A method and system for end point detection of a high hardness brittle material polishing process

Abstract

The application belongs to the field of mechanical processing, and discloses an endpoint detection method and system for polishing of high-hard-brittle materials, which comprises the following steps: step 1, constructing a mathematical model of chemical mechanical polishing; step 2, building a test platform to collect force and torque signals; step 3, designing a circuit of an edge computing platform; step 4, performing a parameter control experiment; step 5, developing a host computer system; step 6, finally obtaining a training set of a high-hard-brittle material processing state prediction model based on force and torque signals; step 7, training the processing state prediction model; and step 8, using the trained processing state prediction model to perform endpoint detection. The application uses polishing torque to perform online real-time detection on the polishing state and uses a state feedback method, so that the polishing processing efficiency can be improved, the polishing quality and the polishing yield can be improved, and the cost can be reduced.

Description

Technical Field

[0001] The present invention relates to the field of mechanical processing, and in particular to an endpoint detection method and system for a polishing process of a highly hard and brittle material. Background Art

[0002] Highly hard and brittle materials, represented by silicon carbide, have extremely superior physical properties. The current industrial field has an increasing demand for the processing of highly hard and brittle materials, especially in the fields of aviation technology, semiconductor chips, and energy. High-precision processing requirements and greater processing difficulty. However, this also means that its processing requirements are extremely high and difficult. Taking silicon carbide as an example, when processing it into a semiconductor substrate, its surface must be ultra-precise, defect-free, and damage-free, and the surface roughness must reach the nanometer level. The characteristics of silicon carbide materials include high hardness, stable chemical properties, higher compressive strength than bending strength, and greater hardness and brittleness. These characteristics increase the difficulty of processing.

[0003] During the polishing process, the material is in full contact with the polishing pad, making endpoint detection challenging. Traditionally, this approach has relied on estimating the polishing time required for a specific product through experimentation, then empirically controlling the removal rate and uniformity based on the polishing rate and polishing thickness. However, variations in the material's metal pattern, buildup ratio, and process can lead to variations in polishing removal volume and time. Furthermore, factors such as polishing pad wear, temperature fluctuations, and material pressure distribution contribute to the inability of traditional methods to effectively address the issues of over- and under-polishing. Summary of the Invention

[0004] The purpose of the present invention is to provide an endpoint detection method and system for the polishing process of high-hardness and brittle materials to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for detecting an endpoint of a polishing process of a highly hard and brittle material, comprising:

[0007] Step 1, constructing a mathematical model of chemical mechanical polishing;

[0008] Step 2: Build a test platform to collect force and torque signals;

[0009] Step 3: Design the circuit of the edge computing platform;

[0010] Step 4: Conduct parameter control experiments to control the polishing liquid concentration, polishing liquid temperature, polishing speed, polishing time, and adjust the polishing pressure and polishing torque to obtain corresponding parameter data. At the same time, remove the workpiece at regular intervals to test its MRR and Ra; at the same time, establish a state space observation system to obtain the current polishing friction coefficient;

[0011] Step 5: Develop the host computer system to display the experimental data obtained previously and the friction coefficient obtained by the state observation system, and design the user operation interface to provide the corresponding user operation interface;

[0012] Step 6: Establish the relationship between the force and torque signals and the processing state of the hard and brittle material based on the experimental data obtained above. Normalize the collected force and torque signals, standardize the sample data, and convert all data into values ​​between [0, 1]. The normalization method used is the maximum and minimum normalization method:

[0013]

[0014] Among them, x represents a data element in the data set, x k Indicates the normalized value of the data element, x max Indicates the maximum value of the data set in the data set, x min The minimum value of this data set in this data set;

[0015] Finally, a training set of the high-hardness and brittle material processing state prediction model based on force and torque signals is obtained;

[0016] Step 7: Using the normalized data set to train the processing state prediction model;

[0017] Step 8: Use the trained processing state prediction model to perform endpoint detection, where the reference basis of the endpoint detection standard of the processing state prediction model is the material removal rate MRR and the surface roughness Ra. The polishing endpoint is reached when Ra is less than the standard threshold and MRR is within the set threshold range.

