Information processing method, semiconductor manufacturing method, information processing apparatus, and information processing program

Abstract

Even with a small amount of image data associated with the experimental conditions used for training, the system accurately predicts images showing the processed shape of the material layer formed on the substrate under those experimental conditions. [Solution] The information processing method of the present disclosure constructs a VAE model (153) that takes a plurality of training images (D0) showing a material layer (12) on a substrate (11) as input and outputs a latent variable (Zi), calculates a first latent variable (Z1i) by inputting a plurality of first images (D1) of the material layer (12) associated with actual experimental conditions (X(r)), generates a specific numerical model (154) that represents the relationship between the actual experimental conditions (X(r)) and the first latent variable (Z1i), calculates a second latent variable (Z2i) by inputting a plurality of virtual experimental conditions (X(e)) into the VAE model (153), and reconstructs a second image (D2) by inputting the second latent variable (Z2i) into the VAE model (153).

Description

[Technical Field]

[0001] The present disclosure relates to an information processing method, a semiconductor manufacturing method, an information processing apparatus, and an information processing program. [Background Art]

[0002] Generally, in the manufacture of semiconductors such as semiconductor devices and semiconductor packages, the processed shape of a material layer formed on a base material such as a substrate can greatly affect the device characteristics of these products. Therefore, predicting such a processed shape in advance is extremely important for process evaluation.

[0003] Here, as materials used for manufacturing semiconductor devices and semiconductor packages, photosensitive resin compositions are widely known, for example, which cover the surface of printed circuit boards to protect circuit patterns, or are used for circuit formation on substrates. Such a photosensitive resin composition is laminated on the entire surface of a substrate, exposed through a mask corresponding to a circuit pattern, or exposed without a mask, and then soluble portions are developed with a developer, whereby fine patterns can be formed. The photosensitive resin composition patterned in this way not only plays an important role in preventing solder from adhering to unnecessary portions during component mounting and preventing plating from adhering to unnecessary portions during circuit formation, but also the shape of the photosensitive resin composition formed by exposure and development affects how solder and plating are formed. Therefore, techniques for predicting the processed shape of such a photosensitive resin composition in advance are highly desired.

[0004] However, this requires expert knowledge regarding material composition design and adjustment of process conditions such as exposure and development, and data collection through destructive analysis which is time-consuming and costly is required for evaluation. For these reasons, predicting the processed shape in advance has not been easy. In addition, among processed shapes, particularly the shape of the bottom portion in the cross-section after processing can be a factor that determines the laminated structure formed thereon. Even if a relationship representing the correlation between process conditions and processed shape can be established to some extent, it has been difficult to predict the detailed shape of such a bottom portion. [Prior art documents] [Patent Documents]

[0005] [Patent Document 1] Japanese Patent Application Publication No. 06-112293 [Patent Document 2] Japanese Patent Publication No. 2020-43270 [Overview of the project] [Problems that the invention aims to solve]

[0006] In response to this situation, for example, Patent Document 1 proposes a method for predicting the pattern shape after development of a general photoresist used as a mask material. However, the current situation is that the shape prediction theory used for photoresists and the like cannot be directly applied to the pattern shape after development of photosensitive resin compositions containing scattering materials such as solder resists.

[0007] Furthermore, for example, Patent Document 2 describes a learning device that sets environmental information and process conditions within a processing container used for semiconductor manufacturing, and simulates a semiconductor manufacturing process using predetermined parts (such as gas pipe shapes) installed in the processing container as variables, and calculates a feature vector that associates the simulation data with in-plane (XY) coordinates of the substrate. In this learning device, a learning model is trained that takes the obtained feature vector as input and outputs the processing result of the substrate (in-plane image of the substrate). Thus, in such a learning device, a large amount of simulation data linked to process conditions is required to construct the learning model. However, in many cases, only a small amount of image data linked to environmental information and process conditions can be obtained as learning data, and in many cases, cross-sectional images are preferred over in-plane images from the viewpoint of the degree of influence on the characteristics of the equipment. The learning device in Patent Document 2 had difficulty meeting these requirements.

[0008] Therefore, this disclosure has been made in view of these circumstances, and aims to provide an information processing method, an information processing device, and an information processing program that can accurately predict an image showing the processed shape (particularly the cross-sectional shape) of a material layer formed on a substrate by an experimental condition, even if only a small amount of training images associated with the experimental conditions are available. [Means for solving the problem]

[0009] [1] In order to achieve the above objective, the information processing method according to the present disclosure involves a computer constructing a Variational Autoencoder (VAE) model that includes (1) an encoder that performs machine learning and takes a plurality of training images showing the processed shape of a material layer formed on a substrate as input, extracts dimensionally compressed low-dimensional features, and outputs latent variables, and a decoder that takes the latent variables as input and outputs an image showing the processed shape; (2) inputting a plurality of first images that are associated with actual experimental conditions and show the processed shape of a material layer formed on a substrate into the encoder, and calculating (outputting) a first latent variable (multidimensional) corresponding to each; (3) constructing a specific numerical model that represents the relationship between the actual experimental conditions and the first latent variables; (4) inputting virtual experimental conditions into the specific numerical model, and calculating (outputting) a second latent variable corresponding thereto; and (5) inputting the second latent variables into the decoder, and reconstructing a second image corresponding to those virtual experimental conditions and the second latent variables.

[0010] In this disclosure, the term "VAE model" refers to a type of model that consists of a neural network infrastructure comprising an encoder and a decoder, and a latent environment (latent variables) intervening between them, enabling data generation and structural learning. Specifically, the "VAE model" is a type of deep generative model that performs feature extraction (encoding) to transform input data into a low-dimensional latent environment, and then performs machine learning to approximate a specific regularized probability distribution (e.g., a normal distribution) on that latent environment. When latent variables are given, it reconstructs (decodes) output data in the same format as the input data. In other words, the "encoder" has the function of mapping input data to a probability distribution (including mean and variance) in the latent space as a dimensionality reduction through feature extraction, and the "decoder" has the function of reconstructing higher-dimensional output data using latent variables sampled from or likely to follow that probability distribution.

