CN118691483A - Pattern modeling system and pattern modeling method - Google Patents

Abstract

A pattern modeling system and a pattern modeling method are provided. The pattern modeling method comprises the following steps: generating first image data based on a sample pattern learned by a Deep Neural Network (DNN), generating second image data by measuring the first image data, determining a region of the second image data to which a weight filter is to be applied, training the DNN by applying the weight filter to the determined region of the second image data, and predicting at least one pattern image based on a result of training the DNN.

Description

Pattern modeling system and pattern modeling method

This application is based on and claims the benefit of priority from Korean Patent Application No. 10-2023-0039293 filed on March 24, 2023, in the Korean Intellectual Property Office, the disclosure of which is incorporated herein in its entirety by reference.

Technical Field

The present disclosure relates to a pattern modeling system and a pattern modeling method, and in particular, to a system and a method capable of performing training of a pattern modeling system by applying a weight filter.

Background Art

Photomasks can be used to print integrated circuit (IC) layouts on wafers in photolithography processes in the manufacture of semiconductor devices. Photolithography processes generally use a method of transferring a mask pattern formed on a photomask to a wafer through an optical lens. The photomask may include a transparent area and an opaque area. The transparent area may be formed by etching a metal layer on the photomask, and light may pass through the transparent area. On the other hand, the opaque area may not allow light to pass through. The mask pattern may be formed by a transparent area and an opaque area. The light emitted by a light source may be emitted onto the wafer through the mask pattern of the photomask, and therefore, the IC layout may be printed on the wafer.

As the integration of semiconductor devices increases, the distance between mask patterns of photomasks may become shorter, and the width of transparent regions may become very narrow. Due to this proximity, interference and diffraction of light may occur, and thus, a distorted layout different from the desired layout may be printed on the wafer.

The information disclosed in this background technology section is known to the inventor or derived by the inventor before or during the process of implementing the embodiments of the present application, or is technical information acquired during the process of implementing the embodiments. Therefore, it may contain information that does not constitute prior art known to the public.

Summary of the invention

A pattern modeling system is provided that can improve the prediction of measurement data.

Additional aspects will be set forth in part in the description which follows and, in part, will be apparent from the description, or may be learned by practice of the presented embodiments.

According to an aspect of an example embodiment, a pattern modeling method for predicting image data may include: generating first image data based on a sample pattern learned by a deep neural network (DNN); generating second image data by measuring the first image data; determining an area of the second image data to which a weight filter will be applied; training the DNN by applying the weight filter to the determined area of the second image data; and predicting at least one pattern image based on a result of the training DNN.

According to one aspect of an example embodiment, a pattern modeling system for predicting image data may include: a memory storing instructions; and at least one processor configured to execute the instructions to: generate first image data based on a sample pattern learned by a DNN, the DNN including multiple layers, generate second image data by measuring the first image data, determine an area of the second image data to which a weight filter will be applied, train the DNN by applying the weight filter to the determined area of the second image data, and predict at least one pattern image based on a result of the training DNN.

According to one aspect of an example embodiment, a non-transitory computer-readable storage medium may store instructions that, when executed by at least one processor, cause the at least one processor to: generate first image data based on a sample pattern learned by a DNN, the DNN comprising a plurality of layers; generate second image data by measuring the first image data; determine an area of the second image data to which a weight filter will be applied; train the DNN by applying the weight filter to the determined area of the second image data; and predict at least one pattern image based on a result of the training DNN, the weight filter being applied to an area corresponding to a feature portion of the second image data.

BRIEF DESCRIPTION OF THE DRAWINGS

Additional aspects will be set forth in part in the description which follows and, in part, will be apparent from the description, or may be learned by practice of the presented embodiments in conjunction with the accompanying drawings, in which:

FIG1 is a block diagram of a pattern modeling system according to an embodiment;

FIG2 is a block diagram of a pattern modeling system according to an embodiment;

3 is a flow chart illustrating a modeling method performed by a pattern modeling system according to an embodiment;

FIG4 is a diagram illustrating a method of generating first image data according to an embodiment;

5 is a diagram illustrating a method of determining a weight filter area (WFA) in second image data according to an embodiment;

FIG6A is a diagram illustrating a method of determining a WFA according to an embodiment;

FIG6B is a diagram of an image to which WFA is applied according to an embodiment;

6C illustrates an example of generating first image data and second image data from a sample pattern according to an embodiment;

FIG7 shows simulation results of applying a weight filter according to an embodiment;

FIG8 illustrates an example of a loss function module according to an embodiment;

FIG9 illustrates an example of a deep neural network (DNN) according to an embodiment;

10 is a diagram showing an application example of the pattern modeling system according to the embodiment;

11 is a flowchart illustrating a method of manufacturing a semiconductor device using a pattern modeling system according to an embodiment; and

FIG. 12 is a flowchart illustrating a method of manufacturing an integrated circuit (IC) according to an embodiment.

DETAILED DESCRIPTION

Hereinafter, the disclosed exemplary embodiments will be described in detail with reference to the accompanying drawings. The same reference numerals are used for the same components in the accompanying drawings, and their redundant descriptions will be omitted. The embodiments described herein are exemplary embodiments, and thus the disclosure is not limited thereto, and can be implemented in various other forms.

