Method for predicting and compensating edge breakage defect in wafer cutting process

Abstract

The invention relates to the technical field of semiconductor manufacturing, and discloses an edge breakage defect prediction and compensation method in a wafer cutting process, and the method comprises the steps: obtaining a wafer cutting initial scheme and historical cutting data, extracting edge breakage key parameters, defining an edge vulnerability function, and building a high-risk region space mapping model; according to the real-time cutting position, multi-dimensional data are collected through a three-level monitoring mode, the abnormal deviation state of the track is judged, and an edge breakage risk assessment model is constructed and early warning is carried out; a compensation scheme is formulated after early warning is triggered, cutting is implemented, and the effect is verified; according to the method, multi-dimensional data are fused, the risk area is dynamically delimited, self-adaptive adjustment of the cutting parameters is achieved, and the wafer cutting yield is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of semiconductor manufacturing technology, and more particularly to a method for predicting and compensating edge chipping defects during wafer cutting. Background Art

[0002] In semiconductor manufacturing, wafer dicing is a critical process for separating wafers into individual chips, and its accuracy directly impacts chip yield. However, chip edge chipping caused by grinding wheel wear, wafer stress concentration, and wafer path deviation during dicing remains a technical challenge that hinders improving yield.

[0003] The Chinese patent with publication number CN119115243A proposes a method for correction and early warning of cuts during wafer cutting, and a wafer cutting device. The method calculates the cutting line deviation and corrects it through the Mark template image, and detects the position and width of the cut to identify anomalies. However, this existing technology mainly relies on a single data source of visual inspection, and only performs static correction before cutting. It lacks real-time monitoring of dynamic parameters such as grinding wheel wear and wafer edge stress distribution during the cutting process, and cannot quantify the coupling risk of trajectory deviation and material fragility. For example, when the grinding wheel wear causes a sudden change in cutting force or the wafer produces micro-displacement due to unstable clamping, this existing technology cannot evaluate the risk of edge collapse in real time through multi-dimensional data fusion, and the warning lag may lead to untimely compensation.

[0004] The Chinese patent application with publication number CN117495840A provides a method, equipment and medium for detecting poor cutting of products after wafer cleavage, and realizes quantitative analysis of edge collapse defects through image segmentation and defect recognition. However, this prior art belongs to post-detection, and it is impossible to dynamically predict the risk of edge collapse and implement compensation during the cutting process. Its core limitation is that it only detects the edge collapse defects that have occurred, and does not establish a real-time correlation model between cutting parameters (such as cutting speed, cutting angle) and edge collapse risk, and cannot solve the core problem of "prediction lag". For example, when cutting highly brittle materials, the edge stress concentration caused by improper cutting angle of the grinding wheel needs to be adjusted in real time during the cutting process, and the detection mechanism of this prior art cannot intervene in process control, resulting in insufficient defect prevention capabilities.

[0005] The key defect of existing technologies is the lack of real-time fusion analysis of multi-dimensional data such as grinding wheel wear, wafer stress distribution, and trajectory deviation, which makes it impossible to accurately locate high-risk areas with dynamic changes; the compensation strategy relies on historical experience or static correction, which makes it difficult to deal with nonlinear risks caused by equipment vibration and material property differences during the cutting process. Summary of the Invention

[0006] This invention is suitable for the precision cutting of brittle materials such as silicon wafers and gallium arsenide in semiconductor manufacturing. It can be integrated into fully automatic wafer saws and is suitable for cutting wafers of various sizes, such as 8-inch and 12-inch wafers. It is particularly suitable for high-end chip manufacturing that is highly sensitive to edge chipping defects, such as the cutting process of power semiconductor and MEMS sensor wafers, effectively solving the problem of edge chipping risk control under complex working conditions.