[0011] Furthermore, in this disclosure, "machine learning" refers to the entire process by which a computer learns specific patterns or rules from given data and makes judgments such as predictions and classifications based on them. Such "machine learning" is generally performed based on supervised learning, unsupervised learning, semi-supervised learning, reinforcement learning, or a combination thereof. In addition, it includes cases where algorithms such as neural networks, support vector machines, decision trees, random forests, clustering methods, Bayesian inference, and gradient boosting, and their variations are used, as well as cases where artificial intelligence (AI) technologies, including deep learning, are applied. In particular, in this disclosure, in order to perform machine learning of a "VAE model," it is preferable to construct a generative model that applies deep learning to a neural network and is based on a mathematical framework for handling probability distributions (e.g., variational Bayesian methods).

[0012] In this configuration, an encoder of a VAE model is constructed that is machine-trained to extract low-dimensional features from the pixel information of a training image by inputting the training image and outputting latent variables. Furthermore, a decoder of the VAE model is constructed that is machine-trained to output an image that reproduces the training image from the latent variables obtained by inputting the training image into the encoder. A specific numerical model is constructed from the first latent variables obtained by inputting the first image into the encoder of the VAE model and the actual experimental conditions associated with the first image. By inputting the second latent variables obtained by inputting virtual experimental conditions into that numerical model into the decoder of the VAE model, a second image, which is the processing shape predicted when those virtual experimental conditions are used, is reconstructed. Using this method with a VAE model and numerical model, even if there are only a small number of images associated with the actual experimental conditions used for training, it is possible to accurately predict the second image showing the processing shape of the material layer formed on the substrate by the experimental conditions.

[0013] [2] In addition, in the above configuration, the computer may pre-process at least one of the following with respect to the training images: flattening, binarization, and morphological processing. In this way, even if the training images contain noise components or it is difficult to distinguish between the material layer and the substrate, the contour of the processed shape of the material layer on the substrate can be made clearer. As a result, feature extraction by the VAE model can be performed with greater accuracy and effectiveness, and as a result, the prediction accuracy of the second image by the VAE model can be further improved.

[0014] [3] More specifically, it is preferable that the images showing the processed shape of the material layer (i.e., the training image, the first image, and the second image, etc.) are cross-sectional images perpendicular to the substrate surface. As mentioned above, among the processed shapes, the shape of the bottom in the cross-section after processing can be a factor that influences the layered structure formed thereon, so by using such cross-sectional images, important predictive information can be obtained in the manufacturing of semiconductors such as semiconductor devices and semiconductor packages.

[0015] [4 and 5] Furthermore, although not particularly limited, when the substrate is, for example, made of resin, glass, or metal, and the material is, for example, a photosensitive resin composition containing a scattering material, more specifically, solder resist, it is difficult to directly apply the shape prediction theory used for photoresists and the like as described above, and therefore this information processing method becomes particularly useful.

[0016] [6] In this case, specific experimental conditions include material conditions such as first optical spectral information for the substrate and second optical spectral information for the material layer, and process conditions such as exposure conditions for the material layer. When the material is a photosensitive resin composition, these experimental conditions tend to affect the processed shape of the material layer, so this information processing method becomes even more useful.

[0017] [7] Furthermore, the process conditions may also include development conditions for the material layer. When the material is a photosensitive resin composition, the development process also tends to affect the processed shape of the material layer, so in this respect as well, the information processing method is particularly useful.

[0018] [8] Furthermore, when a light-transmitting protective material layer is laminated on the material layer, the material conditions may also preferably include a third optical spectral information relating to the protective material layer.

[0019] [9] In the above configuration, the computer may be configured to (6) calculate the similarity between a third image showing a desired processed shape for a material layer formed on a substrate and each of the plurality of second images, (7) select a specific second image that is similar to the third image from among the plurality of second images based on the similarity, and (8) output virtual process conditions associated with the selected specific second image (as recommended process conditions). In this way, process conditions that can realize the desired processed shape can be proposed, which ultimately contributes to the manufacture of semiconductors such as semiconductor devices and semiconductor packages that have the desired characteristics.

[0020]

[10] The semiconductor manufacturing method according to the present disclosure also includes a step of processing a material layer using virtual process conditions output by the information processing method according to the present disclosure.

[0021]

[11] Furthermore, the information processing device according to the present disclosure is a device for effectively implementing the information processing method according to the present disclosure, and comprises: (1) a first learning unit that constructs a VAE model including an encoder that performs machine learning, takes a plurality of learning images showing the processed shape of a material layer formed on a substrate as input, extracts dimensionally compressed low-dimensional features and outputs latent variables, and a decoder that takes the latent variables as input and outputs an output image showing the processed shape; (2) a first calculation unit that inputs a plurality of first images that are associated with actual experimental conditions and show the processed shape of a material layer formed on a substrate to the encoder and calculates a first latent variable corresponding to each of the plurality of first images; (3) a second learning unit that constructs a specific numerical model representing the relationship between the actual experimental conditions and the first latent variable; (4) a second calculation unit that inputs a plurality of virtual experimental conditions to the specific numerical model and calculates a second latent variable corresponding to each of the plurality of virtual experimental conditions; and (5) a reconstruction unit that inputs the second latent variable to the decoder and reconstructs a second image corresponding to the second latent variable and the virtual experimental conditions.