[0018] Furthermore, the step 1 specifically includes:

[0019] Assuming that the oxidant concentration and abrasive in the polishing solution are sufficient and evenly distributed, the abrasive is evenly distributed on the SiC substrate and polishing pad surface. The mathematical model of the material removal rate (MRR) per unit time of SiC is expressed as:

[0020] MRR=N a (RR) V

[0021] Among them, (RR) V represents the material removal rate of a single abrasive. Since the abrasive is assumed to be uniformly distributed in the polishing solution, the amount of abrasive Na in the actual contact area Ar between the polishing pad and SiC can be expressed as:

[0022]

[0023] Wherein, A represents the actual contact area, χ represents the volume concentration of abrasive particles in the polishing solution, D represents the average particle size of the abrasive, assuming that the actual pressure on the polishing pad surface is P, then A r Can be expressed as:

[0024]

[0025] Wherein, A0 represents the contact area between the polishing pad and the workpiece, E p , E w , v p , v w The Young's modulus of the polishing pad, the Young's modulus of the SiC substrate, the Poisson's ratio of the polishing pad, the Poisson's ratio of the SiC substrate, C is a constant, σ, ρ respectively the standard deviation of the roughness height distribution of the polishing pad and the average radius of curvature of the roughness peak of the polishing;

[0026] Because in the chemical mechanical polishing process, the abrasive particles slide on the surface with a relative speed V, and the material removal is realized after being pressed into the workpiece, the material removal rate (RR) V Is expressed as:

[0027]

[0028] Wherein β and γ respectively represent the probability of chemical reaction on the substrate surface and the probability of surface oxide layer being removed by mechanical effect, 0≤β≤1, 0≤γ≤1; d m Is the thickness of the oxide layer, δ w Indicates the cutting depth of the abrasive on the SiC substrate surface;

[0029] The final material removal rate MRR formula per unit time is:

[0030]

[0031] In addition, the physical relationship between torque, pressure and friction force in the polishing process is:

[0032]

[0033] Wherein, τ is the polishing torque, P is the actual pressure on the polishing pad surface, K1 is the correction coefficient, μ is the surface friction coefficient, r is the radius of the SiC substrate, and here it is assumed that the friction coefficient μ is linearly related to the surface roughness Ra, so the size of Ra is:

[0034]

[0035] Wherein, k is the slope after K1 correction, and b is a constant term.

[0036] Further, the step 2 specifically comprises:

[0037] Since the filtering circuit uses a notch filter, it is necessary to sample the force and torque signal information and then perform spectrum analysis. The spectrum analysis method mainly uses Fourier transform, fast Fourier transform and wavelet transform to analyze the noise information that regularly appears during the processing and signal acquisition process for subsequent filtering circuit design.

[0038] Furthermore, in step 4, MRR is measured using a high-resolution electronic scale to measure the weight loss of silicon carbide before and after polishing, and is calculated based on the processing time and processing radius. Ra is measured using a high-precision Taylor Hobson white light interferometer.

[0039] Furthermore, the step 7 further includes:

[0040] The neural network model used in the processing state prediction model is PINN. According to the MRR mathematical model established in step 1, MRR=N a (RR) V and the Ra mathematical model Choose two objective functions as:

[0041]

[0042]

[0043] The input values ​​are P and τ, the output values ​​are MRR and Ra, and the remaining parameters are considered constants;

[0044] For the real measurement value MRR true and Ra true and the observed MRR of the neural network pred and Ra pred , the data error term DataLoss is expressed as:

[0045]

[0046] The physical information error term physicsLoss is expressed as:

[0047]

[0048] The total loss function is the weighted sum of the data error term and the physical information error term:

[0049] Total Loss=α·Data Loss+β·Physics Loss

[0050] Among them, α and β are weight parameters of data error and physical information error, respectively, and appropriate values ​​can be selected based on experience;

[0051] Finally, the network weights are adjusted using an optimization algorithm to train the processing state prediction model and minimize the overall loss function.

