WO2023231226A1

WO2023231226A1 - Automatic proportioning system for photoresist cleaning liquid production and proportioning method therefor

Info

Publication number: WO2023231226A1
Application number: PCT/CN2022/119002
Authority: WO
Inventors: 黄斌斌; 袁瑞明; 罗霜; 林金华; 罗永春; 赖志林
Original assignee: 福建天甫电子材料有限公司
Priority date: 2022-05-30
Filing date: 2022-09-15
Publication date: 2023-12-07
Also published as: CN115018068B; CN115018068A

Abstract

The present application relates to the field of photoresist cleaning liquids under intelligent manufacturing, and particularly discloses an automatic proportioning system for photoresist cleaning liquid production and a proportioning method therefor. The automatic proportioning system for photoresist cleaning liquid production intelligently determines the local optimal qualities of formula ingredients one by one by using a deep-learning-based neural network technique, and determines, in this way, the optimum ratio of a photoresist cleaning liquid for a specific object, such that automatic proportioning is performed on the basis of the optimum ratio.

Description

Automatic batching system and batching method for photoresist cleaning liquid production

Technical field

The present invention relates to the field of photoresist cleaning fluid under intelligent manufacturing, and more specifically, to an automatic batching system for the production of photoresist cleaning fluid and a batching method thereof.

Background technique

Generally, in the process of flat panel displays such as liquid crystal, organic EL, plasma displays, or semiconductors and printed circuit boards, in order to obtain fine images, general photolithography technology is used to form patterns of photosensitive components, using radiation-sensitive components such as photoresists. The material is coated to form a thin film on the substrate. After radiation irradiation, use alkaline cleaning solution to remove unnecessary parts of the coating to obtain a good pattern.

Since the photoresist layer on the edges of the substrate is more uneven than in the center of the substrate, the uneven photoresist layer or beads in the chip must be removed and the substrate must be cleaned. It is known that photoresist cleaning agents are usually methyl isobutyl ketone. Although photoresist cleaning agents with this component have relatively satisfactory photoresist cleaning capabilities, they are toxic to people and the environment, so their use is restricted by ISO14000 environmental management certification.

Therefore, it is necessary to replace the use of methyl isobutyl ketone with other substances. In recent years, a variety of photoresist cleaning agents have appeared. For example, US Patent 4983490 discloses a composition including 1-10 parts by weight of propylene glycol monomethyl ether and 1-10 parts by weight of propylene glycol monomethyl ether acetate. Chinese patent CN1987663B discloses a photoresist cleaning agent, which is composed of propylene glycol monomethyl ether acetate and cyclohexanone, in which the weight percentage ratio of propylene glycol monomethyl ether acetate and cyclohexanone is 70% to 30 %.

The photoresist cleaning agent sold on the market is a cleaning agent with formula data, but in fact, the photoresist cleaning agent with the same formula will have different performance in different application scenarios. That is, in the actual industry, the same photoresist cleaning agent A photoresist cleaning agent with a formula may work well in scenario A, but not in scenario B. Therefore, corresponding to different application scenarios, the proportions of different ingredients in the formula need to be adjusted to adaptively obtain scene-adaptive photoresist cleaning agents. That is to say, for actual demanders, it is expected to provide customized photoresist cleaning solutions based on the particularity of the objects they want to clean (for example, substrates).

Therefore, we look forward to an automatic dispensing solution for photoresist cleaning fluid based on specific application requirements in specific scenarios.

Contents of the invention

In order to solve the above technical problems, this application is proposed. Embodiments of the present application provide an automatic batching system and batching method for the production of photoresist cleaning fluid, which uses neural network technology based on deep learning to intelligently determine the local optimal quality of the formula ingredients one by one. Through this The method is used to determine the optimal ratio of the photoresist cleaning fluid for a specific object, and then perform automatic batching based on this optimal ratio.

According to one aspect of the present application, an automatic batching system for photoresist cleaning liquid production is provided, which includes:

The formula data unit is used to obtain multiple formula data of the formula of the photoresist cleaning liquid, wherein the formula of the photoresist cleaning liquid includes water, hydroxyl quaternary ammonium salt compounds, alcoholamine compounds and non- Ionic surfactant, the weights of the water, the quaternary ammonium hydroxide salt compound, and the alcoholamine compound in the multiple formula data of the photoresist cleaning liquid formula are consistent, and the weights are consistent There is a difference in the weight of the non-ionic surfactant; a formula data semantic encoding unit is used to pass each formula data in the plurality of formula data through a trained context encoder including an embedding layer to generate multiple formulas respectively. Ingredient feature vectors, and respectively concatenate the multiple recipe ingredient feature vectors to obtain multiple first feature vectors corresponding to the multiple recipe data; a first convolution coding unit, used to convert the multiple recipe ingredient feature vectors into The plurality of first feature vectors of the plurality of recipe data are two-dimensionally arranged into a feature matrix and then passed through the trained first convolutional neural network to obtain the first feature matrix; the cleaning test data acquisition unit is used to obtain the said The plurality of formula data of the formula of the photoresist cleaning liquid are directed to the cleaning test video of the object to be cleaned and the plurality of time values of the formula data of the formula of the photoresist cleaning liquid to achieve a predetermined clear effect; the video encoding unit, A video encoder for passing the cleaning test video through a trained joint encoder to obtain a first feature vector, the video encoder encoding the cleaning test video using a trained three-dimensional convolutional neural network ; Temporal encoding unit, used to construct the plurality of time values as input vectors and obtain the second feature vector through a temporal encoder including a one-dimensional convolutional layer and a fully connected layer of the trained joint encoder. ; a joint encoding unit for using the joint encoder to fuse the first feature vector and the second feature vector to generate a second feature matrix; a feature matrix fusion unit for fusing the first feature matrix and The second characteristic matrix is used to generate a decoding characteristic matrix; a decoding unit is used to pass the decoding characteristic matrix through a decoder to generate a decoding value, and the decoding value is the local optimal weight value of the nonionic surfactant. ; And a batching unit, used to generate a batching plan based on the decoded value.

In the above-mentioned automatic batching system for the production of photoresist cleaning fluid, a training module is also included. The training module is used to train the context encoder including the embedding layer, the first convolutional neural network and the joint The encoder performs training, wherein the training module includes: a training formula data unit for obtaining multiple formula data of the formula of the photoresist cleaning fluid, wherein the formula of the photoresist cleaning fluid includes water, hydrogen Oxygen quaternary ammonium salt compounds, alcohol amine compounds and nonionic surfactants, the water and the hydroxide quaternary ammonium salts mentioned in the multiple formula data of the photoresist cleaning liquid formula The weights of the compound and the alcoholamine compound are consistent, and the weight of the non-ionic surfactant is different; a training formula data semantic encoding unit is used to pass each formula data in the multiple formula data by including Embedding the context encoder of the layer to respectively generate a plurality of recipe ingredient feature vectors, and respectively concatenating the multiple recipe ingredient feature vectors to obtain a plurality of first feature vectors corresponding to the multiple recipe data; training the first A convolutional coding unit for two-dimensionally arranging the plurality of first feature vectors corresponding to the plurality of recipe data into a feature matrix and then passing the first convolutional neural network to obtain the first feature matrix; training and cleaning A test data acquisition unit, configured to acquire multiple formula data of the formula of the photoresist cleaning fluid, a cleaning test video of the object to be cleaned, and multiple formula data of the formula of the photoresist cleaning fluid to achieve a predetermined cleaning effect. multiple time values; train the video encoding unit to pass the cleaning test video through the video encoder of the joint encoder to obtain the first feature vector, the video encoder uses a three-dimensional convolutional neural network to perform the cleaning test The video is encoded; a temporal encoding unit is trained to construct the multiple temporal values as input vectors and then pass the temporal encoder including a one-dimensional convolutional layer and a fully connected layer of the joint encoder to obtain the second feature vector. ;Training a joint encoding unit for using the joint encoder to fuse the first feature vector and the second feature vector to generate a second feature matrix; Training a feature matrix fusion unit for fusing the first features matrix and the second feature matrix to generate a decoding feature matrix; a first loss calculation unit for calculating a dimensional distribution similarity constraint for feature manifolds between the first feature matrix and the second feature matrix. The loss function value, wherein the loss function value for the dimensional distribution similarity constraint of the feature manifold is based on the cosine distance between the first feature matrix and the second feature matrix and the first feature matrix and the second feature matrix; a second loss calculation unit for passing the decoding feature matrix through a decoder to generate a decoding loss function value; a training unit for using the feature flow as described The context encoder including the embedding layer, the first convolutional neural network and the joint encoder are trained by using a weighted sum of the loss function value constrained by the shape of the dimension distribution similarity and the decoding loss function value.

