Learning device, inference device and learning completion model

Abstract

A learning device for improving the simulation accuracy of a manufacturing process. The learning device includes: an acquisition unit that acquires image data of an object and data on a process for the object; and a learning unit that inputs the image data of the object and the data on the process into a learning model, and causes the learning model to learn in such a way that an output of the learning model approaches the image data of the object after the process.

Description

Technical Field

[0001] The invention relates to a learning device, an inference device and a learning completion model. Background Art

[0002] Traditionally, semiconductor manufacturers generate physical models for each manufacturing process (such as dry etching, deposition, etc.) and perform simulations to explore the best recipe and adjust process parameters.

[0003] On the other hand, since the operation of semiconductor manufacturing process is complex, there are also phenomena that cannot be expressed by physical models, and there is a limit to the simulation accuracy. Therefore, the application of learning models that have been machine-learned to replace simulation based on physical models is being studied recently. Summary of the invention

[0004] <Problems to be Solved by the Invention>

[0005] Here, in the case of a learned model, there is an advantage that it is not necessary to define various phenomena of the semiconductor manufacturing process using physical equations, etc. as in a physical model, and it is expected to achieve simulation accuracy that cannot be achieved by simulation based on a physical model.

[0006] The purpose of the present disclosure is to improve the simulation accuracy of a manufacturing process.

[0007] <Methods used to solve the problem>

[0008] The learning device in one embodiment of the present disclosure has, for example, the following structure, that is, it includes: an acquisition unit that acquires image data of an object and data on processing of the object; and a learning unit that inputs the image data of the object and the data on the processing into a learning model, and causes the learning model to learn in such a way that the output of the learning model approaches the image data of the object after the processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 is a diagram showing an example of the overall structure of a simulation system.

[0010] Figure 2 : is a diagram showing an example of the hardware configuration of each device constituting the simulation system.

[0011] Figure 3 is a diagram showing an example of learning data.

[0012] Figure 4 This is a diagram showing an example of the functional configuration of a learning unit of the learning device in the first embodiment.

[0013] Figure 5This is a diagram showing an example of the functional configuration of the data configuration unit of the learning device in the first embodiment.

[0014] Figure 6 This is a diagram showing a specific example of the processing performed by the data configuration unit of the learning device in the first embodiment.

[0015] Figure 7 This is a diagram showing a specific example of the processing performed by the dry etching learning model of the learning device in the first embodiment.

[0016] Figure 8 : is a flowchart showing the flow of the learning process.

[0017] Fig. 9 It is a diagram showing an example of the functional structure of an execution unit of an estimation device.

[0018] Fig.10 This is a diagram showing the simulation accuracy of the learned model for dry etching.

[0019] Fig.11 It is a diagram showing the simulation accuracy of the learned model for deposition.

[0020] Fig.12 This is a diagram showing an example of the functional configuration of a data configuration unit of the learning device in the second embodiment.

[0021] Fig.13 It is a diagram showing a specific example of the processing performed by the dry etching learning model of the learning device in the second embodiment.

[0022] Fig.14 This is a diagram showing an example of the functional configuration of a learning unit of a learning device in the third embodiment.

[0023] Fig.15 This is a diagram showing an example of the functional configuration of a data configuration unit of a learning device in the fourth embodiment.

[0024] Fig.16 It is a diagram showing an application example of the estimation device. DETAILED DESCRIPTION

[0025] Hereinafter, each embodiment will be described with reference to the drawings. It should be noted that in this specification and the drawings, components having substantially the same functional structure are denoted by the same reference numerals and repeated descriptions are omitted.

[0026] [First embodiment]

[0027] <Overall structure of the simulation system>

[0028] First, the overall configuration of a simulation system for simulating a semiconductor manufacturing process will be described. Figure 1 FIG. 1 is a diagram showing an example of the overall structure of a simulation system. Figure 1 As shown, the simulation system 100 includes a learning device 120 and an inference device 130. It should be noted that various data and various information used in the simulation system 100 are obtained from a database or the like of a semiconductor manufacturer or a semiconductor manufacturing equipment manufacturer.

[0029] like Figure 1 As shown in the upper part of , in the semiconductor manufacturing device 110, by setting predetermined parameter data (the details will be described later) and transporting multiple pre-processed wafers (objects), processing corresponding to each manufacturing process (such as dry etching, deposition) is performed.

[0030] It should be noted that a portion of the pre-processed wafers among the plurality of pre-processed wafers are transported to the measuring device 111, and the measuring device 111 performs shape measurement at various positions. Thus, in the measuring device 111, for example, pre-processed image data (two-dimensional image data) for representing the cross-sectional shape of the pre-processed wafer at various positions is generated. It should be noted that the measuring device 111 includes a scanning electron microscope (SEM), a length measurement scanning electron microscope (CD-SEM), a transmission electron microscope (TEM), an atomic force microscope (AFM), and the like. In addition, it is assumed that the pre-processed image data generated by the measuring device 111 is associated with various metadata such as the magnification of the microscope.

[0031] exist Figure 1 The example of exemplifies a case where the measuring device 111 generates pre-processing image data such as file names "shape data LD001", "shape data LD002", "shape data LD003", etc. as pre-processing image data.

[0032] On the other hand, after executing the processing corresponding to each manufacturing process, the processed wafer is unloaded from the semiconductor manufacturing apparatus 110. In the semiconductor manufacturing apparatus 110, the environment during the processing when executing the processing corresponding to each manufacturing process on the unprocessed wafer is measured and stored as environmental information.

[0033] A portion of the unprocessed wafers among the plurality of processed wafers unloaded from the semiconductor manufacturing apparatus 110 as processed wafers are transported to the measuring device 112, and the shape is measured at various positions by the measuring device 112. Thus, in the measuring device 112, for example, processed image data (two-dimensional image data) for representing the cross-sectional shape of the processed wafer at various positions is generated. It should be noted that, like the measuring device 111, the measuring device 112 includes a scanning electron microscope (SEM), a length measurement scanning electron microscope (CD-SEM), a transmission electron microscope (TEM), an atomic force microscope (AFM), and the like.

[0034] exist Figure 1 The example of exemplifies a case where the measuring device 112 generates processed image data such as file names "shape data LD001'", "shape data LD002'", "shape data LD003'", etc. as processed image data.

[0035] The pre-processed image data generated by the measuring device 111, the parameter data set in the semiconductor manufacturing device 110 and the environmental information stored, and the processed image data generated by the measuring device 112 are collected as learning data by the learning device 120. In addition, in the learning device 120, the collected learning data is stored in the learning data storage unit 123. It should be noted that the parameter data set in the semiconductor manufacturing device 110 and the environmental information stored are arbitrary data about the processing corresponding to the manufacturing process when the semiconductor manufacturing device 110 performs the processing corresponding to the manufacturing process on the pre-processed wafer (object). In this way, by using arbitrary data about the processing when performing the processing corresponding to the manufacturing process as learning data, factors related to various phenomena of the manufacturing process can be reflected in machine learning.

[0036] By installing a data shaping program and a learning program in the learning device 120 and executing the programs, the learning device 120 functions as a data shaping unit 121 and a learning unit 122 .

[0037] The data configuration unit 121 is an example of a processing unit. The data configuration unit 121 reads out the learning data stored in the learning data storage unit 123 and processes part of the read learning data into a predetermined format suitable for input to the learning model by the learning unit 122.

[0038] The learning unit 122 performs machine learning on the learning model using the read learning data (including the learning data processed by the data configuration unit 121 ) and generates a learned model of the semiconductor manufacturing process. The learned model generated by the learning unit 122 is provided to the inference device 130 .

[0039] By installing the data configuration program and the execution program in the inference device 130 and executing the programs, the inference device 130 functions as the data configuration unit 131 and the execution unit 132 .