[0052] The present invention also provides an endpoint detection system for the polishing process of highly hard and brittle materials, which is used to implement the endpoint detection method described above. The system includes an upper transmission shaft, a conductive slip ring, a force and torque sensor, a wafer carrier, a polishing pad, and a lower polishing plate. The upper transmission shaft, the force and torque sensor, and the wafer carrier are arranged in sequence from top to bottom. The conductive slip ring is sleeved on the upper transmission shaft. The wafer carrier and the lower polishing plate are arranged up and down. The polishing workpiece is installed at the bottom of the wafer carrier. The polishing pad is arranged at the upper end of the lower polishing plate. The upper transmission shaft is driven to rotate by a first motor, and the lower polishing plate is driven to rotate by a second motor. The conductive slip ring is used to connect a signal filtering and transmission module.

[0053] Furthermore, the signal filtering and transmission module includes a notch filter circuit, a wireless transmission device, and an edge computing module equipped with a differential tracker. The signal filtering and transmission module filters the acquired signal and uploads it to the computer through the wireless transmission device. The computer processes the collected signal and then monitors the polishing status, and displays the data through the host computer software.

[0054] Compared with the prior art, the present invention has the following beneficial effects:

[0055] Compared to traditional methods that require stopping processing, repeatedly removing the workpiece, inspecting the workpiece's processing results, and evaluating whether rework is necessary, traditional methods are inefficient for polishing highly hard and brittle materials, affecting polishing quality and assembly accuracy, resulting in low yields. The present invention utilizes polishing torque to perform real-time online detection of the polishing status and provides status feedback, thereby improving polishing efficiency, enhancing polishing quality and yield, and reducing costs.

[0056] Compared to traditional methods that analyze the current processing state based on mathematical formulas and personal experience based on the physical quantities measured by sensors before proceeding to the next step, this present invention utilizes machine learning and deep learning methods to propose a polishing process state prediction model based on torque signals. Using this prediction model, during the polishing process, by sampling the torque and friction signals during the process, the polishing state is predicted and the processing parameters are adjusted in real time, eliminating the need for manual intervention or personal experience to select the appropriate processing parameters. This method implements a torque signal-based polishing state prediction method for highly hard and brittle materials. While maintaining high-precision processing state identification, this present invention reduces the sampling requirements of the equipment and ensures control stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flow chart of an endpoint detection method for a polishing process of a highly hard and brittle material according to the present invention;

[0058] Figure 2 This is a signal processing flow chart of an endpoint detection method for a polishing process of a high-hardness and brittle material according to the present invention;

[0059] Figure 3 This is a workflow diagram of a processing state prediction model in an endpoint detection method for a high-hardness and brittle material polishing process according to the present invention.

[0060] Figure 4 The diagram is a structural diagram of an endpoint detection system for a polishing process of a highly hard and brittle material according to the present invention.

[0061] In the figure: 1 is the upper transmission shaft, 2 is the conductive slip ring and signal filtering and transmission module, 3 is the force and torque sensor, 4 is the wafer carrier, 5 is the polishing workpiece, 6 is the polishing pad, and 7 is the lower polishing plate. DETAILED DESCRIPTION

[0062] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0063] See also Figure 1-Figure 3 , a method for detecting the endpoint of a polishing process of a highly hard and brittle material, comprising:

[0064] Step 1: Construct a mathematical model for chemical mechanical polishing. This method assumes that the oxidant concentration and abrasive in the polishing solution are sufficient and evenly distributed. Therefore, the abrasive is evenly distributed on the SiC substrate and polishing pad surface. The mathematical model for the material removal rate (MRR) per unit time of SiC can be expressed as:

[0065] MRR=N a (RR) V

[0066] Among them, (RR) V represents the material removal rate of a single abrasive. Since the abrasive is assumed to be uniformly distributed in the polishing solution, the amount of abrasive Na in the actual contact area Ar between the polishing pad and SiC can be expressed as:

[0067]

[0068] Among them, A r Indicates the actual contact area. Considering that the surface of the polishing pad is distributed with many micro protrusions and the surface is relatively rough. Therefore, the actual contact area A between the SiC substrate and the polishing pad is rSmaller than the surface area of ​​the substrate, assuming the actual pressure on the polishing pad surface is P, then A r It can be expressed as:

[0069]

[0070] Where A0 represents the contact area between the polishing pad and the workpiece, E p 、E w 、v p 、v w are Young’s modulus of the polishing pad, Young’s modulus of the SiC substrate, Poisson’s ratio of the polishing pad, and Poisson’s ratio of the SiC substrate, respectively. C is a constant. σ and ρ are the standard deviation of the roughness height distribution of the polishing pad and the average curvature radius of the polished roughness peak, respectively.

[0071] During the chemical mechanical polishing process, the abrasive particles slide on the surface at a relative speed V and are pressed into the workpiece to remove material. The material removal rate (RR) of a single abrasive is V It can be expressed as an equation:

[0072]

[0073] Where β and γ represent the probability of chemical reaction on the substrate surface and the probability of the surface oxide layer being removed by mechanical effect, respectively, 0≤β≤1, 0≤γ≤1; d m is the thickness of the oxide layer, δ w Indicates the cutting depth of the abrasive on the SiC substrate surface.

[0074] The final formula for the material removal rate per unit time (MRR) is:

[0075]

[0076] Furthermore, the physical relationship between torque, pressure, and friction during polishing is:

[0077]

[0078] Where τ is the polishing torque, P is the actual pressure on the polishing pad surface, K1 is the correction factor, and μ is the surface friction coefficient. Due to the large number of interference factors, a state observer is required to implement the specific μ function.

[0079] Here it is assumed that the friction coefficient μ is linearly related to the surface roughness Ra, so the value of Ra is:

[0080]

[0081] Where k is the slope corrected by K1, and b is the constant term.

[0082] Step 2: Build a test platform to collect force and torque signals. Due to the notch filter used in the filtering circuit, the force and torque signal information needs to be sampled and then subjected to spectrum analysis. The spectrum analysis method mainly uses Fourier transform, fast Fourier transform and wavelet transform to analyze the noise information that appears regularly during the processing and signal acquisition process.

[0083] Step 3: Design the circuitry for the edge computing platform. This primarily includes a 50Hz twin-T notch filter circuit, a single-chip microcontroller module for implementing the differential tracker, a power supply regulator module, a computing chip, and a wireless communication module. These modules are responsible for powering the circuit board and sensors, performing ADC conversion on the signals, and uploading the data, respectively. This circuit design is well-known and will not be detailed here.

[0084] Step 4: Conduct parameter control experiments. Since the main factors affecting polishing effect and efficiency include parameters such as polishing liquid concentration, polishing liquid temperature, polishing torque, polishing pressure, polishing speed, and polishing time, the polishing liquid concentration, polishing liquid temperature, polishing speed, and polishing time are controlled, and the polishing pressure and polishing torque are adjusted to conduct experiments to obtain corresponding parameter data. At the same time, the workpiece is removed at regular intervals to test its MRR and Ra. The MRR is measured using a high-resolution electronic scale (within 0.1 mg) to measure the weight loss of silicon carbide before and after polishing and is calculated based on the processing time and processing radius. The Ra is measured using a high-precision Taylor Hobson white light interferometer. At the same time, a state-space observation system is established to obtain the friction coefficient of the current polishing.