As used herein, when a statement such as "at least one of..." follows a list of elements, the statement modifies the entire list of elements and does not modify the individual elements in the list. For example, the statement "at least one of a, b, and c" should be understood to include: only a, only b, only c, both a and b, both a and c, both b and c, or all of a, b, and c.

FIG1 is a block diagram of a pattern modeling system 100 according to an embodiment. As shown in FIG1 , the pattern modeling system 100 may include a memory 110 and a processor 120. However, the configuration shown in FIG1 is an example for implementing the embodiment, and as will be understood by a person of ordinary skill in the art from the disclosure herein, other hardware and software configurations may be additionally included in the pattern modeling system 100. According to an example, the pattern modeling system 100 may be implemented in the form of an electronic device.

According to the disclosure, the pattern modeling system 100 can be applied to the simulation operation of the semiconductor patterning process. According to an example, the semiconductor patterning process can be a photolithography process. According to the disclosure, the pattern modeling system 100 can learn the simulation results of the sample pattern used in the semiconductor patterning process to generate image data with high precision prediction about the critical dimension (CD), etc. According to the disclosure, the pattern modeling system 100 may include a learning model and can be trained by applying an image to which a weight filter is applied to a specific area. Thus, the pattern modeling system 100 can generate image data with high precision prediction about the specific area.

The memory 110 may store commands or data related to at least one other component of the pattern modeling system 100. In addition, the memory 110 may be accessed by the processor 120, and reading/writing/modifying/deleting/updating of data may be performed by the processor 120.

In the disclosure, the term memory may include the memory 110, a read-only memory (ROM) in the processor 120, or a random access memory (RAM) in the processor 120, or a memory card (e.g., a micro secure digital (SD) card or a memory stick) installed in the pattern modeling system 100. In addition, the memory 110 may store programs and data for configuring various screens to be displayed on a display area of the display.

According to an example, the memory 110 may include a non-volatile memory capable of maintaining stored information even if power supply is interrupted, and a volatile memory requiring a continuous power supply to maintain stored information. For example, the non-volatile memory may be implemented as at least one of a one-time programmable ROM (OTPROM), a programmable ROM (PROM), an erasable programmable ROM (EPROM), an electrically erasable programmable ROM (EEPROM), a mask ROM, or a flash ROM, and the volatile memory may be implemented as at least one of a dynamic RAM (DRAM), a static RAM (SRAM), or a synchronous dynamic RAM (SDRAM).

According to an example, the memory 110 may store a pre-trained neural network model. The pre-trained neural network model may be a model that is sent to the pattern modeling system 100 after being trained in an external server. Alternatively, the pre-trained neural network model may be a model trained within the pattern modeling system 100. The neural network model may include a plurality of layers and may include weight data learned by the server. According to an example, the neural network model may be a generative adversarial network (GAN) model. In the disclosure, the expression of the neural network model and the expression of the learning model may be used interchangeably.

The processor 120 may be electrically connected to the memory 110 to control all operations and functions of the pattern modeling system 100. The processor 120 may obtain a loss function using output data and labels, the output data being obtained by inputting training data into a pre-trained neural network model stored in the memory, the labels corresponding to the training data. The labels may represent actual values to be output when the training data is input into the neural network model, and the loss function may represent a function that quantifies the difference between the output data and the labels.

The processor 120 may obtain the magnitude of the weight change of each of the multiple layers included in the neural network model based on the loss function. The weight change of each of the multiple layers may represent the value that the weight of each of the multiple layers needs to be changed in order to minimize the value of the loss function. The magnitude of the weight change may be expressed as the magnitude of the weight loss, the magnitude of the anti-derivative, or the magnitude of the differential value (e.g., the L2 norm value of the derivative).

According to the disclosure, the processor 120 can calculate a loss function based on an image in which a weight filter is applied to a specific area thereof, train a neural network model stored in the memory 110 based on the loss function, update the magnitude of the weight change, and obtain a prediction of the specific area.

Functions related to artificial intelligence (AI) disclosed herein may be performed by the processor 120 and the memory 110. The processor 120 may include one processor or multiple processors. The one processor or multiple processors may include a general-purpose processor (such as a central processing unit (CPU), an application processor (AP), a digital signal processor (DSP), etc.), a graphics-specific processor (such as a graphics processing unit (GPU), a visual processing unit (VPU), etc.) or an AI-specific processor (such as a neural processing unit (NPU)).

The processor 120 or the processors 120 may process the input data according to predefined operating rules or AI models stored in the memory 110. Alternatively, when the processor or the processors include an AI dedicated processor, the AI dedicated processor may be designed to have a hardware structure dedicated to processing a specific AI model.

Predefined operating rules or AI models may be generated through training. When an AI model is generated through training, this may indicate that a base AI model is trained based on a learning algorithm using a plurality of training data sets, so that a predefined operating rule or AI model configured to perform a desired characteristic (or purpose) is generated. Such training may be performed by a device on which the disclosed AI is implemented, or by a separate server and/or system.

