A method for replacing developer in a coating developer based on neural network

Through neural network analysis of temperature and three-dimensional model data during the development process, the effectiveness changes of the developer are judged, which solves the problem of low developer replacement efficiency in traditional development processes, and realizes more accurate developer management and more efficient development process.

CN118519320BActive Publication Date: 2024-11-19JIANGSU LEIBO MICROELECTRONICS EQUIP CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410993724.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-07-24
Publication Date
2024-11-19
Estimated Expiration
2044-07-24

AI Technical Summary

Technical Problem

The effectiveness changes of the developer in the traditional development process are not fully utilized, resulting in low efficiency of the developer replacement and long development time, and the developer is over-consuming when the effectiveness is good.

Method used

Using the neural network-based uniform developer developer solution replacement method, the temperature data and three-dimensional model during the development process are collected, voxelization processing and depth gradient analysis are carried out, and the development depth and effectiveness consumption estimates are calculated, combined with the temperature characteristic frequency changes. , generate a characteristic enhancement temperature signal to determine whether the developer needs to be replaced.

Benefits of technology

The efficiency of the developer is improved, ensuring that the developer is replaced when the effectiveness reaches the threshold, avoid unnecessary consumption, and optimize the development process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118519320B_ABST
    Figure CN118519320B_ABST
Patent Text Reader

Abstract

The present application relates to the technical field of developer detection for a developer, and in particular to a developer replacement method for a photosensitive developer based on a neural network, the method comprising: for each development, collecting a development duration, a temperature detection vector and a detection three-dimensional model; obtaining a first detection three-dimensional voxelized model and a standard voxelized model; calculating an average development depth of each development according to the three-dimensional coordinates of voxel blocks in the first detection three-dimensional voxelized model and the standard voxelized model, and obtaining a development consumption estimate in combination with the average development depths of adjacent developments; obtaining a characteristic enhanced temperature signal of each development in combination with the development consumption estimate and the temperature detection vector; obtaining a developer replacement result of the current development according to the characteristic enhanced temperature signals, the average development depth and the development consumption estimate of all developments, and improving the developer replacement efficiency.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of developer detection for a developer, and in particular to a developer replacement method for a photoresist developer based on a neural network. Background Art

[0002] Etching, the core process in the chip manufacturing process, is to gradually remove the material on the chip surface by chemical or physical corrosion, thereby forming the required fine structures and circuits. The pattern on the mask is transferred to the photoresist layer on the wafer through light passing through the mask, and the wafer is further corroded by the developer, thereby leaking the material layer under the wafer, making the pattern of the material layer the same as the mask pattern, and completing the etching operation on the chip. The developer is the solution that corrodes the wafer pattern during the etching process, which directly affects the final structure and performance of the chip. Therefore, the effectiveness of the developer needs to be strictly controlled to ensure the success of chip manufacturing.

[0003] During the development process, the concentration of chemical substances in the developer is constantly changing, and the corrosion effect on the wafer is also constantly changing. In traditional development processes, a fixed development time is usually used during the development process, without taking into account the changes in the developer's effectiveness. This results in a longer development time when the developer's effectiveness is better, which in turn leads to excessive consumption of the developer when the effectiveness is better, resulting in the developer's effectiveness being consumed too quickly, so that the developer's effectiveness cannot be fully utilized, and the developer replacement efficiency is low. Summary of the invention

[0004] In order to solve the above technical problems, the present application provides a method for replacing the developer of a coating developer based on a neural network to solve the existing problems.

[0005] The method for replacing developer of a coating developer based on a neural network in the present application adopts the following technical solution:

[0006] One embodiment of the present application provides a method for replacing developer in a coating developer based on a neural network, the method comprising the following steps:

[0007] Step S10, for each development, collecting the development duration and the temperature of the wafer surface at each moment during the development process to form a temperature detection vector; collecting a wafer surface image of the wafer surface after development to obtain a detection three-dimensional model; obtaining a standard wafer surface three-dimensional model, and voxelizing the detection three-dimensional model and the standard wafer surface three-dimensional model respectively to obtain a first detection three-dimensional voxelized model and a standard voxelized model;

[0008] Step S20, obtaining the depth gradient direction according to the three-dimensional coordinates of the voxel block in the first detection three-dimensional voxelized model; obtaining the residual liquid removal interference height according to the depth gradient direction; calculating the average development depth of each development according to the standard voxelized model and the residual liquid removal interference height, and obtaining the development consumption estimation value in combination with the average development depth of adjacent developments;

[0009] Step S30, combining the estimated development consumption value and the temperature detection vector to obtain a characteristic enhanced temperature signal of each development;

[0010] Step S31, performing signal decomposition on the temperature detection vector of each development to obtain each temperature decomposition signal, and calculating the development temperature characteristic frequency according to the temperature decomposition signal; obtaining the temperature characteristic frequency change tendency according to the development temperature characteristic frequency corresponding to all temperature detection vectors; obtaining the effectiveness consumption change tendency according to the development consumption estimation value, and dividing the frequency change level and effectiveness consumption change level of the temperature decomposition signal in combination with the temperature characteristic frequency change tendency;

[0011] Step S32, obtaining a probability matrix of each signal decomposition according to the frequency change level and the efficacy consumption change level of the temperature decomposition signal; obtaining a characteristic enhancement temperature signal of each development according to the probability matrix;

[0012] Step S40, obtaining a developer replacement result of the current development according to the characteristic enhanced temperature signal of all developments, the average development depth and the estimated development consumption value.

