Fig. 5.

[0036] First, the forward model generator 11 acquires a plurality of actual data sets from the storage device 140 or the like (step S1). Next, the forward model generator 11 generates a forward model based on the plurality of actual data sets acquired in step S1 (step S2). Furthermore, the pseudo data generator 12 acquires pseudo processing conditions for a plurality of substrates W based on the processing conditions for any of the substrates W in the plurality of actual data sets acquired in step S1 (step S3). Step S3 may be executed before step S2 or in parallel with step S2.

[0037] Next, the pseudo data generator 12 uses the forward model generated in step S2 to predict pseudo processing results for the plurality of substrates W from the pseudo processing conditions for the plurality of substrates W acquired in step S3 (step S4). Thereafter, the pseudo data generator 12 generates a plurality of pseudo data sets by associating the pseudo processing conditions for the plurality of substrates W acquired in step S3 with the pseudo processing results for the plurality of substrates W predicted in step S4 (step S5).

[0038] Finally, the inverse model generation unit 13 generates an inverse model based on the real data sets acquired in step S1 and the pseudo data sets generated in step S5 (step S6), thereby completing the learning model generation process.

[0039] (5) Effects In the learning model generation device 100 according to this embodiment, multiple pseudo data sets are generated by the pseudo data generation unit 12 using a forward model generated by the forward model generation unit 11 based on multiple actual data sets. Therefore, even if the number of actual data sets obtained by actual measurement is small, a sufficient number of pseudo data sets are generated. In this case, the inverse model can learn a sufficient number of pseudo data sets. Therefore, even if the processing conditions include a large number of parameters, the inverse model generation unit 13 can easily generate an inverse model, which is a trained model that estimates the processing conditions from the processing results.

[0040] In this embodiment, the inverse model generation unit 13 generates the inverse model further based on a plurality of actual data sets. In this case, the number of data sets used to generate the inverse model is further increased. Therefore, the inverse model can be generated more easily. Here, the inverse model generation unit 13 trains the inverse model with a weight greater than that of the pseudo data set. Therefore, it is possible to generate an inverse model that can more accurately predict the processing conditions of the substrate W from the processing results of the substrate W.

[0041] The pseudo data generator 12 generates a plurality of pseudo data sets, the number of which is greater than the number of actual data sets obtained by actual measurement. In this case, a large number of pseudo data sets are generated. This makes it possible to more easily generate an inverse model. Here, the pseudo data generator 12 obtains pseudo processing conditions for the substrate W by performing a calculation on any of the parameters in the processing conditions for the substrate W. In this case, pseudo processing conditions for a large number of substrates W can be easily obtained from the processing conditions for the substrate W depending on the content of the calculation and the parameters to be calculated. This makes it possible to easily generate a large number of pseudo data sets.

[0042] 2. Second Embodiment (1) Learning Model Generation Process The following describes the learning model generation process in the second embodiment, focusing on differences from the learning model generation process in the first embodiment. Figure 7 is a flowchart showing an example of the flow of the learning model generation process in the second embodiment. As shown in Figure 7, in this embodiment, steps S1 to S6 are executed, which are similar to the learning model generation process in the first embodiment.

[0043] Thereafter, the pseudo data generator 12 uses the inverse model generated in step S6 to predict pseudo processing conditions for the substrate W from the processing results for the substrate W in the multiple actual data sets acquired in step S1 and the pseudo processing results for the substrate W predicted in step S4 (step S7). Next, the pseudo data generator 12 uses the forward model generated in step S2 to predict pseudo processing results for the substrate W from the pseudo processing conditions for the substrate W predicted in step S7 (step S8).

[0044] Next, the pseudo data generator 12 generates (adds) a further pseudo data set (step S9) by associating the pseudo processing conditions for the substrate W predicted in step S7 with the pseudo processing results for the substrate W predicted in step S8. In step S9, a plurality of pseudo data sets are basically added, but one pseudo data set may also be added.