Examples of learning algorithms may include, but are not limited to, supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning.

The AI model may include multiple neural network models, and the neural network model may include multiple layers. Each of the multiple neural network layers may have multiple weight values, and the neural network operation may be performed by the operation between the operation result of the previous layer and the multiple weight values. The multiple weight values of the multiple neural network layers may be optimized by the training results of the AI model. For example, the multiple weight values may be updated to reduce or minimize the weight loss value (e.g., the magnitude of the weight change) or cost value obtained in the AI model during the training process.

FIG. 2 is a block diagram of a pattern modeling system according to an embodiment.

2, the pattern modeling system 200 may include a deep neural network (DNN) 210, a loss function module 220, and a data preprocessor 230. According to an example, the DNN 210 may include a learning model. According to an example, the DNN 210 may include a GAN model. The DNN 210 may include a generator that generates a pseudo image from input data and a classifier that identifies the pseudo image. For example, the generator may output a pseudo image and the classifier may output a probability of being a real image (or a probability of being a pseudo image). The classifier may be trained to identify the pseudo image based on the real image and the pseudo image, and the generator may be trained so that the classifier identifies the pseudo image generated by the generator as a real image. Therefore, the trained generator can generate a pseudo image that is very similar to the real image.

For example, the input data received by the generator may be an image extracted from layout data obtained by designing a pattern to be formed by a semiconductor process design. The generator may generate a pattern expected to be formed on a semiconductor substrate by performing a semiconductor process based on the layout data as an output image by using the input data. The classifier may compare the output image with the pattern image. The pattern image may be an image of an actual pattern formed by performing a semiconductor process based on layout data to which an optical proximity correction (OPC) model is applied, and may be a scanning electron microscope (SEM) image or a transmission electron microscope (TEM) image.

DNN 210 may include multiple layers, and weight data between multiple layers may be updated by loss function module 220. Loss function module 220 may be configured as a processing component that implements various operations corresponding to the loss function as described in detail throughout the specification (for example, this is described in detail below with reference to FIG. 9). In the disclosure, DNN and deep learning network may be used interchangeably. According to an example, data learned by DNN 210 may be represented as training data.

According to an example, the loss function module 220 may calculate a difference between the image data learned by the DNN 210 and the image data processed by the data preprocessor 230, and transmit the difference to the DNN 210. According to an example, the image data to be calculated in the loss function module 220 may be image data to which a weight filter will be applied by the data preprocessor 230. This is described in detail below with reference to FIG. 8.

According to an example, the data preprocessor 230 may generate and preprocess data about a sample pattern that is a learning target of the pattern modeling system 200. According to an example, the data preprocessor 230 may generate image data corresponding to the sample pattern, and determine an area of the generated image data to which a weight filter will be applied. The data preprocessor 230 may determine an area to which a weight filter will be applied, and apply the weight filter to the generated image data. This will be described in detail with reference to FIGS. 4A to 7.

According to an example, the image data as the learning target of the pattern modeling system 200 may include at least one of a design layout, a resist image, an aerial image, a slope map, a density map, or a photon map. The design layout may represent a bitmap image or an image of any other appropriate format configured to implement a target pattern on a wafer. The resist image may be an image of a photoresist derived from the design layout by simulation. The aerial image may be an image representing the intensity distribution of the exposure light reaching the photoresist derived from the design layout. The slope map may be an image in which the value of each pixel included in the slope map is the gradient of each pixel of the aerial image. The density map may be an image in which the bit value of a specific pixel is determined by the pattern density near the specific pixel. The photon map may be an image obtained by simulating the number of photons reaching each pixel during the exposure process.

In some cases, the design layout, resist image, aerial image, slope map, density map, or photon map in the training operation of DNN 210 may be referred to as a training data set. The training data set may be related to the design layout that has been transferred to the wafer and may include SEM images.

According to an example, a semiconductor process may be performed based on layout data. Various patterns may be formed on a semiconductor substrate by transmitting layout data to generate an exposure process of a mask and an etching or deposition process using the mask generated in the exposure process. In order to minimize the difference between the layout data and the pattern formed on the semiconductor substrate, a proximity correction may be applied. According to the disclosure, the pattern modeling system 200 may learn image data that can minimize the difference and generate prediction data to reduce errors when performing proximity correction. When receiving a sample pattern, the pattern modeling system 200 according to the disclosure may predict a nano-geometry research (NGR) measurement image corresponding to the corresponding pattern. According to the disclosure, in order to generate a model for better predicting the CD of a specific main pattern while using a deep learning model training method, the pattern modeling system 200 may perform training by applying a weight filter described in the disclosure so that the deep learning model may increase the predictive power of the CD of the measurement data. The measurement data may be data measured by a sample pattern included in the pattern modeling system 200. The measurement data may reflect the results that the pattern modeling system 200 has learned. The disclosed pattern modeling system 200 can generate a deep learning model that maintains accurate CD values of anchor patterns or specific locations.

FIG. 3 is a flowchart illustrating a modeling method performed by a pattern modeling system according to an embodiment.