In the above-mentioned automatic batching system for the production of photoresist cleaning fluid, the training recipe data semantic encoding unit is further used to: use the embedding layer of the encoder model containing the context of the embedding layer to respectively separate the plurality of Each recipe data in the recipe data is converted into an input vector to obtain a sequence of input vectors; using the converter of the encoder model containing the context of the embedded layer to perform global context-based semantic encoding on the sequence of input vectors to obtain the sequence of input vectors. A plurality of formula ingredient feature vectors; the plurality of formula ingredient feature vectors are respectively concatenated to obtain a plurality of first feature vectors corresponding to the plurality of recipe data.

In the above-mentioned automatic batching system for the production of photoresist cleaning fluid, the training of the first convolutional coding unit is further used: each layer of the first convolutional neural network performs input processing in the forward transmission of the layer. The data is subjected to convolution processing, mean pooling processing along the channel dimension and activation processing to generate the first feature matrix from the last layer of the first convolutional neural network, wherein The input to the first layer is the feature matrix.

In the above-mentioned automatic batching system for the production of photoresist cleaning fluid, the training video encoding unit is further used: the video encoder of the convolutional neural network using a three-dimensional convolution kernel uses the following formula to encode the cleaning test video Processing is performed to generate the first feature vector; wherein the formula is:

Among them, H _j , W _j and R _j represent the length, width and height of the three-dimensional convolution kernel respectively, m represents the number of the (l-1)th layer feature map,

is the convolution kernel connected to the m-th feature map of layer (l-1), b _lj is the bias, and f(·) represents the activation function.

In the above-mentioned automatic batching system for the production of photoresist cleaning fluid, the training timing encoding unit includes: a construction subunit for arranging the multiple time values into a one-dimensional input vector; a fully connected subunit , used to use the fully connected layer of the temporal encoder to fully connect the input vector obtained by the construction subunit with the following formula to extract the high-dimensional hidden features of the eigenvalues of each position in the input vector, Among them, the formula is:

where X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector,

Represents matrix multiplication; a one-dimensional convolution subunit, used to perform one-dimensional convolution encoding on the input vector obtained by the construction subunit using the following formula using the one-dimensional convolution layer of the temporal encoder to extract the The high-dimensional implicit correlation features between the eigenvalues of each position in the input vector, where the formula is:

Among them, a is the width of the convolution kernel in the x direction, F is the convolution kernel parameter vector, G is the local vector matrix that operates with the convolution kernel function, and w is the size of the convolution kernel.

In the above-mentioned automatic batching system for the production of photoresist cleaning fluid, the first loss calculation unit is further used to calculate the relationship between the first characteristic matrix and the second characteristic matrix according to the following formula: The loss function value used for the dimensional distribution similarity constraint of the feature manifold; where the formula is:

Where cos(M ₁ , M ₂ ) represents the cosine distance between the first feature matrix M ₁ and the second feature matrix M ₂ , and d(M ₁ , M ₂ ) represents the Euclidean distance therebetween.

In the above-mentioned automatic batching system for photoresist cleaning liquid production, the second loss calculation unit is further used to: use a decoder to perform decoding regression on the decoding feature matrix using the following formula to obtain the decoding value; Among them, the formula is:

where X is the decoding feature matrix, Y is the decoding value, and W is the weight matrix,

Represents matrix multiplication; calculate the cross entropy value between the decoded value and the real value as the decoding loss function value.

According to another aspect of the present application, a batching method of an automatic batching system for photoresist cleaning liquid production includes: obtaining multiple recipe data of the photoresist cleaning liquid recipe, wherein the photoresist cleaning liquid The formula of the cleaning liquid includes water, quaternary ammonium hydroxide salt compounds, alcohol amine compounds and non-ionic surfactants. The water, the The weights of the hydroxyl quaternary ammonium salt compounds and the alcoholamine compounds are consistent, and the weights of the nonionic surfactants are different;

Pass each of the multiple recipe data through a trained context encoder including an embedding layer to generate multiple recipe ingredient feature vectors, and concatenate the multiple recipe ingredient feature vectors to obtain A plurality of first feature vectors corresponding to the plurality of recipe data; after the plurality of first feature vectors corresponding to the plurality of recipe data are two-dimensionally arranged into a feature matrix, the first volume completed through training Accumulate the neural network to obtain the first feature matrix; obtain multiple formula data of the formula of the photoresist cleaning fluid for the cleaning test video of the object to be cleaned and the multiple formula data of the formula of the photoresist cleaning fluid to achieve Predetermine multiple time values of the clear effect; pass the cleaning test video through the video encoder of the trained joint encoder to obtain the first feature vector, the video encoder uses the trained three-dimensional convolutional neural network to obtain the first feature vector The cleaning test video is encoded; after constructing the plurality of time values as input vectors, the second feature is obtained through the temporal encoder of the trained joint encoder including a one-dimensional convolutional layer and a fully connected layer. vector; using the joint encoder to fuse the first feature vector and the second feature vector to generate a second feature matrix; fusing the first feature matrix and the second feature matrix to generate a decoding feature matrix; The decoded feature matrix is passed through a decoder to generate a decoded value that is a local optimal weight value of the nonionic surfactant; and a dosing plan is generated based on the decoded value.

Compared with the existing technology, the automatic batching system and batching method provided by this application for the production of photoresist cleaning fluid uses neural network technology based on deep learning to intelligently determine the local optimal quality of the formula ingredients one by one. In this way, the optimal ratio of the photoresist cleaning fluid for a specific object is determined, and then automatic batching is performed based on the optimal ratio.

Description of the drawings

The above and other objects, features and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The drawings are used to provide further understanding of the embodiments of the present application, and constitute a part of the specification. They are used to explain the present application together with the embodiments of the present application, and do not constitute a limitation of the present application. In the drawings, like reference numbers generally represent like components or steps.

Figure 1 is an application scenario diagram of an automatic batching system for the production of photoresist cleaning fluid according to an embodiment of the present application.

FIG. 2 is a block diagram of an automatic batching system for producing photoresist cleaning fluid according to an embodiment of the present application.

FIG. 3A is a flow chart of the training phase in the batching method of the automatic batching system for photoresist cleaning liquid production according to an embodiment of the present application.

3B is a flow chart of the inference stage in the batching method of the automatic batching system for photoresist cleaning liquid production according to an embodiment of the present application.

FIG. 4A is a schematic diagram of the architecture of the training phase in the batching method of the automatic batching system for photoresist cleaning liquid production according to an embodiment of the present application.

4B is a schematic diagram of the architecture of the inference stage in the batching method of the automatic batching system for photoresist cleaning liquid production according to an embodiment of the present application.

Detailed ways

Hereinafter, example embodiments according to the present application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. It should be understood that the present application is not limited by the example embodiments described here.

Scenario overview

As mentioned above, in order to obtain fine images in flat panel displays such as liquid crystal, organic EL, plasma displays, semiconductors, and printed circuit boards, general photolithography technology is used to form photosensitive patterns, using photoresist. Radiation-sensitive compositions such as agents are coated to form a thin film on the substrate. After radiation irradiation, use alkaline cleaning solution to remove unnecessary parts of the coating to obtain a good pattern.

Since the photoresist layer on the edges of the substrate is more uneven than in the center of the substrate, the uneven photoresist layer or beads in the chip must be removed and the substrate must be cleaned. Known photoresist cleaning agents are usually methyl isobutyl ketone (Methyl isobutyl ketone, MIBK). Although the photoresist cleaning agent with this composition has a relatively satisfactory ability to clean photoresist, it is toxic to people and the environment, so its use is restricted by ISO14000 environmental management certification.

Therefore, it is necessary to replace the use of methyl isobutyl ketone with other substances. A variety of photoresist cleaning agents have appeared in recent years. For example, U.S. Patent 4983490 discloses a composition including 1-10 parts by weight of propylene glycol mono-methyl ether (PGME) and 1-10 parts by weight of acetic acid. Propylene glycol mono-methyl ether acetate (PGMEA). Chinese patent CN1987663 B discloses a photoresist cleaning agent, which is composed of propylene glycol monomethyl ether acetate and cyclohexanone, in which the weight percentage ratio of propylene glycol monomethyl ether acetate and cyclohexanone is 70%. 30%.