[0045] Thereafter, the inverse model generation unit 13 updates the inverse model by re-learning the pseudo data set added to the inverse model in step S9 (step S10). This completes the learning model generation process. If the inverse model is configured with multiple layers, in step S10, it is possible to re-learn only the bottom layer of the inverse model without re-learning all layers of the inverse model. This makes it easy to prevent a decrease in the performance of the inverse model even when re-learning is performed.

[0046] (2) Effects In the learning model generation device 100 according to this embodiment, a pseudo data set is added by the pseudo data generation unit 12 using the forward model generated by the forward model generation unit 11 and the backward model generated by the backward model generation unit 13. Therefore, by using the added pseudo data set, the backward model can be easily updated by the backward model generation unit 13 so as to improve the accuracy of prediction.

[0047] 3. Third Embodiment (1) Learning Model Generation Device The following describes the differences between the learning model generation device 100 according to the third embodiment and the learning model generation device 100 according to the first embodiment. Fig. 8 is a block diagram showing the functional configuration of the learning model generation device 100 according to the third embodiment. As shown in Fig. 8, the learning model generation device 100 further includes a plurality of simplex model generation units 14 as the functional unit 10.

[0048] Each simplex model generation unit 14 generates a simplex model that predicts the processing conditions of the substrate W from the processing results of the substrate W, based on the multiple actual data sets generated by the forward model generation unit 11 and the multiple pseudo data sets generated by the pseudo data generation unit 12. The multiple simplex model generation units 14 use different machine learning models or learning algorithms to generate the simplex models. Therefore, the multiple simplex models generated by the multiple simplex model generation units 14 have different characteristics.

[0049] The inverse model generation unit 13 generates an inverse model based on a plurality of simplex models generated by the plurality of simplex model generation units 14. That is, in this embodiment, the inverse model is generated by ensemble learning. In this case, the inverse model is composed of a plurality of simplex models, and by averaging or combining a plurality of prediction results output by the plurality of simplex models, it becomes possible to predict the processing conditions for the substrate W from the processing results of the substrate W more accurately than the prediction results from each simplex model.

[0050] (2) Learning Model Generation Process Fig. 9 is a flowchart showing an example of the flow of the learning model generation process in the third embodiment. As shown in Fig. 9, in this embodiment, steps S1 to S5 similar to those in the learning model generation process in the first embodiment are executed. Then, a plurality of steps S5a are executed. The plurality of steps S5a correspond to the plurality of simplex model generation units 14 in Fig. 8, respectively.

[0051] Specifically, in each step S5a, the corresponding simplex model generator 14 generates a simplex model based on the multiple real data sets acquired in step S1 and the multiple pseudo data sets generated in step S5 (step S5a). In the example of Fig. 9, multiple steps S5a are executed in parallel, but may also be executed sequentially in series.

[0052] Finally, the backward model generation unit 13 generates a backward model based on the plurality of simplex models generated in the plurality of steps S5a (step S6), thereby completing the learning model generation process.

[0053] (3) Effects The learning model generating device 100 according to this embodiment is provided with a plurality of simplex model generating units 14. A plurality of simplex models that predict the processing conditions of the substrate W from the processing results of the substrate W are generated by the plurality of simplex model generating units 14, respectively. An inverse model is generated by the inverse model generating unit 13 based on the plurality of simplex models generated by the plurality of simplex model generating units 14, respectively. In this case, an inverse model that can more accurately predict the processing conditions of the substrate W from the processing results of the substrate W can be generated.

[0054] 4. Other Embodiments (1) In the above embodiment, the substrate processing is etching, but the embodiment is not limited to this. The substrate processing may be other processing such as cleaning, application of a processing liquid, or development. Furthermore, the processing condition for the substrate W is the position of the nozzle 243, but the embodiment is not limited to this. The processing condition for the substrate W may be other parameters such as the concentration of the processing liquid, the temperature of the processing liquid, the supply amount of the processing liquid, or the rotation speed of the substrate.