In operation S310, first image data may be generated based on imaging information about a sample pattern. The sample pattern may be a pattern for generating an image that is a learning target of a pattern modeling system. The sample pattern may be a pattern of a photomask that is a prediction target of a pattern modeling system. The shape of the sample pattern is not limited to a specific shape and may be provided in various shapes. The first image data may represent image data that may be formed by the sample pattern. According to an example, the first image data may be image data obtained by simulating the sample pattern.

In operation S320, the second image data may be generated by measuring the first image data. According to an example, the first image data may be image data obtained by simulating a photolithography process using a sample pattern. The second image data may be image data obtained as a result of performing simulation measurement by applying various conditions to the first image data. According to an example, the second image data may be image data obtained as a result of measuring the first image data by applying process conditions, focus position, dose, etc. differently to the first image data.

According to an example, in operation S310 or operation S320, the first image data and the second image data may be image data obtained by converting a target layout or a measurement profile image into a dithered image.

In operation S330, the area of the second image data to which the weight filter will be applied may be determined. In the disclosure, the area to which the weight filter will be applied may be described interchangeably with terms such as weight filter area (WFA). The area to which the weight filter will be applied may represent an area with a higher weight to be learned by the disclosed pattern modeling system. According to an example, the area to which the weight filter will be applied may be different according to the characteristics of the second image data. According to an example, the area to which the weight filter will be applied may be an anchor pattern. Alternatively, the area to which the weight filter will be applied may be any one of the main patterns included in the modeling system. Alternatively, the area to which the weight filter will be applied may be an area in which a large distribution is displayed. The area to which the weight filter will be applied is not standardized, and may be a portion of the image data that is desired to be emphasized and learned.

In operation S340, DNN training may be performed by applying a weight filter to the second image data and the reference image data. According to an example, when the area to which the weight filter is applied is determined, a weight filter image in which the weight is applied to the corresponding area may be generated. According to an example, the reference image data may be the first image data or pre-stored sample image data. Alternatively, the reference image data may be contrast target data output by training. According to an example, a loss function may be determined by applying a weight filter to each of the reference image data and the second image data and comparing the difference therebetween, and DNN training may be performed. Thus, training may be performed by applying the weight filter to the area.

In operation S350, a pattern image of a pattern of a semiconductor device may be predicted based on the DNN training result. According to an example, the pattern image obtained as a result of the DNN training on the sample pattern may be an image with higher accuracy than before training on the region to which the weight filter is applied.

3, with respect to the image data obtained by simulation of the sample pattern, the area to which the weight filter will be applied can be determined, and the loss function calculation and DNN training can be performed based on the image data to which the weight filter is applied. Thus, the prediction data with high accuracy with respect to the sample pattern can be output, and the prediction of the pattern image can be fully ensured.

FIG. 4 is a diagram illustrating a method of generating first image data according to an embodiment.

4 , (a) shows an example of a sample pattern SP, and (b) and (c) show first image data D1 and D1' generated from the sample pattern SP.

According to an example, the sample pattern SP may be provided by a designed mask layout. The mask layout may include a sample pattern SP required to print an integrated circuit (IC) on a wafer. The sample pattern SP may define a planar shape of a cell pattern to be formed in a cell array region of the wafer.

According to an example, the image data generated by the sample pattern SP may be an image obtained by SEM. According to an example, the SEM image may be generated from an NGR device or an SEM device manufactured by NGR Corporation. The SEM image may be an image of a photoresist pattern generated by after-development inspection (ADI) or an image of an actual circuit pattern generated by after-cleaning inspection (ACI).

In the sample pattern SP in (a), it can be confirmed that the CD is measured to be 86.24. According to the example, in (b), an ACI (ADI) contour image for training can be generated from the sample pattern SP. According to the example, the ACI (ADI) contour image for training in (b) can be transmitted as a layout file in an NGR (SEM) device. The layout file may be a graphic design system (GDS) file. For deep learning training, it may be necessary to convert the layout file into a dithering process of image data. Thus, the first image data D1' in (c) can be generated. That is, the first image data D1' in (c) may be image data obtained by dithering the GDS file.

According to the example, when dithering is performed from the GDS file to the image data, the coordinate information as the position of the measured CD disappears. The GDS file displays the information of all polygons as coordinates, but the image data is expressed in pixels, so the coordinate information disappears.

Therefore, the measured coordinate information may not be included in the generated first image data D1 and D1'. According to an example, when the coordinate information that has disappeared is an anchor pattern responsible for process reference, accurately predicting the CD of the corresponding position is an important performance indicator of the model's predictive power, and training may be required by emphasizing the CD of the corresponding position.

FIG. 5 is a diagram illustrating a method of determining a WFA in second image data according to an embodiment.

According to an example, the second image data D2 of FIG. 5 may be an image data result obtained by applying the measurement value to the first image data. According to the result of the second image data D2, an area to which a weight filter is applied may be selected. In the example of FIG. 5 , WFA is shown. Referring to the example of FIG. 5 , WFA may represent a CD area that may be measured in the second image data D2.

