Method and system for dynamically adjusting silicon carbide cmp polishing solution based on multi-parameter feedback control

Abstract

The application belongs to the technical field of semiconductor manufacturing, and discloses a silicon carbide CMP polishing liquid dynamic adjustment method and system based on multi-parameter feedback control, which comprises the following steps: collecting the oxidation-reduction potential, pH value and turbidity data of the polishing liquid in real time; inputting the sensor data into an edge computing platform for filtering and standardization processing; establishing a dynamic adjustment model based on a fuzzy PID control algorithm, taking a process state index as a control parameter switching basis; driving a high-precision actuator to complete dynamic replenishment of oxidizing agents, pH regulators and dispersants, and realizing closed-loop control of the polishing liquid composition. The application can monitor and automatically adjust the polishing liquid state in real time, effectively solve the problems of parameter drift and manual intervention lag under the traditional open-loop control mode, ensure the stability and consistency of the silicon carbide polishing process, improve the polishing quality, and reduce the waste and maintenance cost of the polishing liquid.

Description

Technical Field

[0001] This invention relates to the field of semiconductor manufacturing technology, specifically to a method and system for dynamically adjusting silicon carbide CMP polishing slurry based on multi-parameter feedback control. Background Technology

[0002] In the manufacturing process of high-hardness and brittle materials, chemical mechanical polishing (CMP) is a key process for achieving ultra-smooth surfaces, and polishing slurry is its core consumable material. Material removal rate (MRR) and surface roughness (Ra) are typically used as core indicators to evaluate the performance of polishing slurry. During the polishing process of high-hardness and brittle materials, as polishing time increases, the chemical properties (such as redox potential and pH value) and physical states (such as turbidity) of the polishing slurry change, leading to a decrease in MRR and an increase in Ra. When the MRR falls below the process threshold or Ra exceeds the acceptable range due to the deterioration of the polishing slurry performance, the polishing slurry needs to be replaced to ensure processing quality. Currently, the status of the polishing slurry is judged by manual offline detection (such as timed sampling and analysis of MRR) or static rule control (such as preset quantitative replenishment based on the number of polishing pads). This method suffers from drawbacks such as response lag, insufficient accuracy, and high maintenance costs, easily leading to waste of polishing slurry or affecting processing quality. Therefore, developing a method for real-time closed-loop adjustment of polishing slurry components is crucial for maintaining the stability of silicon carbide CMP processes and reducing consumable costs. Summary of the Invention

[0003] The purpose of this invention is to provide a method and system for dynamic adjustment of silicon carbide CMP polishing slurry based on multi-parameter feedback control, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A method for online control of multiple parameters of silicon carbide CMP polishing slurry includes the following steps:

[0006] Step 1: Preload the baseline values ​​of polishing slurry parameters from the historical process database;

[0007] Step 2: Real-time data collection of the redox potential, pH value, and turbidity of the polishing slurry, as well as the polishing head pressure and rotation speed parameters of the polishing platform, is achieved through multi-source sensors.

[0008] Step 3: The edge computing platform performs timestamp alignment and sliding window filtering on the multi-source sensor data, and standardizes it to a unified dimension to obtain standardized data.

[0009] Step 4: Compare the preprocessed multi-source sensor data with the baseline values ​​to calculate the Process State Deviation Index (PDI).

[0010] Step 5: Based on the comparison between the real-time process status deviation index (PDI) and the preset threshold, output the control command.

[0011] Furthermore, in step 2, the multi-source sensor includes:

[0012] Redox potential sensor to detect the redox potential of polishing slurry;

[0013] pH sensor to detect the pH value of polishing fluid;

[0014] Laser turbidity sensor to detect the dispersion state of abrasive in polishing slurry;

[0015] Pressure sensor detects the pressure between the polishing head and the wafer;

[0016] Polishing head speed sensor detects the rotational speed of the polishing head;

[0017] Polishing pad speed sensor detects the rotational speed of the polishing pad.

[0018] Furthermore, step 3 includes: the edge computing platform performs timestamp synchronization on the data of each sensor, and then eliminates noise in the redox potential, pH value, polishing head rotation speed, polishing pad rotation speed and pressure value between the polishing head and the wafer through recursive averaging filtering, and performs normalization processing to obtain standardized data.