[0055] (2) In the above embodiment, the inverse model is generated based on the real data set and the pseudo data set. However, the embodiment is not limited to this. The inverse model may be generated based on at least a portion of the real data set and at least a portion of the pseudo data set. Furthermore, the real data set does not have to be used to generate the inverse model. That is, the inverse model may be generated based on at least a portion of the pseudo data set.

[0056] (3) In the above embodiment, the inverse model trains the real data set with a larger weight than the pseudo data set, but the embodiment is not limited to this. If the relationship between the simulated processing conditions for the substrate W and the simulated processing results for the substrate W in each pseudo data set is as accurate as the relationship between the processing conditions for the substrate W and the processing results for the substrate W in each real data set, the inverse model does not need to train the real data set with a larger weight than the pseudo data set. In other words, the inverse model may train the real data set with a weight similar to that of the pseudo data set.

[0057] (4) In the above embodiment, the number of pseudo data sets generated is greater than the number of actual data sets obtained by actual measurements, but the embodiment is not limited to this. As long as a sufficient number of data sets can be prepared to generate an inverse model, the number of pseudo data sets generated may be approximately the same as the number of actual data sets, or may be less than the number of actual data sets.

[0058] (5) In the above embodiment, the pseudo data generating unit 12 obtains the pseudo processing conditions for the substrate W by performing calculations on any of the parameters in the processing conditions for the substrate W, but the embodiment is not limited to this. The pseudo data generating unit 12 may obtain the pseudo processing conditions for the substrate W by receiving input of the pseudo processing conditions for the substrate W from a user or the like.

[0059] (6) In the second embodiment, the pseudo processing conditions for the substrate W are predicted from the processing results for the substrate W in the plurality of actual data sets and the pseudo processing results for the substrate W in the plurality of pseudo data sets, but the embodiment is not limited to this. The pseudo processing conditions for the substrate W may be predicted from at least some of the processing results for the substrate W in the plurality of actual data sets and the pseudo processing results for the substrate W in the plurality of pseudo data sets.

[0060] (7) In the second embodiment, the inverse model is updated only once, but the embodiment is not limited to this. After step S10 of the learning model generation process, steps S7 to S10 may be repeated again, thereby updating the inverse model two or more times.

[0061] (8) The second and third embodiments may be implemented in combination. For example, in the third embodiment, the simplex models generated by each simplex model generation unit 14 may be updated in the same manner as in the second embodiment. In this case, the inverse model generation unit 13 generates an inverse model based on the simplex models updated by the simplex model generation units 14.

[0062] 5. Correspondence Between Each Component of the Claims and Each Part of the Embodiments Below, examples of the correspondence between each component of the claims and each element of the embodiments are described, but the present invention is not limited to the following examples. Various other elements having the configurations or functions described in the claims may also be used as each component of the claims. In the above embodiment, the substrate W is an example of a substrate, the forward model generation unit 11 is an example of a forward model generation unit, the pseudo data generation unit 12 is an example of a pseudo data generation unit, the inverse model generation unit 13 is an example of an inverse model generation unit, the learning model generation device 100 is an example of a learning model generation device, and the simplex model generation unit 14 is an example of a simplex model generation unit.

[0063] 6. Summary of Embodiments (Item 1) A learning model generation method according to item 1 includes: generating a forward model that predicts a substrate processing result from a substrate processing condition, based on a plurality of actual data sets that indicate the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing result; generating a plurality of pseudo data sets that indicate the relationship between the pseudo substrate processing conditions and the pseudo substrate processing result by predicting a pseudo substrate processing result from each of the plurality of pseudo substrate processing conditions using the forward model; and generating a reverse model that predicts a substrate processing condition from the substrate processing result, based on at least a portion of the plurality of pseudo data sets, and is executed by a computer.