[0022] [

[12] ] Further, the information processing program according to the present disclosure causes a computer to function as: (1) a first learning unit that constructs a VAE model including an encoder that performs machine learning, inputs a plurality of learning images indicating processing shapes of a material layer formed on a base material, extracts dimension-compressed low-dimensional feature quantities and outputs latent variables, and a decoder that inputs latent variables and outputs an output image indicating a processing shape; (2) a first calculation unit that inputs a plurality of first images associated with actual experimental conditions and indicating processing shapes of a material layer formed on a base material to the encoder, and calculates a first latent variable corresponding to each of the plurality of first images; (3) a second learning unit that constructs a specific numerical model representing the relationship between actual experimental conditions and the first latent variables; (4) a second calculation unit that inputs a plurality of virtual experimental conditions to the specific numerical model, and calculates a second latent variable corresponding to each of the plurality of virtual experimental conditions; and (5) a reconstruction unit that inputs the second latent variable to the decoder and reconstructs a second image corresponding to the second latent variable and the virtual experimental conditions.

Effects of the Invention

[0023] According to the information processing method, information processing apparatus, and information processing program of the present disclosure, even when a small number of learning images associated with experimental conditions are available, it is possible to accurately predict an image indicating the processing shape (particularly the cross-sectional shape) of a material layer formed on a substrate under the experimental conditions from the experimental conditions.

Brief Description of Drawings

[0024] [Figure 1] It is a conceptual diagram showing a schematic configuration of an information processing system 1 including an information processing apparatus 100 according to an embodiment of the present disclosure. [Figure 2] It is a block diagram showing an outline of the hardware configuration and functional configuration of the information processing apparatus 100. [Figure 3] It is a schematic cross-sectional view showing an example of a laminate in which a material layer is formed on a base material in the present embodiment, where (A) shows an example of a state before processing, and (B) shows an example of a state after processing. [Figure 4]This flowchart shows an example of a processing procedure in an information processing method effectively implemented using Information Processing System 1. [Figure 5] This is a conceptual diagram illustrating the configuration of the VAE model 153 and an example of a specific process performed by the first learning unit 111. [Figure 6] This photograph shows the results of constructing VAE model 153 using unprocessed SEM images. [Figure 7] This photograph shows the results of constructing VAE model 153 using SEM images after image processing. [Figure 8] This is a conceptual diagram showing the configuration of the VAE model 153 and its relationship to a specific numerical model 154, and is intended to explain an example of a specific process performed by the calculation unit 112 and the second learning unit 113. [Figure 9] (A) through (D) are graphs showing examples of material and process conditions. [Figure 10] This document includes a photograph showing the processed shape of a material layer processed under predetermined process conditions, and an image showing the processed shape predicted by the information processing method of this disclosure, using the predetermined process conditions as virtual process conditions. [Modes for carrying out the invention]

[0025] This embodiment will be described below with reference to the attached drawings. To facilitate understanding of the description, the same reference numerals are used for identical components in each drawing whenever possible, and redundant descriptions are omitted. The following embodiments are illustrative examples for illustrating the present disclosure and are not intended to limit the present disclosure to these embodiments only. Furthermore, the present disclosure can be modified in various ways without departing from its essence. Moreover, those skilled in the art can adopt embodiments in which each of the elements described below is replaced with equivalent components, and such embodiments are also included within the scope of the present disclosure.

[0026] [Example of an information processing device configuration] Figure 1 is a conceptual diagram showing the schematic configuration of an information processing system 1 including an information processing device 100 according to one embodiment of the present disclosure. The information processing system 1 has a configuration in which the information processing device 100, which functions as a server, and a user terminal 200 are connected to each other via a wired or wireless network N. The network N is not particularly limited and may include, for example, a public network including the Internet, a private network such as a LAN, WAN, PAN, VPN, sensor network, cloud network, or any combination thereof.

[0027] The information processing device 100 is implemented by a computer and a specific information processing program, and generally performs the following processing: (1) trains an encoder and decoder of a VAE model using multiple training images showing the processed shape of a material layer formed on a substrate, (2) inputs a first image associated with actual experimental conditions into the encoder to calculate a first latent variable, (3) constructs a numerical model from those actual experimental conditions and the first latent variable, (4) calculates a second latent variable from that numerical model and virtual experimental conditions, and (5) inputs that second latent variable into the decoder to reconstruct a second image as a predicted image when using virtual experimental conditions. The user terminal 200 is an information processing device used by the user of the information processing system 1, and is implemented by, for example, a smartphone, tablet terminal, personal computer, or workstation. The user can use this user terminal 200 to access the information processing device 100, provide training images and the first image to the information processing device 100, and refer to the second image predicted from the virtual experimental conditions.

[0028] Next, Figure 2 is a block diagram illustrating the schematic hardware and functional configuration of the information processing device 100. As shown in the figure, the information processing device 100 includes, for example, a processor 110, a communication interface 120, an input / output interface 130, a memory 140, storage 150, and one or more communication buses 160 for interconnecting these components.

[0029] The processor 110 executes processes, functions, or methods implemented by the code or instructions contained in the information processing program 151 stored in the storage 150. The processor 110 is not particularly limited and includes, for example, one or more central processing units (CPUs), MPUs, GPUs, etc. By using the storage area or executing the information processing program 151 through logic circuits (hardware) formed in an integrated circuit or the like, the processor 110 realizes various functional units (first learning unit 111, calculation unit 112, second learning unit 113, reconstruction unit 114, and display control unit 115, etc.), as well as the processes and methods executed by them.

[0030] The communication interface 120 transmits and receives various types of data with other devices via the network N. This communication may be performed via wired, wireless, or a combination thereof, and any appropriate communication protocol can be used as long as communication between the devices is possible. For example, the communication interface 120 can be implemented as hardware such as a network adapter, various types of communication software, or a combination thereof.

[0031] The input / output interface 130 includes an input device for inputting various operations to the information processing device 100, and an output device for outputting processing results processed by the information processing device 100. For example, the input / output interface 130 includes information input devices such as a keyboard, mouse, touch panel, touch screen, microphone, camera, sensor, barcode reader, and scanner, as well as information output devices such as a display, speaker, printer, projector, and other display devices, or a device that serves as both an input and output device (e.g., a touchscreen display).