[0040] The data configuration unit 131 is an example of a processing unit. The data configuration unit 131 obtains the pre-processed image data generated by the measuring device 111, and the parameter data and environmental information input to the inference device 130. In addition, the data configuration unit 131 processes the obtained parameter data and environmental information into a predetermined format suitable for input to the learning completed model by the execution unit 132.

[0041] The execution unit 132 inputs the pre-processed image data, the parameter data and the environmental information processed into a predetermined format in the data configuration unit 131 into the learned model, and executes simulation to output (infer) the processed image data (simulation result).

[0042] The user of the estimation device 130 verifies the learned model by comparing the processed image data output by causing the execution unit 132 to execute simulation using the learned model with the corresponding processed image data generated by the measurement device 112 .

[0043] Specifically, the user of the inference device 130

[0044] The processed image data, parameter data, and environmental information are input to the data configuration unit 131, and the processed image data is output by the execution unit 132.

[0045] Processed image data generated by causing the semiconductor manufacturing apparatus 110 to process the pre-processed wafer and causing the measuring apparatus 112 to measure the processed wafer

[0046] Thus, the user of the inference device 130 can calculate the simulation error of the learned model and verify the simulation accuracy.

[0047] It should be noted that after the verification is completed, any pre-processing data, any parameter data, and any environmental information are input to the inference device 130 to execute various simulations.

[0048] <Hardware configuration of each device constituting the simulation system>

[0049] Next, use Figure 2 The hardware configuration of each device (learning device 120 , inference device 130 ) constituting the simulation system 100 will be described. Figure 2 : is a diagram showing an example of the hardware configuration of each device constituting the simulation system.

[0050] It should be noted that since the hardware structures of the learning device 120 and the inference device 130 are substantially the same, the hardware structure of the learning device 120 will be described here.

[0051] Figure 2 is a diagram showing an example of the hardware structure of a learning device. Figure 2 As shown, the learning device 120 has a CPU (Central Processing Unit) 201 and a ROM (Read Only Memory) 202. In addition, the learning device 120 has a RAM (Random Access Memory) 203 and a GPU (Graphics Processing Unit) 204. It should be noted that the processors (processing circuits, processing circuits, processing circuitry) such as the CPU 201 and GPU 204, and the memories such as the ROM 202 and RAM 203 form a so-called computer.

[0052] Furthermore, the learning device 120 includes an auxiliary storage device 205 , an operation device 206 , a display device 207 , an I / F (Interface) device 208 , and a driver device 209 . It should be noted that the hardware components of the learning device 120 are connected to each other via a bus 210 .

[0053] The CPU 201 is a computing device that executes various programs (for example, a data configuration program, a learning program, etc.) installed in the auxiliary storage device 205 .

[0054] ROM 202 is a non-volatile memory and functions as a main storage device. ROM 202 is used to store various programs and data required for executing various programs installed in auxiliary storage device 205. Specifically, ROM 202 stores boot programs such as BIOS (Basic Input / Output System) or EFI (Extensible Firmware Interface).

[0055] The RAM 203 is a volatile memory such as a DRAM (Dynamic Random Access Memory) or an SRAM (Static Random Access Memory), and functions as a main storage device. The RAM 203 provides a work area for various programs installed in the auxiliary storage device 205 to be executed by the CPU 201 .

[0056] The GPU 204 is a computing device for image processing, and performs high-speed computing for parallel processing on various image data when the CPU 201 executes various programs.

[0057] The auxiliary storage device 205 is a storage unit for storing various programs and various image data for which the GPU 204 performs image processing when the various programs are executed by the CPU 201 . For example, the learning data storage unit 123 is implemented in the auxiliary storage device 205 .

[0058] The operation device 206 is an input device used when the administrator of the learning device 120 inputs various instructions to the learning device 120. The display device 207 is a display device that displays the internal state of the learning device 120. The I / F device 208 is a connection device for connecting to other devices and communicating.

[0059] The drive device 209 is a device for setting the recording medium 220. The recording medium 220 mentioned here includes a medium that records information optically, electrically or magnetically, such as a CD-ROM, a floppy disk, a magneto-optical disk, etc. In addition, the recording medium 220 may include a semiconductor memory that records information electrically, such as a ROM, a flash memory, etc.

[0060] It should be noted that the various programs installed in the auxiliary storage device 205 are installed by, for example, placing the distributed recording medium 220 in the drive device 209 and reading the various programs recorded in the recording medium 220 using the drive device 209. Alternatively, the various programs installed in the auxiliary storage device 205 can be installed by downloading via a network (not shown).

[0061] <Explanation of learning data>

[0062] Next, the learning data stored in the learning data storage unit 123 will be described. Figure 3 is a diagram showing an example of learning data. Figure 3 As shown, the learning data 300 includes “process”, “job ID”, “image data before processing”, “parameter data”, “environmental information”, and “image data after processing” as information items.

[0063] In the "process", names indicating semiconductor manufacturing processes are stored. Figure 3 The example shows a case where two names, “dry etching” and “deposition”, are stored as “processes”.

[0064] In “Job ID”, an identifier for identifying a job executed by the semiconductor manufacturing apparatus 110 is stored.

[0065] Figure 3 The example shows a case where "PJ001" and "PJ002" are stored as "job IDs" of dry etching. In addition, Figure 3 The example shows a case where "PJ101" is stored as the deposited "job ID".

[0066] In “pre-processing image data”, the file name of the pre-processing image data generated by the measuring device 111 is stored. Figure 3 The example shows that when job ID="PJ001", the measurement device 111 generates pre-processing image data with a file name="shape data LD001" for one pre-processing wafer in the lot (wafer group) of the job.

[0067] in addition, Figure 3 The example shows that when job ID = "PJ002", the measurement device 111 generates pre-processing image data with a file name = "shape data LD002" for one pre-processing wafer in the batch (wafer group) of the job. Figure 3 The example shows that when job ID="PJ101", the measurement device 111 generates pre-processing image data with a file name="shape data LD101" for one pre-processing wafer in the lot (wafer group) of the job.

[0068] In the “parameter data”, parameters indicating predetermined processing conditions set when the pre-processing wafer is processed in the semiconductor manufacturing apparatus 110 are stored. Figure 3 The example shows a case where “parameter 001_1”, “parameter 001_2”, “parameter 001_3”, . . . are set when a process of job ID=“PJ001” is executed in the semiconductor manufacturing apparatus 110 .

[0069] It should be noted that “parameter 001_1”, “parameter 001_2”, “parameter 001_3”, etc. include, for example, the following data or information.

[0070] Data set as set values ​​in the semiconductor manufacturing apparatus 110, such as Pressure (pressure in the chamber), Power (power of the high-frequency power supply), Gas (gas flow rate), and Temperature (temperature in the chamber or temperature of the surface of the wafer),

[0071] CD (Critical Dimension), Depth (depth), Taper (taper angle), Tilting (tilt angle), Bowing (bend), etc., which are data set as target values ​​in the semiconductor manufacturing apparatus 110,

[0072] Information on the hardware configuration of the semiconductor manufacturing apparatus 110

[0073] In the “environmental information”, information indicating the environment during the processing of the pre-processed wafer, which is measured when the pre-processed wafer is processed in the semiconductor manufacturing apparatus 110 , is stored. Figure 3 The example shows a case where “environmental data 001_1”, “environmental data 001_2”, “environmental data 001_3”, . . . , which are environmental information, are measured when the semiconductor manufacturing apparatus 110 executes the process of job ID="PJ001".

[0074] It should be noted that “environmental data 001_1”, “environmental data 001_2”, “environmental data 001_3”, etc. include, for example, the following data.