[0013] Further, the obtaining of the depth gradient direction according to the three-dimensional coordinates of the voxel block in the first detected three-dimensional voxelized model includes:

[0014] The z-axis coordinate of the voxel block corresponding to each projection voxel block is marked as the height of each projection voxel block;

[0015] Any projected voxel block is recorded as a voxel block to be analyzed, and a neighborhood window of a preset size is constructed with the voxel block to be analyzed as the center; within the neighborhood window, the projected voxel blocks in each preset direction centered on the voxel block to be analyzed are arranged from small to large according to the sum of the x-axis coordinate and the y-axis coordinate, so as to obtain a sequence of projected voxel blocks in each preset direction; a sequence composed of the heights of all projected voxel blocks in the sequence of projected voxel blocks in each preset direction is recorded as a depth feature vector of the voxel block to be analyzed in each preset direction;

[0016] For the projected voxel block sequences in each preset direction, obtain the Euclidean distance between each projected voxel block in the projected voxel block sequence and the voxel block to be analyzed; obtain the maximum value of the Euclidean distances of all projected voxel blocks in the projected voxel block sequence; calculate the difference between the maximum value and the Euclidean distance; obtain the sum of the difference values ​​of all projected voxel blocks in the projected voxel block sequence; calculate the ratio of the difference value to the sum value; obtain the first-order derivative of the corresponding element of each projected voxel block in the depth feature vector; calculate the product of the first-order derivative and the ratio; and use the sum of the products of all projected voxel blocks in the projected voxel block sequence as the depth gradient direction of the voxel block to be analyzed in each preset direction.

[0017] Further, the step of obtaining the residual liquid interference height according to the depth gradient direction includes:

[0018] Calculating the absolute value of the depth gradient direction; taking the sum of the absolute values ​​of the voxel block to be analyzed in all preset directions as the non-wafer surface weight of the voxel block to be analyzed;

[0019] Obtain the maximum value of the non-wafer surface weight of all projected voxel blocks; calculate the difference between the maximum value and the non-wafer surface weight of the voxel block to be analyzed, recorded as the first difference; obtain the sum of the first difference values ​​of all projected voxel blocks, recorded as the first sum; calculate the ratio of the first difference value to the first sum, recorded as the first ratio; obtain the product of the height of the voxel block to be analyzed and the first ratio, recorded as the first product; use the sum of the first products of all projected voxel blocks as the residual liquid interference removal height.

[0020] Further, the method of calculating the average development depth of each development according to the standard voxelization model and the interference height of the residual liquid removal, and obtaining the development consumption estimation value in combination with the average development depth of adjacent developments, includes:

[0021] For each development, the mean of the z-axis coordinates of all voxel blocks in the standard voxelized model is calculated; the absolute value of the difference between the residual liquid interference height and the mean is recorded as the average development depth of each development; the ratio of the average development depth of each development to the development duration is calculated, recorded as the second ratio; the ratio of the average development depth of the previous development of each development to the development duration is calculated, recorded as the third ratio; the absolute value of the difference between the second ratio and the third ratio is used as the estimated value of the development consumption of each development.

[0022] Furthermore, the step of performing signal decomposition on the temperature detection vector of each development to obtain each temperature decomposition signal, and calculating the characteristic frequency of the development temperature according to the temperature decomposition signal, comprises:

[0023] The temperature detection vector of each development is subjected to signal decomposition to decompose into a first preset number of temperature decomposition signals; the temperature decomposition signal is evenly divided into a second preset number of window signals; each window signal is subjected to fast Fourier transform to obtain a spectrum diagram, and the center frequency of the spectrum diagram is calculated; the center frequencies corresponding to all window signals are arranged in the order in which the window signals are obtained to obtain a development temperature frequency variation vector of each temperature decomposition signal; and the mean value of all elements in the development temperature frequency variation vector is used as the development temperature characteristic frequency of each temperature decomposition signal.

[0024] Further, obtaining the temperature characteristic frequency change tendency according to the development temperature characteristic frequencies corresponding to all temperature detection vectors includes:

[0025] For any two adjacent temperature detection vectors of development, the DTW distance between the development temperature frequency variation vectors of each temperature decomposition signal of the two temperature detection vectors is calculated, which is recorded as the temperature cycle difference between each temperature decomposition signal of the two developments; the mean of all the temperature cycle differences is obtained, which is recorded as the significance of the development temperature cycle of each signal decomposition;

[0026] The mean value of the characteristic frequency of the developing temperature of each temperature decomposition signal of all temperature detection vectors is used as the characteristic frequency scale of the developing temperature of each signal decomposition;

[0027] For each temperature detection vector of development, the ratio of the development temperature characteristic frequency of each temperature decomposition signal of the temperature detection vector to the development temperature characteristic frequency scale is recorded as the temperature characteristic frequency change tendency of each temperature decomposition signal of the temperature detection vector.

