A smart control system for a space-type atomic layer deposition apparatus

Abstract

This invention discloses an intelligent control system for a space-type atomic layer deposition equipment, belonging to the field of intelligent control technology. The system of this invention includes: an acquisition module that acquires the pressure difference and flow rate values ​​of the gas curtain pipeline and performs low-pass filtering to obtain a pure signal matrix; a feature extraction module that calculates the flow field impedance and impedance residual based on this matrix and converts them into an empirical probability distribution vector; a model building module that constructs a degenerate distribution uncertainty set and a scheduling objective cost function based on this vector, and converts them into a mixed integer second-order cone programming model; an offline solution module that discretizes and classifies the dynamic distribution divergence tolerance and solves the model, and pre-generates a scheduling strategy library; and an online control module that looks up tables in the strategy library according to the real-time divergence tolerance and outputs substrate transmission speed control and channel switching commands. This invention transforms the minimization-maximization probability model into a convex optimization problem for offline solution, breaking the computing power bottleneck and balancing control response speed and scheduling decision-making through efficient table lookup.

Description

Technical Field

[0001] This invention relates to the field of intelligent control technology, and in particular to an intelligent control system for a space-type atomic layer deposition apparatus. Background Technology

[0002] Spatial atomic layer deposition (SALD) equipment plays a crucial role in semiconductor microprocessing and continuous thin film manufacturing. During the continuous operation of SALD equipment, the substrate is usually continuously translated by a motor below, passing through different precursor gas regions in sequence. In order to effectively isolate different reactive gases and prevent cross-contamination, the equipment needs to rely on nitrogen to purge the gas curtain pipeline to maintain a stable and pure gas curtain isolation zone. However, as the process continues, the unavoidable parasitic deposition phenomenon will cause the gap at the bottom of the gas curtain nozzle to become smaller and smaller, which will increase the physical resistance of gas flow. The degree of degradation of this gas curtain flow field often shows a high degree of uncertainty. Cross-contamination caused by the deterioration of the flow field will not only seriously affect the film quality, but may even lead to a large number of wafer scraps. In existing technologies for monitoring and scheduling gas curtain flow fields in atomic layer deposition (ALD) equipment, the extraction of flow field degradation characteristics is hampered by several factors. The continuous movement of the substrate causes periodic changes in the free volume within the chamber, resulting in low-frequency pressure fluctuations. Simultaneously, the operating environment is filled with complex external mechanical high-frequency vibration interference. True flow field degradation is an extremely slow, gradual process. Existing systems often rely on raw, isolated physical measurement data for judgment, making it difficult to effectively remove high-frequency mechanical noise and retain the true flow field degradation signal. Furthermore, the vast transient data dimension cannot directly characterize future degradation trends. Based on physical principles, the flow field impedance caused by irreversible nozzle gap reduction should only increase unidirectionally. However, due to sensor measurement noise, negative fluctuations with decreasing impedance frequently appear in the monitoring data. Conventional control algorithms typically assume symmetrical spatial relationships when constructing predictive models, lacking a truncation and discrimination mechanism that incorporates prior physical knowledge such as unidirectional irreversibility. This easily leads to the inclusion of cases that violate physical principles, such as the gap increasing in reverse, resulting in distorted control benchmarks. Summary of the Invention

[0003] The technical problem solved by this invention is that when monitoring the air curtain flow field and scheduling production of atomic layer deposition equipment, existing systems often directly use raw, isolated physical measurement point data for judgment, which makes it difficult to effectively remove high-frequency mechanical noise and retain the real flow field degradation signal. Conventional control algorithms usually assume symmetric space when building prediction models and lack a truncation and discrimination mechanism that combines physical prior knowledge such as unidirectional irreversibility. This easily leads to the inclusion of situations that violate physical common sense, such as the gap increasing in reverse, in the calculation, resulting in the distortion of the control benchmark.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: an intelligent control system for a space-type atomic layer deposition equipment, comprising: an acquisition module, a feature extraction module, a model building module, an offline solving module, and an online control module; The acquisition module is used to acquire the differential pressure and flow rate values ​​of the gas curtain pipeline of the space atomic layer deposition equipment, and to perform low-pass filtering on the differential pressure and flow rate values ​​to obtain a clean signal matrix. The feature extraction module is used to calculate the flow field impedance based on the pure signal matrix, extract the impedance residual, and convert the impedance residual into an empirical probability distribution vector. The model building module is used to construct a degenerate distribution uncertainty set based on the empirical probability distribution vector and the statistical characteristics of the impedance residual, construct a production scheduling objective cost function containing the degenerate distribution uncertainty set, and transform the production scheduling objective cost function into a mixed integer second-order cone programming model. The offline solution module is used to discretely classify the value range of the dynamic distribution divergence tolerance, solve the mixed integer second-order cone programming model based on the discretely classified dynamic distribution divergence tolerance, and generate a production scheduling strategy library. The online control module is used to look up the table in the production scheduling strategy library according to the dynamic distribution divergence tolerance obtained in real time, and output substrate drive speed adjustment command and channel switching command.

[0005] Preferably, the specific process of the acquisition module performing low-pass filtering includes: The differential pressure value and the flow rate value are collected with a set time window length and sampling frequency, and the collected data are integrated into an initial joint signal matrix; The initial joint signal matrix is ​​filtered by calling a fourth-order Butterworth low-pass filter algorithm, the cutoff frequency is set, and the clean signal matrix is ​​output.

[0006] Preferably, the specific process by which the feature extraction module extracts the impedance residual and converts it into an empirical probability distribution vector includes: The ratio of the pressure difference value in the first column to the flow rate value in the second column of each row of the pure signal matrix is ​​used as the flow field impedance of the corresponding sampling point. The difference between the flow field impedance and the preset calibration impedance is calculated as the impedance residual. If the impedance residual is negative, the impedance residual is forced to be zero, and the residual column vector composed of the impedance residuals of all sampling points is obtained. Remove invalid data points with zero values ​​from the residual column vector, and extract the maximum and minimum residuals from the remaining valid data; The closed interval formed by the maximum residual and the minimum residual is divided into a predetermined number of consecutive equal-length sub-intervals, and the number of residuals falling into each sub-interval is counted. The ratio of the number of residuals falling into each sub-interval to the number of remaining elements in the residual column vector after removing invalid data points is calculated as the probability that the impedance residual falls into the corresponding sub-interval. The probabilities of all sub-intervals are concatenated according to the order of sub-interval division to construct the empirical probability distribution vector.