[0075] Data output from the semiconductor manufacturing apparatus 110 during processing, such as Vpp (potential difference), Vdc (direct current self-bias voltage), OES (luminous intensity based on emission spectrum analysis), and Reflect (reflected wave power) (mainly data on current or voltage)

[0076] Plasma density, ion energy, ion flux, and other data measured during the process (mainly data on light, as well as data on temperature and pressure).

[0077] In “processed image data”, the file name of the processed image data generated by the measuring device 112 is stored. Figure 3 The example of FIG. 1 shows a case where, in the case of job ID="PJ001", the measurement device 112 generates processed image data having a file name="shape data LD001'".

[0078] in addition, Figure 3 The example shows that when the work ID is "PJ002", the measured device 112 generates processed image data with a file name of "shape data LD002'". Figure 3 The example of FIG. 1 shows a case where, in the case of job ID="PJ101", the measurement device 112 generates processed image data having a file name="shape data LD101'".

[0079] <Functional structure of learning device>

[0080] Next, the functional configuration of each unit (data configuration unit 121, learning unit 122) of the learning device 120 will be described in detail.

[0081] (1) Details of the functional structure of the learning department

[0082] First, the functional configuration of the learning unit 122 of the learning device 120 will be described in detail. Figure 4 FIG. 1 is a diagram showing an example of the functional configuration of the learning unit of the learning device in the first embodiment. Figure 4 As shown, the learning unit 122 of the learning device 120 includes a dry etching learning model 420 , a deposition learning model 421 , a comparison unit 430 , and a change unit 440 .

[0083] The pre-processed image data, parameter data, and environmental information of the learning data 300 stored in the learning data storage unit 123 are read out by the data configuration unit 121 and input into the corresponding learning model. It should be noted that in the present embodiment, the parameter data and environmental information are input into the corresponding learning model after being processed into a predetermined form by the data configuration unit 121. However, the parameter data and environmental information may also be processed into a predetermined form in advance, and the data configuration unit 121 reads out the data and information processed into a predetermined form in advance and inputs them into the corresponding learning model.

[0084] The dry etching learning model 420 is input with the image data before processing, and the parameter data and environment information processed into a predetermined format by the data configuration unit 121 (however, only the data and information related to "process" = "dry etching" are input). After the image data before processing, and the parameter data and environment information processed into a predetermined format are input, the dry etching learning model 420 outputs an output result. In addition, the dry etching learning model 420 inputs the output result to the comparison unit 430.

[0085] Similarly, the image data before processing, and the parameter data and environmental information processed into a predetermined format by the data configuration unit 121 (however, only data and information related to "process" = "deposition") are input to the deposition learning model 421. After the image data before processing, and the parameter data and environmental information processed into a predetermined format are input, the deposition learning model 421 outputs an output result. In addition, the deposition learning model 421 inputs the output result to the comparison unit 430.

[0086] The comparison unit 430 compares the output result input from the dry etching learning model 420 with the processed image data of the learning data 300 (processed image data associated with “process”=“dry etching”), and notifies the change unit 440 of the difference information.

[0087] Similarly, the comparison unit 430 compares the output result input from the deposition learning model 421 with the processed image data of the learning data 300 (the processed image data associated with “process”=“deposition”), and notifies the change unit 440 of the difference information.

[0088] The changing unit 440 updates the model parameters of the dry etching learning model 420 or the deposition learning model 421 based on the difference information notified from the comparing unit 430. It should be noted that the difference information used to update the model parameters may be a square error or an absolute error.

[0089] In this way, in the learning unit 122, the image data before processing, as well as the parameter data and environmental information processed into a predetermined form are input to the learning model, and the model parameters are updated through machine learning in such a way that the output result output by the learning model is close to the processed image data.

[0090] Thus, in the learning unit 122 , the processed image data showing various phenomena in the semiconductor manufacturing process can be reflected in the machine learning, and the relationship between these phenomena and the parameter data and environmental information can be machine learned.

[0091] (2) Details of the functional structure of the data configuration unit

[0092] Next, the functional configuration of the data configuration unit 121 of the learning device 120 will be described in detail. Figure 5 FIG. 2 is a diagram showing an example of the functional structure of the data configuration unit of the learning device in the first embodiment. Figure 5 As shown, the data configuration unit 121 includes a shape data acquisition unit 501 , a channel identification data generation unit 502 , a one-dimensional data acquisition unit 511 , a one-dimensional data development unit 512 , and a connection unit 520 .

[0093] The shape data acquisition unit 501 reads out the pre-processed image data of the learning data 300 from the learning data storage unit 123 , and notifies the channel identification data generation unit 502 of the data.

[0094] The channel-specific data generating unit 502 is an example of a generating unit. The channel-specific data generating unit 502 obtains the pre-processing image data notified from the shape data acquiring unit 501 (here, data represented by pixel values ​​corresponding to the composition ratio (or content ratio) of each material). In addition, the channel-specific data generating unit 502 generates image data of multiple channels corresponding to the types of materials based on the obtained pre-processing image data. Hereinafter, the image data of multiple channels corresponding to the types of materials will be referred to as channel-specific data. For example, the channel-specific data generating unit 502 generates channel-specific data including a layer of air and four channel-specific data including layers of four materials, respectively, based on the pre-processing image data.

[0095] In addition, the channel specific data generation unit 502 notifies the connection unit 520 of the generated plurality of channel specific data. It should be noted that in this embodiment, although the channel specific data is generated by the channel specific data generation unit 502, the channel specific data may be generated in advance. In this case, the channel specific data generation unit 502 reads the pre-generated channel specific data and notifies the connection unit 520 of the pre-generated channel specific data.

[0096] The one-dimensional data acquisition unit 511 reads parameter data and environmental information of the learning data 300 from the learning data storage unit 123 , and notifies the one-dimensional data expansion unit 512 of the same.

[0097] The one-dimensional data expansion unit 512 processes the parameter data and environmental information notified from the one-dimensional data acquisition unit 511 into a format corresponding to the pre-processed image data (a two-dimensional arrangement format corresponding to the vertical and horizontal dimensions of the pre-processed image data).

[0098] Here, the parameter data is formed by one-dimensionally arranging the numerical values ​​of each parameter, such as "parameter 001_1", "parameter 001_2", "parameter 001_3", etc. Specifically, the parameter data is formed by one-dimensionally arranging the numerical values ​​of N types of parameters.

[0099] Therefore, in the one-dimensional data expansion unit 512, the numerical values ​​of the N parameters included in the parameter data are extracted one by one, and the extracted numerical values ​​are arranged two-dimensionally according to the longitudinal and lateral dimensions of the image data before processing. Thus, in the one-dimensional data expansion unit 512, N parameter data arranged two-dimensionally are generated.

[0100] In addition, the one-dimensional data expansion unit 512 notifies the connection unit 520 of the N parameter data arranged two-dimensionally.

[0101] Similarly, the environmental information is formed by one-dimensionally arranging the numerical values ​​of each environmental information such as "environmental data 001_1", "environmental data 001_2", "environmental data 001_3", etc. Specifically, the environmental information is formed by one-dimensionally arranging the numerical values ​​of M types of environmental data.

[0102] Therefore, in the one-dimensional data expansion unit 512, the numerical values ​​of the M types of environmental data included in the environmental information are extracted one by one, and the extracted numerical values ​​are arranged two-dimensionally according to the longitudinal and lateral dimensions of the image data before processing. Thus, in the one-dimensional data expansion unit 512, M environmental information arranged two-dimensionally is generated.

[0103] In addition, the one-dimensional data expansion unit 512 notifies the connection unit 520 of the M pieces of environment information arranged two-dimensionally.

[0104] The connection unit 520 connects the N parameter data and the M environmental information arranged two-dimensionally notified from the one-dimensional data expansion unit 512 as a new channel on the plurality of channel specific data notified from the channel specific data generation unit 502, and generates connection data. It should be noted that in the present embodiment, although the connection data is generated by the connection unit 520, the connection data may be generated in advance. In this case, the connection unit 520 reads the pre-generated connection data and inputs it into the learning model.