Fig. 6A is a diagram illustrating a method of determining a WFA according to an embodiment. Fig. 6B is a diagram of an image to which a WFA is applied according to an embodiment.

6A, the second image data to which WFA is applied is shown on the left, and image data WFI_1 obtained by applying WFA to the second image data is shown on the right. In FIG6B, the second image data to which WFA is applied is shown on the left, and image data WFI_2 obtained by applying WFA to the second image data is shown on the right.

6A , there is shown an example in which, when measuring the CD in the second image data, a region corresponding to the CD is determined as the WFA.

6B , there is shown an example in which, when the distribution is measured in the second image data, a region in which a plurality of distributions are distributed is determined as a WFA.

6A and 6B, when a WFA is determined in the second image data, weight filter images WFI_1 and WFI_2 to which the determined WFA is applied may be generated. The examples of setting the WFA shown in FIGS. 6A and 6B may correspond to some embodiments. According to an example, when the WFA is set, the sample pattern SP may be set to have different positions and backgrounds. According to another example, the WFA may be set to be the same as the position of the measurement point. According to another example, the WFA may be set to correspond to the shape of a specific pattern of the second image data. According to another example, the weight to be applied to the set WFA may be adjusted.

FIG. 6C illustrates an example of generating first image data and second image data from a sample pattern according to an embodiment.

6C , the sample pattern SP may be provided in a designed form, and a plurality of first image data D1 may be generated through the sample pattern SP. When the plurality of first image data D1 are generated, the second image data D2 may be generated by measuring the plurality of first image data D1.

FIG. 7 shows simulation results of applying a weight filter according to an embodiment.

7 , by generating an image in which 100% of the weight is reflected in the center of the image and only 5% of the weight is reflected in the background, image data in which a weight is applied to a desired portion may be generated.

According to an example, in order to reflect the weight on a specific position of the image data, the coordinates of the GDS may be first changed to pixel positions. To automate this, the measured image and the jittered image may be indexed. Thereafter, the jittered image having the same index and measured coordinates may be changed to pixel positions.

According to the disclosure, a weight layer image corresponding to the weight filter can be generated. The weight layer image can be generated in the form of grayscale 0 to 255. This is to multiply the measured jitter image with each pixel. The grayscale can store the value of 0 to 255 (256) in each pixel as an integer, and the value of 0 to 255 (256) can be easily normalized to 0 to 1.

That is, according to the disclosure, the weight layer image can be used to convolve and normalize the measured jitter image. According to the example, when the user intuitively sets the weight value, the weight value can be set to a normalized value between 0 and 1. Finally, when the weight filter is applied only to the measured pixels of the weight layer to multiply the jitter image of the measured contour with the weight layer image, only the image of the measured portion can be retained as the training result. Therefore, when generating the weight layer image, it may be necessary to reflect the background value.

7 , a weight filter image may be generated by setting a weight to 100% and a background value to 5% in WFA, and a specific value may be adjusted.

Then, the corresponding image data can be used to perform deep convolutional GAN (DCGAN) training. According to the disclosure, the trained data can be used to perform optical proximity correction (OPC) and process proximity correction (PPC). That is, deep learning that effectively matches the CD of the anchor pattern and the CD of the important pattern can be used to generate the OPC prediction model and the PPC prediction model.

Fig. 8 shows an example of a loss function module according to an embodiment. Referring to the loss function module 221 of Fig. 8, the difference value may be transmitted to the DNN 211. The loss function equation implemented by the loss function module 221 of Fig. 8 may be represented by equation (1).

Loss(NetG _target ,NetG _sim )×Weight (1)

According to an example, NetG _target may be second image data obtained as a result of measurement, and NetG _sim may be reference image data. Weight may represent a weight layer image to which WFA is applied. According to an example, a weight filter may be applied to each of the second image data and the reference image data. According to an example, a difference value (or difference image) diff_B obtained by convolving the second image data with the weight filter image and convolving the reference image data with the weight filter image may be transmitted to DNN 211. According to an example, the weight filter image applied to the second image data may be the same as the weight filter image applied to the reference image data.

According to the disclosure, a weight filter image of a specific portion of the extracted area can be generated according to the input size and used to calculate the loss function of the back propagation of the DNN 211 to reflect the weight of the specific pixel area. According to an example, when changing the image of the contour layout used for image-to-image training, the engineer can select an area that has been measured or a desired area, and generate a weight filter image of the same size as the additional input image with respect to the selected area.

When comparing a simulated image generated as an output after performing forward training of DNN 211 with a discriminator image, the generated weight filter image can be used. According to an example, the simulated image can be a pseudo image, and the discriminator image can be a real image. According to an example, the simulated image can be reference image data, and the discriminator image can be second image data. When there is no weight filter image, a difference image diff_A can be generated, but the weight filter image can be used to generate a difference value diff_B of the area of interest to the user. As a result of the loss function module 221 applying the loss function, the weight data between the layers can be updated by GAN back propagation.