[0028] Further, the method of obtaining the effect consumption change tendency according to the development consumption estimation value and dividing the frequency change level and the effect consumption change level of the temperature decomposition signal in combination with the temperature characteristic frequency change tendency includes:

[0029] Constructing a third preset number of preset discrete intervals, arranging them in order according to the upper limit of the preset discrete intervals, and obtaining the sequence number of each preset discrete interval; using the sequence number of each preset discrete interval as the frequency change level of the temperature decomposition signal corresponding to the temperature characteristic frequency change tendency contained in each preset discrete interval;

[0030] Calculate the mean of the estimated development consumption values ​​of all developments; record the ratio of the estimated development consumption value of each development to the mean as the effectiveness consumption change tendency of each development; construct a fourth preset number of preset effectiveness discrete intervals, arrange them in sequence according to the upper limit of the preset effectiveness discrete intervals, and obtain the sequence number of each preset effectiveness discrete interval; use the sequence number of each preset effectiveness discrete interval as the effectiveness consumption change level of the effectiveness consumption change tendency contained in each preset effectiveness discrete interval.

[0031] Further, the probability matrix of each signal decomposition is obtained according to the frequency change level and the effectiveness consumption change level of the temperature decomposition signal, including:

[0032] For each development, the number pairs consisting of the frequency change level and the effectiveness consumption change level of the temperature decomposition signal of each signal decomposition of the temperature detection vector are recorded as temperature-effectiveness isovariant pairs; all the temperature-effectiveness isovariant pairs of the temperature detection vector in each signal decomposition are arranged in the order of acquisition to obtain the temperature-effectiveness isovariant data set of each signal decomposition; the Markov transfer matrix algorithm is used for the temperature-effectiveness isovariant data set to obtain the probability matrix of each signal decomposition.

[0033] Further, the step of obtaining the characteristic enhanced temperature signal of each development according to the probability matrix includes:

[0034] For the probability matrix of each signal decomposition, record the sequence composed of elements in any row of the probability matrix as the sequence to be analyzed, calculate the standard deviation of all elements of the sequence to be analyzed; obtain the product of each element in the sequence to be analyzed and the standard deviation; obtain the sum of the products of all elements in the sequence to be analyzed; obtain the sum of the sums of all rows in the probability matrix, record it as a second sum; use the product of the second sum and the significance of the development temperature cycle of each signal decomposition as the prediction weight of the development effect of each signal decomposition;

[0035] For the temperature detection vector of each development, the normalized value of the development effectiveness prediction weight is multiplied by the temperature decomposition signal of each signal decomposition of the temperature detection vector to obtain the weighted temperature decomposition signal of each signal decomposition; for the weighted temperature decomposition signals of all signal decompositions, a signal reconstruction algorithm is used to obtain the characteristic enhanced temperature signal of each development.

[0036] Further, the step of obtaining a developer replacement result of the current development according to the characteristic enhancement temperature signal, the average development depth and the development consumption estimation value of all developments includes:

[0037] The average development depth after the last development is recorded as the expected development depth; the ratio of the average development depth of each development to the expected development depth is used as the development effectiveness of each development; the characteristic enhancement temperature signal and the development consumption estimation value of all developments are used as input feature values, and the development effectiveness is used as an input label to input the neural network training to obtain the trained neural network function;

[0038] The characteristic enhanced temperature signal of the current development and the estimated value of the development consumption are input into the neural network function to obtain the development effectiveness of the current development; if the development effectiveness is less than a preset threshold, the development is stopped and the developer is replaced; otherwise, the developer does not need to be replaced.

[0039] This application has at least the following beneficial effects:

[0040] The present application analyzes the correlation between the change of temperature and the effectiveness of the developer. First, the effectiveness consumption estimate is calculated according to the approximate time required for each development and the wafer corrosion condition, the developer effectiveness consumption is estimated, and then the change of the developer effectiveness is analyzed; further, the significance of the development temperature cycle is calculated by the frequency change of the decomposition signal of the temperature detection vector, and the temperature VMD decomposition signal reflecting the periodic temperature change with the development process is reflected, so as to amplify the information of the temperature change and pave the way for further reflecting the change correlation between the temperature and the effectiveness of the developer; the correlation strength between the development process and the temperature decomposition signal is characterized by the Markov state probability transfer matrix, and finally the development effectiveness prediction weight is calculated, the temperature vector is data enhanced to obtain the feature enhanced temperature signal, and the temperature signal is enhanced according to the correlation strength between the temperature and the effectiveness of the developer, so that the temperature signal can better reflect the effectiveness consumption of the developer; finally, the LSTM neural network is trained by the feature enhanced temperature signal to obtain the effectiveness of the developer during development, and then it is determined whether the developer needs to be replaced according to the effectiveness of the developer, thereby improving the replacement efficiency of the developer. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present application or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0042] Figure 1 A flowchart of the steps of the method for replacing the developer of the coating developer machine based on the neural network provided in the present application;

[0043] Figure 2 Schematic diagram of the Fourier transform result of the window signal;

[0044] Figure 3 Schematic diagram of the specific process of acquiring feature-enhanced temperature signals. DETAILED DESCRIPTION

[0045] In order to further explain the technical means and effects adopted by the present application to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation method, structure, features and effects of the developer replacement method of the neural network-based coating developer proposed in the present application in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.

[0046] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0047] The specific scheme of the developer replacement method of the coating developer machine based on neural network provided by the present application is described in detail below with reference to the accompanying drawings.