[0105] <Specific example of processing performed by each unit of the learning device>

[0106] Next, among the processes performed by the various units of the learning device 120 , specific examples of the processes performed by the data configuration unit 121 and the processes performed by the dry etching learning model 420 in the learning unit 122 will be described.

[0107] (1) Specific example of processing performed by the data configuration unit

[0108] Figure 6 FIG. 2 is a diagram showing a specific example of the processing performed by the data configuration unit. Figure 6 In the example, the pre-processing image data 600 is pre-processing image data having a file name="shape data LD001".

[0109] like Figure 6 As shown, the pre-processing image data 600 includes a layer of air, a layer of material A, a layer of material B, a layer of material C, and a layer of material D. In this case, in the channel specific data generating unit 502, channel specific data 601, 602, 603, 604, and 605 are generated.

[0110] In addition, if Figure 6 As shown, the parameter data 610 is formed by, for example, arranging the numerical values ​​of the parameters ("parameter 001_1", "parameter 001_2", "parameter 001_3", . . . , etc.) in one dimension.

[0111] In addition, if Figure 6 As shown, the environmental information 620 is formed by, for example, arranging numerical values ​​of each environmental data such as "environmental data 001_1", "environmental data 001_2", "environmental data 001_3", . . . in a one-dimensional manner.

[0112] In this case, the one-dimensional data expansion unit 512 arranges the parameter 001_1 two-dimensionally (arranges the same value in the vertical and horizontal directions) according to the vertical and horizontal dimensions of the image data 600 before processing. Similarly, the one-dimensional data expansion unit 512 arranges the parameter 001_2 two-dimensionally according to the vertical and horizontal dimensions of the image data 600 before processing. Similarly, the one-dimensional data expansion unit 512 arranges the parameter 001_3 two-dimensionally according to the vertical and horizontal dimensions of the image data 600 before processing.

[0113] In addition, the one-dimensional data expansion unit 512 arranges the environment data 001_1 two-dimensionally (arranges the same values ​​in the vertical and horizontal directions) according to the vertical and horizontal dimensions of the pre-processed image data 600. Similarly, the one-dimensional data expansion unit 512 arranges the environment data 001_2 two-dimensionally according to the vertical and horizontal dimensions of the pre-processed image data 600. Similarly, the one-dimensional data expansion unit 512 arranges the environment data 001_3 two-dimensionally according to the vertical and horizontal dimensions of the pre-processed image data 600.

[0114] The two-dimensionally arranged parameter data 611 , 612 , 613 , etc. and the two-dimensionally arranged environment information 621 , 622 , 623 , etc. are connected to the channel identification data 601 to 605 as new channels by the connection unit 520 , and the connection data 630 is generated.

[0115] (2) Specific Example of Processing by the Learning Model for Dry Etching

[0116] Next, a specific example of the processing performed by the dry etching learning model 420 in the learning unit 122 will be described. Figure 7 FIG. 1 is a diagram showing a specific example of processing performed by the dry etching learning model of the learning device in the first embodiment. Figure 7 As shown, in this embodiment, a learning model based on a U-shaped convolutional neural network (CNN) (so-called UNET) is used as the dry etching learning model 420 .

[0117] In the case of UNET, image data is usually input and output. Therefore, by using the UNET as a learning model of the learning unit 122, image data before and after processing in the semiconductor manufacturing process can be input and output.

[0118] On the other hand, in the case of UNET, for data in a form other than image data, it is necessary to process the form of image data in advance. The one-dimensional data expansion unit 512 of the data configuration unit 121 is configured to arrange parameter data and environmental information two-dimensionally in order to convert the data input to UNET into the form of image data. By enabling the input of parameter data and environmental information to UNET, machine learning can be performed using factors related to various phenomena of dry etching.

[0119] Figure 7 The example of FIG. 1 shows a case where connection data 630 is input to the dry etching learning model 420 using UNET and an output result 700 including a plurality of channel specific data is output.

[0120] It should be noted that in Figure 7 The example of FIG. 4 shows a specific example of the process performed by the dry etching learning model 420 , and the same is true for a specific example of the process performed by the deposition learning model 421 .

[0121] <Flow of learning process>

[0122] Next, the overall flow of the learning process will be described. Figure 8 : is a flowchart showing the flow of the learning process.

[0123] In step S801 , the measuring device 111 measures the shape of a pre-processed wafer at various positions before being processed by the semiconductor manufacturing apparatus 110 , and generates pre-processed image data.

[0124] In step S802 , the measuring device 112 measures the shape of the processed wafer processed by the semiconductor manufacturing apparatus 110 at various positions, and generates processed image data.

[0125] In step S803 , the learning device 120 acquires parameter data set in the semiconductor manufacturing apparatus 110 when the semiconductor manufacturing apparatus 110 executes processing corresponding to each manufacturing process, and environmental information obtained by measuring the environment during the processing.

[0126] In step S804 , the learning device 120 stores the pre-processed image data generated by the measuring device 111 , the processed image data generated by the measuring device 112 , the acquired parameter data, and the environmental information in the learning data storage unit 123 as learning data.

[0127] In step S805 , the data configuration unit 121 of the learning device 120 reads out the pre-processed image data, parameter data, and environmental information from the learning data storage unit 123 , and generates connection data.

[0128] In step S806 , the learning unit 122 of the learning device 120 takes the connection data as input and the processed image data as output, performs machine learning on the learning model, and generates a learning completed model.

[0129] In step S807 , the learning unit 122 of the learning device 120 sends the generated learning model to the inference device 130 .

[0130] <Functional structure of inference device>

[0131] Next, the functional structure of the inference device 130 will be described in detail. It should be noted that the functional structure of the data configuration unit 131 among the various units (data configuration unit 131, execution unit 132) of the inference device 130 is the same as the functional structure of the data configuration unit 121 of the learning device 120. Therefore, the functional structure of the data configuration unit 131 will be omitted here, and the functional structure of the execution unit 132 will be described in detail.

[0132] Fig. 9 FIG. 1 is a diagram showing an example of a functional structure of an execution unit of an inference device. Fig. 9 As shown, the execution unit 132 of the estimation device 130 includes a learned model 920 for dry etching, a learned model 921 for deposition, and an output unit 930 .

[0133] After obtaining the pre-processed image data (e.g., data not used for machine learning) generated by the measuring device 111 and inputting the parameter data and environmental information to the inference device 130, the data configuration unit 131 generates the connection data. In addition, the data configuration unit 131 inputs the generated connection data into each learning completed model. It should be noted that Fig. 9 The example of exemplifies a case where file names "shape data SD001", "shape data SD002", . . . etc. are obtained as pre-processed image data generated by the measuring device 111.

[0134] In the dry etching learned model 920, simulation is executed after the connection data is input from the data configuration unit 131. In addition, the dry etching learned model 920 notifies the output unit 930 of the output result output by executing the simulation.

[0135] Similarly, in the learned model for deposition 921, after the connection data is input from the data configuration unit 131, the simulation is executed. In addition, the learned model for deposition 921 notifies the output unit 930 of the output result output by executing the simulation.

[0136] It should be noted that, although the pre-processing image data generated by the measuring device 111 is input here, any pre-processing image data may be input to the learned model 920 for dry etching and the learned model 921 for deposition.

[0137] The output unit 930 generates processed image data (e.g., file name = "shape data SD001"") based on the output results notified from the learning completion model 920 for dry etching, and outputs it as a simulation result. Similarly, the output unit 930 generates processed image data (e.g., file name = "shape data SD101"") based on the output results notified from the learning completion model 921 for deposition, and outputs it as a simulation result.