According to the disclosure, a model training method is provided for generating a weight filter about the position of an image pixel and using the weight filter for training during image deep learning modeling of a semiconductor patterning process. According to the disclosure, in the process of training deep learning modeling based on image data, a weight filter about a specific pattern or a specific area can be generated and used to adjust the pixel weight of each area, which can be used to improve model accuracy and expand data mining.

9 shows an example of a DNN according to an embodiment. According to an example, the DNN 211 may include a plurality of layers ch1, ch2, ch3, ... The input image in may be input to the DNN 211. The weight data W may be reflected between the plurality of layers ch1, ch2, ch3, ... The weight data W of the disclosed DNN 211 may be updated by applying the weight filter image to the result of the output data out of the DNN 211.

According to the disclosed DNN 211, a structure that allows pixel correlation over a long distance may be provided by including a plurality of down sample layers (e.g., down1 and down2). Because the down sample layer implies pattern information, the input image may be reduced by half in the output layer each time the down sample layer passes through a pixel. However, the reduced image still corresponds to the same area as the input image, and the information represented may correspond to twice the input image (or 4 times with respect to the area). Therefore, even when the kernel of the same size is used, the kernel acting on the image that has passed through more down sample layers may represent pixel correlation about a wider area. For example, when the image of the first down sample layer is about 1/4 of the input image (1/16 in the area concept), the kernel corresponding to the second down sample layer may cover a very small area in the input image, but may cover most of the area in the image of the second down sample layer. According to some embodiments, DNN211 may also include other layers, such as up sample layers (up1 and up2), scaling layers (Scale(S1) and Scale(S2)), and the like.

According to the disclosure, the DNN 211 can reflect the influence of a pattern located far away, and the output image can generate a high-precision image that can ensure sub-pixel-level accuracy. According to the disclosure, the pattern complexity is different for each layer of the semiconductor device, so that modeling can be achieved by changing the structure.

The image-based model according to the disclosed DNN 211 may be able to be trained one-to-one and may be trained to predict one answer. According to the disclosure, in order to guide the training to match a specific CD or better predict an area with a poor distribution, a weight filter may be used to configure the loss function, which is an additional tool between the real image and the generated image. According to an example, when training with newly processed data, the disclosed DNN 211 may use only areas suitable for data consistency for training.

Fig. 10 is a diagram for explaining an application example of a pattern modeling system according to an embodiment. Referring to Fig. 10 , the pattern modeling system according to the disclosure may be able to perform deep learning modeling on a pattern with few measurement images.

10, a deep learning model can be generated in the presence of only measurement results and a small number of measurement images. The pattern modeling system according to the disclosure can be applied to use simulated images for specific patterns without measurement images or to generate a model that increases the prediction accuracy of CD or pattern shape in a specific area.

1002 of FIG. 10 shows a case where only a contour image exists, and 1004 of FIG. 10 shows a case where a measurement image exists but the number of the measurement images is small. When the measurement image exists but the number of the measurement images is insufficient for training, repeated training may be performed on the area of the measurement image using the weight filter. According to an example, training may be performed by setting a high weight and a high background value for the measurement image and a low background value for the contour image.

According to an example, assuming that there are 259 measurement images and 1984 CD values are measured by simulation during modeling, when training is performed by setting all pixels of the CD values configured by simulation to the same weight, an overfitting problem may occur. In this case, according to the disclosure, training can be performed on the generated pattern for the measured CD values, and model training can be performed in a mixed form with samples having existing measurement images.

FIG. 11 is a flowchart illustrating a method of manufacturing a semiconductor device using a pattern modeling system according to an embodiment.

In operation S1110, data of a sample pattern to be learned may be input. In this case, the input data may be first image data generated based on the sample pattern.

In operation S1120, the area to which the weight filter is applied may be determined by preprocessing the data. According to an example, the preprocessing of the data may include generating first image data and generating second image data by measuring the first image data. The area to which the weight filter is applied may be determined according to characteristics of the second image data.

In operation S1130, the pattern data may be learned by applying the weight filter. In operation S1130, the image data to which the weight filter is applied may be applied to a loss function.

In operation S1140, OPC and PPC may be performed by applying the learned data. The learned data has higher accuracy with respect to the region to which the weight filter is applied, and thus, OPC and PPC may be more effectively performed on the learned data.

OPC and PPC may be performed based on lithography simulation, which predicts the profile image and CD that will be formed on the wafer according to the designed mask layout.

According to the disclosure, in order to change (reflect) the weight of a specific position of an image during deep learning modeling, a weight filter image in pixels can be set. The weight filter image can be image data for training to which a weight filter is applied. According to an example, the weight filter image can be used to calculate a loss function during deep learning training and to update weights during back propagation calculations.

FIG. 12 is a flowchart illustrating a method of manufacturing an IC according to an embodiment.

12, the standard cell library D10 may include information about the standard cell (e.g., function information, characteristic information, layout information, etc.). The standard cell library D10 may include data DC defining the layout of the standard cell. The data DC may include data defining the structure of the standard cell that performs the same function and has different layouts. The data DC may include first data DC1 that performs a first function and defines the structure of the standard cell with different layouts, and nth data DCn (n is a natural number of 2 or greater) that performs an nth function and defines the structure of the standard cell with different layouts.