[0048] An embodiment of the present application provides a method for replacing developer solution in a coating developer based on a neural network. Specifically, the following method for replacing developer solution in a coating developer based on a neural network is provided. Figure 1 , the method comprises the following steps:

[0049] Step S001 , collecting temperature detection vectors and detection three-dimensional models after each development, and obtaining a first detection three-dimensional voxelized model and a standard voxelized model.

[0050] The wafer after etching is subjected to depth detection. During the detection, the surface of the wafer after each development is photographed at different angles by a white light interferometer. Specifically, the wafer surface images taken at 45° and -45° directions are obtained, with the center point of the wafer as the center, and the upward direction perpendicular to the wafer placement plane as the 0° direction. The stereoscopic vision 3D reconstruction algorithm is used for all wafer surface images to obtain a detection 3D model of the wafer surface after etching. The specific implementation process of the stereoscopic vision 3D reconstruction algorithm is a well-known technology and will not be repeated here. It should be noted that when photographing the wafer surface, the horizontal placement of the wafer must be ensured. The center point of the wafer surface is used as the coordinate origin. The upward direction perpendicular to the plane where the wafer is placed is marked as the z-axis direction, the horizontal right direction in the plane where the wafer is placed is the x-axis direction, and the vertical downward direction in the plane is the y-axis direction. At this time, the chips on the wafer are neatly arranged along the x-axis and y-axis directions, which is convenient for subsequent data processing. The 3D coordinates of each 3D coordinate point on the 3D model are obtained.

[0051] First, obtain the three-dimensional model of the standard wafer surface in the database. It should be noted that the implementer can construct the database of the standard wafer surface three-dimensional model by himself, and can obtain the three-dimensional model of the standard wafer surface by artificially selecting the existing standard wafer surface for three-dimensional reconstruction to form a database of the standard wafer surface three-dimensional model. This embodiment does not limit this. The detection three-dimensional model and the standard wafer surface three-dimensional model are voxelized, and the parameter voxelization grid side length is 0.1 nanometers. The output is the voxelization result, which is recorded as the first detection three-dimensional voxelized model and the standard voxelized model respectively. The voxelization result is to evenly divide the three-dimensional space into voxel blocks with a side length of 0.1 nanometers. The voxelization process is a well-known technology in the art and will not be repeated.

[0052] The function of the constant temperature control system in the coating developer is to make the temperature fluctuate within a very small range. However, due to the change of chemical energy during the development process, the wafer will absorb or generate heat, which will cause uncontrollable temperature fluctuations during the development process, and then cause the consumption rate of the developer to fluctuate. Therefore, the temperature change of the developer is obtained through an infrared temperature sensor. The temperature change is recorded as the temperature detection vector TC, and the sampling interval is 0.1 seconds.

[0053] Step S002, obtaining the average development depth and the development consumption estimation value according to the three-dimensional coordinates of the voxel blocks in the first detection three-dimensional voxelized model and the standard voxelized model; and obtaining the characteristic enhanced temperature signal of each development by combining the development consumption estimation value and the temperature detection vector.

[0054] Project all voxel blocks in the first detection three-dimensional voxelized model to the plane where the wafer surface is located to obtain each projected voxel block; mark the z-axis coordinate of the voxel block corresponding to each projected voxel block as the height of each projected voxel block. With the pth projected voxel block as the center, construct a neighborhood window of size N×N, N=55, and obtain N number of projected voxel blocks on the straight lines in the four directions of 0°, 45°, 90° and 135° in the neighborhood window, respectively, and arrange them from small to large according to the sum of the x-axis coordinate and the y-axis coordinate to form 4 groups of projected voxel block sequences. The sequence composed of the z-axis coordinates of all projected voxel blocks in the projected voxel block sequence is recorded as the depth feature vector. Calculate the residual liquid interference height on the wafer surface:

[0055] ;

[0056] ;

[0057] ;

[0058] In the formula, td w,p is the depth gradient direction of the p-th projected voxel block in the w direction; Ms p is the non-wafer surface weight of the p-th projected voxel block, PH is the residual liquid interference height of the wafer surface, N is the length of the projected voxel block sequence, at n,w,p is the first-order derivative of the nth element in the depth feature vector in the w direction of the pth projected voxel block, δ(n,p) w is the Euclidean distance between the p-th projected voxel block and the n-th projected voxel block in the sequence of its projected voxel blocks in the w direction, maxδ(p) w is the maximum value of the Euclidean distance between the p-th projected voxel block and all projected voxel blocks in its w-direction projected voxel block sequence, 4 is the number of directions, P is the number of projected voxel blocks, z pis the height of the p-th projected voxel block, and maxMs is the maximum value of all non-wafer surface weights.

[0059] In the formula, the first-order derivatives of the elements in the depth feature vector of the projection voxel block are weighted and summed with the inverse normalized Euclidean distance as the weight. The obtained depth gradient direction represents the depth gradient information of the p-th projection voxel block in the w direction. The positive or negative value represents whether the height at the p-th projection voxel block is rising or falling along the w direction, and the absolute value represents the slope of the rise or fall. The larger the absolute value of the depth gradient direction, the less likely it is that the p-th projection voxel block is the projection voxel block obtained from the three-dimensional model of the wafer surface. Therefore, the non-wafer surface weight is obtained, which represents the possibility that the projection voxel block is the projection voxel block on the wafer surface. The non-wafer surface weight is further inversely normalized and multiplied by the height of the projection voxel block, and the value obtained can represent the overall height of the projection voxel block on the current wafer surface.