[0007] Preferably, the specific process of the model building module constructing the degenerate distribution uncertainty set includes: A constrained support domain is constructed by combining the minimum residual with a unidirectional support domain truncation operation. The mathematical expression for the constrained support domain is: ; in, The random variable representing the unknown real flow field impedance residual is... This represents the minimum residual.

[0008] Preferably, the specific process of constructing the degenerate distribution uncertainty set by the model building module further includes: Calculate the statistical variance of the valid residual data after removing invalid data points; The product of the statistical variance and the preset confidence amplification factor is used as the dynamic distribution divergence tolerance. A distance metric is constructed based on a quadratic transmission cost function, wherein the quadratic transmission cost function is the squared difference between any two residual values ​​within the constrained support domain. Centered on the empirical probability distribution vector, within the restricted support domain, the set of all legal probability distributions that satisfy the transfer deviation between the unknown distribution calculated based on the quadratic transmission cost function and the center distribution being less than or equal to the dynamic distribution divergence tolerance is taken as the degenerate distribution uncertainty set.

[0009] Preferably, the specific process of the model building module constructing the production scheduling target cost function includes: Initialize the production scheduling decision vector, which includes substrate drive speed control instructions and channel switching instructions; Construct a production scheduling target cost function, the mathematical expression of which is: ; in, Let the production scheduling objective cost function be... This is a column vector for production scheduling decisions, containing elements. and , The true impedance random variable of the future unknown flow field. This is the preset capacity delay cost coefficient. This refers to the specific value of the substrate drive speed control command. The preset cross-contamination penalty coefficient, This is a channel switching command. To minimize residuals, Set a fixed sunk cost coefficient for the preset channel switching.

[0010] Preferably, the specific process of transforming the production scheduling objective cost function into a mixed-integer second-order cone programming model includes: Under the constraint of the degenerate distribution uncertainty set, a minimization-maximization model is constructed with the objective of minimizing the expected value of the production scheduling target cost function; By introducing Lagrange multipliers and auxiliary variables, the minimization-maximization model is transformed into a dual model. Auxiliary variables are introduced, and boundary constraints are constructed by combining the preset production delay cost coefficient, the substrate transmission speed control command, the channel switching command, the center representative value of each sub-interval, the Lagrange multiplier, and the preset cross-contamination penalty coefficient, so that the boundary constraints constitute a second-order rotating cone inequality. Through the dual transformation and the boundary constraints, the minimization-maximization model is equivalently transformed into the mixed integer second-order cone programming model.

[0011] Preferably, the specific process by which the offline solution module generates the production scheduling strategy library includes: The effective engineering interval of the dynamic distribution divergence tolerance is set, and the effective engineering interval is divided into a preset number of degradation level intervals with a fixed step size; For each degradation level interval, the right boundary value of the degradation level interval is substituted as a fixed parameter into the mixed integer second-order cone programming model; The mixed-integer optimization solver is invoked to solve the mixed-integer second-order cone programming model, and the optimal production scheduling decision vector for the corresponding degradation level interval is obtained; The optimal production scheduling decision vectors obtained by solving all degradation level intervals are combined row by row to generate a production scheduling strategy library.

[0012] Preferably, the specific process of the online control module outputting substrate drive speed adjustment commands and channel switching commands includes: During equipment operation, the dynamic distribution divergence tolerance is calculated in real time at the current moment; The row index number is calculated based on the dynamic distribution divergence tolerance and the fixed step size. The row index number is the rounded-up value of the dynamic distribution divergence tolerance divided by the fixed step size.

[0013] Preferably, the specific process of the online control module outputting substrate drive speed adjustment commands and channel switching commands further includes: A safety clamping determination is performed on the row index number: if the row index number is greater than the preset number of the degradation level range, the row index number is assigned the preset number; if the row index number is less than the set lower limit value, the row index number is assigned the set lower limit value. Based on the row index number determined by the safety clamping, the row vector of the corresponding row is extracted from the production scheduling strategy library; The optimal substrate transmission speed value in the first column of the row vector is converted into a substrate transmission speed control command and sent to the programmable logic controller. At the same time, the optimal channel switching value in the second column is converted into a channel switching command and sent to the manufacturing execution system.

[0014] The beneficial effects of this invention are as follows: By independently filtering the collected differential pressure and flow rate values ​​using a low-pass filtering algorithm, this invention effectively preserves the true low-frequency flow field degradation signal, which includes the equipment's baseline operating state and the slow degradation trend. Furthermore, by implementing a unidirectional support domain truncation operation in the context of irreversible sediment deposition, this invention not only effectively identifies and eliminates invalid physical noise caused by sensor measurements but also cuts off the reverse impedance fluctuation path, ensuring that the worst-case scenario of the model search matches the real-world scenario. Finally, by introducing a quadratic transmission cost function to construct a distance metric, this invention assigns a heavy penalty weight to extremely deteriorating operating conditions. In terms of production scheduling, the production target generation constructed by this invention... The valence function comprehensively weighs the costs of capacity delays, cross-contamination penalties, and fixed sunk costs of channel switching. Using fixed sunk costs as a decision barrier, it absorbs minor disturbances by prioritizing speed reduction, thereby effectively suppressing frequent channel switching and equipment status fluctuations caused by minor impedance fluctuations. This invention, through dual transformation and the setting of boundary constraints, equivalently transforms the original minimization-maximization non-convex probabilistic model, which contains infinite dimensions and is difficult to solve directly, into a standard mixed-integer second-order cone programming model. Based on the mathematical property that convex geometry has no local minimum trap, it fundamentally ensures that the obtained production scheduling decision column vector is the globally optimal solution, eliminating the hidden danger of the algorithm getting trapped in local optima. Attached Figure Description

[0015] Figure 1 This is a basic flowchart illustrating an intelligent control system for a spatial atomic layer deposition apparatus, as provided in one embodiment of the present invention. Detailed Implementation