[0138] Here, it is assumed that the user of the inference device 130 inputs the same parameter data and environmental information as those set in the semiconductor manufacturing device 110 and the environmental information stored. In this case, the user of the inference device 130 can compare the processed image data output from the output unit 930 with the processed image data (e.g., file name = "shape data SD101'") generated by the measurement device 112. Therefore, the user of the inference device 130 can verify the simulation accuracy of the inference device 130.

[0139] Fig.10 : is a graph showing the simulation accuracy of the dry etching learning model. Fig.10 10a shows the pre-processed image data (file name = "shape data SD001") generated by the measuring device 111 and the post-processed image data (file name = "shape data SD001'") generated by the measuring device 112 as comparison objects.

[0140] on the other hand, Fig.10 10 b shows image data before processing (file name = “shape data SD001”) and image data after processing (file name = “shape data SD001”) when simulation is performed by the dry etching learned model 920 .

[0141] In the right Fig.10 The processed image data of 10a (file name = "shape data SD001") and Fig.10 After comparing the processed image data of 10b (file name = "shape data SD001"), there is no difference between the two. In this way, based on the dry etching learning model 920, high-precision simulation can be achieved for dry etching performed by the semiconductor manufacturing device 110.

[0142] Likewise, Fig.11 is a diagram showing the simulation accuracy of the learning model for deposition. Fig.11 11a shows the pre-processed image data (file name = "shape data SD101") generated by the measuring device 111 and the post-processed image data (file name = "shape data SD101'") generated by the measuring device 112 as comparison objects.

[0143] on the other hand, Fig.11 11b shows image data before processing (file name = "shape data SD101") and image data after processing (file name = "shape data SD101") when simulation is performed by the deposition learning completion model 921.

[0144] In the right Fig.11 The processed image data of 11a (file name = "shape data SD101'") and Fig.11 After comparing the processed image data of 11b (file name = "shape data SD101"), there is no difference between the two. In this way, according to the deposition learning completion model 921, a high-precision simulation can be achieved for the deposition performed by the semiconductor manufacturing device 110.

[0145] It should be noted that, even when compared with a general physical model (a model for verifying a semiconductor manufacturing process based on physical laws), each learned model of the execution unit 132 can improve simulation accuracy.

[0146] The reason for this is that, in the case of a general physical model, various phenomena that cannot be expressed using physical equations cannot be reflected in the simulation, whereas in the case of a learning model, various phenomena that have an impact on the processed image data can be reflected in machine learning. In addition, the reason for this is that, in the case of the learning model in this embodiment, since factors (parameter data, environmental information) related to various phenomena in the semiconductor manufacturing process are input, it is possible to perform machine learning on the relationship between each phenomenon and the factors.

[0147] It should be noted that, as various phenomena that cannot be expressed using physical equations, in the case of dry etching, for example, the phenomenon that the gas in the chamber becomes uneven, etc., or the phenomenon that the etched particles adhere as deposits, etc., can be cited. In addition, in the case of deposition, for example, the phenomenon that the particles adhere after rebounding more than once, etc., can be cited.

[0148] In general physical models, these phenomena are treated as the gas in the chamber being uniform during dry etching, and as particles being attached to the first contact position during deposition.

[0149] In contrast, in the case of a learned model, processed image data showing the effects of these phenomena can be reflected in machine learning, and the relationship between these phenomena and parameter data and environmental information can be machine learned. Therefore, compared with a general physical model, the learned model can improve simulation accuracy.

[0150] In this way, the learned model can achieve simulation accuracy that cannot be achieved in simulation based on the physical model. In addition, the learned model can shorten the simulation time compared to the simulation based on the physical model. Furthermore, the learned model also has the advantage of not having to manually create rules or construct appropriate physical equations as in the simulation based on the physical model.

[0151] Summary

[0152] As can be seen from the above description, the learning device in the first embodiment performs the following steps.

[0153] In a semiconductor manufacturing apparatus, parameter data set when processing an unprocessed wafer and environmental information indicating an environment during processing of the unprocessed wafer measured when processing the unprocessed wafer are acquired.

[0154] Acquiring pre-processed image data, which is image data showing the pre-processed shape of a pre-processed wafer processed in a semiconductor manufacturing apparatus.

[0155] The obtained parameter data and environmental information are processed into the form of image data by arranging the obtained parameter data and environmental information two-dimensionally according to the longitudinal size and the lateral size of the obtained pre-processed image data. In addition, the processed parameter data and environmental information are connected to the pre-processed image data to generate connected data.

[0156] The generated connection data is input into a learning model based on a U-shaped convolutional neural network, and machine learning is performed so that the output result approaches the processed image data for representing the shape of the processed wafer.

[0157] Thus, according to the learning device in the first embodiment, factors related to various phenomena in the semiconductor manufacturing process can be reflected in machine learning, and a learned model that realizes high-precision simulation can be generated.

[0158] In addition, the estimation device in the first embodiment performs the following steps.

[0159] Obtain pre-processed image data, parameter data, and environmental information.

[0160] The acquired parameter data and environmental information are arranged two-dimensionally according to the longitudinal and lateral dimensions of the acquired pre-processed image data, thereby processing them into the form of image data. In addition, the processed parameter data and environmental information are connected to the pre-processed image data to generate connected data.

[0161] The generated connection data is input into the learned model and the simulation is performed.

[0162] Therefore, according to the inference device in the first embodiment, it is possible to generate a learned model using machine learning of factors related to various phenomena in the semiconductor manufacturing process, and to realize high-precision simulation.

[0163] As described above, according to the first embodiment, it is possible to improve the simulation accuracy in the simulation of the semiconductor manufacturing process.

[0164] [Second embodiment]

[0165] In the above-mentioned first embodiment, a case is described in which parameter data and environmental information are processed into the form of image data according to the longitudinal and lateral dimensions of the pre-processed image data, and it is linked to the pre-processed image data and input into the learning model (or the learning completed model).

[0166] However, the method of processing parameter data and environmental information, and the method of inputting the processed parameter data and environmental information to the learning model (or the learning completed model) are not limited to this. For example, it can be configured so that the processed parameter data and environmental information are input into each layer of the learning model (or the learning completed model). In addition, it can be configured so that when the parameter data and environmental information are input into each layer of the learning model (or the learning completed model), the parameter data and environmental information are processed into a predetermined form used when converting the image data convolved by each layer of the learning model (or the learning completed model). The following is an explanation of the second embodiment, centering on the differences between it and the above-mentioned first embodiment.

[0167] <Functional structure of data configuration unit>

[0168] First, the functional configuration of the data configuration unit of the learning device in the second embodiment will be described in detail. Fig.12 FIG. 2 is a diagram showing an example of the functional configuration of the data configuration unit of the learning device in the second embodiment. Figure 5 The functional structure of the data configuration unit 121 shown is different in that: Fig.12 The data shaping unit 1200 shown includes a concatenation unit 1201 and a normalization unit 1202 .

[0169] The concatenation unit 1201 concatenates a plurality of channel identification data notified from the channel identification data generation unit 502 to generate concatenated data.

[0170] The normalization unit 1202 normalizes the parameter data and the environment information notified from the one-dimensional data acquisition unit 511 , and generates normalized parameter data and normalized environment information.

[0171] <Specific example of processing performed by the learning model>

[0172] Next, a specific example of the process performed by the dry etching learning model will be described. Fig.13 It is a diagram showing a specific example of the processing performed by the dry etching learning model of the learning device in the second embodiment.

[0173] like Fig.13 As shown, in the case of the learning device in the second embodiment, the connection data 1310 generated by the connection unit 1201 of the data configuration unit 1200 is input to the dry etching learning model 130 .

[0174] In addition, if Fig.13 As shown, in the case of the learning device in the second embodiment, the normalized parameter data and the normalized environment information generated by the normalization unit 1202 of the data configuration unit 1200 are input to the dry etching learning model 1300 .