In recent years, deep learning and neural networks have been widely used in computer vision, natural language processing, text signal processing and other fields. In addition, deep learning and neural networks have also shown that they are close to or even beyond human performance in areas such as image classification, object detection, semantic segmentation, and text translation.

In recent years, the development of deep learning and neural networks has provided solutions and solutions for the automatic batching of photoresist cleaning fluid.

When selecting an adaptive formula, it is obviously impossible to obtain the best formula through exhaustive methods, because this approach is too wasteful of manpower and material resources. On the other hand, there is a relationship between the various formula components of the formula, and the effect of one formula component is difficult to judge apart from the formula as a whole. In order to adapt to the complex non-linear mapping relationship between different formula data corresponding to cleaning effects, the inventor of the present application tried to use neural network technology based on deep learning to intelligently infer a relatively better batching solution.

Specifically, in the technical solution of the present application, multiple formula data of the formula of the photoresist cleaning liquid are first obtained. Taking the formula of the photoresist cleaning solution as water, hydroxyl quaternary ammonium salt compounds, alcohol amine compounds and non-ionic surfactants as an example, in the multiple formula data, the The weights of water, the quaternary ammonium hydroxide salt compound, and the alcoholamine compound are consistent, and the weight of the nonionic surfactant is different, that is, the weight of other components is controlled to remain unchanged, The weight of the nonionic surfactant was adjusted to generate multiple formulation data. This uses the idea of controlling variables, that is, setting the weights of other formula components in the formula to constant values to solve for the local optimal value of a single formula component, and then obtaining the photoresist cleaning method through step-by-step iteration. The best recipe data for the cleansing liquid recipe.

In order to extract the high-dimensional implicit correlation between each formula component in the formula data of the photoresist cleaning solution formula, a context encoder including an embedding layer is used to perform full-text-based high-dimensional semantic encoding on each formula data to generate the corresponding A plurality of first feature vectors of the plurality of recipe data. In order to screen out the local optimal recipe, a plurality of first feature vectors of the multiple recipe data are further two-dimensionally arranged into a feature matrix, wherein each row vector in the feature matrix is one of the first features. The vector, that is, the feature matrix constructs the association between each recipe data at the data level.

Further, a first convolutional neural network is used to encode the feature matrix to extract high-dimensional local implicit correlation features in the feature matrix, that is, extract each recipe ingredient, all ingredients in the multiple recipe data Each recipe data in the plurality of recipe data and the high-dimensional implicit association of each recipe ingredient in one recipe data are used to generate the first feature matrix.

Next, cleaning test videos of multiple formula data of the photoresist cleaning solution are obtained. That is, a specific object to be cleaned is obtained, and then the photoresist cleaning liquid with different formula data is used to clean it, and the cleaning process is video-monitored. It should be understood that the cleaning test video data contains dynamic cleaning process characteristics and information of the object to be cleaned using photoresist cleaning fluids with different formula data. Correspondingly, in order to evaluate the correlation between the cleaning effects of different formula data and the cleaning effects of different formula data, in the embodiment of the present application, a joint encoder is used to jointly encode the cleaning test video to generate a second feature matrix, Wherein, the joint encoder includes a video encoder for encoding video data and a timing encoder for encoding tag data. The tag data is a plurality of formulas of the photoresist cleaning solution. Multiple time values for the data to achieve the predetermined cleaning effect.

Then, by fusing the high-dimensional latent features used to represent different formula data and the high-dimensional latent features used to represent the clear effects of different formula data, decoding regression can be performed to obtain the local optimal weight value of the nonionic surfactant.

When fusing the first feature matrix and the second feature matrix, since the first feature matrix expresses the inter-sample correlation features of semantic context coding, and the second feature matrix expresses the image semantic feature coding of temporal direction constraints, the corresponding feature flow pattern There will be a large offset in the high-dimensional feature space, resulting in sparse features of the fused feature matrix.

Based on this, a loss function for the dimensional distribution similarity constraint of the feature manifold is introduced, expressed as:

Where cos(M ₁ , M ₂ ) represents the cosine distance between the first characteristic matrix M ₁ and the second characteristic matrix M ₂ , and d(M ₁ , M ₂ ) represents the Euclidean distance therebetween.

By training the respective encoding branches of the first feature matrix and the second feature matrix with this loss function, the distribution similarity of the feature manifold observed from different dimensional perspectives in the high-dimensional feature space can be constrained, that is, through the geometric similarity of the distribution property constraints, the local feature description of the association between the first feature matrix and the second feature matrix can be optimized to alleviate the feature sparseness of the fused feature matrix caused by the spatial complexity of the high-dimensional feature space. In this way, the accuracy of decoding regression is improved.

Further, after determining the local optimal weight value of the nonionic surfactant, the weight of the nonionic surfactant in the photoresist cleaning solution is set to the local optimal weight value, and The local optimal quality of other formula components is determined one by one in the method described above. In this way, the optimal ratio of the photoresist cleaning liquid for a specific object is determined, and then automatic processing can be performed based on this optimal ratio. Ingredients.

Based on this, this application proposes an automatic batching system for the production of photoresist cleaning fluid, which includes: a formula data unit used to obtain multiple formula data of the formula of the photoresist cleaning fluid, wherein the photoresist cleaning fluid The formula of the photoresist cleaning fluid includes water, quaternary ammonium hydroxide salt compounds, alcohol amine compounds and non-ionic surfactants. Water, The weights of the hydroxyl quaternary ammonium salt compounds and the alcoholamine compounds are consistent, and the weights of the nonionic surfactants are different; the formula data semantic encoding unit is used to combine the multiple Each recipe data in the recipe data is passed through the trained context encoder including the embedding layer to respectively generate multiple recipe ingredient feature vectors, and the multiple recipe ingredient feature vectors are respectively cascaded to obtain the corresponding feature vectors of the multiple recipe ingredients. A plurality of first feature vectors of a plurality of recipe data; a first convolution coding unit, configured to two-dimensionally arrange the plurality of first feature vectors corresponding to the plurality of recipe data into a feature matrix and then complete the training through The first convolutional neural network obtains the first feature matrix; a cleaning test data acquisition unit is used to acquire multiple formula data of the formula of the photoresist cleaning liquid for the cleaning test video of the object to be cleaned and the light A plurality of time values for the formula data of the blocking cleaning liquid to reach a predetermined clear effect; a video encoding unit for passing the cleaning test video through the video encoder of the trained joint encoder to obtain the first feature Vector, the video encoder uses a trained three-dimensional convolutional neural network to encode the cleaning test video; a temporal encoding unit is used to construct the multiple time values into input vectors and then pass all the trained A temporal encoder including a one-dimensional convolutional layer and a fully connected layer of the joint encoder to obtain the second feature vector; a joint encoding unit for using the joint encoder to fuse the first feature vector and the third feature vector two feature vectors to generate a second feature matrix; a feature matrix fusion unit, used to fuse the first feature matrix and the second feature matrix to generate a decoding feature matrix; a decoding unit, used to decode the decoding feature matrix a device to generate a decoded value, the decoded value being the local optimal weight value of the nonionic surfactant; and a batching unit for generating a batching plan based on the decoded value.

FIG. 1 illustrates an application scenario diagram of an automatic batching system for photoresist cleaning liquid production according to an embodiment of the present application. As shown in Figure 1, in this application scenario, first, the photoresist cleaning liquid (for example, as shown in Figure 1) is obtained from an automatic batching system (for example, H as shown in Figure 1) for photoresist cleaning liquid production. Multiple recipe data of the recipe P) shown in Figure 1, and the multiple recipe data of the recipe of the photoresist cleaning liquid is acquired through a camera (for example, C shown in Figure 1) for the target to be cleaned The cleaning test video of the object (for example, B as shown in Figure 1) and the multiple recipe data of the photoresist cleaning liquid formula obtained through a timer (for example, T as shown in Figure 1) reach a predetermined Multiple time values for the cleaning effect. Then, the plurality of recipe data, the cleaning test video and the plurality of time values are input into a server deployed with an automatic batching algorithm for photoresist cleaning liquid production (for example, as illustrated in Figure 1 Cloud server S), wherein the server can process the plurality of formula data, the cleaning test video and the plurality of time values with an automatic batching algorithm for photoresist cleaning liquid production to generate a Based on the decoded value representing the local optimal weight value of the nonionic surfactant, a batching plan is then generated based on the decoded value.