In operation S10, a logic synthesis operation may be performed to generate netlist data D20 from ready-to-learn (RTL) data D11. For example, a semiconductor design tool (e.g., a logic synthesis module) may perform a logic synthesis operation according to RTL data D11 with reference to a standard cell library D10, thereby generating netlist data D20 including a bitstream or a netlist, the RTL data D11 being written in a very high speed IC (VHSIC) hardware description language (VHDL) and a hardware description language (HDL) (such as Verilog). An RTL design may be developed according to a design verification method and a computing system, and the RTL data D11 may be generated.

The standard cell library D10 may include data DC that perform the same function and define structures of standard cells having different layouts, and the standard cells may be included in the IC with reference to such information during logic synthesis.

In operation S20, a place and route (P&R) operation of generating layout data D30 according to netlist data D20 may be performed. Layout data D30 may have a format such as GDSII, and may include geometric information of standard cells and interconnections.

For example, a semiconductor design tool (e.g., a P&R module) may lay out a plurality of standard cells with reference to the standard cell library D10 from the netlist data D20. The semiconductor design tool may select one of the layouts of the standard cells defined by the layout data D103 with reference to the data DC, and may lay out the selected layout of the standard cell.

In operation S20, an operation of generating an interconnection may also be performed. The interconnection may electrically connect an output pin and an input pin of the standard cell, and may include, for example, at least one through-hole and at least one conductive pattern.

In operation S30, OPC may be performed. OPC may refer to an operation of forming a pattern of a desired shape by correcting a distortion phenomenon such as refraction caused by the characteristics of light in photolithography included in a semiconductor process for manufacturing an IC, and the pattern on the mask may be determined by applying OPC to the layout data D30. In some embodiments, the layout of the IC may be limitedly modified in operation S30, and the limited modification of the IC in operation S30 is a post-processing for optimizing the structure of the IC, and may be referred to as design polishing. According to an embodiment, the layout data D30 may be layout data learned and predicted by the pattern modeling system 100 of FIG. 1. Using the layout data learned and predicted by the pattern modeling system 100, a result with higher pattern accuracy may be obtained.

In operation S40, an operation of manufacturing a mask may be performed. For example, when OPC is applied to layout data D30, a pattern on a mask may be defined to form patterns formed on a plurality of layers, and at least one mask (or photomask) may be manufactured to form respective patterns of the plurality of layers.

In operation S50, an operation of manufacturing an IC may be performed. For example, the IC may be manufactured by patterning a plurality of layers using at least one mask manufactured in operation S40. Operation S50 may include operations S51 and S52.

In operation S51, a front-end-of-line (FEOL) process may be performed. FEOL may refer to a process of forming individual elements (e.g., transistors, capacitors, resistors, etc.) on a substrate during a manufacturing process of an IC. For example, the FEOL process may include planarizing and cleaning a wafer, forming a trench, forming a well, forming a gate line, and forming a source and a drain.

In operation S52, a back-end-of-line (BEOL) process may be performed. BEOL may refer to a process of interconnecting individual components (e.g., transistors, capacitors, resistors, etc.) during the manufacturing process of an IC. For example, the BEOL process may include silicidating gate, source and drain regions, adding dielectrics, planarizing, forming holes, forming metal layers, forming vias, and forming passivation layers. The IC may then be packaged in a semiconductor package and used as a component in various applications.

As used in conjunction with the various embodiments disclosed, the term "module" may include units implemented in hardware, software, or firmware, and may be used interchangeably with other terms (e.g., logic, logic block, component, or circuit). A module may be a single integrated component or its smallest unit or portion adapted to perform one or more functions. For example, according to an embodiment, a module may be implemented in the form of an application specific integrated circuit (ASIC).

The various embodiments set forth herein may be implemented as software (e.g., a program) including one or more instructions stored in a storage medium (e.g., an internal memory or an external memory) readable by a machine (e.g., an electronic device). For example, a processor of a machine (e.g., an electronic device) may call at least one of the one or more instructions stored in the storage medium, and execute at least one instruction under the control of the processor with or without the use of one or more other components. This allows the machine to be operated to perform at least one function according to at least one instruction called. One or more instructions may include code generated by a compiler or code executable by a parser. The machine-readable storage medium may be provided in the form of a non-transitory storage medium. Among them, the term "non-transitory" may indicate that the storage medium is a tangible device and does not include signals (e.g., electromagnetic waves), but this term does not distinguish between a case where data is semi-permanently stored in a storage medium and a case where data is temporarily stored in a storage medium.

According to an embodiment, the method according to various disclosed embodiments may be included and provided in a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be distributed in the form of a machine-readable storage medium (e.g., a compact disc read-only memory (CD-ROM)), or distributed online (e.g., downloaded or uploaded) through an application store (e.g., PlayStore ^TM ), or distributed directly between two user devices (e.g., smart phones). If distributed online, at least a portion of the computer program product may be temporarily generated or at least temporarily stored in a machine-readable storage medium (such as a memory of a manufacturer's server, a server of an application store, or a relay server).