[0085] Step 5: Develop the host computer system to display the experimental data obtained previously and the friction coefficient obtained by the state observation system. At the same time, design the user operation interface and provide the corresponding user operation interface, mainly including the start and stop of the polishing equipment, end point detection manual shutdown and recovery interfaces, etc.

[0086] Step 6: Establish the relationship between the force and torque signals and the processing state of the hard and brittle material based on the experimental data obtained above, and perform data normalization on the collected force and torque signals. In order to improve the convergence speed and generalization of the neural network during the training process, it is necessary to first normalize the sample data and convert all data into values ​​between [0, 1]. The normalization method used in this invention is the maximum and minimum normalization method:

[0087]

[0088] Among them, x represents a data element in the data set, x k Indicates the normalized value of the data element, x max Indicates the maximum value of the data set in the data set, x min The minimum value of this data set in this dataset.

[0089] Finally, a training set of the high-hardness and brittle material processing state prediction model based on force and torque signals is obtained.

[0090] In step 7, the processing state prediction model is trained using the normalized data set. The neural network model used is PINN. PINN is a machine learning model that combines deep learning and physics knowledge. It uses physical laws to guide the model during the learning process, which can greatly improve the model's generalization ability, especially when there is less data or more noise. Based on the mathematical model established in step 1, two objective functions are selected:

[0091]

[0092]

[0093] The input values ​​are P and τ, the output values ​​are MRR and Ra, and the remaining parameters are considered constants.

[0094] The data error term measures the difference between the predicted output of the neural network and the actual observed data. For a given input P and τ, there are two target outputs MRR and Ra. For the actual measured value MRR true and Ra true and the observed MRR of the neural network pred and Ra pred , the data error term DataLoss can be expressed as:

[0095]

[0096] The physical information error term physicsLoss can be expressed as:

[0097]

[0098] The total loss function is the weighted sum of the data error term and the physical information error term:

[0099] Total Loss=α·Data Loss+β·Physics Loss

[0100] Among them, α and β are the weight parameters of data error and physical information error respectively, and appropriate values ​​can be selected based on experience.

[0101] Finally, gradient descent or other conventional optimization algorithms are used to adjust the network weights and train the processing state prediction model to minimize the overall loss function.

[0102] Step 8: Use the trained processing state prediction model to perform endpoint detection, where the reference basis of the endpoint detection standard of the processing state prediction model is the material removal rate MRR and the surface roughness Ra. The polishing endpoint is reached when Ra is less than the standard threshold and MRR is within the set threshold range.

[0103] The signal processing flow is as follows: Figure 2 As shown in the figure, after filtering by notch filter and differential tracker, the signal is processed by Fourier transform, fast Fourier transform and wavelet transform, and the change of μ is observed by state observer. At the same time, the processed information is used as the model training data set, and finally the high hardness and brittle material processing state prediction model is obtained.

[0104] Regarding the use of the prediction model for the processing state of high hardness and brittle materials during polishing, such as Figure 3 As shown in the figure, after polishing starts, real-time force and torque signals are collected. After passing through the signal filtering circuit, the data is uploaded to the computer for data analysis and processing, and enters the processing state prediction model. The processing is stopped after the polishing end point is detected. If the processing end point is not reached, the polishing state detection is carried out. If the MRR value decreases significantly and the Ra value increases significantly, the processing is stopped. If no abnormal processing state is detected, the polishing process continues.

[0105] like Figure 4 As shown, the present invention also provides an endpoint detection system for the polishing process of highly hard and brittle materials, which is used to implement the endpoint detection method described above, comprising an upper transmission shaft 1, a conductive slip ring 2, a force and torque sensor 3, a wafer carrier 4, a polishing pad 6, a lower polishing plate 7, an external power supply, and a computer. The upper transmission shaft 1, the force and torque sensor 3, and the wafer carrier 4 are arranged in sequence from top to bottom. The conductive slip ring 2 is mounted on the upper transmission shaft 1, and the wafer carrier 4 and the lower polishing plate 7 are arranged vertically. A polishing workpiece 5 is mounted on the bottom of the wafer carrier 4, and a polishing pad 6 is mounted on the top of the lower polishing plate 7. The upper transmission shaft 1 is driven to rotate by a first motor, and the lower polishing plate is driven to rotate by a second motor. The conductive slip ring 2 is connected to a signal filtering and transmission module.