[0032] Memory 140 temporarily stores the information processing program 151 loaded from storage 150 and provides a workspace for the processor 110. Various data generated while the processor 110 is executing the information processing program 151 are also temporarily stored in memory 140. Furthermore, there are no particular restrictions on the memory 140, and examples include high-speed random access memory such as DRAM, SRAM, DDR RAM, DDR SDRAM, LPDDR, and other random access solid-state storage devices (ROM, flash memory, NVMe, MRAM, FRAM, etc.), or any combination thereof.

[0033] The storage 150 stores the information processing program 151, the VAE model 153, and a specific numerical model 154, as well as the calculation results of each functional unit (such as the latent variable Zi), and various other data. The storage 150 is not particularly limited and may include, for example, one or more magnetic disk storage devices (such as hard disk drives), optical disk storage devices (such as CDs, DVDs, Blu-ray® discs), flash memory devices (such as USB memory, SD cards, SSDs), or other non-volatile solid-state storage devices (such as NAND flash memory, NOR flash memory, 3D NAND), or non-volatile memory such as next-generation memory technologies (such as MRAM, PRAM, ReRAM, FRAM®), or any combination thereof.

[0034] The communication bus 160 is not particularly limited as long as it is a known dedicated communication path for exchanging data and control information between hardware configurations, and the communication method may be wired, wireless, or a combination thereof. Examples of such a communication bus 160 include a parallel bus, serial bus, memory bus, input / output bus, or a high-performance bus that integrates these, and furthermore, PCI, PCIe, I 2 This includes buses based on C, SPI, CAN bus, USB, or similar or compliant communication standards. Furthermore, these communication buses are not limited to wired communication methods, but also include cases where they are implemented using wireless communication methods.

[0035] Furthermore, among the various functional units realized by the processor 110 of the information processing device 100, the first learning unit 111 constructs a VAE model 153 that includes an encoder for extracting dimensionally compressed low-dimensional features from multiple learning images showing the processed shape of a material layer formed on a substrate, and a decoder for outputting an image showing the processed shape from latent variables. The calculation unit 112 first uses the encoder 41 of the VAE model 153 to calculate a first latent variable corresponding to each of the multiple first images that are associated with actual experimental conditions and show the processed shape of a material layer formed on a substrate. In addition, the calculation unit 112 uses a specific numerical model described later to calculate a second latent variable corresponding to a virtual experimental condition. Thus, the calculation unit 112 corresponds to an example of the "first calculation unit" and "second calculation unit" in this disclosure. Furthermore, the second learning unit 113 constructs a specific numerical model 154 that represents the relationship between the actual experimental conditions and the first latent variable. In addition, the reconstruction unit 114 uses the decoder 42 of the VAE model 153 to reconstruct a second image corresponding to the second latent variable. Furthermore, the display control unit 115 controls the display of the reconstructed second image.

[0036] Here, Figures 3(A) and 3(B) are schematic cross-sectional views showing an example of a laminate in which a material layer is formed on a substrate in this embodiment. Figure 3(A) shows an example of the state before processing, and Figure 3(B) shows an example of the state after processing. These cross-sectional views also show a cross-section perpendicular to the substrate surface. As shown in Figure 3(A), the laminate 10 before processing has a material layer 12 formed on a substrate 11, and if necessary, a light-transmitting protective material layer 13 is further laminated on top of the material layer 12. As shown in Figure 3(B), in the laminate 10' after processing, via holes 121 and dams 122 are processed and formed in the material layer 12 on the substrate 11, and the protective material layer 13 has been removed.

[0037] The substrate 11 is not particularly limited and includes, for example, a resin substrate such as BT resin, a glass substrate, or a metal substrate. "Resin substrate" means a substrate containing resin as one of its main components, and may also include substrates containing other components, such as glass epoxy substrates. The type of material layer 12 is not particularly limited as long as it contains resin, and includes, for example, a photosensitive resin composition which may contain a scattering agent, more specifically, an alkali-developable photosensitive resin composition or a solvent-developable photosensitive resin composition. A solder resist containing an alkali-soluble resin, a scattering agent, a photopolymerization initiator, and, as needed, epoxy resin, a colorant, a curing accelerator, an inorganic filler, an additive, etc., is particularly preferred. The material layer 12 made of a photosensitive resin composition may be a negative type in which the hardened portion remains after exposure and is developed, or a positive type in which the decomposed portion is removed after exposure. A thermosetting resin composition is also an example of a type of material layer 12. When the material layer 12 is a thermosetting resin composition, the information processing method of this disclosure can be applied to the thermosetting resin composition or its cured product, for example, to a processed shape obtained by laser processing. Furthermore, the type of protective material layer 13 is not particularly limited and includes, for example, polyolefin films such as polypropylene films, polyester films such as polyethylene terephthalate films, polycarbonate films, polyacrylate films, and other films.

[0038] [Example of processing procedure for information processing methods] Figure 4 is a flowchart showing an example of a processing procedure in an information processing method effectively implemented using the information processing system 1. In this embodiment, first, prior to step S41, a large number of training images D0 used for machine learning of the VAE model 153 are transmitted from the user terminal 200 to the information processing device 100 via the network N. These training images D0 are images showing the processed shape of the material layer 12 as shown in Figure 3(B), and the display data and pixel information data related to them are stored in the storage 150 via the communication interface 120 and the input / output interface 130. Note that the training images D0 shown in Figure 5 may include training images that are not associated with experimental conditions. That is, the training images D0 do not necessarily have to be training images that can be associated with experimental conditions.

[0039] In this disclosure, "experimental conditions" include material conditions (such as optical spectra) which are information regarding the material properties of the substrate 11, material layer 12, and protective material layer 13, and / or process conditions (such as exposure conditions and development conditions) which are information regarding the process of processing the material layer 12. Details of these material conditions and process conditions will be described later.