[0007] In order to overcome the above-mentioned defects of the prior art, the present invention provides a method for predicting and compensating edge chipping defects during wafer cutting. By defining an edge fragility function to dynamically delineate high-risk areas, and combining a three-level monitoring mode to collect data such as cutting force and edge images in real time, a risk assessment model that includes trajectory deviation and regional fragility is constructed to achieve accurate prediction of edge chipping risks. After the early warning is triggered, parameters such as cutting speed, trajectory micro-displacement, and cutting angle are adaptively adjusted, and the compensation effect is ensured through closed-loop verification. This solution significantly improves the real-time prediction of edge chipping defects and the accuracy of the compensation strategy, effectively reduces the edge chipping rate during wafer cutting, and ensures the stability of the yield of high-end chip manufacturing.

[0008] To achieve the above object, the present invention provides the following technical solutions:

[0009] The prediction and compensation methods for edge chipping defects during wafer dicing include:

[0010] Obtain the initial wafer cutting plan and historical cutting data, extract the key parameters when edge chipping defects occur in the historical cutting data, define the edge fragility function F(x,y) based on the key parameters when edge chipping defects occur, and establish a high-risk area spatial mapping model based on the edge fragility function F(x,y); obtain the real-time cutting position, and based on the high-risk area spatial mapping model, adopt a three-level monitoring mode to collect multi-dimensional cutting process data in real time and determine the abnormal trajectory deviation status; construct a wafer edge chipping risk assessment model, and issue a wafer edge chipping warning based on the wafer edge chipping risk assessment model and the results of the abnormal trajectory deviation status determination;

[0011] After the wafer edge collapse warning is triggered, a compensation plan is formulated for the initial wafer cutting plan; after compensating the initial wafer cutting plan according to the compensation plan, a compensated wafer cutting plan is obtained; according to the compensated wafer cutting plan, wafer cutting is carried out to verify the compensation effect.

[0012] Furthermore, the wafer cutting initial plan includes at least wafer clamping pre-pressure F0; the key parameters when edge chipping defects occur in the historical cutting data include at least grinding wheel wear W(t) and wafer edge stress distribution S(x, y).

[0013] Furthermore, the method further comprises:

[0014] In the no-load test, the potential response signal U(t) of the wafer clamping device is synchronously monitored to determine the potential response signal reference value U0;

[0015] Whether the clamping device is slightly vibrating is determined based on the potential response signal U(t) and the potential response signal reference value U0. If it is slightly vibrating, the pre-pressure compensation algorithm is triggered to generate a corrected clamping force F1.

[0016] Furthermore, the method for generating the corrected clamping force F1 is: according to the wafer clamping pre-pressure F0, the potential response signal U(t) and the potential response signal reference value U0, the corrected clamping force F1 is obtained by a pre-pressure compensation algorithm.

[0017] Furthermore, the method for defining the edge fragility function F(x,y) is: linearly superimposing the grinding wheel wear W(t) and the wafer stress S(x,y) to define the edge fragility function F(x,y).

[0018] Furthermore, the establishment of a high-risk area spatial mapping model based on the edge vulnerability function F(x,y) includes:

[0019] Generate a two-dimensional fragility heat map of the wafer based on the edge fragility function;

[0020] According to the two-dimensional vulnerability heat map of the wafer, the wafer is divided into three risk areas; the three risk areas of the wafer include red area, orange area and green area.

[0021] Furthermore, the method for generating a two-dimensional fragility heat map of a wafer includes:

[0022] The wafer surface is divided into n2 grid cells, and the coordinates of the center point of each grid cell are (x j' ,y j' ), calculate the edge fragility value F(x j' ,y j' ), according to the edge fragility value F(x j' ,y j' ), generate a two-dimensional fragility heat map of the wafer; where (x j' ,y j' ) is the coordinate of the center point of the j'th grid cell, F(x j' ,y j' ) is the edge vulnerability value of the j'th grid cell, 1≤j'≤n2.

[0023] Furthermore, the three-level monitoring mode is:

[0024] If the real-time cutting position is within the red area, the first monitoring mode is activated;

[0025] If the real-time cutting position is within the orange area, the second monitoring mode is activated;

[0026] If the real-time cutting position is within the green area, the third monitoring mode is maintained;

[0027] The first monitoring mode is a high-frequency monitoring mode, the second monitoring mode is a medium-frequency monitoring mode, and the third monitoring mode is a low-frequency monitoring mode.