[0106] Specifically, a polishing pad 6 is placed on the upper surface of the lower polishing plate 7 for polishing. A force and torque sensor 3 is installed between the wafer carrier 4 and the upper transmission shaft 1 to obtain axial torque and pressure signals. A conductive slip ring 2 is installed above the torque sensor 3 and outside the upper transmission shaft 1. The conductive slip ring 2 is used to connect the signal filtering and transmission module and the power module. The signal filtering and transmission module contains a notch filter circuit, a wireless transmission device, and an edge computing module equipped with a differential tracker. The acquired signal is filtered and uploaded to the computer via the wireless transmission device. The computer monitors the polishing status after signal processing, and the data is displayed by the host computer software.

[0107] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the endpoint of a polishing process of a highly hard and brittle material, characterized in that: include: Step 1, constructing a mathematical model of chemical mechanical polishing; Step 2: Build a test platform to collect force and torque signals; Step 3: Design the circuit of the edge computing platform; Step 4: Conduct parameter control experiments to control the polishing liquid concentration, polishing liquid temperature, polishing speed, polishing time, and adjust the polishing pressure and polishing torque to obtain corresponding parameter data. At the same time, remove the workpiece at regular intervals to test its MRR and Ra; at the same time, establish a state space observation system to obtain the current polishing friction coefficient; Step 5: Develop the host computer system to display the experimental data obtained previously and the friction coefficient obtained by the state observation system, and design the user operation interface to provide the corresponding user operation interface; Step 6: Establish the relationship between the force and torque signals and the processing state of the hard and brittle material based on the experimental data obtained above. Normalize the collected force and torque signals, standardize the sample data, and convert all data into values ​​between [0, 1]. The normalization method used is the maximum and minimum normalization method: Among them, x represents a data element in the data set, x k Indicates the normalized value of the data element, x max Indicates the maximum value of the data set in the data set, x min The minimum value of this data set in this data set; Finally, a training set of the high-hardness and brittle material processing state prediction model based on force and torque signals is obtained; Step 7: Using the normalized data set to train the processing state prediction model; Step 8: Use the trained processing state prediction model to perform endpoint detection, where the reference basis of the endpoint detection standard of the processing state prediction model is the material removal rate MRR and the surface roughness Ra. The polishing endpoint is reached when Ra is less than the standard threshold and MRR is within the set threshold range.

2. The endpoint detection method for a polishing process of a high-hardness and brittle material according to claim 1, characterized in that: The step 1 specifically includes: Assuming that the oxidant concentration and abrasive in the polishing solution are sufficient and evenly distributed, the abrasive is evenly distributed on the SiC substrate and polishing pad surface. The mathematical model of the material removal rate (MRR) per unit time of SiC is expressed as: MRR=N a (RR) V Among them, (RR) V represents the material removal rate of a single abrasive particle. Since the abrasive is assumed to be uniformly distributed in the polishing solution, the amount of abrasive Na in the actual contact area Ar between the polishing pad and SiC can be expressed as; Among them, A r represents the actual contact area, X represents the volume concentration of abrasive particles in the polishing solution, D represents the average particle size of the abrasive, and assuming that the actual pressure on the polishing pad surface is P, then A r It can be expressed as: Where A0 represents the contact area between the polishing pad and the workpiece, E p 、E w 、v p 、v w are the Young's modulus of the polishing pad, the Young's modulus of the SiC substrate, the Poisson's ratio of the polishing pad, and the Poisson's ratio of the SiC substrate, respectively. C is a constant. σ and ρ are the standard deviation of the roughness height distribution of the polishing pad and the average curvature radius of the polished roughness peak, respectively. During the chemical mechanical polishing process, the abrasive particles slide on the surface at a relative speed V and are pressed into the workpiece to remove material. The material removal rate (RR) of a single abrasive is V Expressed as: Where β and γ represent the probability of chemical reaction on the substrate surface and the probability of the surface oxide layer being removed by mechanical effect, respectively, 0≤β≤1, 0≤γ≤1; d m is the thickness of the oxide layer, δ w Indicates the cutting depth of the abrasive on the surface of the SiC substrate; The final formula for the material removal rate MRR per unit time is: Furthermore, the physical relationship between torque, pressure, and friction during polishing is: Where τ is the polishing torque, P is the actual pressure on the polishing pad surface, K1 is the correction factor, μ is the surface friction coefficient, and r is the radius of the SiC substrate. It is assumed here that the friction coefficient μ is linearly related to the surface roughness Ra, so the size of Ra is: Where k is the slope corrected by K1, and b is the constant term.