[0019] Furthermore, step 4 specifically includes:

[0020] a) Compare the standardized data output from step 3 with the benchmark value to generate the following feature deviation values:

[0021] Oxidative degradation rate deviation Based on the percentage deviation from the rated oxidation degradation rate in real time;

[0022] pH offset The absolute difference between the real-time pH and the target pH;

[0023] Reunion Index Deviation ,real time The positive deviation relative to the nominal aggregation index;

[0024] b) Calculate the process state deviation index using a dynamic weighted formula:

[0025]

[0026] in, For deviations from the process status index, For deviations in oxidative degradation rate, This is the pH offset. This is due to the reunification index deviation. This is the weighting coefficient for the oxidant activity deviation. This is the pH stability weighting coefficient. The weighting coefficients for the stability of suspended particles satisfy the following conditions: ,and Decreasing over time Increases over time.

[0027] Further, step 5 specifically includes: comparing the Process State Deviation Index (PDI) with a threshold:

[0028] ① When the process state deviation index PDI > the threshold, the fuzzy PID controller generates control commands for oxidant, pH adjuster and dispersant compensation based on the process state deviation index PDI and its rate of change to control the actuator to adjust the polishing fluid parameters.

[0029] ② When the process status deviates from the exponent PDI ≤ threshold, maintain the current output value of all actuators constant;

[0030] ③ When the fluctuation range of the Process Status Deviation Index (PDI) exceeds the platform warning baseline value over multiple consecutive sampling periods, the alarm unit is triggered to prompt the polishing fluid to be replaced. The setting range of the warning baseline value is: , ,in is the average index value during the steady-state operation of the equipment, and k is the process sensitivity factor.

[0031] Furthermore, the fuzzy PID controller is implemented using a three-level architecture, including: a fuzzification layer, a fuzzy inference decision layer, and a parameter self-tuning layer;

[0032] The fuzzification layer is responsible for receiving the process state deviation index PDI and its first derivative, and converting them into fuzzy linguistic variables through triangular membership functions;

[0033] Input variable definition:

[0034] Deviation E = PDI - threshold, fuzzy linguistic variable set is Wherein, NB, NS, ZO, PS, and PB represent the degree of deviation of the process state: severe degradation, mild degradation, stable state, mild exceedance, and severe exceedance, respectively.

[0035] Deviation change rate The fuzzy language variable set is ;

[0036] The mapping from continuous universes of discourse to discrete linguistic variables is achieved using trigonometric membership functions. The expression for the trigonometric membership function is as follows:

[0037]

[0038] in, Input variable actual value Membership degree of a specific fuzzy linguistic variable, parameter The left boundary point of the membership function defines the starting point of the effective scope of the linguistic variable; parameters The right boundary point of the membership function defines the endpoint of the effective scope of the linguistic variable; parameters The peak point of the membership function identifies the core region to which the linguistic variable is fully subordinate;

[0039] The fuzzy inference decision layer stores a two-dimensional fuzzy rule base based on two input variables and performs logical inference according to a predefined decision matrix. This fuzzy inference decision layer performs fuzzy inference operations and outputs the adjustment amount of the PID controller parameters. Its core rule structure paradigm is as follows:

[0040] If E is A and EC is B, then

[0041] The number of fuzzy rules is determined by the number of linguistic values ​​of the input variables E and EC, where E represents the deviation and EC represents the rate of change of deviation. These typically constitute a set of fuzzy rules. The decision matrix of order n, where n and m represent the number of linguistic variables in E and EC, respectively; These are the adjustment amounts for proportional, integral, and derivative parameters, respectively; A is the linguistic value of the deviation E, and B is the linguistic value of the deviation change rate EC; This represents a control action vector for specific fuzzy conditions;

[0042] The parameter self-tuning layer consists of a defuzzification unit, a decoupling operation unit, and a dynamic correction unit. Through multi-level processing, it transforms the fuzzy output into precise compensation instructions.

[0043] The centroid method is used to defuzzify the unit:

[0044]

[0045] in, To precisely adjust the PID parameters that need to be solved, These correspond to the proportional term p, the integral term i, and the differential term d, respectively. This represents the total number of valid fuzzy rules actually triggered at that moment; Refers to the rule number currently being processed; This represents the activation strength of the i-th rule; This represents the original fuzzy adjustment component of the j-type parameter preset in the i-th rule;

[0046] Decoupling the operation unit enables multi-compensation quantity coordination:

[0047]

[0048] in, , , The output compensation vector corresponds to the compensation amounts for oxidant, pH value, and degree of aggregation, respectively; K is the decoupling matrix, which affects the PID parameter correction vector. Achieve cross compensation and satisfy: For matrix The element in the i-th row and j-th column, For predefined coupling coefficients;