[0175] It should be noted that if Fig.13 As shown, the dry etching learning model 1300 includes, in addition to UNET which is a CNN-based learning model, a neural network unit 1301 which is a fully connected learning model.

[0176] After receiving the normalized parameter data and the normalized environment information, the neural network unit 1301 outputs a value in a predetermined format (e.g., coefficients γ and β of a first-order formula) used when converting the value of each pixel of each image data subjected to convolution processing by each layer of the UNET. In other words, the neural network unit 1301 functions to process the normalized parameter data and the normalized environment information into a predetermined format (e.g., the format of coefficients of a first-order formula).

[0177] exist Fig.13 In the example of , since UNET consists of 9 layers, the neural network unit 1301 outputs (γ1, β1) to (γ9, β9) as coefficients of the first-order formula. Fig.13 In the example, due to space reasons, one set of coefficients of the first-order formula is input in each layer, whereas multiple sets of coefficients of the first-order formula are input in each layer according to each channel-specific data.

[0178] In each layer of UNET, for example, the first-order formula = h×γ+β (h×γ1+β1 in the case of the first layer) is used to convert the value of each pixel (here set to "h") of each image data of each channel specific data undergoing convolution processing.

[0179] Here, the coefficients (γ1, β1) to (γ9, β9) of the first-order formula can be regarded as an index for indicating which image data is important among the image data of each channel specific data subjected to convolution processing in each layer of UNET. In other words, in the neural network unit 1301, a process for calculating an index for indicating the importance of each image data processed in each layer of the learning model is performed based on the normalized parameter data and the normalized environment information.

[0180] According to the above structure, after the connection data 1310, the normalized parameter data and the normalized environment information are input into the dry etching learning model 1300, an output result 700 including a plurality of channel specific data is output. It should be noted that the output result 700 is compared with the processed image data in the comparison unit 430, and the difference information is calculated. In the case of the learning device in the second embodiment, the change unit 440 updates the model parameters of the UNET in the dry etching learning model 1300 and the model parameters of the neural network unit 1301 based on the difference information.

[0181] As described above, according to the learning device in the second embodiment, when performing machine learning on the dry etching learning model 1300, it is possible to extract image data with high importance based on the normalized parameter data and the normalized environment information in each layer of the UNET.

[0182] Summary

[0183] As can be seen from the above description, the learning device in the second embodiment performs the following steps.

[0184] In a semiconductor manufacturing apparatus, parameter data set when processing an unprocessed wafer and environmental information indicating an environment during processing of the unprocessed wafer measured when processing the unprocessed wafer are acquired.

[0185] Acquiring pre-processed image data, which is image data showing the pre-processed shape of a pre-processed wafer processed in a semiconductor manufacturing apparatus.

[0186] The acquired parameter data and environmental information are normalized and processed into the form of coefficients of a first-order formula used when converting the value of each pixel of each image data convolutionally processed by each layer of the learning model.

[0187] When performing machine learning, the learning unit converts the value of each pixel of the image data subjected to convolution processing in each layer using a first-order formula.

[0188] Therefore, according to the learning device in the second embodiment, factors related to various phenomena in the semiconductor manufacturing process can be reflected in machine learning, and a learned model that realizes high-precision simulation can be generated.

[0189] It should be noted that, in the second embodiment, the learning device is described, but the same processing is performed when the execution unit executes the simulation in the estimation device.

[0190] [Third embodiment]

[0191] In the first and second embodiments, when the learning unit performs machine learning, the constraints inherent in the semiconductor manufacturing process are not specifically mentioned. On the other hand, there are inherent constraints in the semiconductor manufacturing process, and by reflecting them in the machine learning performed by the learning unit (that is, by reflecting domain knowledge in the machine learning performed by the learning unit), the simulation accuracy can be further improved. The following describes the third embodiment that reflects domain knowledge, focusing on the differences between it and the first and second embodiments.

[0192] <Details of the functional structure of the learning model>

[0193] Fig.14 is a diagram showing an example of the functional configuration of the learning unit of the learning device in the third embodiment. Figure 4 The functional configuration of the learning unit 122 shown is different from the internal configuration of the learning model. It should be noted that, although the internal configuration of the learning model is described here using the dry etching learning model 1410 , the deposition learning model also has the same internal configuration.

[0194] like Fig.14 As shown, the dry etching learning model 1410 of the learning unit 1400 includes a sigmoid function unit 1412 and a multiplication unit 1413 in addition to the UNET 1411 .

[0195] The sigmoid function section 1412 is an example of a processing section, and Fig.14 As shown, by multiplying the first output result as the output of UNET 1411 by a sigmoid function 1420, a second output result 1421 is output.

[0196] The multiplication calculation unit 1413 obtains the second output result 1421 from the sigmoid function unit 1412. In addition, the multiplication calculation unit 1413 obtains the pre-processing image data from the data configuration unit 121. Furthermore, the multiplication calculation unit 1413 multiplies the obtained pre-processing image data by the obtained second output result 1421, thereby notifying the comparison unit 430 of the final output result 1422.

[0197] In this way, by configuring so as to multiply the pre-processed image data and output the final output result 1422 , the UNET 1411 in which the dry etching learning model 1410 is subjected to machine learning outputs image data indicating the cutting rate as the first output result.

[0198] Here, the cutting rate is a value indicating the rate of change of how much the layers of each material included in the pre-processing image data are cut in the post-processing image data. The cutting rate is made close to the value obtained by dividing the post-processing image data by the pre-processing image data by performing machine learning on the dry etching learning model 1410. However, the first output result outputted from the UNET 1411 takes an arbitrary value in the process of machine learning.

[0199] On the other hand, in the case of dry etching, there is a constraint condition (domain knowledge) regarding the change in shape that "the raw material does not increase before and after the process." Therefore, in the case of dry etching, the cutting rate is within the range of 0 to 1.

[0200] Here, the sigmoid function unit 1412 is a function that converts an arbitrary value into a value of 0 to 1, and by converting the first output result into the second output result using the sigmoid function unit 1412, the above-mentioned domain knowledge can be reflected.

[0201] It should be noted that although Fig.14 Although not shown in the figure, the same processing can be performed by configuring an S-type function unit and a multiplication unit in the deposition learning model. Specifically, the UNET outputs image data representing the adhesion rate as the first output result when the deposition learning model is subjected to machine learning.

[0202] Here, the adhesion rate is a value indicating the rate of change of the degree of thin film adhesion in the processed image data relative to the layers of each material included in the pre-processed image data. By making the deposition learning model perform machine learning, the adhesion rate is made close to the value obtained by dividing the difference between the pre-processed image data and the post-processed image data by the pre-processed image data. However, the first output result outputted from the UNET takes an arbitrary value during the machine learning process.

[0203] On the other hand, in the case of deposition, there is a constraint condition (domain knowledge) regarding the change in shape: "the raw material does not decrease before and after the treatment." Therefore, in the case of deposition, the adhesion rate is within the range of 0 to 1.

[0204] As described above, the sigmoid function part is a function that converts an arbitrary value into a value of 0 to 1, and by converting the first output result into the second output result using the sigmoid function part, the above-mentioned domain knowledge can be reflected.

[0205] As described above, according to the learning unit 1400 of the learning device 120 in the third embodiment, domain knowledge can be reflected in machine learning, and simulation accuracy can be further improved.

[0206] [Fourth embodiment]

[0207] In the first to third embodiments, the data configuration unit generates the concatenated data of the longitudinal and lateral dimensions corresponding to the longitudinal and lateral dimensions of the pre-processed image data. However, the longitudinal and lateral dimensions of the concatenated data generated by the data configuration unit may be any dimensions, and the concatenated data may be generated by compressing the pre-processed image data. Hereinafter, the fourth embodiment will be described with the focus on the differences between the first to third embodiments described above.