[0040] Processing begins, and in step S41, the first learning unit 111 loads the training image D0 stored in the storage 150 and holds it in the memory 140. Here, Figure 5 is a conceptual diagram illustrating the configuration of the VAE model 153 and an example of a specific process performed by the first learning unit 111. As defined earlier, the VAE model 153 consists of an encoder 41 and a decoder 42, and a latent environment 43 intervening between them.

[0041] The first learning unit 111 constructs a VAE model 153 using machine learning. In doing so, the first learning unit 111 inputs multiple training images D0 into the encoder 41. The encoder 41 extracts low-dimensional (e.g., several to tens of dimensions) features that are dimensionally compressed from the information content of the training images D0 (which have a maximum of the same number of dimensions as the number of pixels) through encoding. These features are represented by variables mapped to a latent environment 43, which is expressed as a normal probability distribution, and the encoder 41 outputs multiple (i-dimensional) latent variables Zi. The first learning unit 111 performs machine learning by repeatedly performing this feature extraction process on the training images D0. The first learning unit 111 further inputs each obtained latent variable Zi into the decoder 42, and through decoding, reconstructs a restored image D0' with the same dimensions as the training images D0. The first learning unit 111 performs machine learning while evaluating the error between the training image D0 and the reconstructed image D0', and ultimately constructs a VAE model 153 that can reconstruct the reconstructed image D0' from the training image D0 with sufficient accuracy.

[0042] Figure 6 is a photograph showing an example of the results of constructing a VAE model 153 using machine learning with a 5-dimensional latent environment 43 (latent variable Zi(i=5)) and approximately 1800 SEM images showing a part of the processed shape of the material layer 12 shown in Figure 3(B). (A) to (D) in the same figure show examples of different training images D0, and (A') to (D') show the reconstructed images D0'. The training image D0 here is a cross-sectional image perpendicular to the surface of the substrate 11 and is the original image without any image processing (unprocessed). As can be seen, it was confirmed that the VAE model 153 obtained at this time was able to faithfully reconstruct the processed shape of the material layer 12 (the overall rectangular shape and the shapes of the different fine bottom structures) in the training image D0. It was also confirmed that sufficient reproduction accuracy could be obtained even when the latent environment 43 was 10-dimensional (latent variable Zi(i=10)).

[0043] Incidentally, a closer examination of the results shown in Figure 6 reveals that the reconstructed images D0' shown in (A') to (D') have a slightly fainter outline compared to the training images D0 shown in (A) to (D). Therefore, Figure 7 shows the results of performing the same machine learning using processed images obtained by applying outline-enhancing image processing to the training images D0 in Figures 6(A) to (D). Specifically, Figure 7 is a photograph showing an example of the results of constructing a VAE model 153 using machine learning with a 5-dimensional latent environment 43 (latent variable Zi (i=5)) using approximately 1800 image-processed SEM images. In the same figure, (A) to (D) each show examples of different training images D0 after image processing, and (A') to (D') show the respective reconstructed images D0'. As shown above, it was confirmed that when using the image-processed training images D0, the processed shape of the material layer 12 can be reconstructed more faithfully and with greater clarity.

[0044] Here, such image processing is not particularly limited and includes, for example, flattening, binarization, and morphological processing. Of these, "flattening" is a process that adjusts the overall appearance and characteristics of an image by making the distribution of specific features (luminance, hue, grayscale, texture, edges, etc.) in the image uniform or smooth. Examples include redistribution based on the image histogram (histogram equalization), spatial filtering, and smoothing techniques using frequency transformation. "Binarization" is a process that determines the value of each pixel in an image based on a predetermined threshold, for example, converting values ​​above the predetermined threshold to a primary value (e.g., white) and values ​​below the threshold to a secondary value (e.g., black). Examples of threshold settings include fixed thresholding, adaptive thresholding, or any combination thereof. Furthermore, "morphological processing" is generally a method that emphasizes, extracts, or removes geometric or structural features in an image and is performed based on a predetermined structural element (kernel, etc.). More specifically, these include erosion (shrinking of objects in an image), dilation (expansion of objects in an image), opening (removal of minute noise from objects), closing (filling in minute gaps), or any combination thereof. In the training image D0 shown in Figures 7(A) to (D), binarization was performed first, followed by three opening processes using a 5x5 cross kernel, and then three closing processes (however, the number of times each process is performed is not limited to these).

[0045] Furthermore, Figure 8 is a conceptual diagram showing the relationship between the configuration of the VAE model 153 and a specific numerical model 154, and is also a conceptual diagram to explain an example of a specific process performed by the calculation unit 112 and the second learning unit 113. Following step S41 shown in Figure 4, in step S42, the calculation unit 112 inputs a plurality of first images D1 showing the processed shape of the material layer 12 associated with the actual experimental conditions X(r) to the encoder 41 of the VAE model 153, and calculates a first latent variable Z1i corresponding to each first image D1 through encoding processing. Except that it is essential that these first images D1 are associated with the actual experimental conditions X(r), SEM images in the same format as the learning image D0 (see Figures 6 and 7) can be used.

[0046] Next, in step S43, the second learning unit 113 generates a specific numerical model 154 (Z=f(X)) that represents the relationship between the actual experimental conditions X(r) and the calculated first latent variable Z1i. When the material layer 12 is made of a photosensitive resin composition, such actual experimental conditions X(r) may include material conditions such as first optical spectral information for the substrate 11, second optical spectral information for the material layer 12, and third optical spectral information for the protective material layer 13; process conditions such as exposure conditions for the material layer 12 and development conditions for the material layer 12; or any combination thereof. Figures 9(A) to (D) show examples of such experimental conditions.

[0047] Of these experimental conditions, the "first optical spectral information" of the substrate 11 is not particularly limited and can include, for example, the reflectance spectrum as shown in Figure 9(A) (spectrums for Cu substrate, BT resin substrate, and glass substrate are shown as examples). As shown in the same figure, the reflectance spectrum may differ depending on the type of substrate. In the case of a glass substrate, the "first optical spectral information" may include the absorbance spectrum, total light transmittance spectrum, or direct transmittance spectrum.