[0049] Dynamic correction unit with historical trend adjustment:

[0050] Based on recent The process state deviation from the exponential sequence for each control cycle is used to calculate the historical trend factor. ;

[0051] Historical trend analysis is performed independently for each compensation amount, and three corrected compensation amounts are output in the end. :

[0052]

[0053] in, , , The final compensation amounts are the corrected amounts, corresponding to the final compensation amounts for oxidant, pH value, and degree of aggregation, respectively. , , These correspond to the adjustment sensitivity coefficients for pre-configured oxidant, pH value, and degree of aggregation, respectively. , , These correspond to historical trend factors for oxidant, pH value, and degree of aggregation, respectively.

[0054] Furthermore, the fuzzy reasoning decision layer possesses a dynamic knowledge base update mechanism:

[0055] Based on the system energy gradient mode Time series variance of deviation from process state index Establish rule weights that match real-time operating conditions. The applicability of the quantification rule, g(·) refers to the weight function, which is applied to the same rule group in N consecutive activations. The renormalization operator is triggered at a certain time; the activated renormalization operator automatically updates the rules according to the following feedback mechanism:

[0056]

[0057] in, The learning rate is configurable. For boundary saturation functions, This represents the actual compensation effect vector. Predict compensation vectors for the rules. For the updated rule group;

[0058] By comparing the difference between the actual compensation effect and the rule prediction output, a closed-loop feedback loop is formed to achieve online adaptive correction of the rule base parameters.

[0059] Furthermore, the historical trend factor Based on sliding window construction:

[0060]

[0061] Where i is the variable used for traversal within the window, and t is the current reference point. This is the trend balance coefficient. Here, n is the width of the dynamic sliding window, and n is the sign function. This is the current PDI value. This is a historical benchmark value. This is the mean of the PDI window.

[0062] This invention also provides an adaptive control system for silicon carbide CMP polishing slurry based on multi-source data fusion, used to implement the control method described above, characterized in that it includes:

[0063] The data acquisition module includes an oxidation-reduction potential sensor, a pH sensor, a laser turbidity sensor, a pressure sensor, a polishing head speed sensor, and a polishing pad speed sensor. The oxidation-reduction potential sensor detects the oxidation-reduction potential of the polishing slurry; the pH sensor detects the pH value of the polishing slurry; the laser turbidity sensor detects the dispersion state of the abrasive in the polishing slurry; the pressure sensor detects the pressure between the polishing head and the wafer; the polishing head speed sensor detects the speed of the polishing head; and the polishing pad speed sensor detects the speed of the polishing pad.

[0064] The edge computing platform is communicatively connected to the data acquisition module. It performs time alignment and recursive averaging filtering on the raw data acquired by the data acquisition module to eliminate high-frequency noise; it normalizes the data of each parameter; and it calculates the process state deviation index by comparing the preprocessed data with the baseline value.

[0065] The actuator module includes an oxidant metering pump, a dispersant feeding device, and a pH adjuster feeding device, which are used to automatically adjust the composition of the polishing slurry according to control commands to achieve dynamic component control of the polishing slurry.

[0066] The host computer system includes an adaptive control module with a three-layer fuzzy controller architecture, which processes the pre-processed process state deviation index (PDI), generates polishing fluid state assessment results and control instructions through threshold comparison, and feeds them back to the operator through the human-machine interface.

[0067] Compared with existing technologies, the advantages of this invention are as follows: Traditional silicon carbide polishing slurry control modes are limited by offline detection and experience-based replacement strategies, resulting in inherent defects such as response lag, failure to detect dynamic operating conditions, and resource waste. This invention innovatively integrates multi-source data, including process dynamic characteristics, and incorporates a fuzzy PID control algorithm model to achieve real-time capture of the polishing slurry state and component adaptation. Simultaneously, it avoids surface defect risks by dynamically adjusting the concentration of key additives before the abrasive performance decays to the critical point, improving the effective utilization rate of the polishing slurry and reducing the processing cost per wafer. This invention avoids the process instability and excessive resource consumption caused by response lag and extensive management in traditional technologies, thereby maintaining the stability of the silicon carbide CMP process and reducing production costs. Attached Figure Description

[0068] Figure 1 This is a flowchart of a method for dynamically adjusting silicon carbide CMP polishing slurry based on multi-parameter feedback control according to the present invention.

[0069] Figure 2 This is a schematic diagram of the polishing platform and polishing fluid circulation system in this invention.