According to various embodiments, each of the above-mentioned components (e.g., a module or a program) may include a single entity or multiple entities, and some of the multiple entities may be separately set in different components. According to various embodiments, one or more of the above-mentioned components may be omitted, or one or more other components may be added. Alternatively or additionally, multiple components (e.g., modules or programs) may be integrated into a single component. In this case, according to various embodiments, the integrated component may still perform one or more functions of each of the multiple components in the same or similar manner as performed by the corresponding components of the multiple components before integration. According to various embodiments, the operations performed by a module, a program or another component may be performed sequentially, in parallel, repeatedly or heuristically, or one or more of the operations may be performed or omitted in a different order, or one or more other operations may be added.

Each embodiment provided in the above description is not exclusive of association with one or more features of additional examples or additional embodiments that may or may not be provided herein but are consistent with the disclosure.

While the disclosure has been particularly shown and described with reference to disclosed embodiments, it will be understood that various changes in form and details may be made therein without departing from the spirit and scope of the appended claims.

Claims (20)

1. A pattern modeling method for predicting image data, the pattern modeling method comprising: generating first image data based on a sample pattern learned by a deep neural network DNN; generating second image data by measuring the first image data; determining a region of the second image data to which the weight filter is to be applied; training the DNN by applying a weighted filter to the determined regions of the second image data; and At least one pattern image is predicted based on the result of training the DNN. 2 . The pattern modeling method according to claim 1 , wherein the second image data comprises image data generated by applying a process condition to the first image data. 3 . The pattern modeling method according to claim 1 , wherein the step of determining the area of the second image data to which the weight filter is to be applied comprises determining an area corresponding to a critical size of the second image data. 4 . The pattern modeling method according to claim 1 , wherein the step of determining the region of the second image data to which the weight filter is to be applied comprises determining a region corresponding to a distribution of the second image data. 5 . The pattern modeling method according to claim 1 , wherein the step of determining the region of the second image data to which the weight filter is to be applied comprises determining a region corresponding to a pattern shape of the second image data. 6. The pattern modeling method according to any one of claims 1 to 5, wherein the DNN is trained based on: third image data obtained by convolving the weight filter with the second image data; and Fourth image data obtained by convolving the weight filter with the reference image data. 7 . The pattern modeling method according to claim 6 , wherein the step of training the DNN comprises calculating a loss function based on a difference between the third image data and the fourth image data. 8. The pattern modeling method according to claim 7, wherein the step of training the DNN further comprises: updating the weight data of the DNN based on the calculation result of the loss function. 9. A pattern modeling system for predicting image data, the pattern modeling system comprising: a memory for storing instructions; and at least one processor configured to execute the instructions to: generating first image data based on a sample pattern learned by a deep neural network DNN, the DNN comprising a plurality of layers, generating second image data by measuring the first image data, determining a region of the second image data to which the weight filter is to be applied, training the DNN by applying a weighted filter to the determined regions of the second image data; and At least one pattern image is predicted based on the result of training the DNN. 10. The pattern modeling system of claim 9, wherein the at least one processor comprises at least one data preprocessor, The at least one data preprocessor is configured to execute the instructions to preprocess the image data based on the sample pattern. 11. The pattern modeling system of claim 10, wherein the first image data comprises image data obtained at least in part based on a sample pattern, The second image data includes image data obtained by applying the process conditions to the first image data. 12 . The pattern modeling system of claim 11 , wherein the region to which the weight filter is applied includes a region corresponding to a critical size of the second image data. 13 . The pattern modeling system of claim 11 , wherein the region to which the weight filter is to be applied includes a region corresponding to a distribution of the second image data. 14 . The pattern modeling system of claim 11 , wherein the region to which the weight filter is applied includes a region corresponding to a pattern shape of the second image data. 15. The pattern modeling system of claim 11, wherein the at least one data pre-processor is further configured to execute the instructions to: applying a weight filter corresponding to the region to the second image data; and The image with the weighted filter applied is sent to the loss function module. 16. The pattern modeling system according to claim 15, wherein the loss function module is configured to: determining a difference between third image data generated based on applying the weighted filter to the second image data and fourth image data generated based on applying the weighted filter to the reference image data, and Minimize differences. 17. The pattern modeling system of claim 16, wherein the reference image data comprises output image data of the DNN. 18. A non-transitory computer-readable storage medium storing instructions that, when executed by at least one processor, cause the at least one processor to: generating first image data based on a sample pattern learned by a deep neural network (DNN), the DNN comprising a plurality of layers; generating second image data by measuring the first image data; determining a region of the second image data to which the weight filter is to be applied; training the DNN by applying a weighted filter to the determined regions of the second image data; and Predict at least one pattern image based on the result of training the DNN, Therein, the weight filter is applied to the area corresponding to the feature portion of the second image data. 19 . The non-transitory computer-readable storage medium of claim 18 , wherein the feature portion includes an area corresponding to a critical dimension of the second image data. 20 . The non-transitory computer-readable storage medium of claim 18 , wherein the characteristic portion includes a region corresponding to a distribution of the second image data.