3. The endpoint detection method for a polishing process of a high-hardness and brittle material according to claim 1, characterized in that: The step 2 specifically includes: Since the filtering circuit uses a notch filter, it is necessary to sample the force and torque signal information and then perform spectrum analysis. The spectrum analysis method uses Fourier transform, fast Fourier transform and wavelet transform to analyze the noise information that regularly appears during the processing and signal acquisition process for subsequent filtering circuit design.

4. The endpoint detection method for a polishing process of a high-hardness and brittle material according to claim 1, characterized in that: In step 4, MRR is measured by using a high-resolution electronic scale to measure the weight loss of silicon carbide before and after polishing, and is calculated based on the processing time and processing radius. Ra is measured using a high-precision Taylor Hobson white light interferometer.

5. The endpoint detection method for a polishing process of a high-hardness and brittle material according to claim 2, characterized in that: The step 7 further comprises: The neural network model used in the processing state prediction model is PINN. According to the MRR mathematical model established in step 1, MRR=N a (RR) V and the Ra mathematical model Choose two objective functions as: The input values ​​are P and τ, the output values ​​are MRR and Ra, and the remaining parameters are considered constants; For the real measurement value MRR true and Ra true and the observed MRR of the neural network pred and Ra pred , the data error term DataLoss is expressed as: The physical information error term physicsLoss is expressed as: The total loss function is the weighted sum of the data error term and the physical information error term: Total Loss=α·Data Loss+β·Physics Loss Where α and β are the weight parameters of data error and physical information error respectively; Finally, the network weights are adjusted using an optimization algorithm to train the processing state prediction model and minimize the overall loss function.

6. An endpoint detection system for a polishing process of a hard and brittle material, used to implement the endpoint detection method according to any one of claims 1 to 5, characterized in that: It includes an upper transmission shaft, a conductive slip ring, a force and torque sensor, a wafer carrier, a polishing pad, and a lower polishing plate. The upper transmission shaft, the force and torque sensor, and the wafer carrier are arranged in sequence from top to bottom. The conductive slip ring is sleeved on the upper transmission shaft. The wafer carrier and the lower polishing plate are arranged up and down. The polishing workpiece is installed at the bottom of the wafer carrier, and the polishing pad is set at the upper end of the lower polishing plate. The upper transmission shaft is driven to rotate by a first motor, and the lower polishing plate is driven to rotate by a second motor. The conductive slip ring is used to connect the signal filtering and transmission module.

7. The endpoint detection system for the polishing process of highly hard and brittle materials according to claim 6, characterized in that: The signal filtering and transmission module includes a notch filter circuit, a wireless transmission device, and an edge computing module equipped with a differential tracker. The signal filtering and transmission module filters the acquired signal and uploads it to the computer through the wireless transmission device. The computer processes the collected signal and then monitors the polishing status, while displaying the data through the host computer software.

Images