[0070] Figure 3 This is a schematic diagram of a dynamic adjustment system for silicon carbide CMP polishing slurry based on multi-parameter feedback control according to the present invention.

[0071] Labeling descriptions: 1. Polishing head speed sensor, 2. Polishing head pressure sensor, 3. Polishing pad speed sensor, 4. Polishing fluid circulation pipe, 5. Primary recovery tank, 6. First centrifugal pump, 7. Recovery filter, 8. Supply tank, 9. Second centrifugal pump, 10. Third centrifugal pump, 11. Supply filter, 12. Multi-parameter sensing unit (pH sensor, laser turbidity sensor, oxidation-reduction potential sensor), 13. Fourth centrifugal pump, 14. Wastewater tank, 15. Communication line, 16. Three-channel actuator, 17. Fifth centrifugal pump, 18. Electric valve. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] like Figures 1-2 As shown, this invention provides a method for dynamically adjusting silicon carbide CMP polishing slurry based on multi-parameter feedback control, comprising the following steps:

[0074] Step 1: Preload the baseline values ​​of polishing slurry parameters from the historical process database, including: rated oxidation degradation rate, target pH value, and rated agglomeration index. These values ​​are automatically loaded into the host computer memory when the system starts up, serving as a reference for real-time control.

[0075] Step 2 involves real-time acquisition of multi-source data on the polishing slurry and polishing platform during the processing using multi-source sensors. This specifically includes:

[0076] First, a polishing slurry supply-side detection subsystem is installed in the polishing slurry circulation pipeline. This subsystem includes a multi-parameter sensing unit 12 (pH sensor, laser turbidity sensor, and oxidation-reduction potential sensor). The laser turbidity sensor detects the dispersion state of the abrasive in the polishing slurry, the pH sensor detects the pH value of the polishing slurry, and the oxidation-reduction potential sensor detects the oxidation-reduction potential of the polishing slurry. During detection, a second centrifugal pump 9 supplies slurry to the multi-parameter sensor unit, and a fourth centrifugal pump 13 transports the waste liquid after detection to a wastewater tank 14. Next, a polishing machine-side detection subsystem is installed. This subsystem includes a polishing head speed sensor 1, a polishing head pressure sensor 2, and a polishing pad speed sensor 3. In this subsystem, the polishing head speed sensor 1 detects the polishing head speed, the polishing head pressure sensor 2 detects the polishing head pressure, and the polishing pad speed sensor 3 detects the polishing pad speed.

[0077] Step 3: The detection subsystem transmits the collected data to the edge computing platform, where the edge computing platform performs data preprocessing, including:

[0078] Step 3.1: Unify time labels and align timestamps. Use the sensor with the lowest sampling frequency as the time reference to unify time labels for sensor data with different sampling frequencies.

[0079] Step 3.2, noise reduction: a sliding window filtering method is used to reduce noise and smooth the input data curve.

[0080] Step 3.3, standardization: normalize all sensor data to the 0-1 range to facilitate model input processing.

[0081] Step 4: Subsequently, the edge computing platform transmits the preprocessed data to the host computer control system via wireless communication. The host computer then executes the following PDI calculation process:

[0082] Step 4.1, Feature deviation calculation: The input layer receives the standardized sensor data stream (oxidative degradation rate / pH / agglomeration index) and generates three key deviation quantities (oxidative degradation rate deviation, pH offset, and agglomeration index deviation) by comparing the baseline values ​​in real time.

[0083] Step 4.2, Dynamically weighted PDI synthesis, calculates the process state deviation index using the following dynamic weighting formula:

[0084]

[0085] in, For deviations from the process status index, For deviations in oxidative degradation rate, This is the pH offset. This is due to the reunification index deviation. This is the weighting coefficient for the oxidant activity deviation. This is the pH stability weighting coefficient. The weighting coefficients for the stability of suspended particles satisfy the following conditions: ,and Decreasing over time Increases over time.

[0086] Its temporal regulation mechanism is: oxidative decay weighting The weighting decreases exponentially with polishing time; agglomeration control weighting Increasing over time; pH shift weighting Keep it constant.

[0087] Step 5: Based on the comparison between the real-time process status deviation index (PDI) and the preset threshold, output control commands, including:

[0088] The process state deviation index (PDI) is compared with a threshold:

[0089] ① When the process state deviation index PDI > the threshold, the fuzzy PID controller generates control commands for oxidant, pH adjuster and dispersant compensation based on the process state deviation index PDI and its rate of change to control the actuator to adjust the polishing fluid parameters.