After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be specifically introduced below with reference to the accompanying drawings.

Example system

FIG. 2 illustrates a block diagram of an automatic dispensing system for photoresist cleaning liquid production according to an embodiment of the present application. As shown in Figure 2, an automatic batching system 200 for photoresist cleaning liquid production according to an embodiment of the present application includes: a training module 210 and an inference module 220. Among them, the training module 210 includes: a training formula data unit 2101, which is used to obtain multiple formula data of the formula of the photoresist cleaning liquid, wherein the formula of the photoresist cleaning liquid includes water, hydroxyl quaternary ammonium base Salt compounds, alcohol amine compounds and non-ionic surfactants, the water, the hydroxyl quaternary ammonium salt compound, the alcohol mentioned in the multiple formula data of the photoresist cleaning liquid formula The weight of the amine compound is consistent, and the weight of the non-ionic surfactant is different; the training recipe data semantic encoding unit 2102 is used to pass each of the multiple recipe data through the context including the embedded layer. The encoder generates a plurality of recipe ingredient feature vectors respectively, and respectively concatenates the multiple recipe ingredient feature vectors to obtain a plurality of first feature vectors corresponding to the multiple recipe data; trains the first convolutional encoding Unit 2013 is used to two-dimensionally arrange the plurality of first feature vectors corresponding to the plurality of recipe data into a feature matrix and then obtain the first feature matrix through the first convolutional neural network; training and cleaning test data acquisition Unit 2104 is used to obtain multiple formula data of the formula of the photoresist cleaning liquid, a cleaning test video of the object to be cleaned, and multiple formula data of the formula of the photoresist cleaning liquid to achieve the predetermined cleaning effect. time values; train the video encoding unit 2105 for passing the cleaning test video through the video encoder of the joint encoder to obtain the first feature vector. The video encoder uses a three-dimensional convolutional neural network to encode the cleaning test video. Carry out encoding; train the temporal encoding unit 2106, which is used to construct the plurality of time values into input vectors and then pass the temporal encoder including a one-dimensional convolutional layer and a fully connected layer of the joint encoder to obtain the second feature vector ; Training joint encoding unit 2107, for using the joint encoder to fuse the first feature vector and the second feature vector to generate a second feature matrix; Training feature matrix fusion unit 2108, for fusing the first feature vector A feature matrix and the second feature matrix to generate a decoding feature matrix; a first loss calculation unit 2109 for calculating the dimensional distribution for the feature manifold between the first feature matrix and the second feature matrix The loss function value of the similarity constraint, wherein the loss function value of the dimensional distribution similarity constraint for the feature manifold is based on the cosine distance between the first feature matrix and the second feature matrix and the third feature matrix. The Euclidean distance between a feature matrix and the second feature matrix is generated; the second loss calculation unit 2110 passes the decoding feature matrix through the decoder to generate a decoding loss function value; the training unit 2111 is used to use the The context encoder including the embedding layer, the first convolutional neural network and the joint encoder are trained by the weighted sum of the loss function value constrained by the dimensional distribution similarity of the feature manifold and the decoding loss function value. . Among them, the inference module 220 includes: a formula data unit 2201, which is used to obtain multiple formula data of the formula of the photoresist cleaning liquid, wherein the formula of the photoresist cleaning liquid includes water, quaternary ammonium hydroxide base salt compounds, alcoholamine compounds and nonionic surfactants, the water, the hydroxyl quaternary ammonium salt compound, the alcoholamine in the multiple formula data of the photoresist cleaning liquid formula The weights of the similar compounds are consistent, and the weights of the non-ionic surfactants are different; the formula data semantic encoding unit 2202 is used to pass each formula data in the plurality of formula data through the trained embedding layer. a context encoder to respectively generate a plurality of recipe ingredient feature vectors, and respectively concatenate the multiple recipe ingredient feature vectors to obtain a plurality of first feature vectors corresponding to the multiple recipe data; a first convolution Encoding unit 2203, configured to two-dimensionally arrange the plurality of first feature vectors corresponding to the plurality of recipe data into a feature matrix and then obtain the first feature matrix through the trained first convolutional neural network; The cleaning test data acquisition unit 2204 is used to acquire multiple formula data of the formula of the photoresist cleaning liquid for the cleaning test video of the object to be cleaned and the multiple formula data of the formula of the photoresist cleaning liquid reaches a predetermined level. Multiple time values of the clear effect; the video encoding unit 2205 is used to pass the cleaning test video through the video encoder of the trained joint encoder to obtain the first feature vector, the video encoder uses the trained joint encoder to obtain the first feature vector The three-dimensional convolutional neural network encodes the cleaning test video; the temporal encoding unit 2206 is used to construct the multiple time values into input vectors and then pass the one-dimensional convolution layer of the trained joint encoder and a temporal encoder of the fully connected layer to obtain the second feature vector; a joint encoding unit 2207 for using the joint encoder to fuse the first feature vector and the second feature vector to generate a second feature matrix; The feature matrix fusion unit 2208 is used to fuse the first feature matrix and the second feature matrix to generate a decoding feature matrix; the decoding unit 2209 is used to pass the decoding feature matrix through a decoder to generate a decoding value, the The decoded value is the local optimal weight value of the nonionic surfactant; and, a batching unit 2210 is used to generate a batching plan based on the decoded value.

Specifically, in the embodiment of the present application, the training recipe data unit 2101 and the training recipe data semantic encoding unit 2102 are used to obtain multiple recipe data of the photoresist cleaning liquid recipe, wherein the photoresist cleaning solution The formula of the cleaning liquid includes water, quaternary ammonium hydroxide salt compounds, alcohol amine compounds and non-ionic surfactants. The water, all the ingredients mentioned in the multiple formula data of the photoresist cleaning liquid formula are The weights of the hydroxyl quaternary ammonium salt compounds and the alcoholamine compounds are consistent, and the weights of the nonionic surfactants are different, and each formula data in the multiple formula data is passed through A context encoder including an embedding layer is used to respectively generate a plurality of recipe ingredient feature vectors, and respectively concatenate the plurality of recipe ingredient feature vectors to obtain a plurality of first feature vectors corresponding to the plurality of recipe data. It should be understood that when selecting an adaptive formula, it is obviously impossible to obtain the best formula through exhaustive methods, because this approach is too wasteful of manpower and material resources. On the other hand, there is a relationship between the various formula components of the formula, and the effect of one formula component is difficult to judge apart from the formula as a whole. Therefore, in order to adapt to the complex nonlinear mapping relationship between different formula data corresponding to cleaning effects, in the technical solution of this application, neural network technology based on deep learning is used to intelligently infer a relatively better batching solution.

That is, specifically, in the technical solution of the present application, multiple formula data of the formula of the photoresist cleaning liquid are first obtained. Taking the formula of the photoresist cleaning solution as water, hydroxyl quaternary ammonium salt compounds, alcohol amine compounds and non-ionic surfactants as an example, in the multiple formula data, the The weights of water, the quaternary ammonium hydroxide salt compound, and the alcoholamine compound are consistent, and the weight of the nonionic surfactant is different, that is, the weight of other components is controlled to remain unchanged, The weight of the nonionic surfactant was adjusted to generate multiple formulation data. This uses the idea of controlling variables, that is, setting the weights of other formula components in the formula to constant values to solve for the local optimal value of a single formula component, and then obtaining the photoresist cleaning method through step-by-step iteration. The best recipe data for the cleansing liquid recipe.

Then, it should be understood that in order to extract the high-dimensional implicit correlation between the various formula components in the formula data of the photoresist cleaning liquid formula, a context encoder including an embedding layer is further used to perform a full-text analysis on each formula data respectively. high-dimensional semantic encoding to generate a plurality of first feature vectors corresponding to the plurality of recipe data.

More specifically, in the embodiment of the present application, the training recipe data semantic encoding unit is further configured to: use the embedding layer of the encoder model containing the context of the embedding layer to respectively convert each recipe in the plurality of recipe data. The data is converted into input vectors to obtain a sequence of input vectors; the sequence of input vectors is globally-based contextual semantic encoding using a converter of the encoder model containing the context of the embedding layer to obtain the multiple recipe ingredient features. Vector; respectively concatenate the plurality of formula ingredient feature vectors to obtain a plurality of first feature vectors corresponding to the plurality of formula data.