[0208] <Details of the functional structure of the data configuration unit>

[0209] Fig.15 FIG. 1 is a diagram showing an example of the functional structure of the data configuration unit of the learning device in the fourth embodiment. Fig.15 15a shows a data configuration unit 1510 obtained by adding a compression unit 1511 to the data configuration unit 121 of the learning device in the first embodiment.

[0210] The compression unit 1511 compresses the pre-processed image data acquired by the shape data acquisition unit 501. In the compression unit 1511, for example, an average value is calculated for the pixel values ​​of n adjacent pixels (n is an integer greater than or equal to 2. For example, in the case of 2 pixels in the vertical direction and 2 pixels in the horizontal direction, n=4) and the average value is used as the pixel value of one pixel that is a collection of the n pixels. Thus, in the compression unit 1511, the pre-processed image data can be compressed to 1 / n times.

[0211] Thus, in the compression unit 1511, since the image data before processing is image data indicating the composition ratio (or content ratio) of each material, the compression processing is performed in a manner that the composition ratio (or content ratio) of the materials before and after compression is maintained as much as possible. It should be noted that the compression rate of the compression processing performed by the compression unit 1511 is not limited to an integer multiple, and the compression processing of any compression rate can be performed in the compression unit 1511.

[0212] Likewise, Fig.15 15b shows a data configuration unit 1520 obtained by adding a compression unit 1511 to the data configuration unit 1200 of the learning device in the second embodiment.

[0213] The compression unit 1511 included in the data configuration unit 1520 has the same function as the compression unit 1511 included in the data configuration unit 1510. Therefore, a detailed description thereof will be omitted here.

[0214] Thus, by adding the compression unit 1511 to the data configuration unit 1510 or 1520, the size of the concatenated data input to the learning unit 122, 1400 (or the execution unit 132) can be reduced. Therefore, according to the fourth embodiment, the learning time when the learning unit 122, 1400 performs machine learning or the simulation time when the execution unit 132 performs simulation can be shortened.

[0215] [Other embodiments]

[0216] In the first embodiment described above, the dry etching learning model 420 and the deposition learning model 421 are provided separately in the learning unit 122 , and machine learning is performed separately using different learning data.

[0217] However, in the semiconductor manufacturing process, dry etching and deposition may occur simultaneously. In anticipation of such a situation, a learning model may be provided in the learning unit 122 to perform machine learning on the situation where dry etching and deposition occur simultaneously.

[0218] In this case, the learning unit 122 performs machine learning for the one learning model using learning data including pre-processing image data before dry etching and deposition and post-processing image data after dry etching and deposition.

[0219] As a result, in a general physical model, it is necessary to set up simulators separately for dry etching and deposition, but in a learning model, the settings can also be combined.

[0220] In addition, in the first embodiment described above, the data configuration unit 121 processes both parameter data and environmental information into a predetermined format and inputs them into the corresponding learning model. However, the data configuration unit 121 may process only parameter data into a predetermined format and input it into the corresponding learning model. In other words, when the learning unit 122 performs machine learning of the learning model, only parameter data may be used without using environmental information.

[0221] Similarly, in the first embodiment described above, the data configuration unit 131 processes both parameter data and environmental information into a predetermined format and inputs them into the corresponding learning model. However, the data configuration unit 131 may only process parameter data into a predetermined format and input it into the corresponding learning model. In other words, when the learning model is used for simulation in the execution unit 132, only parameter data may be used without using environmental information.

[0222] In the first embodiment, the image data before and after processing are two-dimensional image data. However, the image data before and after processing are not limited to two-dimensional image data, but may be three-dimensional image data (so-called voxel data).

[0223] It should be noted that, although when the image data before processing is two-dimensional image data, the connection data is an arrangement of (channel, longitudinal size, lateral size), when the image data before processing is three-dimensional image data, the connection data is an arrangement of (channel, longitudinal size, lateral size, depth size).

[0224] In the first embodiment, although the two-dimensional image data is processed as it is, it is also possible to deform the two-dimensional image data or the three-dimensional image data. For example, three-dimensional image data may be acquired, two-dimensional image data of a predetermined section may be generated, and the two-dimensional image data may be input as the image data before processing. Alternatively, three-dimensional image data may be generated based on two-dimensional image data of continuous predetermined sections, and the two-dimensional image data may be input as the image data before processing.

[0225] In the first embodiment, the channel specific data generating unit 502 generates channel specific data according to the air layer and each material layer. However, the method of generating channel specific data is not limited to this, and the channel specific data may be generated based on a larger classification such as oxide, silicon, organics, nitride, etc. instead of a specific film type.

[0226] In addition, in the above-mentioned first to fourth embodiments, the case where the processed image data is output and the processing is terminated when the inference device 130 is input with the pre-processed image data, parameter data and environmental information is described. However, the structure of the inference device 130 is not limited to this. For example, it is also possible to configure the processed image data output by inputting the pre-processed image data, parameter data and environmental information, together with the corresponding parameter data and environmental information, to be input again into the inference device 130. Thus, in the inference device 130, the change of shape can be continuously output. It should be noted that when the processed image data is input again into the inference device 130, the corresponding parameter data and environmental information can be arbitrarily changed.

[0227] In the first to fourth embodiments, no specific application examples of the estimation device 130 are described. However, the estimation device 130 can be applied to a service that searches for an optimal recipe, optimal parameter data, or optimal hardware configuration and provides it to a semiconductor manufacturer.

[0228] Fig.16 16 is a diagram showing an application example of the inference device, and shows an example in which the inference device 130 is applied to a service providing system 1600 .

[0229] The service providing system 1600 is connected to each office of the semiconductor manufacturer via the network 1640 , for example, and acquires the pre-processing image data. In addition, in the service providing system 1600 , the acquired pre-processing image data is stored in the data storage unit 1602 .

[0230] In addition, the estimation device 130 reads the pre-processed image data from the data storage unit 1602 and executes the simulation while changing the parameter data and the environmental information. Thus, the user of the estimation device 130 can search for the optimal recipe, the optimal parameter data, or the optimal hardware shape.

[0231] In the information providing device 1601, the optimal recipe and the optimal parameter data found by the user of the estimation device 130 are provided to each office of the semiconductor manufacturer.

[0232] In this way, by applying the inference device 130 to the service providing system 1600 , it is possible to provide the semiconductor manufacturer with the optimal recipe and the optimal parameter data in the service providing system 1600 .

[0233] In the first to fourth embodiments, the object is described as a wafer before processing. However, the object is not limited to the wafer before processing, and may be, for example, the inner wall of the chamber of the semiconductor manufacturing apparatus 110 or the surface of a part.

[0234] In the first to fourth embodiments described above, the case where the image data before processing (or the image data after processing) is generated by the measuring device 111 (or the measuring device 112) is described. However, the present invention is not limited to the case where the image data before processing (or the image data after processing) is generated by the measuring device 111 (or the measuring device 112). For example, it is also possible to configure the measuring device 111 (the measuring device 112) to generate multidimensional measurement data for representing the shape of the object, and the learning device 120 to generate the image data before processing (or the image data after processing) based on the measurement data.

[0235] It should be noted that the measurement data generated by the measuring device 111 (or the measuring device 112) includes, for example, data having position information and film type information. Specifically, it includes data combining position information generated by CD-SEM and CD measurement length data. Alternatively, it includes data combining two-dimensional or three-dimensional shape and film type information generated by X-ray or Raman method. In other words, the multidimensional measurement data for representing the shape of the object includes various representation forms corresponding to the type of the measuring device.

[0236] Furthermore, in the above-described first to fourth embodiments, the learning device 120 and the estimation device 130 are shown as separate entities, but they may be configured integrally.