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0017] Example, refer to Figure 1 A smart control system for a space-type atomic layer deposition apparatus is provided, comprising: an acquisition module, a feature extraction module, a model building module, an offline solution module, and an online control module; The acquisition module is used to acquire the differential pressure and flow rate values ​​of the gas curtain pipeline of the space atomic layer deposition equipment, and to perform low-pass filtering on the differential pressure and flow rate values ​​to obtain a clean signal matrix. The feature extraction module is used to calculate the flow field impedance based on the pure signal matrix, extract the impedance residual, and transform the impedance residual into an empirical probability distribution vector. The model building module is used to construct a degenerate distribution uncertainty set based on the statistical characteristics of the empirical probability distribution vector and impedance residuals, construct a production scheduling objective cost function containing the degenerate distribution uncertainty set, and transform the production scheduling objective cost function into a mixed integer second-order cone programming model. The offline solution module is used to discretize and classify the value range of the dynamic distribution divergence tolerance, and solve the mixed integer second-order cone programming model based on the discretized dynamic distribution divergence tolerance to generate a production scheduling strategy library. The online control module is used to look up tables in the production scheduling strategy library based on the real-time dynamic distribution divergence tolerance and output substrate drive speed adjustment instructions and channel switching instructions.

[0018] This invention successfully transforms the highly uncertain physical degradation process of the air curtain flow field into a stable mathematical model by organically integrating low-pass filtering, probability distribution mapping, robust optimization, and offline solution and online table lookup strategies. It can output substrate transmission speed control commands and channel switching commands in real time and intelligently, meeting the scheduling requirements of continuous manufacturing processes.

[0019] The specific process of the acquisition module performing low-pass filtering includes: Pressure difference and flow rate values ​​are collected within a set time window and sampling frequency, and the collected data are integrated into an initial joint signal matrix. The fourth-order Butterworth low-pass filter algorithm is used to perform filtering calculations on the initial joint signal matrix, the cutoff frequency is set, and the clean signal matrix is ​​output.

[0020] In a specific embodiment of the present invention, during the operation of the spatial atomic layer deposition equipment, the original source of the data is a high-precision dynamic differential pressure sensor and a mass flow meter installed in the nitrogen purge gas curtain pipeline; During the data acquisition phase, a fixed time window length of 10 seconds and a sampling frequency of 1000Hz are set. Within each time window, the acquired data is integrated into an initial joint signal matrix. In the initial joint signal matrix, each row corresponds to a sampling point. The elements in the first column represent the pressure difference value before and after the air curtain nozzle at that moment, that is, the difference between the upstream air supply pressure of the nozzle and the back pressure inside the chamber, in kilopascals (kPa). The elements in the second column represent the flow rate value at that sampling moment, in liters per minute. After obtaining the initial joint signal matrix, interference signals need to be removed. Because the substrate of the spatial atomic layer deposition equipment is continuously translated by a motor below, the movement of the substrate causes periodic changes in the free volume inside the chamber, resulting in basic low-frequency pressure fluctuations in the gas curtain pipeline. External high-frequency mechanical vibration interference also exists during equipment operation. Since the actual flow field degradation signal is an extremely slow, gradual process, it is also contained in the low-frequency band. To retain the actual low-frequency flow field degradation signal, which includes the equipment's baseline operating state and the slow degradation trend, and to remove high-frequency mechanical vibration noise, this embodiment uses a fourth-order Butterworth low-pass filter algorithm to independently filter the first and second columns of the initial joint signal matrix. In the fourth-order Butterworth low-pass filter algorithm, the cutoff frequency is set to 10Hz. After filtering, high-frequency mechanical noise above the cutoff frequency in the initial joint signal matrix is ​​filtered out, outputting a clean signal matrix.

[0021] This invention effectively filters out interference caused by high-frequency mechanical vibrations by calling a fourth-order Butterworth low-pass filter algorithm and setting a specific time window and cutoff frequency. At the same time, it completely preserves the real flow field degradation signal contained in the low-frequency band, thus ensuring the purity and reliability of the subsequent feature extraction and degradation evaluation benchmark from the source.

[0022] The specific process by which the feature extraction module extracts impedance residuals and transforms them into empirical probability distribution vectors includes: The ratio of the pressure difference value in the first column to the flow rate value in the second column of each row in the pure signal matrix is ​​used as the flow field impedance of the corresponding sampling point. The difference between the flow field impedance and the preset calibration impedance is calculated as the impedance residual. If the impedance residual is negative, the impedance residual is forced to be zero, and the residual column vector composed of the impedance residuals of all sampling points is obtained. Remove invalid data points with zero values ​​from the residual column vector, and extract the maximum and minimum residuals from the remaining valid data; Divide the closed interval consisting of the maximum and minimum residuals into a predetermined number of consecutive equal-length subintervals, and count the number of residuals falling into each subinterval. The ratio of the number of residuals falling into each sub-interval to the number of remaining elements in the residual column vector after removing invalid data points is used as the probability that the impedance residual falls into the corresponding sub-interval. The probabilities of all subintervals are concatenated according to the order of subinterval division to construct an empirical probability distribution vector.