Specifically, in the embodiment of the present application, the training first convolutional encoding unit 2013 is used to two-dimensionally arrange the plurality of first feature vectors corresponding to the plurality of recipe data into a feature matrix and then pass The first convolutional neural network obtains the first feature matrix. It should be understood that in order to screen out the local optimal recipe, in the technical solution of the present application, the plurality of first feature vectors of the multiple recipe data are further two-dimensionally arranged into a feature matrix. Here, in the feature matrix Each row vector is one of the first feature vectors, that is, the feature matrix constructs a correlation between each recipe data at the data level.

Further, the first convolutional neural network is used to encode the feature matrix to extract high-dimensional local implicit correlation features in the feature matrix, that is, to extract each recipe component in the multiple recipe data , each recipe data in the plurality of recipe data, and a high-dimensional implicit association of each recipe ingredient in one recipe data to generate the first feature matrix. Correspondingly, in a specific example, each layer of the first convolutional neural network performs convolution processing, mean pooling processing along the channel dimension and activation processing on the input data in the forward pass of the layer to be processed by the The last layer of the first convolutional neural network generates the first feature matrix, wherein the input of the first layer of the first convolutional neural network is the feature matrix.

Specifically, in this embodiment of the present application, the training cleaning test data acquisition unit 2104 and the training video encoding unit 2105 are used to obtain multiple formula data of the photoresist cleaning liquid formula for the object to be cleaned. The cleaning test video and the multiple formula data of the photoresist cleaning liquid formula reach multiple time values of the predetermined cleaning effect, and the cleaning test video is passed through the video encoder of the joint encoder to obtain the first feature vector , the video encoder uses a three-dimensional convolutional neural network to encode the cleaning test video. That is to say, in the technical solution of the present application, a plurality of formula data of the formula of the photoresist cleaning liquid and a cleaning test video of the object to be cleaned are further obtained through a camera and the photoresist cleaning liquid is obtained through a timer. The multiple recipe data of the recipe achieve multiple time values for the predetermined cleaning effect. That is, a specific object to be cleaned is obtained, and then the photoresist cleaning liquid with different formula data is used to clean it, and the cleaning process is video-monitored.

It should be understood that the cleaning test video data contains dynamic cleaning process characteristics and information of the object to be cleaned using photoresist cleaning fluids with different formula data. Correspondingly, in order to evaluate the correlation between the cleaning effects of different formula data and the cleaning effects of different formula data, in the embodiment of the present application, a joint encoder is used to jointly encode the cleaning test video to generate a second feature matrix, Wherein, the joint encoder includes a video encoder for encoding video data and a timing encoder for encoding tag data. The tag data is a plurality of formulas of the photoresist cleaning solution. Multiple time values for the data to achieve the predetermined cleaning effect. That is, specifically, in the technical solution of the present application, first, the cleaning test video is encoded through a video encoder with a three-dimensional convolutional neural network of a joint encoder to extract the different recipe data The photoresist cleaning fluid determines the dynamic cleaning implicit characteristics of the object to be cleaned, thereby obtaining the first feature vector.

More specifically, in the embodiment of the present application, the training video encoding unit is further used to: use the video encoder of the convolutional neural network of the three-dimensional convolution kernel to process the cleaning test video according to the following formula to generate the The first eigenvector;

Among them, the formula is:

Specifically, in this embodiment of the present application, the training temporal encoding unit 2106 and the training joint encoding unit 2107 are used to construct the multiple time values into input vectors and then pass the joint encoder to a one-dimensional The temporal encoder of the convolutional layer and the fully connected layer obtains the second feature vector, and the joint encoder is used to fuse the first feature vector and the second feature vector to generate a second feature matrix. That is to say, in the technical solution of the present application, the tag data needs to be further encoded, that is, the multiple formula data of the photoresist cleaning solution formula needs to be sequenced for multiple time values to achieve the predetermined cleaning effect. Encoding to obtain the multiple time values and high-dimensional implicit correlation feature information between the multiple time values, thereby obtaining a second feature vector. In this way, the joint encoder is then used to fuse the first feature vector and the second feature vector to generate a second feature matrix, which can then enhance the encoding of the image along a specific direction of the text to encode the cleaning test video. The corresponding attributes of the image frame highlight the implicit correlation features related to the time value label, thereby improving the accuracy of subsequent decoding and regression.

More specifically, in this embodiment of the present application, the training of the temporal coding unit includes: first, arranging the multiple time values into a one-dimensional input vector. Then, use the fully connected layer of the temporal encoder to perform fully connected encoding on the input vector obtained by the construction subunit with the following formula to extract the high-dimensional hidden features of the feature values of each position in the input vector, where , the formula is:

Represents matrix multiplication. Finally, use the one-dimensional convolution layer of the temporal encoder to perform one-dimensional convolution encoding on the input vector obtained by the construction subunit using the following formula to extract the difference between the feature values of each position in the input vector High-dimensional implicit correlation features, where the formula is:

Specifically, in this embodiment of the present application, the training feature matrix fusion unit 2108 and the first loss calculation unit 2109 are used to fuse the first feature matrix and the second feature matrix to generate a decoding feature matrix, And calculate the loss function value between the first feature matrix and the second feature matrix for the dimensional distribution similarity constraint of the feature manifold, wherein the loss function value for the dimensional distribution similarity constraint of the feature manifold is The loss function value is generated based on the cosine distance between the first feature matrix and the second feature matrix and the Euclidean distance between the first feature matrix and the second feature matrix. That is, in the technical solution of the present application, then, by fusing the high-dimensional implicit features used to represent different formula data and the high-dimensional implicit features used to represent the clear effects of different formula data, decoding regression can be performed to obtain the non-ionic properties. Local optimal weight value of surfactant.

It should be understood that when fusing the first feature matrix and the second feature matrix, the first feature matrix expresses the inter-sample correlation features of semantic context encoding, while the second feature matrix expresses the temporal direction constraints. Image semantic feature encoding, the corresponding feature flow pattern will have a large offset in the high-dimensional feature space, resulting in sparse features of the fused feature matrix. Therefore, in the technical solution of this application, a loss function for the dimensional distribution similarity constraint of the feature manifold is further introduced.

More specifically, in the embodiment of the present application, the first loss calculation unit is further configured to: calculate the characteristic manifold between the first characteristic matrix and the second characteristic matrix using the following formula: The loss function value of the dimension distribution similarity constraint;

Among them, the formula is:

Where cos(M ₁ , M ₂ ) represents the cosine distance between the first feature matrix M ₁ and the second feature matrix M ₂ , and d(M ₁ , M ₂ ) represents the Euclidean distance therebetween. It should be understood that in this way, by training the respective encoding branches of the first feature matrix and the second feature matrix with the loss function, the feature manifold can be constrained to be observed from different dimensional perspectives in the high-dimensional feature space. Distribution similarity, that is, through the geometric similarity constraint of distribution, the local feature description of the association between the first feature matrix and the second feature matrix can be optimized to alleviate the high-dimensionality problem of the fused feature matrix. The feature sparseness caused by the spatial complexity of the feature space improves the accuracy of decoding regression.

Specifically, in this embodiment of the present application, the second loss calculation unit 2110 and the training unit 2111 pass the decoding feature matrix through a decoder to generate a decoding loss function value, and use the decoding feature matrix for feature manifold The context encoder including the embedding layer, the first convolutional neural network and the joint encoder are trained by using the weighted sum of the loss function value constrained by the dimension distribution similarity and the decoding loss function value. That is, in the technical solution of the present application, the decoding feature matrix is further passed through a decoder to generate a decoding loss function value. Correspondingly, in a specific example, a decoder is used to perform decoding regression on the decoding feature matrix with the following formula to obtain the decoding value; wherein the formula is:

Represents matrix multiplication; calculate the cross entropy value between the decoded value and the real value as the decoding loss function value. Then, the context encoder including the embedding layer and the first convolution can be trained with the weighted sum of the loss function value for the dimensional distribution similarity constraint of the feature manifold and the decoding loss function value. Neural network and the joint encoder.

After training is completed, enter the inference module. That is, the trained context encoder including the embedding layer, the first convolutional neural network and the joint encoder are used in actual inference.