[0090] ② When the process status deviates from the exponent PDI ≤ threshold, maintain the current output value of all actuators constant;

[0091] ③ When the fluctuation range of the Process Status Deviation Index (PDI) exceeds the platform warning baseline value over multiple consecutive sampling periods, the alarm unit is triggered to prompt the polishing fluid to be replaced. The setting range of the warning baseline value is: , ,in is the average index value during the steady-state operation of the equipment, and k is the process sensitivity factor.

[0092] The fuzzy PID controller is implemented using a three-level architecture, including: a fuzzification layer, a fuzzy inference decision layer, and a parameter self-tuning layer.

[0093] In the fuzzification layer of the fuzzy PID controller, the system receives the process state deviation index (PDI) and its first derivative in real time, namely the deviation amount (E, defined as the difference between PDI and the threshold) and the deviation change rate (EC, defined as the rate of change of PDI over time). The fuzzy linguistic variable set for the deviation amount E is {NB, NS, ZO, PS, PB}, corresponding to the process state changes from severe degradation, mild degradation, stable state, mild exceedance, and severe exceedance. The fuzzy linguistic variable set for the deviation change rate EC is {NB, ZO, PB}, representing the trends of rapid deterioration, stable state, and rapid improvement.

[0094] And use trigonometric membership functions This achieves the mapping from continuous domains to discrete linguistic variables, with parameters rationally configured according to the process tolerance range to ensure coverage of fluctuations under normal operating conditions. The trigonometric membership functions are as follows:

[0095] in, Input variable actual value Membership degree of a specific fuzzy linguistic variable, parameter The left boundary point of the membership function defines the starting point of the effective scope of the linguistic variable; parameters The right boundary point of the membership function defines the endpoint of the effective scope of the linguistic variable; parameters The peak point of the membership function identifies the core region to which the linguistic variable is fully subordinate.

[0096] The fuzzy inference decision layer of the control model stores a two-dimensional fuzzy rule base based on two input variables and performs logical inference according to a predefined 5×3 order decision matrix, outputting the adjustment amount of the PID controller parameters. Each rule follows a unified logical structure: "If E is A and EC is B, then..." "The number of fuzzy rules is determined by the number of linguistic values ​​of the input variables E and EC, where E represents the deviation and EC represents the rate of change of deviation, typically constituting a..." The decision matrix of order n, where n and m represent the number of linguistic variables in E and EC, respectively; These are the adjustment amounts for proportional, integral, and derivative parameters, respectively; A is the linguistic value of the deviation E, and B is the linguistic value of the deviation change rate EC; This represents the control action vector for specific fuzzy conditions. Simultaneously, the system possesses online learning capabilities, utilizing a dynamic knowledge base update mechanism in the fuzzy inference decision layer, based on the system's energy gradient... Time series variance of deviation from process state index Establish rule weights The applicability of the quantified rules is assessed, and the renormalization operator is triggered after 10 consecutive activations of the same set of rules. The activated renormalization operator is automatically updated according to the following feedback mechanism:

[0097] in, The learning rate is configurable. For boundary saturation functions, This represents the actual compensation effect vector. Predict compensation vectors for the rules. This is the updated rule group.

[0098] By comparing the difference between the actual compensation effect and the rule prediction output, a closed-loop feedback loop is formed to achieve online adaptive correction of the rule base parameters, ensuring that the control rules continuously closely match the actual process characteristics.

[0099] Finally, the parameter self-tuning layer consists of a defuzzification unit, a decoupling operation unit, and a dynamic correction unit. Upon entering the top layer of the parameter self-tuning layer, the defuzzification unit first performs defuzzification using the centroid method:

[0100] in, To precisely adjust the PID parameters that need to be solved, These correspond to the proportional term p, the integral term i, and the differential term d, respectively. This represents the total number of valid fuzzy rules actually triggered at that moment; Refers to the rule number currently being processed; This represents the activation strength (membership calculation result) of the i-th rule. This represents the original fuzzy adjustment component of the j-type parameter preset in the i-th rule.

[0101] This method calculates the precise parameter adjustment by integrating the contributions of all activation rules.