[0023] In one specific embodiment of the present invention, the reduced gap at the bottom of the nozzle leads to an increase in the physical resistance to gas outflow. This resistance is quantified by the ratio of pressure difference to flow rate, i.e., the air curtain flow field impedance. This embodiment further extracts the impedance fluctuation residual to represent the degree of deterioration of the air curtain flow field impedance caused by parasitic deposition. The impedance fluctuation residual is the difference between the flow field impedance and a preset calibration impedance. In this embodiment, the preset calibration impedance is set to 1.5. ; Since the impedance caused by the irreversible shrinkage of the nozzle gap is physical, it will only increase in one direction. Any negative air curtain flow field impedance fluctuation residual caused by sensor measurement noise, that is, the calculated air curtain flow field impedance is lower than the calibrated impedance, is considered as invalid physical noise. In the continuous manufacturing process of spatial atomic layer deposition equipment, the degree of degradation of the air curtain flow field is highly uncertain, and the trial and error cost caused by cross-contamination is high. This invention introduces the core architecture of the sub-Bruker optimization. Its core idea is to construct a degradation distribution uncertainty set that covers all flow field degradation probability distributions that conform to physical reality, and to find the optimal production scheduling decision under the worst case of air curtain flow field degradation. After obtaining the residual column vector, it needs to be input into the model building module containing the sub-Brubars optimization algorithm. However, the residual column vector is only a discrete record of the transient fluctuations of the flow field within a fixed historical time window. Its huge data dimension cannot meet the real-time and speed requirements of scheduling solution. The isolated physical measurement points cannot directly characterize the future degradation trend of the air curtain flow field, and cannot be directly used as the calculation benchmark of the minimization-maximization model in the sub-Brubars optimization algorithm. Therefore, this embodiment constructs a connection mechanism from physical fluctuations to probability space mapping. By introducing a statistical histogram processing method, the residual column vector is transformed into a discrete empirical probability distribution that can be used for sub-Bruker optimization. Specifically, after traversing the residual column vector, invalid data points with a value of 0 are removed. If the residual column vector is empty after removal, the current air curtain flow field is determined to be in a baseline healthy state. The full-load transmission speed command and the current channel maintenance command are directly issued, and the subsequent calculation of the current control cycle is terminated. If the residual column vector is not empty after removal, the maximum and minimum residuals in the remaining valid data are extracted. Since the previous physical truncation excludes negative values, the minimum residual extracted at this time is greater than or equal to 0. The maximum residual will be and minimum residual The closed interval formed The interval is divided into q consecutive subintervals of equal length. In this embodiment, q is 50. For each subinterval, the center value of the interval is extracted as the representative value of the subinterval. Counting those falling into the first place in turn The number of residuals within each subinterval, of which The value of is in the range of 1 to q, and it will fall into the q-th ... The ratio of the number of residuals within each sub-interval to the number of remaining elements in the residual column vector after removing invalid data points with values ​​of 0 is used as the impedance residual falling within the th The probability of each sub-interval.

[0024] This invention effectively overcomes the computational obstacles caused by the massive transient data dimension by calculating the flow field impedance residual, eliminating invalid data, and using a statistical histogram processing method to transform it into an empirical probability distribution vector. It successfully transforms isolated physical measurement points into discrete distribution benchmarks that can be directly utilized by the Bruker optimization algorithm.

[0025] The specific process of constructing the degenerate distribution uncertainty set by the model building module includes: A constrained support domain is constructed by combining minimum residual with unidirectional support domain truncation operation. The mathematical expression for the constrained support domain is: ; in, Represents the unknown real flow field impedance residual random variable, with units of , This represents the minimum residual.

[0026] In one specific embodiment of the present invention, in order to make the general split-Brow bar algorithm suitable for the physical characteristics of spatial atomic layer deposition equipment and meet the control requirements, the present invention improves the split-Brow bar algorithm, specifically including: In the construction of fuzzy sets, the conventional assumption of symmetric space was abandoned. Combining the physical prior knowledge of the one-way irreversibility of sediments, a one-way support domain truncation operation was implemented to ensure that the worst case of model search does not violate physical common sense. The process of unidirectional support domain truncation includes: This embodiment defines a continuous set of real numbers. This space set represents the range of values ​​for all possible impedance fluctuation residuals in the future; The lower boundary value of the residual interval, i.e., the minimum residual, is determined by using the row vector of the discrete empirical probability distribution. Constructing a constrained support domain ; This truncation operation cuts off the random variable. The path with a value less than the minimum residual excludes the case where the actual impedance is lower than the current historical benchmark, which is a case where the gap increases in reverse, which is contrary to the laws of physics.

[0027] This invention implements a unidirectional support domain truncation operation by combining the physical principle of unidirectional irreversible miniaturization of sediments. This not only eliminates invalid negative physical noise caused by sensor measurements, but also cuts off the value path that violates the principle of inverse gap enlargement. This ensures that the worst-case degradation of the model search fully matches the real physical evolution law and avoids distortion of the control benchmark.

[0028] The specific process of constructing the degenerate distribution uncertainty set by the model building module also includes: Calculate the statistical variance of the valid residual data after removing invalid data points; The product of the statistical variance and the preset confidence level amplification factor is used as the dynamic distribution divergence tolerance. A distance metric is constructed based on a quadratic transmission cost function, which is the squared difference between any two residual values ​​within the constrained support domain. Centered on the empirical probability distribution vector, within the constrained support domain, the set of all legal probability distributions that satisfy the transfer deviation between the unknown distribution calculated based on the quadratic transmission cost function and the central distribution is less than or equal to the dynamic distribution divergence tolerance is taken as the degenerate distribution uncertainty set.

[0029] In a specific embodiment of the present invention, after determining the value space of impedance fluctuation residuals, this embodiment generates the boundary of the fuzzy set by extracting the dynamic features of the impedance fluctuation residuals. Specifically, invalid data points in the residual column vector that are assigned a value of 0 due to the preceding physical truncation are removed, and for the remaining valid residual data points, their arithmetic mean and statistical variance are calculated. The product of the statistical variance and the preset confidence amplification factor of 1.5 is used as the dynamic distribution divergence tolerance of the degradation distribution uncertainty set. This value is selected as the empirical optimal balance point obtained through statistical analysis of historical operating data and orthogonal experiments on spatial atomic layer deposition equipment: when the amplification factor is lower than 1.5, the generated dynamic distribution divergence tolerance is too small, which makes it unable to cover about 5% of the real extreme flow field deterioration conditions, and there is a risk of underreporting cross-contamination; when the amplification factor is higher than 1.5, the assessment of degradation risk is too conservative, which leads to the production scheduling decision to frequently reduce the speed, resulting in unnecessary loss of equipment capacity of more than 15%. Therefore, using 1.5 as the coefficient can achieve the optimal engineering balance between the contamination interception rate and the maximization of capacity. The more unstable the air curtain flow field, the larger the calculated variance, the larger the final generated dynamic distribution divergence tolerance, and the wider the range of degradation probability distributions that the degradation distribution uncertainty set can cover. After determining the restricted support domain and dynamic distribution divergence tolerance Subsequently, existing technologies typically use conventional fixed numerical threshold comparison methods to define the degradation boundaries that need to be prevented. However, this invention bypasses the dimensionality curse and lack of physical penalty weights for extreme conditions caused by conventional fixed numerical thresholds when dealing with high-dimensional probability distribution features. It proposes to construct a distance metric method based on a quadratic transmission cost function to quantify the transfer deviation between the unknown distribution and the known center distribution. This results in the technical effect of assigning realistic heavy penalty weights to extreme deterioration conditions at the physical level, and providing an indispensable geometric convexity bridge for subsequent dimensionality reduction to mixed integer second-order cone programming. The quadratic transmission cost function is a constrained support domain. The squared difference between any two residual values ​​is used to quantify the transfer penalty cost between the two probability distributions. The introduction of this quadratic transfer cost function not only gives the extremely deteriorated working conditions a realistic heavy penalty weight at the physical level, but also provides an indispensable geometric convexity bridge for the subsequent dimensionality reduction of infinite dimension Bruker optimization into mixed integer second-order cone programming. Based on the distance metric constructed using this transmission cost function, a degenerate distribution uncertainty set is constructed. The mathematical expression for the degenerate distribution uncertainty set is: ; in, Represents the constrained support domain The set of all legal probability distributions defined above; This represents the unknown, highly probable true distribution variable in the set that is likely to deteriorate. The row vector of the discrete empirical probability distribution; For dynamic distribution divergence tolerance; Represents the unknown distribution calculated based on the quadratic transmission cost function. With known central distribution The square of the second-order Wasserstein distance between them.