[0060] The absolute value of the difference between the height PH of the wafer surface disturbed by the residual liquid and the mean z-axis coordinate of all voxel blocks in the standard voxelized model is calculated and recorded as the average development depth μDH.

[0061] Calculate developer consumption estimates:

[0062] ;

[0063] Where KT is the estimated value of developer consumption, D is the number of developer acquisitions, which is 5000 in this embodiment, and μDH m , μDH m-1 are the average depths of the mth and m-1th developments, Dr m 、Dr m-1 are the development durations of the m-th and m-1-th developments, respectively.

[0064] In the formula, the ratio of the average development depth to the development duration represents the intensity of the corrosion effect of the mth development on the wafer. The larger the value, the stronger the corrosion effect of the mth development. The absolute value of the difference in the development corrosion effect per unit time between adjacent development times is further calculated to represent the difference in the size of the developer efficiency consumption caused by the mth development and the previous developer efficiency consumption. The larger the value, the greater the developer efficiency consumption of the development operation.

[0065] For development operations, every time the wafer development model is changed, data must be collected again to calculate the estimated development consumption value, because it represents the estimated amount of developer effectiveness consumption of a model, and the developer effectiveness consumption of different wafer development models is different.

[0066] Further, the temperature detection vector Tc of the mth development is mThe VMD decomposition algorithm is used for calculation, the parameter VMD decomposition times is recorded as G=4, and the output is 4 temperature decomposition signals. The VMD decomposition algorithm is a well-known technology in the art and will not be described in detail.

[0067] Each temperature decomposition signal is further divided evenly into S window signals of the same length, S=10, such as Figure 2 As shown, the sth window signal is used as the original data, and the fast Fourier transform algorithm is used for calculation, and the output is a fast Fourier transform image with the horizontal axis as the frequency and the vertical axis as the energy intensity, that is, a spectrum diagram, and the center frequency is calculated for the spectrum diagram. The S center frequencies are arranged in the order in which the window signals are obtained to form the development temperature frequency variation vector of each temperature decomposition signal; the fast Fourier transform algorithm and the center frequency calculation are well-known technologies in the art and will not be described in detail.

[0068] In the coating developer, the temperature fluctuates within a very small range due to the production environment control, and the change of the ambient temperature has a negligible effect on the temperature of the developer. Therefore, the temperature of the developer is mainly caused by the chemical energy generated and absorbed during the corrosion process. Therefore, the temperature change will reflect the corrosion ability of the developer, and the effectiveness of the developer can be estimated in real time. However, since this temperature change is too subtle, it is necessary to enhance the temperature data to obtain a characteristic enhanced temperature signal. The specific process is as follows: Figure 3 shown.

[0069] Due to the complete process of developing a wafer, the intensity of chemical energy generated and absorbed is different at different stages of a development, which leads to a relatively fixed shape of the temperature change curve during a development, and the increase and decrease rules of the elements in the development temperature frequency variation vector are relatively similar. In addition, as the number of developments progresses, the development efficiency gradually decreases, which causes the frequency in the development temperature frequency variation vector to gradually decrease as a whole.

[0070] Therefore, the development temperature frequency variation vector Fv of the g-th temperature decomposition signal developed at the m-th and m+1-th times is calculated: g,m and Fv g,m+1 The DTW algorithm is used as input, and the output is the DTW distance of two vectors, which is recorded as the temperature cycle difference between the g-th temperature decomposition signal of the m-th development and the m+1-th development. The DTW algorithm is a well-known technology and will not be described in detail. After D developments, there are a total of D-1 temperature cycle differences. The average of all temperature cycle differences is obtained to obtain the development temperature cycle significance Tf of the g-th VMD decomposition. g The larger the value, the more fixed the shape of the temperature change curve of the temperature frequency variation vector of the g-th VMD decomposition is, and the more information about the development progress is contained in the corresponding g-th VMD decomposition signal.

[0071] Further, the mean value of all elements in the development temperature frequency variation vector of the g-th temperature decomposition signal of the temperature detection vector of the m-th development is recorded as the development temperature characteristic frequency fm of the g-th temperature decomposition signal of the temperature detection vector of the m-th development g,m The characteristic frequency fm of the developing temperature of the g-th temperature decomposition signal of the M-th developing operation is g,m The average value is recorded as the characteristic frequency scale FM of the development temperature of the g-th VMD decomposition g , and the characteristic frequency fm of the developing temperature g,m The characteristic frequency scale FM of the development temperature g The ratio of the temperature detection vector of the mth development is recorded as the temperature characteristic frequency change tendency Bf of the gth temperature decomposition signal g,m , discretize the temperature characteristic frequency change tendency, set the discrete intervals to (-∞,-2], (-2,-0.8], (-0.8,0.8), [0.8,2), [0.8,+∞), marked as frequency change levels 1, 2, 3, 4, 5 respectively. When Bf g,m When it belongs to the kth interval, the frequency change level of the gth VMD decomposition of the mth development is recorded as k. When k=1, it means that the characteristic frequency drops sharply, when k=2, the characteristic frequency drops slightly, when k=3, the characteristic frequency remains stable, when k=4, the characteristic frequency increases slightly, and when k=5, the characteristic frequency increases sharply.