[0237] In addition, in the above-mentioned first to fourth embodiments, although the case where the learning device 120 is composed of one computer is described, it may be composed of a plurality of computers. Similarly, in the above-mentioned first to fourth embodiments, although the case where the inference device 130 is composed of one computer is described, it may be composed of a plurality of computers.

[0238] In addition, in the first to fourth embodiments, although the learning device 120 and the inference device 130 are applied to the semiconductor manufacturing process, it is obvious that they can also be applied to other processes other than the semiconductor manufacturing process. The other processes other than the semiconductor manufacturing process mentioned here include other manufacturing processes other than the semiconductor manufacturing process and non-manufacturing processes.

[0239] In the first to fourth embodiments described above, the learning device 120 and the inference device 130 are implemented by causing a general-purpose computer to execute various programs. However, the implementation method of the learning device 120 and the inference device 130 is not limited thereto.

[0240] For example, it can be realized by a dedicated electronic circuit (i.e., hardware) such as an IC (integrated circuit) equipped with a processor, memory, etc. Multiple components can be realized by one electronic circuit, one component can be realized by multiple electronic circuits, and components and electronic circuits can be realized in a one-to-one manner.

[0241] It should be noted that other elements may be combined with the configurations listed in the above embodiments, and the present invention is not limited to the configurations shown here. In this regard, changes may be made within the scope of the present invention and may be appropriately determined according to its application.

[0242] This application claims priority based on Japanese invention patent application No. 2018-164931 filed on September 3, 2018, and this application cites the entire contents of the Japanese invention patent application.

[0243] Explanation of symbols

[0244] 100: Simulation system

[0245] 110: Semiconductor manufacturing equipment

[0246] 111: Measurement device

[0247] 112: Measurement device

[0248] 120: Learning device

[0249] 121: Data configuration unit

[0250] 122: Learning Department

[0251] 130: Inference Device

[0252] 131: Data configuration unit

[0253] 132: Execution Department

[0254] 300: Learning data

[0255] 420: Learning Model for Dry Etching

[0256] 421: Deposition Learning Model

[0257] 430: Comparison Department

[0258] 440: Change Department

[0259] 501: Shape data acquisition unit

[0260] 502: Channel specific data generation unit

[0261] 511: One-dimensional data acquisition unit

[0262] 512: One-dimensional data expansion unit

[0263] 520: Connection

[0264] 600: Image data before processing

[0265] 601~605: Channel specific data

[0266] 610: Parameter data

[0267] 611~613: Parameter data in two-dimensional arrangement

[0268] 620: Environmental information

[0269] 621~623: Two-dimensional arrangement of environmental information

[0270] 630: Link data

[0271] 700: Output result

[0272] 920: Learning model for dry etching

[0273] 921: Deposition using learning completion model

[0274] 930: Output

[0275] 1200: Data configuration unit

[0276] 1201: Connection

[0277] 1300: Learning model for dry etching

[0278] 1301: Neural Network Department

[0279] 1310: Link data

[0280] 1400: Learning Department

[0281] 1510, 1520: Data configuration unit

[0282] 1511: Compression

Claims

1. A method for inferring a manufacturing process, wherein: at least one processor inputs pre-processed image data of a second object and data on a process corresponding to a manufacturing process for the second object into a learning completed model, and infers post-processed image data of the second object after the process corresponding to the manufacturing process is performed, The learning completion model updates model parameters based on the output when pre-processed image data of the first object and data about the processing corresponding to the manufacturing process for the first object are input as input data for learning, and difference information of post-processed image data of the first object after the processing corresponding to the manufacturing process is performed to make the output and the post-processed image data of the first object close, thereby performing learning.

2. The inference method according to claim 1, wherein: The at least one processor processes the data related to the process corresponding to the manufacturing process into a form corresponding to the pre-processing image data of the second object, The at least one processor inputs processed data on the process corresponding to the manufacturing process into the learning completed model.

3. The inference method according to claim 2, wherein: The format corresponding to the pre-processed image data of the second object is a two-dimensional arrangement format corresponding to the longitudinal size and the lateral size of the pre-processed image data of the second object.

4. The inference method according to any one of claims 1 to 3, wherein: The pre-processed image data of the second object input to the learned model has a plurality of channels corresponding to materials included in the second object.

5. The inference method according to claim 1, wherein: The at least one processor re-inputs the processed image data of the second object and data on the process corresponding to the manufacturing process into the learning completed model, and infers the processed image data of the second object after the process corresponding to the manufacturing process is re-performed.

6. The inference method according to claim 1 or 5, wherein: The processing corresponding to the manufacturing process is processing of one step among the steps included in the manufacturing process.

7. The inference method according to claim 1 or 5, wherein: The learning completion model includes a neural network.

8. The inference method according to claim 7, wherein: The neural network outputs processed image data of the second object.

9. The inference method according to claim 7, wherein: The at least one processor generates processed image data of the second object based on the pre-processed image data of the second object and an output from the neural network.

10. The inference method according to claim 9, wherein: The neural network outputs information on a rate of change relative to pre-processed image data of the second object.

11. The inference method according to claim 1 or 5, wherein: The process corresponding to the manufacturing process is a process corresponding to the semiconductor manufacturing process.

12. The inference method according to claim 11, wherein: The processing corresponding to the semiconductor manufacturing process includes at least one of etching and deposition.

13. The inference method according to claim 11, wherein: The data on the process corresponding to the manufacturing process includes information on parameters indicating process conditions when the semiconductor manufacturing apparatus executes the process corresponding to the semiconductor manufacturing process.

14. The inference method according to claim 11, wherein: The data regarding the process corresponding to the manufacturing process includes environmental information measured when a semiconductor manufacturing apparatus executes the process corresponding to the semiconductor manufacturing process.

15. The inference method according to claim 13, wherein: The parameter includes at least one of a setting value set in the semiconductor manufacturing apparatus and a hardware configuration of the semiconductor manufacturing apparatus.

16. The inference method according to claim 14, wherein: The environmental information includes at least any one of data on current, data on voltage, data on light, data on temperature, and data on pressure measured in the semiconductor manufacturing apparatus.

17. An inference device for a manufacturing process, comprising at least one memory and at least one processor, wherein: The at least one memory is used to store a learning model, wherein the learning model updates model parameters based on an output when pre-processed image data of a first object and data on a process corresponding to a manufacturing process for the first object are input as input data for learning, and difference information of processed image data of the first object after the process corresponding to the manufacturing process is performed so that the output and the processed image data of the first object are close, thereby learning is performed, The at least one processor is configured to perform the following steps: inputting pre-processed image data of the second object and data about the processing corresponding to the manufacturing process for the second object into the learning completion model, and inferring post-processed image data of the second object after the processing corresponding to the manufacturing process is performed.

18. The inference device according to claim 17, wherein: The at least one processor is configured to further perform the following steps: re-input the processed image data of the second object and the data about the processing corresponding to the manufacturing process into the learning completion model, and infer the processed image data of the second object after the processing corresponding to the manufacturing process is re-performed.

19. A method for generating a model for a manufacturing process, wherein: At least one processor inputs pre-processed image data of an object and data about a processing corresponding to a manufacturing process on the object into a learning model as input data for learning, and updates model parameters based on the output of the learning model and difference information between processed image data of the object after the processing corresponding to the manufacturing process is performed to make the output and the processed image data close, thereby enabling the learning model to learn.

20. A learning device for a manufacturing process, comprising at least one memory and at least one processor, wherein: The at least one processor is configured to perform the following steps: inputting pre-processed image data of an object and data about a processing corresponding to a manufacturing process on the object into a learning model as input data for learning, and updating model parameters based on the output of the learning model and difference information between processed image data of the object after the processing corresponding to the manufacturing process is performed to make the output and the processed image data close, thereby enabling the learning model to learn.

Images