[0030] This invention dynamically generates a divergence tolerance by multiplying the statistical variance by a specific amplification factor, achieving a balance between extreme pollution interception rates and maximizing equipment capacity. By introducing a quadratic transmission cost function to construct a distance metric, it not only assigns realistic heavy penalty weights to extreme flow field deterioration conditions at the physical level, but also provides an indispensable geometric convexity bridge for the subsequent dimensionality reduction of the infinite-dimensional model into second-order cone programming.

[0031] The specific process of constructing the production scheduling target cost function by the model building module includes: Initialize the production scheduling decision vector, which includes substrate drive speed control instructions and channel switching instructions; Construct the production scheduling target cost function. The mathematical expression of the production scheduling target cost function is as follows: ; in, Let the production scheduling objective cost function be... This is a column vector for production scheduling decisions, containing elements. and , The true impedance random variable of the future unknown flow field. The preset capacity delay cost coefficient is 5. , This refers to the specific value of the substrate drive speed control command. The preset cross-contamination penalty coefficient is 200. , This is a channel switching command. To minimize residuals, A fixed sunk cost coefficient is set for preset channel switching to suppress equipment status fluctuations caused by frequent false triggers.

[0032] In a specific embodiment of the present invention, the initial dimension is: The production scheduling decision column vector contains decision variables in two collaborative dimensions. The elements in the first row represent the specific values ​​of the substrate drive speed control commands in the micro-process intervention dimension. These are continuous real number variables with values ​​limited to a range of [value range missing]. Between these values, the unit is millimeters per second, where 100 represents the device's maximum full-load transmission speed; the second row of elements represents channel switching instructions in the macro-scheduling dimension, which are... The integer variable, when the second row of elements is 1, means that the wafer continues to be processed in the channel where there is a risk of degradation; when the second row of elements is 0, it means that the current channel is closed, the anti-fouling flushing command of the inert gas is triggered simultaneously, and the subsequent wafer sequence is routed to the backup healthy nozzle channel. To collaboratively reconfigure wafer scheduling and process resources, this embodiment constructs a scheduling objective cost function containing a quadratic structure. The first half of the scheduling objective cost function... The linear structure indicates that the lower the substrate transmission speed is than the full-load speed, the greater the economic loss caused by production delays. The latter half of the production scheduling objective cost function It presents a quadratic structure with respect to the impedance random variable. The pollution penalty cost takes effect if and only if the decision is made to continue using the current channel. The further the true impedance deviates from the minimum residual, the more quadratic the penalty cost increases. In the actual architecture of spatial atomic deposition equipment, multiple parallel deposition gas curtain channels are typically present. A value of 0 indicates that the current channel is severely degraded, and the system is incurring fixed costs. Closing the channel and switching the subsequent wafer flow to other backup healthy nozzle channels effectively prevents decision instructions from frequently fluctuating between 0 and 1; In this embodiment, c3 is set to 1000 yuan. This value is calculated based on the actual rigid cost generated by the equipment performing a channel switching action. Specifically, it covers the material cost of high-purity inert gas consumed by triggering the anti-fouling flushing command once, as well as the fixed cycle delay cost caused by the wafer changing the transmission route. The fixed penalty of 1000 yuan is introduced as a judgment barrier. Its practical engineering significance is that a switching command will only be issued when the current channel flow field deteriorates severely, causing the expected secondary contamination penalty for continued processing to exceed 1000 yuan. If it is only a small impedance fluctuation, the fluctuation will be absorbed by prioritizing the execution of the substrate transmission speed control command, thereby effectively suppressing frequent channel switching and equipment status jitter caused by small disturbances.

[0033] The cost function constructed in this invention uses the fixed sunk cost of channel switching as a decision barrier. When faced with small impedance fluctuations, it will first absorb the disturbance by executing the substrate drive deceleration command. Switching will only be triggered when the expected contamination penalty exceeds the rigid sunk cost, thus suppressing equipment status jitter and material loss caused by frequent false triggering of switching commands.

[0034] The specific process of transforming the production scheduling objective cost function into a mixed-integer second-order cone programming model includes: Under the constraint of a degenerate distribution uncertainty set, a minimization-maximization model is constructed with the objective of minimizing the expected value of the production scheduling target cost function; By introducing Lagrange multipliers and auxiliary variables, a dual transformation is performed on the minimization-maximization model; Auxiliary variables are introduced, and boundary constraints are constructed by combining preset production delay cost coefficient, substrate transmission speed control command, channel switching command, central representative value of each sub-interval, Lagrange multiplier and preset cross-contamination penalty coefficient, so that the boundary constraints constitute a second-order rotating cone inequality. By employing dual transformation and boundary constraints, the minimization-maximization model is equivalently transformed into a mixed-integer second-order cone programming model.