Specifically, in the embodiment of the present application, first, multiple formula data of the formula of the photoresist cleaning liquid are obtained, wherein the formula of the photoresist cleaning liquid includes water, hydroxyl quaternary ammonium salt compound, Alcoholamine compounds and nonionic surfactants, the water, the hydroxyl quaternary ammonium salt compound, and the alcoholamine compound in the multiple formula data of the photoresist cleaning solution formula The weights are consistent, and there are differences in the weights of the nonionic surfactants. Next, each of the multiple recipe data is passed through a trained context encoder including an embedding layer to generate multiple recipe ingredient feature vectors, and the multiple recipe ingredient feature vectors are cascaded respectively. A plurality of first feature vectors corresponding to the plurality of recipe data are obtained. Then, the plurality of first feature vectors corresponding to the plurality of recipe data are two-dimensionally arranged into a feature matrix and then passed through the trained first convolutional neural network to obtain the first feature matrix. Next, obtain multiple recipe data of the photoresist cleaning liquid recipe, a cleaning test video of the object to be cleaned, and multiple recipe data of the photoresist cleaning liquid recipe to achieve a predetermined cleaning effect. . Then, the cleaning test video is passed through the video encoder of the trained joint encoder to obtain the first feature vector. The video encoder uses the trained three-dimensional convolutional neural network to encode the cleaning test video. . Next, the plurality of time values are constructed as input vectors and then passed through a temporal encoder including a one-dimensional convolutional layer and a fully connected layer of the trained joint encoder to obtain a second feature vector. The joint encoder is then used to fuse the first feature vector and the second feature vector to generate a second feature matrix. Next, the first feature matrix and the second feature matrix are fused to generate a decoding feature matrix. The decoded feature matrix is then passed through a decoder to generate decoded values, which are local optimal weight values of the nonionic surfactant. Finally, a batching recipe is generated based on the decoded values.

In summary, based on the embodiment of the present application, the automatic batching system 200 for the production of photoresist cleaning fluid is clarified, which uses neural network technology based on deep learning to intelligently determine the local optimal quality of the formula ingredients one by one, through In this way, the optimal ratio of the photoresist cleaning fluid for a specific object is determined, and then automatic batching is performed based on the optimal ratio.

As mentioned above, the automatic batching system 200 for the production of photoresist cleaning fluid according to the embodiment of the present application can be implemented in various terminal devices, such as a server of the automatic batching algorithm for the production of photoresist cleaning fluid, etc. In one example, the automatic dispensing system 200 for photoresist cleaning liquid production according to the embodiment of the present application can be integrated into the terminal device as a software module and/or hardware module. For example, the automatic dispensing system 200 for photoresist cleaning liquid production may be a software module in the operating system of the terminal device, or may be an application program developed for the terminal device; of course, the The automatic dispensing system 200 for photoresist cleaning liquid production can also be one of the many hardware modules of the terminal equipment.

Alternatively, in another example, the automatic dispensing system 200 for photoresist cleaning liquid production and the terminal equipment may also be separate devices, and the automatic dispensing system 200 for photoresist cleaning liquid production may be Connect to the terminal device through a wired and/or wireless network, and transmit interactive information according to the agreed data format.

Example methods

3A illustrates a flow chart of the training phase in the batching method of the automatic batching system for photoresist cleaning liquid production according to an embodiment of the present application. As shown in Figure 3A, the batching method of the automatic batching system for photoresist cleaning liquid production according to the embodiment of the present application includes: a training phase, including step: S110, obtaining multiple formulas of photoresist cleaning liquid formulas Data, wherein the formula of the photoresist cleaning liquid includes water, quaternary ammonium hydroxide salt compounds, alcohol amine compounds and non-ionic surfactants, and the formula of the photoresist cleaning liquid includes multiple The weights of the water, the quaternary ammonium hydroxide salt compound, and the alcoholamine compound in the formula data are consistent, and there is a difference in the weight of the nonionic surfactant; S120, add the polyol Each recipe data in the recipe data is passed through a context encoder including an embedding layer to respectively generate multiple recipe ingredient feature vectors, and the multiple recipe ingredient feature vectors are respectively concatenated to obtain corresponding to the multiple recipe data. A plurality of first feature vectors; S130, two-dimensionally arrange the plurality of first feature vectors corresponding to the plurality of recipe data into a feature matrix and then pass the first convolutional neural network to obtain the first feature matrix; S140. Acquire multiple recipe data of the photoresist cleaning liquid recipe, a cleaning test video of the object to be cleaned and multiple time values of the multiple recipe data of the photoresist cleaning liquid recipe to achieve the predetermined cleaning effect. ; S150, pass the cleaning test video through the video encoder of the joint encoder to obtain the first feature vector, and the video encoder uses a three-dimensional convolutional neural network to encode the cleaning test video; S160, pass the multiple After constructing a time value as an input vector, the joint encoder is passed through a temporal encoder including a one-dimensional convolutional layer and a fully connected layer to obtain a second feature vector; S170, use the joint encoder to fuse the first feature vector. feature vector and the second feature vector to generate a second feature matrix; S180, fuse the first feature matrix and the second feature matrix to generate a decoding feature matrix; S190, calculate the first feature matrix and the The loss function value for the dimensional distribution similarity constraint of the feature manifold between the second feature matrices, wherein the loss function value for the dimensional distribution similarity constraint of the feature manifold is based on the first feature matrix and The cosine distance between the second feature matrices and the Euclidean distance between the first feature matrix and the second feature matrix are generated; S200, pass the decoding feature matrix through a decoder to generate a decoding loss function value; S201, train the context encoder including the embedding layer and the first convolutional neural network with the weighted sum of the loss function value for the dimensional distribution similarity constraint of the feature manifold and the decoding loss function value. and the joint encoder.

FIG. 3B illustrates a flow chart of the inference stage in the batching method of the automatic batching system for photoresist cleaning liquid production according to an embodiment of the present application. As shown in Figure 3B, the batching method of the automatic batching system for photoresist cleaning liquid production according to the embodiment of the present application includes: an inference stage, including step: S210, obtaining multiple formulas of the photoresist cleaning liquid formula. Data, wherein the formula of the photoresist cleaning liquid includes water, quaternary ammonium hydroxide salt compounds, alcohol amine compounds and non-ionic surfactants, and the formula of the photoresist cleaning liquid includes multiple The weights of the water, the hydroxyl quaternary ammonium salt compound, and the alcoholamine compound in the formula data are consistent, and there is a difference in the weight of the nonionic surfactant; S220, add the poly(hydroxide) quaternary ammonium salt compound and the alcoholamine compound to the same weight. Each recipe data in the recipe data is passed through the trained context encoder including the embedding layer to respectively generate multiple recipe ingredient feature vectors, and the multiple recipe ingredient feature vectors are respectively cascaded to obtain the corresponding feature vector. Multiple first feature vectors of multiple recipe data; S230, two-dimensionally arrange the multiple first feature vectors corresponding to the multiple recipe data into a feature matrix and then use the trained first convolutional neural Network to obtain the first feature matrix; S240, obtain multiple formula data of the formula of the photoresist cleaning fluid for the cleaning test video of the object to be cleaned and multiple formula data of the formula of the photoresist cleaning fluid to reach Predetermine multiple time values of the clear effect; S250, pass the cleaning test video through the video encoder of the trained joint encoder to obtain the first feature vector, the video encoder uses the trained three-dimensional convolutional neural network The network encodes the cleaning test video; S260, after constructing the multiple time values as input vectors, the trained joint encoder includes a temporal encoder including a one-dimensional convolution layer and a fully connected layer to Obtain a second feature vector; S270, use the joint encoder to fuse the first feature vector and the second feature vector to generate a second feature matrix; S280, fuse the first feature matrix and the second feature vector. Feature matrix to generate a decoding feature matrix; S290, pass the decoding feature matrix through a decoder to generate a decoding value, the decoding value is the local optimal weight value of the nonionic surfactant; and, S300, based on the The above decoded value generates a batching plan.