[0064] According to this learning model generation method, multiple pseudo data sets are generated using a forward model generated based on multiple actual data sets. Therefore, even if the number of actual data sets obtained by actual measurement is small, a sufficient number of pseudo data sets are generated. In this case, the inverse model can learn from a sufficient number of pseudo data sets. Therefore, even if the processing conditions include a large number of parameters, it is possible to easily generate an inverse model, which is a trained model that estimates processing conditions from processing results.

[0065] (2) In the learning model generation method described in 1, generating the inverse model may include generating the inverse model further based on at least a portion of the plurality of actual data sets.

[0066] In this case, the number of data sets used to generate the inverse model is further increased, which makes it easier to generate the inverse model.

[0067] (3) In the learning model generation method described in paragraph 2, generating the inverse model may include training the inverse model on the real data set with a weight greater than that on the pseudo data set.

[0068] In this case, an inverse model can be generated that can more accurately predict the processing conditions for a substrate from the processing results of the substrate.

[0069] (4) In the learning model generation method described in any one of paragraphs 1 to 3, generating the plurality of pseudo data sets may include generating the plurality of pseudo data sets in a number greater than the number of the plurality of actual data sets obtained by actual measurement.

[0070] In this case, a large number of pseudo data sets are generated, which makes it easier to generate an inverse model.

[0071] (Item 5) In the learning model generation method described in any one of items 1 to 4, the pseudo processing conditions for the substrate may be obtained by performing calculations on any of the parameters in the processing conditions for the substrate.

[0072] In this case, pseudo processing conditions for a large number of substrates can be easily obtained from the processing conditions of the substrate according to the content of the calculation and the parameters to be calculated, thereby easily generating a large number of pseudo data sets.

[0073] (Item 6) The learning model generation method described in any one of Items 1 to 5 may further include: using the inverse model to predict pseudo processing conditions for a substrate from at least some of the processing results for the substrate in the plurality of actual data sets and the pseudo processing results for the substrate in the plurality of pseudo data sets; adding a pseudo data set by using the forward model to predict the pseudo processing results for the substrate from the predicted pseudo processing conditions for the substrate; and updating the inverse model based on the added pseudo data set.

[0074] In this case, the generated forward and inverse models are used to populate the pseudo data set, which can then be used to easily update the inverse model to improve prediction accuracy.

[0075] (Item 7) In the learning model generation method described in any one of Items 1 to 6, generating the inverse model may include generating a plurality of simplex models that predict substrate processing conditions from substrate processing results, and generating the inverse model based on the plurality of simplex models.

[0076] In this case, an inverse model can be generated that can more accurately predict the processing conditions for a substrate from the processing results of the substrate.

[0077] (Item 8) The learning model generation device according to item 8 includes a forward model generation unit that generates a forward model that predicts a substrate processing result from the substrate processing conditions, based on a plurality of actual data sets that indicate the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing result; a pseudo data generation unit that generates a plurality of pseudo data sets that indicate the relationship between the substrate processing conditions and the substrate processing result, by predicting a pseudo substrate processing result from each of the plurality of pseudo substrate processing conditions, using the forward model generated by the forward model generation unit; and an inverse model generation unit that generates an inverse model that predicts a substrate processing condition from the substrate processing result, based on at least a portion of the plurality of pseudo data sets generated by the pseudo data generation unit.

[0078] In this learning model generation device, multiple pseudo data sets are generated using a forward model generated based on multiple actual data sets. Therefore, even if the number of actual data sets obtained by actual measurement is small, a sufficient number of pseudo data sets are generated. In this case, the inverse model can learn a sufficient number of pseudo data sets. Therefore, even if the processing conditions include a large number of parameters, it is possible to easily generate an inverse model, which is a trained model that estimates processing conditions from processing results.

[0079] (Item 9) In the learning model generation device described in Item 8, the backward model generation unit may generate the backward model further based on at least a portion of the plurality of actual data sets.

[0080] In this case, the number of data sets used to generate the inverse model is further increased, which makes it easier to generate the inverse model.