[0035] In a specific embodiment of the present invention, based on the constructed production scheduling objective cost function, it is necessary to solve for the production scheduling decision column vector that minimizes the expected cost under the worst unknown degradation probability distribution. On the solution path of production scheduling decision, the minimization-maximization model contains infinite-dimensional probability distribution constraints, which cannot be directly calculated. Therefore, the present invention calls the dual transformation algorithm to reduce its dimensionality. Specifically, a non-negative scalar Lagrange multiplier is introduced. This represents a breakthrough in the dynamic distribution divergence tolerance. Boundary penalties and auxiliary row vectors; Furthermore, the present invention constructs a production scheduling objective cost function configured with a cross-contamination penalty term that grows quadratically, and the decision variables include 0-1 integer variables for route switching. After dual transformation, this infinite-dimensional minimization-maximization problem is equivalently transformed into a standard mixed integer second-order cone programming model. The transformed equivalent minimization objective function becomes: ; in, The transformed equivalent minimization objective function is... For production scheduling decision column vectors, For non-negative scalar Lagrange multipliers, For auxiliary row vectors, row vector The first in One probability element, For auxiliary row vectors The first in One element, To accommodate the dynamic distribution divergence tolerance, This represents the number of discrete sub-intervals; in this embodiment, it is 50. In order to obtain the production scheduling decision column vector by solving the objective function The equivalent minimization objective function is also subject to boundary constraints used to simultaneously solve for the auxiliary and decision variables. The mathematical expression for the boundary constraints is: ; in, For auxiliary row vectors The first in One element, This is the preset capacity delay cost coefficient. This refers to the specific value of the substrate drive speed control command. This is a channel switching command. The true impedance random variable of the future unknown flow field. To minimize residuals, For non-negative scalar Lagrange multipliers, The preset cross-contamination penalty coefficient is 200. This represents the economic penalty weight for wafer scrapping or yield reduction caused by parasitic deposition due to deterioration of flow field impedance. For a restricted support domain, For the upper bound, For the aforementioned The center of each consecutive equal-length subinterval represents a numerical value. Set a fixed sunk cost coefficient for the preset channel switching; Furthermore, to ensure that the above quadratic form equation containing the supremum is convergent and bounded, the following stable boundary conditions must be satisfied: ; Combined variables Due to the 0-1 discrete nature of the problem, the above model equivalently transforms the original minimization-maximization probability model, which contained infinite dimensions and was difficult to solve directly, into a standard mixed-integer second-order cone programming (MISOCP) problem containing integer variables and a rotating second-order cone constraint. This constraint can losslessly encompass the quadratic penalty term used to quantify the risk of cross-contamination in the original objective function; The equivalent minimization objective function not only eliminates the maximization layer, but also, because the original cost function has a clear quadratic structure, all the transformed constraints meet the convex geometric boundary requirements of the second-order cone. Specifically, this is manifested in the rotating second-order cone inequality constraints formed between the elements of the introduced auxiliary row vector and the quadratic terms of the original cost function. This transformation equivalently transforms the original minimization-maximization non-convex probabilistic model, which contained infinite dimensions and was difficult to solve directly, into a standard convex optimization problem. Based on the mathematical property that convex geometry has no local minimum trap, this transformation ensures that the obtained production scheduling decision column vector is the global optimal solution, eliminating the hidden danger of making suboptimal production scheduling decisions due to the algorithm getting trapped in local optima. This rotating second-order cone constraint can losslessly encompass the quadratic penalty term used in the original objective function to quantify the risk of cross-contamination. While maintaining the strict convexity of the model, it retains the real physical characteristic that the greater the deviation of the impedance residual, the more exponentially the contamination risk increases.

[0036] This invention introduces Lagrange multipliers and auxiliary variables for dual transformation and constructs a rotating second-order cone inequality boundary constraint condition. This transforms the original minimization-maximization non-convex probabilistic model, which has infinite dimensions and is extremely difficult to solve, into a standard strictly convex optimization problem. Based on the characteristic that convex geometry has no local minimum traps, the anti-sticking algorithm avoids getting trapped in local optima and making suboptimal decisions.

[0037] The specific process of the offline solution module generating the production scheduling strategy library includes: Set the effective engineering interval of dynamic distribution divergence tolerance, and divide the effective engineering interval into a preset number of degradation level intervals with a fixed step size; For each degradation level interval, the right boundary value of the degradation level interval is substituted as a fixed parameter into the mixed integer second-order cone programming model; The mixed-integer optimization solver is called to solve the mixed-integer second-order cone programming model, and the optimal production scheduling decision vector for the corresponding degradation level interval is obtained; The optimal production scheduling decision vectors obtained from solving all degradation level intervals are combined row by row to generate a production scheduling strategy library.

[0038] In a specific embodiment of the present invention, in order to absolutely guarantee the response speed required for continuous manufacturing of spatial atomic layer deposition equipment, this embodiment adopts a spatially discretized hierarchical lookup table mechanism; During the offline phase before equipment commissioning, the continuous value space of the dynamic distribution divergence tolerance is discretized and graded. Specifically, the following steps are set: The effective engineering range is and with a fixed step size Divide it into A range of degradation levels; For these 100 degradation level intervals, the right boundary value of each degradation level interval is used as a fixed parameter. Substitute the equivalent minimization objective function generated in step three into the following steps; The Gurobi mixed integer optimization solver, which combines the branch and bound method with the interior point method, is invoked and run 100 times. The transpose of the optimal production scheduling decision column vector obtained from the 100 times is combined into a production scheduling strategy library matrix by row vectors. Each row of the production scheduling strategy library matrix corresponds to the optimal production scheduling decision column vector obtained in one solution process. The first column stores the optimal substrate transmission speed value solved in the corresponding interval, and the second column stores the optimal dynamic routing switch value solved in the corresponding interval.

[0039] This invention pre-discretizes and classifies the effective engineering interval of the dynamic distribution divergence tolerance in the offline stage before equipment production, and calls the solver to traverse and calculate all optimal production scheduling decision vector combinations into a library. This pre-emptively moves the extremely large and time-consuming mathematical iterative calculations, thus removing the heavy computational burden from the online actual processing stage.

[0040] The specific process of the online control module outputting substrate drive speed adjustment commands and channel switching commands includes: During equipment operation, the dynamic distribution divergence tolerance is calculated in real time at the current moment; The row index number is calculated based on the dynamic distribution divergence tolerance and the fixed step size. The row index number is the rounded-up value of the dynamic distribution divergence tolerance divided by the fixed step size.