Figure 4A illustrates an architectural schematic diagram of the training phase in the batching method of the automatic batching system for photoresist cleaning liquid production according to an embodiment of the present application. As shown in Figure 4A, in the training phase, in the network architecture, first, each of the multiple recipe data obtained (for example, P1 as shown in Figure 4A) is passed through the embedding layer. A context encoder (e.g., E1 as illustrated in FIG. 4A ) to respectively generate a plurality of recipe ingredient feature vectors (eg, as VF1 as illustrated in FIG. 4A ), and respectively stage the multiple recipe ingredient feature vectors. concatenate to obtain a plurality of first feature vectors corresponding to the plurality of recipe data (for example, VF2 as illustrated in Figure 4A); then, the plurality of first feature vectors corresponding to the plurality of recipe data are The vectors are two-dimensionally arranged into a feature matrix (for example, MF1 as shown in Figure 4A) and then passed through the first convolutional neural network (for example, CNN1 as shown in Figure 4A) to obtain the first feature matrix (for example, as shown in Figure 4A MF2 shown in Figure 4A); then, the obtained cleaning test video (for example, P2 shown in Figure 4A) is passed through the video encoder of the joint encoder (for example, E2 shown in Figure 4A ) to obtain a first feature vector (for example, VF3 as illustrated in Figure 4A); then, construct the plurality of time values (for example, P3 as illustrated in Figure 4A) as an input vector (for example, as shown in Figure V) shown in Figure 4A) is then passed through a temporal encoder including a one-dimensional convolutional layer and a fully connected layer of the joint encoder (for example, E3 shown in Figure 4A) to obtain a second feature vector (for example, VF4 as illustrated in Figure 4A); then, using the joint encoder (eg, CE1 as illustrated in Figure 4A) to fuse the first feature vector and the second feature vector to generate a second feature matrix (for example, MF as illustrated in Figure 4A); then, fuse the first feature matrix and the second feature matrix to generate a decoding feature matrix (for example, M as illustrated in Figure 4A); then, Calculate the loss function value between the first feature matrix and the second feature matrix for the dimensional distribution similarity constraint of the feature manifold (for example, CLV as illustrated in Figure 4A); then, the The decoded feature matrix is passed through the decoder (for example, D as shown in Figure 4A) to generate a decoding loss function value (for example, as DLV as shown in Figure 4A); finally, the dimension distribution for the feature manifold is The context encoder including the embedding layer, the first convolutional neural network and the joint encoder are trained by using a weighted sum of the similarity constrained loss function values and the decoding loss function value.

4B illustrates an architectural schematic diagram of the inference stage in the batching method of the automatic batching system for photoresist cleaning liquid production according to an embodiment of the present application. As shown in Figure 4B, in the inference phase, in the network architecture, first, each of the multiple recipe data obtained (for example, P1 as shown in Figure 4B) is passed through the trained A context encoder including an embedding layer (e.g., C1 as illustrated in Figure 4B) to respectively generate a plurality of recipe ingredient feature vectors (e.g., VF1 as illustrated in Figure 4B), and respectively The feature vectors are concatenated to obtain a plurality of first feature vectors corresponding to the plurality of recipe data (for example, VF2 as shown in Figure 4B); then, the multiple first feature vectors corresponding to the plurality of recipe data are The first feature vectors are two-dimensionally arranged into a feature matrix (for example, MF1 as shown in Figure 4B) and then passed through the trained first convolutional neural network (for example, CN1 as shown in Figure 4B) to obtain The first feature matrix (for example, MF2 as shown in Figure 4B); then, the obtained cleaning test video (for example, P2 as shown in Figure 4B) is passed through the video encoding of the trained joint encoder processor (for example, C2 as shown in Figure 4B) to obtain a first feature vector (for example, VF3 as shown in Figure 4B); then, the multiple time values obtained (for example, as shown in Figure 4B The illustrated P3) is constructed as a temporal encoder including a one-dimensional convolutional layer and a fully connected layer (eg, C3 as illustrated in Figure 4B) to obtain a second feature vector (eg, VF4 as illustrated in Figure 4B); then, use the joint encoder (eg, CE2 as illustrated in Figure 4A) to fuse The first feature vector and the second feature vector are fused to generate a second feature matrix (for example, MF as illustrated in Figure 4B); then, the first feature matrix and the second feature matrix are fused to generate Decode a feature matrix (eg, M as illustrated in Figure 4B); then, pass the decoded feature matrix through a decoder (eg, as D as illustrated in Figure 4B) to generate a decoded value, the decoded value being the local optimal weight value of the nonionic surfactant; and, finally, generate a batching plan based on the decoded value.

In summary, the batching method of the automatic batching system for photoresist cleaning liquid production based on the embodiment of the present application has been clarified, which uses neural network technology based on deep learning to intelligently determine the local optimal quality of the formula ingredients one by one. In this way, the optimal ratio of the photoresist cleaning fluid for a specific object is determined, and then automatic batching is performed based on this optimal ratio.

The basic principles of the present application have been described above in conjunction with specific embodiments. However, it should be pointed out that the advantages, advantages, effects, etc. mentioned in this application are only examples and not limitations. These advantages, advantages, effects, etc. cannot be considered to be Each embodiment of this application must have. In addition, the specific details disclosed above are only for the purpose of illustration and to facilitate understanding, and are not limiting. The above details do not limit the application to be implemented using the above specific details.

The block diagrams of the devices, devices, equipment, and systems involved in this application are only illustrative examples and are not intended to require or imply that they must be connected, arranged, or configured in the manner shown in the block diagrams. As those skilled in the art will recognize, these devices, devices, equipment, and systems may be connected, arranged, and configured in any manner. Words such as "includes," "includes," "having," etc. are open-ended terms that mean "including, but not limited to," and may be used interchangeably therewith. As used herein, the words "or" and "and" refer to the words "and/or" and are used interchangeably therewith unless the context clearly dictates otherwise. As used herein, the word "such as" refers to the phrase "such as, but not limited to," and may be used interchangeably therewith.

It should also be pointed out that in the device, equipment and method of the present application, each component or each step can be decomposed and/or recombined. These decompositions and/or recombinations shall be considered equivalent versions of this application.

The above description of the disclosed aspects is provided to enable any person skilled in the art to make or use the present application. Various modifications to these aspects will be readily apparent to those skilled in the art, and the general principles defined herein may be applied to other aspects without departing from the scope of the application. Thus, this application is not intended to be limited to the aspects shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

The foregoing description has been presented for the purposes of illustration and description. Furthermore, this description is not intended to limit the embodiments of the present application to the form disclosed herein. Although various example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, changes, additions and sub-combinations thereof.