[0081] (10) In the learning model generation device described in 9, the inverse model generation unit may cause the inverse model to learn the real data set with a weight greater than that of the pseudo data set.

[0082] In this case, an inverse model can be generated that can more accurately predict the processing conditions for a substrate from the processing results of the substrate.

[0083] (11) In the learning model generation device described in any one of paragraphs 8 to 10, the pseudo data generation unit may generate a greater number of the pseudo data sets than the number of the actual data sets obtained by actual measurement.

[0084] In this case, a large number of pseudo data sets are generated, which makes it easier to generate an inverse model.

[0085] (Item 12) In the learning model generation device described in any one of items 8 to 11, the pseudo data generation unit may obtain pseudo processing conditions for the substrate by performing calculations on any of the parameters in the processing conditions for the substrate.

[0086] In this case, pseudo processing conditions for a large number of substrates can be easily obtained from the processing conditions of the substrate according to the content of the calculation and the parameters to be calculated, thereby easily generating a large number of pseudo data sets.

[0087] (Item 13) In the learning model generation device described in any one of Items 8 to 12, the pseudo data generation unit uses the inverse model generated by the inverse model generation unit to predict pseudo processing conditions for a substrate from at least a portion of the processing results for the substrate in the plurality of actual data sets and the pseudo processing results for the substrate in the plurality of pseudo data sets, and adds a pseudo data set by predicting the pseudo processing results for the substrate from the predicted pseudo processing conditions for the substrate using the forward model generated by the forward model generation unit, and the inverse model generation unit may update the inverse model based on the pseudo data set added by the pseudo data generation unit.

[0088] In this case, the generated forward and inverse models are used to populate the pseudo data set, which can then be used to easily update the inverse model to improve prediction accuracy.

[0089] (14) The learning model generation device described in any one of paragraphs 8 to 13 may further include a plurality of simplex model generation units that each generate a plurality of simplex models that predict substrate processing conditions from the substrate processing results, and the reverse model generation unit may generate the reverse model based on the plurality of simplex models generated by the plurality of simplex model generation units, respectively.

[0090] In this case, an inverse model can be generated that can more accurately predict the processing conditions for a substrate from the processing results of the substrate.

[0091] (Item 15) The learning model generation program according to Item 15 causes a computer to perform the following processes: generating a forward model that predicts a substrate processing result from the substrate processing conditions, based on a plurality of actual data sets that indicate the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing result; generating a plurality of pseudo data sets that indicate the relationship between the substrate's pseudo processing conditions and the substrate's pseudo processing result by predicting a pseudo substrate processing result from each of the plurality of pseudo substrate processing conditions using the forward model; and generating a reverse model that predicts a substrate processing condition from the substrate processing result, based on at least a portion of the plurality of pseudo data sets.

[0092] According to this learning model generation program, multiple pseudo data sets are generated using a forward model generated based on multiple actual data sets. Therefore, even if the number of actual data sets obtained by actual measurement is small, a sufficient number of pseudo data sets are generated. In this case, the inverse model can learn from a sufficient number of pseudo data sets. Therefore, even if the processing conditions include a large number of parameters, it is possible to easily generate an inverse model, which is a trained model that estimates processing conditions from processing results.

Claims

1. A computer-implemented learning model generation method, comprising: generating a forward model that predicts a substrate processing result from a substrate processing condition based on a plurality of actual data sets that indicate the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing result; generating a plurality of pseudo data sets that indicate the relationship between the pseudo substrate processing conditions and the pseudo substrate processing result by predicting a pseudo substrate processing result from each of the plurality of pseudo substrate processing conditions using the forward model; and generating a reverse model that predicts a substrate processing condition from the substrate processing result based on at least a portion of the plurality of pseudo data sets.

2. The learning model generation method according to claim 1, wherein generating the inverse model includes generating the inverse model further based on at least a portion of the plurality of real data sets.