[0041] In one specific embodiment of the present invention, during the online phase of actual processing by the equipment, when the dynamic distribution divergence tolerance is received in real time from the upstream module... When performing a table lookup operation, since the radius is obtained by linearly amplifying the statistical variance through a fixed coefficient, judging the level of the radius is equivalent to judging the discrete level interval into which the real-time variance falls. The row index number is calculated according to a fixed mapping formula. The mathematical expression for the row index number is: ; in, The row index number. To provide a tolerance for dynamic distribution divergence received in real time. For a fixed step size of the degradation level interval, This represents the floor function.

[0042] This invention simplifies the originally complex online real-time optimization process into a simple division and mapping operation based on real-time distributed divergence tolerance and fixed step size to calculate row index numbers, and directly extracts and issues instructions in parallel in the production scheduling strategy library.

[0043] The specific process of the online control module outputting substrate drive speed adjustment commands and channel switching commands also includes: Perform a safety clamping judgment on the row index number: if the row index number is greater than the preset number of the degradation level range, then the row index number is assigned the preset number; if the row index number is less than the set lower limit value, then the row index number is assigned the set lower limit value. Based on the row index number after the safety clamping determination, the row vector of the corresponding row is extracted from the production scheduling strategy library; The optimal substrate drive speed value in the first column of the row vector is converted into a substrate drive speed control command and sent to the programmable logic controller. At the same time, the optimal channel switching value in the second column is converted into a channel switching command and sent to the manufacturing execution system.

[0044] In one specific embodiment of the present invention, to prevent index overflow caused by real-time data exceeding the preset engineering range due to extreme abnormal operating conditions, this embodiment executes a safety clamping determination logic, which includes: If the row index number is greater than 100, then the index number will be forced to be 100; if the row index number is less than 1, then the index number will be forced to be 1. Based on the row index number, the row vector of the row containing the row index number in the production scheduling strategy library matrix is ​​extracted. The first column element in the row vector is extracted and converted into a standard digital speed setpoint instruction. This instruction is then synchronously sent to the programmable logic controller (PLC) via industrial Ethernet. The PLC then drives the linear motor to perform speed reduction. Simultaneously, the second column of elements is converted into a digital status code and sent to the manufacturing execution system via a network protocol. Based on this, the manufacturing execution system immediately executes the command to shut down the current wafer channel and switch the subsequent wafer flow to the healthy nozzle channel.

[0045] When the equipment encounters extreme abnormal operating conditions that cause the real-time collected and calculated data to exceed the preset effective engineering range, this forced assignment logic can effectively prevent index overflow and system crash caused by the row index number crossing the boundary of the strategy library matrix, thus comprehensively ensuring the operational stability of the control system and the security of command control in complex and harsh industrial manufacturing environments.

[0046] This invention uses a low-pass filtering algorithm to independently filter the collected differential pressure and flow rate values, effectively preserving the real low-frequency flow field degradation signal that includes the equipment's baseline operating state and slow degradation trend, while eliminating high-frequency mechanical vibration noise. In the feature extraction and degradation distribution uncertainty set construction stages, this invention combines the scenario of unidirectional irreversible sediments to implement unidirectional support domain truncation operation, which not only effectively identifies and eliminates invalid physical noise caused by sensor measurements, but also cuts off the reverse impedance fluctuation path, ensuring that the worst case of model search matches the real scenario. This invention dynamically sets the distribution divergence tolerance based on the statistical variance of historical valid data, achieving a balance between pollution interception rate and capacity maximization. By introducing a quadratic transmission cost function to construct a distance metric, it assigns a heavy penalty weight to extreme deterioration conditions. In terms of production scheduling, the production scheduling target cost function constructed by this invention comprehensively weighs the capacity delay cost, cross-contamination penalty, and fixed sunk cost of channel switching. Using the fixed sunk cost as a decision barrier, it absorbs small disturbances by prioritizing the execution of speed reduction, thereby effectively suppressing frequent channel switching and equipment status jitter caused by small impedance fluctuations. This invention transforms the original minimization-maximization non-convex probabilistic model, which contains infinite dimensions and is difficult to solve directly, into a standard mixed-integer second-order cone programming model through dual transformation and boundary constraint conditions. Based on the mathematical property that convex geometry is free from local minima traps, it fundamentally ensures that the obtained production scheduling decision column vector is the global optimal solution, eliminating the hidden danger of the algorithm getting trapped in local optima. This invention generates a production scheduling strategy library by performing discrete hierarchical solutions on the value interval in the offline stage, and performs fast table lookup based on row index number in the online stage by combining a safety clamping judgment mechanism. This ensures the millisecond-level response speed required for continuous manufacturing of spatial atomic layer deposition equipment and the control safety under extreme abnormal conditions.

[0047] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0048] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the protection scope of the present invention.

Claims

1. An intelligent control system for a spatial atomic layer deposition apparatus, characterized in that, The application relates to a space-type atomic layer deposition equipment gas curtain pipeline differential pressure and flow value acquisition module, a feature extraction module, a model construction module, an offline solution module and an online control module. The acquisition module is used for acquiring the differential pressure value and the flow value of the space-type atomic layer deposition equipment gas curtain pipeline, and performing low-pass filtering processing on the differential pressure value and the flow value to obtain a pure signal matrix. The feature extraction module is used for calculating a flow field impedance according to the pure signal matrix, extracting impedance residuals, and converting the impedance residuals into an empirical probability distribution vector. The model construction module is used for constructing a degenerate distribution uncertainty set based on the empirical probability distribution vector and statistical characteristics of the impedance residuals, constructing a production scheduling target cost function containing the degenerate distribution uncertainty set, and converting the production scheduling target cost function into a mixed integer second-order cone programming model. The offline solution module is used for discretely grading a value interval of a dynamic distribution divergence tolerance, solving the mixed integer second-order cone programming model based on the discretely graded dynamic distribution divergence tolerance, and generating a production scheduling strategy library. The online control module is used for looking up the production scheduling strategy library according to the real-time acquired dynamic distribution divergence tolerance, and outputting wafer transmission speed control instructions and channel switching instructions. The specific process of the acquisition module performing low-pass filtering processing includes the following steps.