[0028] Furthermore, the multi-dimensional cutting process data includes cutting force data and wafer edge images;

[0029] The method for determining the abnormal trajectory deviation state is as follows:

[0030] Perform edge detection on wafer edge images during the cutting process and extract real-time edge contour coordinate sequences;

[0031] Calculate the deviation vector ΔQ between the real-time edge profile coordinate sequence and the theoretical edge profile coordinate sequence;

[0032] Calculate the angle between the real-time cutting path and the theoretical cutting path

[0033] According to the angle and the deviation vector ΔQ to determine the abnormal trajectory deviation state.

[0034] Furthermore, the angle and the deviation vector ΔQ, the determination of the abnormal trajectory deviation state includes:

[0035] When n1 consecutive sampling When |ΔQ| increases, it is determined to be an abnormal trajectory deviation state; is the acute angle threshold, |ΔQ| is the modulus of the deviation vector ΔQ;

[0036] The method for performing wafer edge collapse warning is as follows: obtaining a wafer edge collapse risk prediction value according to a wafer edge collapse risk assessment model; triggering a wafer edge collapse warning when the wafer edge collapse risk prediction value exceeds a wafer edge collapse risk threshold, or when it is determined that the trajectory is in an abnormal deviation state.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] The present invention defines the edge fragility function and constructs a spatial mapping model for high-risk areas by integrating historical cutting data with real-time multi-dimensional process data (such as grinding wheel wear, wafer edge stress distribution, cutting force data and edge images, etc.), thereby overcoming the limitations of traditional methods that rely on single sensor data and achieving accurate analysis of the coupling effect between the dynamic deviation of the cutting trajectory and material properties. The three-level monitoring mode dynamically adjusts the monitoring frequency according to the real-time cutting position, focuses on high-risk areas, and improves data acquisition efficiency and abnormal capture sensitivity. The wafer edge collapse risk assessment model integrates multi-dimensional parameters such as regional fragility, trajectory deviation, and contour deviation, and combines it with the judgment of abnormal trajectory offset status to form a composite early warning mechanism of "status assessment + trend prediction", which effectively solves the problem of prediction lag. After the early warning is triggered, the adaptive compensation scheme for parameters such as cutting speed, trajectory micro-displacement, and entry angle responds to nonlinear risks such as grinding wheel wear and wafer positioning micro-offset in real time, and combines closed-loop control for compensation effect verification to achieve precise regulation of the entire process from defect prediction to dynamic compensation. The overall solution significantly improves the accuracy of edge collapse defect prediction and the real-time performance of the compensation strategy, effectively reducing the risk of edge collapse during wafer cutting and ensuring the stability and reliability of the cutting yield. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 This is a flow chart showing the principle of the method for predicting and compensating edge chipping defects during wafer cutting in the present invention;

[0041] Figure 2 Flowchart of a method for obtaining a corrected clamping force F1 in the method for predicting and compensating edge chipping defects during wafer dicing of the present invention;

[0042] Figure 3 A flow chart of a method for defining an edge chipping sensitive area R in a method for predicting and compensating edge chipping defects during wafer dicing of the present invention;

[0043] Figure 4 A flow chart of a method for establishing a high-risk area spatial mapping model and obtaining a high-risk area spatial mapping matrix M(x, y) in the method for predicting and compensating edge chipping defects during wafer dicing of the present invention;

[0044] Figure 5A flow chart of a method for constructing a wafer edge chipping risk assessment model and performing wafer edge chipping early warning in the method for predicting and compensating edge chipping defects during wafer dicing of the present invention;

[0045] Figure 6 This is a functional module diagram of the system for predicting and compensating edge chipping defects during wafer cutting in the present invention. DETAILED DESCRIPTION

[0046] 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.