[0072] Similarly, the estimated values ​​of development consumption for D development operations are averaged, and the estimated value of development consumption for each development is divided by the mean and recorded as the tendency of effectiveness consumption change for each development. Discrete processing is also performed, and discrete intervals are set to (-∞, 0.2), [0.8, 2), and [0.8, +∞), which are marked as effectiveness consumption change levels 1, 2, and 3, respectively. When the effectiveness consumption change tendency of the m-th development belongs to the l-th interval, the effectiveness consumption change level of the m-th development is l. When l=1, it means that the development effectiveness remains stable. When l=2, it means that the development effectiveness decreases gently. When l=3, it means that the development effectiveness decreases rapidly.

[0073] Then, for the mth development operation, a temperature effect isovariate pair (k g,m ,l m ), where k g,m is the frequency change level of the g-th temperature decomposition signal of the temperature detection vector of the m-th development, l m is the level of change in the effectiveness consumption of the mth development; according to the combination number pairs, there are I=15 number pairs, and the gth signal decomposition can obtain D temperature effectiveness isovariant pairs, and form a temperature effectiveness isovariant data set according to their acquisition order, where the mth data is the temperature effectiveness isovariant pair (k g,m ,l m). With this data set as input, the Markov transfer matrix algorithm is used, and the output is a 15*15 probability matrix, which is recorded as the probability matrix Q of the g-th VMD temperature decomposition signal. g The matrix Q g The element in the i-th row and j-th column represents the probability of the temperature-effectiveness isovariant state transforming from the i-th temperature-effectiveness isovariant pair state to the j-th temperature-effectiveness isovariant pair state. The Markov transfer matrix algorithm is a well-known technology and will not be described in detail.

[0074] The numerical distribution difference in the upper right matrix of the probability matrix is ​​small, while the numerical distribution difference in the lower right matrix is ​​large. When the numerical distribution difference in the matrix is ​​large, it means that the conversion relationship between the developing effect and the temperature is more certain, which means that the temperature information in the g-th VMD temperature decomposition signal is better for predicting the change of the developing solution effect, and more reflects the degree of corrosion of the wafer by the current developing solution. Calculate the developing effect prediction weight:

[0075] ;

[0076] Where Pw g is the prediction weight of the development efficacy of the g-th signal decomposition, Pf g is the significance of the development temperature cycle of the g-th signal decomposition, I is the number of pairs of temperature effects with the same variable, q i,j,g is the i-th row and j-th column element of the probability matrix of the g-th signal decomposition, σQ i,g is the standard deviation of the elements in the i-th row of the probability matrix of the g-th signal decomposition.

[0077] In the formula, the greater the significance of the development temperature cycle, the more periodic the temperature change of the g-th VMD decomposition signal is, and the temperature signal under this decomposition can better reflect the change of the developer effectiveness and the current development progress; the larger the standard deviation of the elements in the j-th row of the probability matrix, the greater the difference in the numerical distribution of the elements in the j-th row, and the clearer the joint relationship between the g-th VMD decomposition signal and the change of the developer effectiveness. The elements in the j-th column of the probability matrix represent the probability that the developer is predicted to change from other states to the j-th row state. Therefore, the elements in the i-th row and j-th column are used as the weights of the standard deviation of the elements in the j-th row of the probability matrix. Finally, the prediction weight of the development effectiveness is obtained. The larger the value, the more the temperature signal under the g-th decomposition can reflect the change of the developer effectiveness and the current development progress.

[0078] Further, the development efficacy prediction weight Pw g Normalized and compared with the g-th temperature decomposition signal Vt of the temperature detection vector of the m-th development g,mThen, all weighted temperature decomposition signals of the temperature detection vector of the mth development are taken as input, and the VMD signal reconstruction algorithm is adopted to output the characteristic enhanced temperature signal of the mth development. The VMD signal reconstruction algorithm is a well-known technology in the art and will not be described in detail.

[0079] The characteristic enhanced temperature signal enhances the temperature change information related to the change of the developer effectiveness through data processing, making the effective temperature information more prominent. Compared with the temperature detection vector, the characteristic enhanced temperature signal is more helpful in characterizing the degree of development, which is beneficial for controlling the duration of each development and reducing the frequency of developer replacement.

[0080] Step S003, obtaining a developer replacement result of the current development according to the characteristic enhanced temperature signal, the average development depth and the development consumption estimation value of all developments.

[0081] First, the average development depth after D developments is obtained, which is recorded as the expected development depth ED. Finally, for the mth development, the input feature value is the feature enhancement temperature signal and the estimated development consumption value, and the input label is the development effectiveness The input feature values ​​and input labels of D developments are combined into a data set.

[0082] The data set is used as input, the LSTM neural network algorithm is used, the activation function is the RLU function, the output layer is Y to represent the current development completion progress, the optimizer is the Adma optimizer, the loss function uses the mean error function as the loss function value, and the output is the trained LSTM neural network function. The training process of the LSTM neural network is a well-known technology in the art and will not be repeated here.

[0083] During the actual development operation, the temperature data vector of the development and the estimated value of the development consumption are obtained in real time as the input of the LSTM neural network function, and the output is the current development efficiency. When the development efficiency is less than 0.5, the development is stopped and the developer is replaced.