[0102] Secondly, the decoupling operation unit, considering the coupling characteristics of the three components of the polishing fluid, is configured with a decoupling transformation matrix K to achieve multi-compensation quantity coordination and output a compensation vector:

[0103] in, , , The output compensation vector corresponds to the compensation amounts for oxidant, pH value, and degree of aggregation, respectively; K is the decoupling matrix, which affects the PID parameter correction vector. Achieve cross compensation and satisfy: , For matrix The element in the i-th row and j-th column, For predefined coupling coefficients.

[0104] Finally, a historical trend adjustment is added to the dynamic correction unit, and the historical trend factor is calculated based on a sliding window. :

[0105] Where i is the variable used for traversal within the window, and t is the current reference point. This is the trend balance coefficient. Here, n is the width of the dynamic sliding window, and n is the sign function. This is the current PDI value. This is a historical benchmark value. This is the mean of the PDI window.

[0106] Then, a historical trend analysis is performed independently on each compensation amount, and finally, three corrected compensation amounts are output. :

[0107] in, , , The final compensation amounts are the corrected amounts, corresponding to the final compensation amounts for oxidant, pH value, and degree of aggregation, respectively. , , These correspond to the adjustment sensitivity coefficients for pre-configured oxidant, pH value, and degree of aggregation, respectively. , , These correspond to historical trend factors for oxidant, pH value, and degree of aggregation, respectively.

[0108] After the model training is completed, the final compensation command is transmitted to the three-channel actuator 16 through the communication bus, and the modulated polishing fluid is injected into the circulation pipeline through the electric valve 18 and delivered to the polishing machine to form a complete closed-loop regulation circuit.

[0109] Step 6: Repeat the polishing process, and use the multi-parameter feedback controlled silicon carbide CMP polishing slurry dynamic adjustment system to perform real-time closed-loop adjustment of the polishing slurry composition to maintain the stability of the silicon carbide CMP process. Figure 2As shown, a polishing head speed sensor 1 and a polishing head pressure sensor 2 are installed on the polishing head of the polishing machine, and a polishing pad speed sensor 3 is installed on the chassis of the polishing machine. The polishing machine is connected to the polishing fluid circulation system. The polishing fluid enters the primary recovery tank 5 through the circulation pipe 4, and is then transported by the first centrifugal pump 6 to the recovery filter 7 for treatment before entering the supply tank 8. Subsequently, it is split into two branches: one branch is transported by the second centrifugal pump 9 to the multi-parameter sensing unit 12 for component detection. The detection data is transmitted to the edge computing platform through the communication line 15, and finally sent to the wastewater tank 14 by the fourth centrifugal pump 13; the other branch returns to the polishing machine through the third centrifugal pump 10, the recovery filter 11, and the electric valve 18. The host computer receives data from the edge computing platform through the communication line 15 and sends instructions to the three-channel actuator 16. The actuator is connected to the circulation loop through the fifth centrifugal pump 17.

[0110] Reference Figure 3 As shown, a dynamic adjustment system for silicon carbide CMP polishing slurry based on multi-parameter feedback control is used to implement the detection method described above. It includes a data acquisition module, an edge computing platform, an actuator module, and a host computer system. The data acquisition module includes a pH sensor, a laser turbidity sensor, a redox potential sensor, a polishing head speed sensor, a polishing head pressure sensor, and a polishing pad speed sensor. The laser turbidity sensor detects the dispersion state of the abrasive in the polishing slurry; the pH sensor detects the pH value of the polishing slurry; the redox potential sensor detects the redox potential of the polishing slurry; the polishing head speed sensor detects the polishing head speed; the polishing head pressure sensor detects the polishing head pressure; and the polishing pad speed sensor detects the polishing pad speed. The edge computing platform preprocesses the raw data collected by the multi-source sensors. The host computer system processes the preprocessed multi-source data to generate a PDI (Process Component Analysis), and uses a loaded fuzzy PID controller to perform closed-loop feedback control of the polishing slurry components. Finally, it sends compensation commands to the three-channel actuator for compensation operation. The system's human-machine interface displays process curves and alarm logs to the operator.