3. The method for generating a learning model according to claim 2, wherein generating the inverse model includes training the inverse model with greater weights on the real data set than on the pseudo data set.

4. A learning model generation method according to any one of claims 1 to 3, wherein generating the plurality of pseudo data sets includes generating a greater number of the pseudo data sets than the number of the plurality of actual data sets obtained by actual measurement.

5. A learning model generation method according to any one of claims 1 to 4, wherein the pseudo processing conditions for the substrate are obtained by performing calculations on any of the parameters in the processing conditions for the substrate.

6. The learning model generation method of any one of claims 1 to 5, further comprising: using the inverse model to predict pseudo processing conditions for a substrate from at least some of the processing results for the substrate in the plurality of real data sets and the pseudo processing results for the substrate in the plurality of pseudo data sets; adding a pseudo data set by using the forward model to predict the pseudo processing results for the substrate from the predicted pseudo processing conditions for the substrate; and updating the inverse model based on the added pseudo data set.

7. A learning model generation method according to any one of claims 1 to 6, wherein generating the inverse model includes: generating a plurality of simplex models that predict substrate processing conditions from substrate processing results; and generating the inverse model based on the plurality of simplex models.

8. A learning model generation device comprising: a forward model generation unit that generates a forward model that predicts a substrate processing result from the substrate processing conditions based on a plurality of actual data sets that indicate the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing result; a pseudo data generation unit that generates a plurality of pseudo data sets that indicate the relationship between the substrate processing conditions and the substrate processing result by predicting a pseudo substrate processing result from each of a plurality of pseudo substrate processing conditions using the forward model generated by the forward model generation unit; and an inverse model generation unit that generates a inverse model that predicts a substrate processing condition from the substrate processing result based on at least a portion of the plurality of pseudo data sets generated by the pseudo data generation unit.

9. The learning model generation device according to claim 8, wherein the inverse model generation unit generates the inverse model based further on at least a portion of the plurality of actual data sets.

10. The learning model generation device according to claim 9, wherein the inverse model generation unit causes the inverse model to learn the real data set with a weight greater than that of the pseudo data set.

11. A learning model generation device according to any one of claims 8 to 10, wherein the pseudo data generation unit generates a greater number of pseudo data sets than the number of actual data sets obtained by actual measurement.

12. A learning model generation device according to any one of claims 8 to 11, wherein the pseudo data generation unit obtains pseudo processing conditions for the substrate by performing calculations on any of the parameters in the processing conditions for the substrate.

13. A learning model generation device according to any one of claims 8 to 12, wherein the pseudo data generation unit uses the inverse model generated by the inverse model generation unit to predict pseudo processing conditions for a substrate from at least some of the processing results for the substrate in the plurality of actual data sets and the pseudo processing results for the substrate in the plurality of pseudo data sets, and adds a pseudo data set by predicting the pseudo processing results for the substrate from the predicted pseudo processing conditions for the substrate using the forward model generated by the forward model generation unit, and the inverse model generation unit updates the inverse model based on the pseudo data set added by the pseudo data generation unit.

14. A learning model generation device according to any one of claims 8 to 13, further comprising a plurality of simplex model generation units that respectively generate a plurality of simplex models that predict substrate processing conditions from substrate processing results, wherein the inverse model generation unit generates the inverse model based on the plurality of simplex models generated by the plurality of simplex model generation units, respectively.

15. A learning model generation program that causes a computer to execute the following processes: a process of generating a forward model that predicts a substrate processing result from the substrate processing conditions based on a plurality of actual data sets that indicate the relationship between the substrate processing conditions obtained by actual measurement and the substrate processing result; a process of generating a plurality of pseudo data sets that indicate the relationship between the substrate's pseudo processing conditions and the substrate's pseudo processing result by predicting a pseudo substrate processing result from each of the plurality of pseudo substrate processing conditions using the forward model; and a process of generating a reverse model that predicts a substrate's processing conditions from the substrate processing result based on at least a portion of the plurality of pseudo data sets.

Images