[0084] It should be noted that the above sequence of the embodiments of the present application is for description only and does not represent the advantages and disadvantages of the embodiments. The above is a description of a specific embodiment of this specification. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0085] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0086] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Modifications to the technical solutions recorded in the aforementioned embodiments, or equivalent replacement of some of the technical features therein, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.

Claims

1. A method for replacing developer in a coating developer based on a neural network, characterized in that: The method comprises the following steps: Step S10, for each development, collecting the development duration and the temperature of the wafer surface at each moment during the development process to form a temperature detection vector; collecting a wafer surface image of the wafer surface after development to obtain a detection three-dimensional model; obtaining a standard wafer surface three-dimensional model, and voxelizing the detection three-dimensional model and the standard wafer surface three-dimensional model respectively to obtain a first detection three-dimensional voxelized model and a standard voxelized model; Step S20, obtaining the depth gradient direction according to the three-dimensional coordinates of the voxel block in the first detection three-dimensional voxelized model; obtaining the residual liquid removal interference height according to the depth gradient direction; obtaining the development duration of each development, and calculating the average development depth and the development consumption estimation value in combination with the standard voxelized model and the residual liquid removal interference height; Step S30, combining the estimated development consumption value and the temperature detection vector to obtain a characteristic enhanced temperature signal of each development; Step S31, performing signal decomposition on the temperature detection vector of each development to obtain each temperature decomposition signal, and calculating the development temperature characteristic frequency according to the temperature decomposition signal; obtaining the temperature characteristic frequency change tendency according to the development temperature characteristic frequency corresponding to all temperature detection vectors; obtaining the effectiveness consumption change tendency according to the development consumption estimation value, and dividing the frequency change level and effectiveness consumption change level of the temperature decomposition signal in combination with the temperature characteristic frequency change tendency; Step S32, obtaining a probability matrix of each signal decomposition according to the frequency change level and the efficacy consumption change level of the temperature decomposition signal; obtaining a characteristic enhancement temperature signal of each development according to the probability matrix; Step S40, obtaining a developer replacement result of the current development according to the characteristic enhanced temperature signal of all developments, the average development depth and the development consumption estimation value; The step of obtaining the depth gradient direction according to the three-dimensional coordinates of the voxel block in the first detected three-dimensional voxelized model includes: The z-axis coordinate of the voxel block corresponding to each projection voxel block is marked as the height of each projection voxel block; Any projected voxel block is recorded as a voxel block to be analyzed, and a neighborhood window of a preset size is constructed with the voxel block to be analyzed as the center; within the neighborhood window, the projected voxel blocks in each preset direction centered on the voxel block to be analyzed are arranged from small to large according to the sum of the x-axis coordinate and the y-axis coordinate, so as to obtain a sequence of projected voxel blocks in each preset direction; a sequence composed of the heights of all projected voxel blocks in the sequence of projected voxel blocks in each preset direction is recorded as a depth feature vector of the voxel block to be analyzed in each preset direction; For the projected voxel block sequences in the preset directions, obtain the Euclidean distance between each projected voxel block in the projected voxel block sequence and the voxel block to be analyzed; obtain the maximum value of the Euclidean distance of all projected voxel blocks in the projected voxel block sequence; calculate the difference between the maximum value and the Euclidean distance; obtain the sum of the difference values ​​of all projected voxel blocks in the projected voxel block sequence; calculate the ratio of the difference value to the sum value; obtain the first-order derivative of the element corresponding to each projected voxel block in the depth feature vector; calculate the product of the first-order derivative and the ratio; and use the sum of the products of all projected voxel blocks in the projected voxel block sequence as the depth gradient direction of the voxel block to be analyzed in each preset direction; The step of obtaining the interference height of the residual liquid according to the depth gradient direction includes: Calculating the absolute value of the depth gradient direction; taking the sum of the absolute values ​​of the voxel block to be analyzed in all preset directions as the non-wafer surface weight of the voxel block to be analyzed; Obtain the maximum value of the non-wafer surface weight of all projected voxel blocks; calculate the difference between the maximum value and the non-wafer surface weight of the voxel block to be analyzed, recorded as a first difference; obtain the sum of the first difference values ​​of all projected voxel blocks, recorded as a first sum; calculate the ratio of the first difference value to the first sum, recorded as a first ratio; obtain the product of the height of the voxel block to be analyzed and the first ratio, recorded as a first product; use the sum of the first products of all projected voxel blocks as the residual liquid interference removal height; The step of obtaining the development duration of each development, and calculating the average development depth and the development consumption estimation value by combining the standard voxelization model and the residual liquid interference height, comprises: For each development, the mean of the z-axis coordinates of all voxel blocks in the standard voxelized model is calculated; the absolute value of the difference between the residual liquid interference height and the mean is recorded as the average development depth of each development; the ratio of the average development depth of each development to the development duration is calculated, recorded as the second ratio; the ratio of the average development depth of the previous development of each development to the development duration is calculated, recorded as the third ratio; the absolute value of the difference between the second ratio and the third ratio is used as the estimated value of the development consumption of each development.