[0111] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for online control of multiple parameters of silicon carbide CMP polishing slurry, characterized in that, Includes the following steps: Step 1: Preload the baseline values ​​of polishing slurry parameters from the historical process database; Step 2: Real-time data collection of the redox potential, pH value, and turbidity of the polishing slurry, as well as the polishing head pressure and rotation speed parameters of the polishing platform, is achieved through multi-source sensors. Step 3: The edge computing platform performs timestamp alignment and sliding window filtering on the multi-source sensor data, and standardizes it to a unified dimension to obtain standardized data. Step 4 involves comparing the preprocessed multi-source sensor data with baseline values ​​to calculate the Process State Deviation Index (PDI); specifically, this includes: a) Compare the standardized data output from step 3 with the benchmark value to generate the following feature deviation values: Oxidative degradation rate deviation Based on the percentage deviation from the rated oxidation degradation rate in real time; pH offset The absolute difference between the real-time pH and the target pH; Reunion Index Deviation ,real time The positive deviation relative to the nominal aggregation index; b) Calculate the process state deviation index using a dynamic weighted formula: ， in, For deviations from the process status index, For deviations in oxidative degradation rate, This is the pH offset. This is due to the reunification index deviation. This is the weighting coefficient for the oxidant activity deviation. This is the pH stability weighting coefficient. The weighting coefficients for the stability of suspended particles satisfy the following conditions: ,and Decreasing over time β(t) remains constant as time increases; Step 5: Based on the comparison between the real-time process state deviation index (PDI) and a preset threshold, output a control command, specifically including: comparing the process state deviation index (PDI) with the threshold. ① When the process state deviation index PDI > the threshold, the fuzzy PID controller generates control commands for oxidant, pH adjuster and dispersant compensation based on the process state deviation index PDI and its rate of change to control the actuator to adjust the polishing fluid parameters. ② When the process status deviates from the exponent PDI ≤ threshold, maintain the current output value of all actuators constant; ③ When the fluctuation range of the Process Status Deviation Index (PDI) exceeds the platform warning baseline value over multiple consecutive sampling periods, the alarm unit is triggered to prompt the polishing fluid to be replaced. The setting range of the warning baseline value is: , ,in is the exponential mean during the steady-state operation of the equipment, and k is the process sensitivity coefficient factor; The fuzzy PID controller is implemented using a three-level architecture, including: a fuzzification layer, a fuzzy inference decision layer, and a parameter self-tuning layer; The parameter self-tuning layer consists of a defuzzification unit, a decoupling operation unit, and a dynamic correction unit. Through multi-level processing, it transforms the fuzzy output into precise compensation instructions. Decoupling the operation unit enables multi-compensation quantity coordination: ， in, , , The output compensation vector corresponds to the compensation amounts for oxidant, pH value, and degree of aggregation, respectively; K is the decoupling matrix, which affects the PID parameter correction vector. Achieve cross compensation and satisfy: For matrix The element in the i-th row and j-th column, For predefined coupling coefficients; Dynamic correction unit with historical trend adjustment: Based on recent The process state deviation from the exponential sequence for each control cycle is used to calculate the historical trend factor. ; Historical trend analysis is performed independently for each compensation amount, and three corrected compensation amounts are output in the end. : ， in, , , The final compensation amounts are the corrected amounts, corresponding to the final compensation amounts for oxidant, pH value, and degree of aggregation, respectively. , , These correspond to the adjustment sensitivity coefficients for pre-configured oxidant, pH value, and degree of aggregation, respectively. , , These correspond to historical trend factors for oxidant, pH value, and degree of aggregation, respectively.

2. The method for online control of multiple parameters of silicon carbide CMP polishing slurry according to claim 1, characterized in that, In step 2, the multi-source sensor includes: Redox potential sensor to detect the redox potential of polishing slurry; pH sensor to detect the pH value of polishing fluid; Laser turbidity sensor to detect the dispersion state of abrasive in polishing slurry; Pressure sensor detects the pressure between the polishing head and the wafer; Polishing head speed sensor detects the rotational speed of the polishing head; Polishing pad speed sensor detects the rotational speed of the polishing pad.

3. The method for online control of multiple parameters of silicon carbide CMP polishing slurry according to claim 1, characterized in that, Step 3 includes: the edge computing platform performs timestamp synchronization on the data of each sensor, and then eliminates noise in the redox potential, pH value, polishing head rotation speed, polishing pad rotation speed and pressure value between the polishing head and the wafer through recursive average filtering, and performs normalization processing to obtain standardized data.