[0048] Furthermore, the "second optical spectral information" of the material layer 12 is not particularly limited and may include, for example, the absorbance spectrum shown in Figure 9(C) (showing examples of spectra of multiple solder resist layers with different compositions and types). As shown in the same figure, the absorbance spectrum tends to differ depending on the composition and type of the material layer 12. The "second optical spectral information" may also include reflectance spectra, total light transmittance spectra, or linear transmittance spectra.

[0049] Furthermore, the "third optical spectral information" of the protective material layer 13 is not particularly limited and may include, for example, the absorbance spectrum shown in Figure 9(D) (showing examples of spectra for multiple PET layers with different compositions). As shown in the same figure, the absorbance spectrum tends to differ depending on the composition of the protective material layer 13. The "third optical spectral information" may also include the reflectance spectrum, total light transmittance spectrum, or direct transmittance spectrum.

[0050] Furthermore, the "exposure conditions" for the material layer 12 are not particularly limited and include, for example, the illuminance spectrum shown in Figure 9(B) (examples of spectra of exposure light sources with different emission colors are shown). In addition, the "exposure conditions" may include the light intensity of the exposure light source, exposure time, positional relationship with the mask, information on defocusing, and other information that affects exposure. Of these, the illuminance spectrum and light intensity of the exposure light source affect the reaction efficiency of the material layer 12, while the exposure time and irradiation angle may affect the accuracy and uniformity of the exposure pattern.

[0051] Furthermore, the "developing conditions" for the material layer 12 are not particularly limited and include, for example, the developer composition, the temperature and development time of the developer, the pressure of the spray used to apply the developer to the material layer 12, the nozzle type, the rinsing time, and various other information that affects development. Of these, the developer composition and temperature affect the development speed, while the spray pressure and nozzle type may affect the development speed from the standpoint of the efficiency of developer exchange near the material layer 12, or they may affect the mechanical removal performance.

[0052] If the material layer 12 is made of a thermosetting composition, process conditions include curing conditions and laser processing conditions for the material layer 12.

[0053] Examples of "curing conditions" for the material layer 12 include curing temperature and curing time. Examples of "laser processing conditions" for the material layer 12 include laser wavelength and laser power.

[0054] Returning to Figure 8, the type of specific numerical model 154 constructed by the second learning unit 113 is not particularly limited, and regression models having a correlation between the experimental conditions X(r) and the first latent variable Z1i can be used. The regression analysis method applied to the generation of such a regression model, the specific numerical model 154, is not particularly limited, and examples include linear analysis methods such as partial least squares (PLS), Ridge regression, LASSO (Least Absolute Shrinkage and Selection Operator), Elastic Net, linear support vector regression (Linear SVR), and Gaussian process regression (GPR); and nonlinear analysis methods such as non-linear support vector regression (Non-Linear SVR), random forest, gradient boosting (XGBoost, Light-GBM), and the application of different kernel functions in Gaussian process regression (GPR).

[0055] Next, in step S44 of Figure 4, the calculation unit 112 inputs a virtual experimental condition X(e) of the same format as the actual experimental condition X(r) into a specific numerical model 154 and calculates a second latent variable Z2i of the same dimension as the first latent variable Z1i that can be input into the decoder 42 of the VAE model 153. Then, in the following step S45, the reconstruction unit 114 inputs the second latent variable Z2i into the decoder 42 of the VAE model 153 and reconstructs the second image D2 by increasing the dimensionality of the second latent variable Z2i through the decoding process of the decoder 42. This second image D2 corresponds to the image predicted as the processed shape of the material layer 12 when the virtual experimental condition X(e) is applied. Then, the reconstruction unit 114 stores the data related to the second image D2 in the memory 140 and / or storage 150. Furthermore, the display control unit 115 creates display data for the reconstructed second image D2 and transmits it to the user terminal 200 via the network N through the communication interface 120 and the input / output interface 130 to provide it to the user.

[0056] Here, Figure 10 is a photograph showing an example of the results of predicting the processed shape from virtual experimental conditions using the decoder 42 of the numerical model 154 and the VAE model 153, which were constructed using approximately 100 first images D1. (A) to (D) in the figure show the actual processed shapes of the material layer 12 obtained under different experimental conditions (with image processing). (A') to (D') in the figure show the processed shapes output by the decoder 42 of the numerical model 154 and the model 153, respectively, when the experimental conditions corresponding to (A) to (D) were input as virtual experimental conditions into the numerical model 154. As shown above, it was confirmed that the VAE model 153 and the numerical model 154 can accurately predict the processed shape of the material layer.

[0057] The series of processes in this embodiment described above may be terminated at this stage, or may be terminated after performing the following steps S46 to S48. In the latter case, in step S46, the calculation unit 112 calculates the similarity St between a third image D3 (in the same format as the second image D2) showing a desired processed shape for a certain material layer 12 on the substrate 11 and each of the plurality of second images D2 obtained in step S45. The calculation unit 112 then selects one or more specific second images D2' from the plurality of second images D2 that are similar to the third image D3 (for example, have a relatively large similarity St) in step S47. Then, in step S48, the display control unit 115 creates display data for the selected specific second image D2 and the corresponding process conditions (virtual process conditions) in a specific virtual experimental condition X(e)', transmits it to the user terminal 200 via the network N through the communication interface 120 and input / output interface 130 to output it to the user, and terminates this example of processing.

[0058] Furthermore, there are no particular limitations on the method for calculating the similarity St between images. Examples include structural similarity evaluation methods including SSIM (Structural Similarity Index Measure), pixel difference evaluation methods using MSE (Mean Squared Error) and PSNR (Peak Signal-to-Noise Ratio), histogram-based similarity evaluation methods (histogram intersection, Earth travel distance, etc.), feature-based evaluation methods (SIFT, ORB, etc.), high-dimensional feature comparison methods using deep learning, and frequency domain comparison methods using phase coherence.