2. The method for replacing developer in a coating developer machine based on a neural network as claimed in claim 1, characterized in that: The step of performing signal decomposition on the temperature detection vector of each development to obtain each temperature decomposition signal, and calculating the development temperature characteristic frequency according to the temperature decomposition signal, comprises: The temperature detection vector of each development is subjected to signal decomposition to decompose into a first preset number of temperature decomposition signals; the temperature decomposition signal is evenly divided into a second preset number of window signals; each window signal is subjected to fast Fourier transform to obtain a spectrum diagram, and the center frequency of the spectrum diagram is calculated; the center frequencies corresponding to all window signals are arranged in the order in which the window signals are obtained to obtain a development temperature frequency variation vector of each temperature decomposition signal; and the mean value of all elements in the development temperature frequency variation vector is used as the development temperature characteristic frequency of each temperature decomposition signal.

3. The method for replacing developer of a coating developer machine based on a neural network as claimed in claim 2, characterized in that: The step of obtaining the temperature characteristic frequency change tendency according to the development temperature characteristic frequencies corresponding to all temperature detection vectors includes: For any two adjacent temperature detection vectors of development, the DTW distance between the development temperature frequency variation vectors of each temperature decomposition signal of the two temperature detection vectors is calculated, which is recorded as the temperature cycle difference between each temperature decomposition signal of the two developments; the mean of all the temperature cycle differences is obtained, which is recorded as the significance of the development temperature cycle of each signal decomposition; The mean value of the characteristic frequency of the developing temperature of each temperature decomposition signal of all temperature detection vectors is used as the characteristic frequency scale of the developing temperature of each signal decomposition; For each temperature detection vector of development, the ratio of the development temperature characteristic frequency of each temperature decomposition signal of the temperature detection vector to the development temperature characteristic frequency scale is recorded as the temperature characteristic frequency change tendency of each temperature decomposition signal of the temperature detection vector.

4. The method for replacing developer of a coating developer based on a neural network as claimed in claim 3, characterized in that: The method of obtaining the effect consumption change tendency according to the development consumption estimation value and dividing the frequency change level and the effect consumption change level of the temperature decomposition signal in combination with the temperature characteristic frequency change tendency includes: Constructing a third preset number of preset discrete intervals, arranging them in order according to the upper limit of the preset discrete intervals, and obtaining the sequence number of each preset discrete interval; using the sequence number of each preset discrete interval as the frequency change level of the temperature decomposition signal corresponding to the temperature characteristic frequency change tendency contained in each preset discrete interval; Calculate the mean of the estimated development consumption values ​​of all developments; record the ratio of the estimated development consumption value of each development to the mean as the effectiveness consumption change tendency of each development; construct a fourth preset number of preset effectiveness discrete intervals, arrange them in sequence according to the upper limit of the preset effectiveness discrete intervals, and obtain the sequence number of each preset effectiveness discrete interval; use the sequence number of each preset effectiveness discrete interval as the effectiveness consumption change level of the effectiveness consumption change tendency contained in each preset effectiveness discrete interval.

5. The method for replacing developer of a coating developer based on a neural network according to claim 1, characterized in that: The method of obtaining a probability matrix of each signal decomposition according to the frequency change level and the effectiveness consumption change level of the temperature decomposition signal includes: For each development, the number pairs consisting of the frequency change level and the effectiveness consumption change level of the temperature decomposition signal of each signal decomposition of the temperature detection vector are recorded as temperature-effectiveness isovariant pairs; all the temperature-effectiveness isovariant pairs of the temperature detection vector in each signal decomposition are arranged in the order of acquisition to obtain the temperature-effectiveness isovariant data set of each signal decomposition; the Markov transfer matrix algorithm is used for the temperature-effectiveness isovariant data set to obtain the probability matrix of each signal decomposition.

6. The method for replacing developer of a coating developer based on a neural network according to claim 1, characterized in that: The step of obtaining the characteristic enhanced temperature signal of each development according to the probability matrix includes: For the probability matrix of each signal decomposition, record the sequence composed of elements in any row of the probability matrix as the sequence to be analyzed, calculate the standard deviation of all elements of the sequence to be analyzed; obtain the product of each element in the sequence to be analyzed and the standard deviation; obtain the sum of the products of all elements in the sequence to be analyzed; obtain the sum of the sums of all rows in the probability matrix, record it as a second sum; use the product of the second sum and the significance of the development temperature cycle of each signal decomposition as the prediction weight of the development effect of each signal decomposition; For the temperature detection vector of each development, the normalized value of the development effectiveness prediction weight is multiplied by the temperature decomposition signal of each signal decomposition of the temperature detection vector to obtain the weighted temperature decomposition signal of each signal decomposition; for the weighted temperature decomposition signals of all signal decompositions, a signal reconstruction algorithm is used to obtain the characteristic enhanced temperature signal of each development.

7. The method for replacing developer of a coating developer based on a neural network as claimed in claim 1, characterized in that: The step of obtaining a developer replacement result of the current development according to the characteristic enhancement temperature signal, the average development depth and the development consumption estimation value of all developments includes: The average development depth after the last development is recorded as the expected development depth; the ratio of the average development depth of each development to the expected development depth is used as the development effectiveness of each development; the characteristic enhancement temperature signal and the development consumption estimation value of all developments are used as input feature values, and the development effectiveness is used as an input label to input the neural network training to obtain the trained neural network function; The characteristic enhanced temperature signal of the current development and the estimated value of the development consumption are input into the neural network function to obtain the development effectiveness of the current development; if the development effectiveness is less than a preset threshold, the development is stopped and the developer is replaced; otherwise, the developer does not need to be replaced.