4. The method for online control of multiple parameters of silicon carbide CMP polishing slurry according to claim 1, characterized in that, The fuzzification layer is responsible for receiving the process state deviation index PDI and its first derivative, and converting them into fuzzy linguistic variables through triangular membership functions. Input variable definition: Deviation E = PDI - threshold, fuzzy linguistic variable set is Wherein, NB, NS, ZO, PS, and PB represent the degree of deviation of the process state: severe degradation, mild degradation, stable state, mild exceedance, and severe exceedance, respectively. Deviation change rate The fuzzy language variable set is ; The mapping from continuous universes of discourse to discrete linguistic variables is achieved using trigonometric membership functions. The expression for the trigonometric membership function is as follows: ， in, Input variable actual value For the membership degree of a specific fuzzy linguistic variable, the parameter The left boundary point of the membership function defines the starting point of the effective scope of the linguistic variable; parameters The right boundary point of the membership function defines the endpoint of the effective scope of the linguistic variable; parameters The peak point of the membership function identifies the core region to which the linguistic variable is fully subordinate; The fuzzy inference decision layer stores a two-dimensional fuzzy rule base based on two input variables and performs logical inference according to a predefined decision matrix. This fuzzy inference decision layer performs fuzzy inference operations and outputs the adjustment amount of the PID controller parameters. Its core rule structure paradigm is as follows: If E is A and EC is B, then , The number of fuzzy rules is determined by the number of linguistic values ​​of the input variables E and EC, where E represents the deviation and EC represents the rate of change of deviation. These typically constitute a set of fuzzy rules. The decision matrix of order n, where n and m represent the number of linguistic variables in E and EC, respectively; These are the adjustment amounts for proportional, integral, and derivative parameters, respectively; A is the linguistic value of the deviation E, and B is the linguistic value of the deviation change rate EC; This represents a control action vector for specific fuzzy conditions; The centroid method is used to defuzzify the unit: ， in, To precisely adjust the PID parameters that need to be solved, These correspond to the proportional term p, the integral term i, and the differential term d, respectively. This represents the total number of valid fuzzy rules actually triggered at that moment; Refers to the rule number currently being processed; This represents the activation strength of the i-th rule; This represents the original fuzzy adjustment component of the j-type parameter preset in the i-th rule.

5. The method for online control of multiple parameters of silicon carbide CMP polishing slurry according to claim 4, characterized in that, The fuzzy reasoning decision layer has a dynamic knowledge base update mechanism: Based on the system energy gradient mode Time series variance of deviation from process state index Establish rule weights that match real-time operating conditions. The applicability of the quantification rule, g(·) refers to the weight function, which is applied to the same rule group in N consecutive activations. The renormalization operator is triggered at a certain time; the activated renormalization operator automatically updates the rules according to the following feedback mechanism: ， in, The learning rate is configurable. For boundary saturation functions, This represents the actual compensation effect vector. Predict compensation vectors for the rules. For the updated rule group; By comparing the difference between the actual compensation effect and the rule prediction output, a closed-loop feedback loop is formed to achieve online adaptive correction of the rule base parameters.

6. The method for online control of multiple parameters of silicon carbide CMP polishing slurry according to claim 5, characterized in that, The historical trend factor Based on sliding window construction: ， Where i is the variable used for traversal within the window, and t is the current reference point. This is the trend balance coefficient. Here, n is the width of the dynamic sliding window, and n is the sign function. This is the current PDI value. This is a historical benchmark value. This is the mean of the PDI window.

7. An adaptive control system for silicon carbide CMP polishing slurry based on multi-source data fusion, used to implement the control method as described in any one of claims 1-6, characterized in that, include: The data acquisition module includes an oxidation-reduction potential sensor, a pH sensor, a laser turbidity sensor, a pressure sensor, a polishing head speed sensor, and a polishing pad speed sensor. The oxidation-reduction potential sensor detects the oxidation-reduction potential of the polishing fluid. A pH sensor detects the pH value of the polishing slurry; a laser turbidity sensor detects the dispersion state of the abrasive in the polishing slurry; a pressure sensor detects the pressure between the polishing head and the wafer; a polishing head speed sensor detects the speed of the polishing head; and a polishing pad speed sensor detects the speed of the polishing pad. The edge computing platform is communicatively connected to the data acquisition module, which performs time alignment and recursive averaging filtering on the raw data acquired by the data acquisition module to eliminate high-frequency noise. Normalize the data for each parameter; The process state deviation index is calculated by comparing the preprocessed data with the baseline value. The actuator module includes an oxidant metering pump, a dispersant feeding device, and a pH adjuster feeding device, which are used to automatically adjust the composition of the polishing slurry according to control commands to achieve dynamic component control of the polishing slurry. The host computer system includes an adaptive control module with a three-layer fuzzy controller architecture, which processes the pre-processed process state deviation index (PDI), generates polishing fluid state assessment results and control instructions through threshold comparison, and feeds them back to the operator through the human-machine interface.

Images