[0059] In addition, a preferred embodiment of the information processing program according to this disclosure is a program that causes a computer to function as the first learning unit 111, calculation unit 112, second learning unit 113, reconstruction unit 114, and, if necessary, display control unit 115.

[0060] Furthermore, one preferred embodiment of the semiconductor manufacturing method according to this disclosure is a semiconductor manufacturing method that includes a step of processing the material layer 12 using the virtual process conditions output in step S48.

[0061] [Effects and Effects of the Embodiment] According to the information processing system 1, information processing device 100, information processing method, and information processing program configured as described above, in step S41, machine learning is performed using multiple training images D0 to extract low-dimensional features that are dimensionally compressed from the pixel information of the training images D0 and output a latent variable Zi, thereby constructing a VAE model 153. Then, by executing each procedure in steps S42 to S45 and using the VAE model 153 and a specific numerical model 154, the second image D2 predicted when the virtual experimental condition X(e) is applied is finally reconstructed. Therefore, by constructing the VAE model 153 in this manner and further constructing the specific numerical model 154 through a combined process, it is not necessary to use images associated with the actual experimental conditions in the machine learning of the VAE model 153, but only to use images associated with the actual experimental conditions in the construction of the numerical model 154. Thus, even if there are only a small number of training images associated with the experimental conditions, it is possible to predict the processed shape of the material layer 12 formed on the substrate 11 by the experimental conditions with sufficient accuracy.

[0062] Furthermore, it was confirmed that by constructing a VAE model 153 using an image obtained by applying a predetermined image processing method to enhance contours to the training image D0 shown in Figures 6(A) to (D), the processed shape of the material layer 12 can be restored more clearly and faithfully, as shown in Figure 7. Therefore, by performing at least one of the following on the training image D0, for example, flattening, binarization, and morphological processing, the prediction accuracy of the second image D2 by the VAE model 153 can be further improved.

[0063] Furthermore, it is preferable that the images showing the processed shape of the material layer 12 (i.e., the learning image D0, the first image D1, and the second image D2, etc.) are cross-sectional images perpendicular to the substrate 11 surface, as in this embodiment. Among the processed shapes, the shape of the bottom in the cross-section after processing can be a factor that influences the laminated structure formed thereon. Therefore, by using such cross-sectional images, important predictive information can be obtained in the manufacturing of semiconductors such as semiconductor devices and semiconductor packages.

[0064] Furthermore, if the substrate 11 is made of resin, glass, or metal, and the material layer 12 is a photosensitive resin composition containing a scattering material, more specifically a solder resist, it is difficult to directly apply the shape prediction theory used in conventional photoresists, etc. Therefore, by using the processing procedure according to this embodiment, it is possible to reliably achieve highly accurate prediction of the second image D2.

[0065] Furthermore, if the actual experimental conditions X(r) used to construct a specific numerical model 154, and the virtual experimental conditions X(e) associated with the second image D2 are material conditions such as first optical spectral information for the substrate 11, second optical spectral information for the material layer 12, and third optical spectral information for the protective material layer 13, process conditions such as exposure conditions for the material layer 12 and development conditions for the material layer 12, or any combination thereof, then the processing procedure according to this embodiment becomes extremely useful because these experimental conditions tend to affect the processed shape of the material layer 12.

[0066] Furthermore, in steps S46 to S48, the similarity St between the third image D3, which shows the desired processed shape of the material layer 12 on the substrate 11, and the second image D2 is calculated. A specific second image D2' that is similar to the third image D3 is selected from among multiple second images D2, and a specific virtual experimental condition X(e)' corresponding to the selected specific second image D2 is output. In this way, process conditions that can realize the desired processed shape in a given material (information on exposure conditions and development conditions for the material layer 12 necessary to achieve the objective when a certain substrate 11, material layer 12, and protective material layer 13 are used) can be proposed to the user. As a result, it is possible to efficiently optimize process conditions that contribute to the manufacture of semiconductors such as semiconductor devices and semiconductor packages that ultimately have the desired characteristics, and it is possible to reduce the time and cost of unnecessary trial and error.

[0067] The embodiments described above with reference to specific examples are provided to facilitate understanding of this disclosure and are not intended to limit its interpretation. In other words, this disclosure is not limited to these specific examples, and modifications made to these examples by those skilled in the art are also included within the technical scope of this disclosure, as long as they retain the features of this disclosure. Furthermore, the elements, arrangements, materials, conditions, shapes, dimensions, scales, etc., of each of the aforementioned specific examples are not limited to those exemplified unless otherwise specified and can be modified as appropriate. Moreover, the elements of each of the aforementioned specific examples can be combined in different ways as appropriate, as long as no technical inconsistencies arise.

[0068] For example, the information processing system 1 is not limited to a client / server system including an information processing device 100 and a user terminal 200. Instead of this system configuration, the user terminal 200 may have the processing functions of the information processing device 100. The information processing device 100 may also accept predetermined inputs or perform predetermined outputs by connecting an external input / output interface 130. Furthermore, the storage 150 may be one or more storage devices installed remotely from the processor 110. Moreover, the laminate 10 may or may not have a protective material layer 13, and exposure to the material layer 12 may be performed via the protective material layer 13 or without the protective material layer 13.