2. An intelligent control system for a spatial atomic layer deposition apparatus as claimed in claim 1, characterized in that, The differential pressure value and the flow value are collected with a set time window length and a sampling frequency, and the collected data are integrated into an initial joint signal matrix. A fourth-order Butterworth low-pass filtering algorithm is called to perform filtering calculation on the initial joint signal matrix, a cutoff frequency is set, and the pure signal matrix is output. The specific process of the feature extraction module extracting impedance residuals and converting the impedance residuals into an empirical probability distribution vector includes the following steps.

3. An intelligent control system for a spatial atomic layer deposition apparatus as claimed in claim 2, characterized in that, The ratio of the differential pressure value in the first column to the flow value in the second column in each row of the pure signal matrix is taken as the flow field impedance of the corresponding sampling point. The difference between the flow field impedance and a preset calibration impedance is taken as the impedance residual, and if the impedance residual is negative, the impedance residual is forcibly assigned as zero, and a residual column vector composed of the impedance residuals of all sampling points is obtained. Invalid data points with a value of zero in the residual column vector are removed, and the maximum residual and the minimum residual in the remaining valid data are extracted. A closed interval composed of the maximum residual and the minimum residual is equally divided into a preset number of continuous equal-length subintervals, and the number of residuals falling into each subinterval is counted. The ratio of the number of residuals falling into each subinterval to the number of remaining elements in the residual column vector after removing the invalid data points is taken as the probability of the impedance residual falling into the corresponding subinterval. The probabilities of all subintervals are spliced in the order of subinterval division to construct the empirical probability distribution vector. The specific process of the model construction module constructing a degenerate distribution uncertainty set includes the following steps.

4. An intelligent control system for a spatial atomic layer deposition apparatus as claimed in claim 3, characterized in that, A limited support domain is constructed by taking the minimum residual and a one-way support domain truncation operation, and the mathematical expression of the limited support domain is as follows: The specific process of the model construction module constructing a degenerate distribution uncertainty set further includes the following steps. ； wherein, represents an unknown real flow field impedance residual random variable, is the minimum residual.

5. An intelligent control system for a spatial atomic layer deposition apparatus as claimed in claim 4, characterized in that, The statistical variance of the valid residual data after removing the invalid data points is calculated. ​ The product of the statistical variance and a preset confidence amplification coefficient is taken as a dynamic distribution divergence tolerance; A distance metric is constructed based on a quadratic transmission cost function, which is the square of the difference between any two residual values in the limited support domain; A set of all legal probability distributions that satisfy the transition deviation between the unknown distribution calculated based on the quadratic transmission cost function and the center distribution being less than or equal to the dynamic distribution divergence tolerance is taken as the degradation distribution uncertainty set in the limited support domain.

6. An intelligent control system for a spatial atomic layer deposition apparatus as claimed in claim 5, characterized in that, The specific process of constructing the scheduling target cost function by the model construction module includes: An initial scheduling decision vector is initialized, which includes wafer driving speed control instructions and channel switching instructions; A scheduling target cost function is constructed, and the mathematical expression of the scheduling target cost function is: ； wherein, is the production scheduling objective cost function, is the production scheduling decision column vector, containing elements and , is the future unknown flow field real impedance random variable, is the preset capacity delay cost coefficient, is the specific value of the substrate transmission speed regulation instruction, is the preset cross-contamination penalty coefficient, is the channel switching instruction, is the minimum residual error, is the preset channel switching fixed sunk cost coefficient.

7. An intelligent control system for a spatial atomic layer deposition apparatus as claimed in claim 6, characterized in that, The specific process of converting the scheduling target cost function into a mixed integer second-order cone programming model includes: Under the constraint of the degradation distribution uncertainty set, a minimax model is constructed to minimize the expected value of the scheduling target cost function; A Lagrange multiplier and an auxiliary variable are introduced to perform dual transformation on the minimax model; An auxiliary variable is introduced, and a boundary constraint condition is constructed by combining a preset capacity delay cost coefficient, the wafer driving speed control instructions, the channel switching instructions, the center representative values of each sub-interval, the Lagrange multiplier, and a preset cross-contamination penalty coefficient, so that the boundary constraint condition constitutes a rotated second-order cone inequality; The minimax model is equivalently converted into the mixed integer second-order cone programming model through the dual transformation and the boundary constraint condition.

8. An intelligent control system for a spatial atomic layer deposition apparatus as claimed in claim 7, characterized in that, The specific process of generating a scheduling strategy library by the offline solving module includes: An effective engineering interval of the dynamic distribution divergence tolerance is set, and the effective engineering interval is divided into a preset number of degradation level intervals at a fixed step length; For each degradation level interval, the right boundary value of the degradation level interval is taken as a fixed parameter and substituted into the mixed integer second-order cone programming model; A mixed integer optimization solver is called to solve the mixed integer second-order cone programming model, and an optimal scheduling decision vector corresponding to the degradation level interval is obtained; The optimal scheduling decision vectors obtained by solving all degradation level intervals are combined in rows to generate a scheduling strategy library.

9. An intelligent control system for a spatial atomic layer deposition apparatus as claimed in claim 8, characterized in that, The specific process of outputting wafer driving speed control instructions and channel switching instructions by the online control module includes: During the operation of the equipment, the dynamic distribution divergence tolerance at the current time is calculated in real time; According to the dynamic distribution divergence tolerance and the fixed step length, a row index number is calculated, which is the upward integer value after the dynamic distribution divergence tolerance is divided by the fixed step length.

10. An intelligent control system for a spatial atomic layer deposition apparatus as claimed in claim 9, characterized in that, The specific process of outputting wafer driving speed control instructions and channel switching instructions by the online control module also includes: A safety clamping determination is performed on the row index number: if the row index number is greater than the preset number of the degradation level range, the row index number is assigned the preset number; if the row index number is less than the set lower limit value, the row index number is assigned the set lower limit value. Based on the row index number determined by the safety clamping, the row vector of the corresponding row is extracted from the production scheduling strategy library; The optimal substrate transmission speed value in the first column of the row vector is converted into a substrate transmission speed control command and sent to the programmable logic controller. At the same time, the optimal channel switching value in the second column is converted into a channel switching command and sent to the manufacturing execution system.

Images