Claims

An automatic batching system for the production of photoresist cleaning fluid, characterized in that it includes: a formula data unit for obtaining multiple formula data of the formula of the photoresist cleaning fluid, wherein the photoresist cleaning fluid The formula includes water, quaternary ammonium hydroxide salt compounds, alcohol amine compounds and non-ionic surfactants. The water, the hydroxide and oxyhydroxide are mentioned in multiple formula data of the formula of the photoresist cleaning solution. The weights of the quaternary ammonium salt compounds and the alcoholamine compounds are consistent, and the weights of the nonionic surfactants are different; a formula data semantic encoding unit is used to combine each of the multiple formula data The recipe data is passed through the trained context encoder including the embedding layer to respectively generate a plurality of recipe ingredient feature vectors, and the multiple recipe ingredient feature vectors are respectively concatenated to obtain a feature vector corresponding to the multiple recipe data. A plurality of first feature vectors; a first convolution coding unit, configured to two-dimensionally arrange the plurality of first feature vectors corresponding to the plurality of recipe data into a feature matrix and then use the trained first volume Accumulate the neural network to obtain the first feature matrix; a cleaning test data acquisition unit, used to acquire multiple formula data of the formula of the photoresist cleaning solution, a cleaning test video of the object to be cleaned and the photoresist cleaning solution The multiple recipe data of the recipe reaches multiple time values of the predetermined clear effect; the video encoding unit is used to pass the cleaning test video through the video encoder of the trained joint encoder to obtain the first feature vector, the A video encoder uses a trained three-dimensional convolutional neural network to encode the cleaning test video; a temporal encoding unit is used to construct the multiple time values into input vectors and then pass the trained joint encoder A temporal encoder including a one-dimensional convolution layer and a fully connected layer to obtain the second feature vector; a joint coding unit for using the joint encoder to fuse the first feature vector and the second feature vector to Generate a second feature matrix; a feature matrix fusion unit, used to fuse the first feature matrix and the second feature matrix to generate a decoding feature matrix; a decoding unit, used to pass the decoding feature matrix through a decoder to generate a decoding value, the decoded value being the local optimal weight value of the nonionic surfactant; and a batching unit configured to generate a batching plan based on the decoded value.
The automatic batching system for the production of photoresist cleaning fluid according to claim 1, further comprising a training module for training the context encoder including the embedding layer, the first convolutional neural network and the joint encoder for training; wherein, the training module includes: a training formula data unit, used to obtain multiple formula data of the formula of the photoresist cleaning liquid, wherein the formula of the photoresist cleaning liquid Including water, quaternary ammonium hydroxide salt compounds, alcohol amine compounds and non-ionic surfactants, the water, the quaternary hydroxyl hydroxide are mentioned in the multiple formula data of the formula of the photoresist cleaning solution The weights of the ammonium salt compounds and the alcoholamine compounds are consistent, and the weights of the non-ionic surfactants are different; the training formula data semantic encoding unit is used to combine each formula in the multiple formula data The data is respectively passed through a context encoder including an embedding layer to respectively generate a plurality of recipe ingredient feature vectors, and the multiple recipe ingredient feature vectors are respectively concatenated to obtain a plurality of first features corresponding to the multiple recipe data. Vector; train a first convolutional coding unit for two-dimensionally arranging the plurality of first feature vectors corresponding to the plurality of recipe data into a feature matrix and then passing the first convolutional neural network to obtain the first feature Matrix; training cleaning test data acquisition unit, used to acquire multiple formula data of the formula of the photoresist cleaning fluid for the cleaning test video of the object to be cleaned and multiple formula data of the formula of the photoresist cleaning fluid A plurality of time values to achieve a predetermined cleaning effect; a training video encoding unit for passing the cleaning test video through a video encoder of a joint encoder to obtain the first feature vector, and the video encoder uses a three-dimensional convolutional neural network to The cleaning test video is encoded; a temporal encoding unit is trained to construct the multiple time values into input vectors and then pass the temporal encoder including a one-dimensional convolutional layer and a fully connected layer of the joint encoder to obtain a second feature vector; training a joint encoding unit for using the joint encoder to fuse the first feature vector and the second feature vector to generate a second feature matrix; training a feature matrix fusion unit for fusing all The first feature matrix and the second feature matrix are used to generate a decoding feature matrix; a first loss calculation unit is used to calculate the dimension for the feature manifold between the first feature matrix and the second feature matrix. A loss function value for a distribution similarity constraint, wherein the loss function value for a dimensional distribution similarity constraint of a feature manifold is based on the cosine distance between the first feature matrix and the second feature matrix and the The Euclidean distance between the first feature matrix and the second feature matrix is generated; a second loss calculation unit is used to pass the decoding feature matrix through the decoder to generate a decoding loss function value; a training unit is used to use the The context encoder including the embedding layer, the first convolutional neural network and the joint encoding are trained by a weighted sum of the loss function value for the dimensional distribution similarity constraint of the feature manifold and the decoding loss function value. device.
The automatic batching system for photoresist cleaning solution production according to claim 2, wherein the training recipe data semantic encoding unit is further configured to: use the embedding layer of the encoder model containing the context of the embedding layer Convert each recipe data in the plurality of recipe data into input vectors to obtain a sequence of input vectors; use the converter of the encoder model that includes the context of the embedded layer to perform global context-based processing on the sequence of input vectors. Semantic coding is performed to obtain the plurality of formula ingredient feature vectors; the plurality of formula ingredient feature vectors are respectively concatenated to obtain a plurality of first feature vectors corresponding to the plurality of recipe data.
The automatic batching system for the production of photoresist cleaning fluid according to claim 3, wherein the training of the first convolutional coding unit is further used to: each layer of the first convolutional neural network is In the forward pass, the input data is subjected to convolution processing, mean pooling processing along the channel dimension, and activation processing to generate the first feature matrix from the last layer of the first convolutional neural network, wherein the The input to the first layer of a convolutional neural network is the feature matrix.
The automatic batching system for the production of photoresist cleaning fluid according to claim 4, wherein the training video encoding unit is further used for: a video encoder using a convolutional neural network with a three-dimensional convolution kernel according to the following formula The cleaning test video is processed to generate the first feature vector; wherein the formula is:
The automatic batching system for photoresist cleaning fluid production according to claim 5, wherein the training time sequence encoding unit includes: a construction subunit for arranging the multiple time values into a one-dimensional input Vector; fully connected subunit, used to use the fully connected layer of the temporal encoder to perform fully connected encoding on the input vector obtained by the construction subunit using the following formula to extract features of each position in the input vector High-dimensional latent features of values, where the formula is:
where X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector,
Represents matrix multiplication; a one-dimensional convolution subunit, used to perform one-dimensional convolution encoding on the input vector obtained by the construction subunit using the following formula using the one-dimensional convolution layer of the temporal encoder to extract the The high-dimensional implicit correlation features between the eigenvalues of each position in the input vector, where the formula is:
Among them, a is the width of the convolution kernel in the x direction, F is the convolution kernel parameter vector, G is the local vector matrix that operates with the convolution kernel function, and w is the size of the convolution kernel.
The automatic batching system for photoresist cleaning liquid production according to claim 6, wherein the first loss calculation unit is further used to: calculate the first characteristic matrix and the second characteristic according to the following formula The loss function value between matrices for the dimensional distribution similarity constraint of the feature manifold; wherein, the formula is:
Where cos(M 1 , M 2 ) represents the cosine distance between the first feature matrix M 1 and the second feature matrix M 2 , and d(M 1 , M 2 ) represents the Euclidean distance therebetween.
The automatic batching system for photoresist cleaning liquid production according to claim 7, wherein the second loss calculation unit is further configured to: use a decoder to perform decoding regression on the decoding feature matrix using the following formula to Obtain the decoded value; wherein, the formula is:
where X is the decoding feature matrix, Y is the decoding value, and W is the weight matrix,
Represents matrix multiplication; calculate the cross entropy value between the decoded value and the real value as the decoding loss function value.
A batching method of an automatic batching system for the production of photoresist cleaning liquid, which is characterized in that it includes: obtaining multiple formula data of the formula of the photoresist cleaning liquid, wherein the formula of the photoresist cleaning liquid includes Water, quaternary ammonium hydroxide salt compounds, alcoholamine compounds and non-ionic surfactants, as described in the multiple formula data of the photoresist cleaning liquid formula, the water, the quaternary ammonium hydroxide The weights of the base salt compound and the alcoholamine compound are consistent, and the weight of the nonionic surfactant is different; each of the multiple formula data is passed through the trained embedding layer. The context encoder is used to respectively generate a plurality of recipe ingredient feature vectors, and respectively concatenate the multiple recipe ingredient feature vectors to obtain a plurality of first feature vectors corresponding to the multiple recipe data; the corresponding After the plurality of first feature vectors of the plurality of recipe data are two-dimensionally arranged into a feature matrix, the first feature matrix is obtained through the trained first convolutional neural network; and the recipe of the photoresist cleaning liquid is obtained The plurality of formula data for the cleaning test video of the object to be cleaned and the multiple formula data of the formula of the photoresist cleaning liquid reach multiple time values of the predetermined clear effect; the cleaning test video is passed through the trained The video encoder of the joint encoder is used to obtain the first feature vector. The video encoder uses a trained three-dimensional convolutional neural network to encode the cleaning test video; after constructing the multiple time values as input vectors The second feature vector is obtained through the temporal encoder of the trained joint encoder including a one-dimensional convolutional layer and a fully connected layer; the joint encoder is used to fuse the first feature vector and the third feature vector. two feature vectors to generate a second feature matrix; fuse the first feature matrix and the second feature matrix to generate a decoding feature matrix; pass the decoding feature matrix through a decoder to generate a decoded value, the decoded value is The local optimal weight value of the nonionic surfactant; and generating a batching plan based on the decoded value.
The batching method of an automatic batching system for photoresist cleaning liquid production according to claim 9, wherein each of the plurality of recipe data is passed through a trained context encoder including an embedding layer. Generating a plurality of formula ingredient feature vectors respectively, and concatenating the plurality of formula ingredient feature vectors respectively to obtain a plurality of first feature vectors corresponding to the plurality of formula data, including: using the embedding layer-containing embedding layer The embedding layer of the context encoder model converts each recipe data in the plurality of recipe data into an input vector to obtain a sequence of input vectors; the converter of the context encoder model including the embedding layer is used to convert the input The sequence of vectors is encoded based on global context semantics to obtain the multiple recipe ingredient feature vectors; the multiple recipe ingredient feature vectors are respectively concatenated to obtain multiple first features corresponding to the multiple recipe data. vector.