[0069] Furthermore, the optical spectra shown in Figures 9(A) to (D) may plot reflectance, absorbance, illuminance, transmittance, or scattering characteristics for each wavelength in a predetermined wavelength range, such as 200 to 800 nm, or they may be numerical values ​​for a specific wavelength. Such wavelength-specific numerical values ​​for reflectance, absorbance, transmittance, or scattering characteristics may be used as a single material condition. Similarly, wavelength-specific numerical values ​​for illuminance may be used as a single process condition. The material conditions and process conditions may be combined to form the experimental conditions. In addition, each optical spectrum may include information regarding the thickness of each layer 11 to 13. For example, absorbance may be measured according to thickness, or it may be normalized per unit thickness (μm). Furthermore, these optical spectra may be subjected to appropriate data processing techniques to reduce noise in the spectral data, clarify trends, extract characteristic peaks, smooth the data, etc. Such data processing techniques are not particularly limited and include, for example, the Savitzky-Golay method, moving average method, principal component analysis (PCA), Fourier transform, wavelet transform, Gaussian fitting, robust smoothing, denoising filter, spline interpolation, envelope extraction, autoregressive moving average model, or any combination thereof. In addition, the information processing program according to this disclosure may be recorded on a temporary or non-temporary (permanent) readable recording medium. [Explanation of Symbols]

[0070] 1... Information processing system, 10, 10'... Laminate, 11... Substrate, 12... Material layer, 13... Protective material layer, 41... Encoder, 42... Decoder, 43... Latent environment, 100... Information processing device, 110... Processor, 111... First learning unit, 112... Calculation unit, 113... Second learning unit, 114... Reconstruction unit, 115... Display control unit, 120... Communication interface, 121... Via hole, 122... Dam, 130... Input / output interface, 140... Memo R, 150...Storage, 151...Information processing program, 153...VAE model, 154...Numerical model, 160...Communication bus, 200...User terminal, D0...Training image, D0'...Restored image, D1...First image, D2, D2'...Second image, D3...Third image, X(e)...Virtual experimental conditions, X(r)...Actual experimental conditions, N...Network, S41~S48...Step, St...Similarity, Z1i...First latent variable, Z2i...Second latent variable, Zi...Latent variable

Claims

1. Computers A VAE model is constructed that includes an encoder that performs machine learning, takes multiple training images showing the processed shape of a material layer formed on a substrate as input, extracts dimensionally compressed low-dimensional features, and outputs latent variables, and a decoder that takes the latent variables as input and outputs an output image showing the processed shape. Multiple first images, which are associated with actual experimental conditions and show the processed shape of a material layer formed on a substrate, are input to the encoder, and a first latent variable corresponding to each of the multiple first images is calculated. A specific numerical model is constructed that represents the relationship between the experimental conditions and the first latent variable. Multiple virtual experimental conditions are input into the aforementioned specific numerical model, and a second latent variable corresponding to each of these multiple virtual experimental conditions is calculated. The decoder is input to the second latent variable and a second image corresponding to the second latent variable and the virtual experimental conditions is reconstructed. Information processing methods.

2. The information processing method according to claim 1, wherein a computer performs at least one of the following on the training image: flattening, binarization, and morphological processing.

3. The information processing method according to claim 1, wherein the image showing the processed shape of the material layer is a cross-sectional image perpendicular to the substrate surface.

4. The aforementioned substrate is made of resin, glass, or metal. The information processing method according to claim 1, wherein the material is a photosensitive resin composition containing a scattering body.

5. The information processing method according to claim 4, wherein the photosensitive resin composition containing the scattering material is a solder resist.

6. The information processing method according to claim 4, wherein the experimental conditions include first optical spectral information relating to the substrate, second optical spectral information relating to the material layer, and exposure conditions for the material layer.

7. The information processing method according to claim 6, wherein the experimental conditions further include development conditions for the material layer.

8. The information processing method according to claim 6, wherein when a light-transmitting protective material layer is laminated on the material layer, the experimental conditions further include third optical spectral information relating to the protective material layer.

9. The aforementioned computer, The degree of similarity between a third image showing a desired processed shape for a material layer formed on a substrate and each of the multiple second images is calculated. Based on the aforementioned similarity, a specific second image that is similar to the third image is selected from among the multiple second images. Output the virtual process conditions associated with the selected specific second image. The information processing method according to claim 1.

10. A semiconductor manufacturing method, The process includes a step of processing the material layer using the virtual process conditions output by the information processing method described in claim 9, A method for manufacturing semiconductors.

11. A first learning unit constructs a VAE model that includes an encoder that performs machine learning, takes multiple training images showing the processed shape of a material layer formed on a substrate as input, extracts dimensionally compressed low-dimensional features, and outputs latent variables, and a decoder that takes the latent variables as input and outputs an output image showing the processed shape. The encoder is input to a plurality of first images that are associated with actual experimental conditions and show the processed shape of a material layer formed on a substrate, and a first calculation unit calculates a first latent variable corresponding to each of the plurality of first images. A second learning unit constructs a specific numerical model that represents the relationship between the experimental conditions and the first latent variable, The aforementioned specific numerical model includes a second calculation unit that inputs multiple virtual experimental conditions and calculates a second latent variable corresponding to each of the multiple virtual experimental conditions, The decoder is provided with a reconstruction unit that inputs the second latent variable and reconstructs a second image corresponding to the second latent variable and the virtual experimental conditions, An information processing device equipped with the following features.

12. Computers, A first learning unit constructs a VAE model that includes an encoder that performs machine learning, takes multiple training images showing the processed shape of a material layer formed on a substrate as input, extracts dimensionally compressed low-dimensional features, and outputs latent variables, and a decoder that takes the latent variables as input and outputs an output image showing the processed shape. The encoder is input to a plurality of first images that are associated with actual experimental conditions and show the processed shape of a material layer formed on a substrate, and a first calculation unit calculates a first latent variable corresponding to each of the plurality of first images. A second learning unit constructs a specific numerical model that represents the relationship between the experimental conditions and the first latent variable, The aforementioned specific numerical model includes a second calculation unit that inputs multiple virtual experimental conditions and calculates a second latent variable corresponding to each of the multiple virtual experimental conditions, The decoder is provided with a reconstruction unit that inputs the second latent variable and reconstructs a second image corresponding to the second latent variable and the virtual experimental conditions, An information processing program that enables a function to work.

Images