[0047] Example 1

[0048] See also Figure 1 As shown, this embodiment provides a method for predicting and compensating edge chipping defects during wafer dicing, including:

[0049] Step S1000: Obtain the initial wafer cutting plan and historical cutting data, extract the key parameters when the edge chipping defect occurs in the historical cutting data, define the edge fragility function F(x,y) based on the key parameters when the edge chipping defect occurs, and establish a high-risk area spatial mapping model based on the edge fragility function F(x,y); obtain the real-time cutting position, and based on the high-risk area spatial mapping model, adopt a three-level monitoring mode to collect multi-dimensional cutting process data in real time and perform trajectory abnormal deviation status determination; construct a wafer edge chipping risk assessment model, and perform wafer edge chipping warning based on the wafer edge chipping risk assessment model and the results of the trajectory abnormal deviation status determination;

[0050] Step S1000 addresses the defects of existing technologies that rely on single sensor data and lack coupled analysis of trajectory and material properties. It collects multi-dimensional data such as cutting force and edge image in real time through a three-level monitoring mode, and combines the chipping sensitive area and edge fragility function defined by historical data to construct a wafer chipping risk assessment model that includes trajectory benchmark, regional fragility, and contour deviation.

[0051] Furthermore, step S1000 includes:

[0052] Step S1100: Obtain an initial plan for wafer cutting, control the grinding wheel to move along a preset cutting path without load, and obtain an initial trajectory deviation E0 based on the actual cutting path and the theoretical cutting path;

[0053] Step S1100 is a pre-calibration step in the wafer cutting process. It aims to solve the trajectory deviation problem caused by inherent errors of the equipment and unclear initial cutting state through systematic pre-calibration and initial parameter setting, thereby suppressing the occurrence of edge chipping defects from the root.

[0054] Furthermore, step S1100 includes:

[0055] Step S1110, obtaining an initial wafer cutting plan, including a first cutting speed v0, a grinding wheel cutting angle θ0, and a wafer clamping pre-pressure F0;

[0056] Specifically, the first cutting speed v0 is defined as the preset speed of the grinding wheel spindle in the lateral direction (X-axis), measured in millimeters per second (mm / s). This parameter is set based on the wafer material (e.g., silicon, gallium arsenide), grinding wheel type (resin / diamond grinding wheel), and cutting depth. For example, for a 500μm thick silicon wafer using a diamond grinding wheel, v0 can be set to 10 mm / s. The grinding wheel engagement angle θ0 is the initial angle between the grinding wheel axis and the tangent direction of the wafer surface. It adjusts the contact area between the grinding wheel and the wafer, thereby controlling the thermal stress distribution during the cutting process. If θ0 is too small, the contact area between the grinding wheel and the wafer increases, potentially leading to localized stress concentration. If θ0 is too large, the cutting action of the grinding wheel edge is weakened, potentially causing vibration. The wafer clamping preload F0 is the static clamping force applied by the wafer fixture (e.g., vacuum chuck or mechanical clamp) before cutting, measured in Newtons (N). This parameter is determined by wafer size (e.g., 8-inch or 12-inch) and thickness.

[0057] The precise setting of the initial cutting parameters is the primary condition for solving the problem of wafer edge chipping. If v0 is too high, the frictional heat between the grinding wheel and the wafer accumulates too quickly, causing local thermal stress to exceed the fracture strength of the material; if θ0 deviates from the reasonable range, the contact mechanical state between the grinding wheel and the wafer is unbalanced, which can easily cause edge cracks; if F0 is insufficient, the wafer will undergo micro-displacement during the cutting process, causing the actual cutting path to deviate from the theoretical trajectory. By setting the initial values ​​of v0, θ0 and F0, a benchmark reference is provided for subsequent real-time monitoring and compensation. For example, in step S1210, the stability judgment of the potential response signal U(t) needs to use F0 as the clamping force benchmark; in step S1710, the dynamic adjustment of the cutting speed needs to use v0 as the compensation starting point. The standardized setting of the initial parameters provides a unified comparison benchmark for the subsequent collection and analysis of multi-dimensional data (such as cutting force and trajectory deviation), avoiding misjudgment due to differences in the initial state.

[0058] Step S1120 , establishing a three-dimensional coordinate system O-XYZ with the wafer geometric center as the origin O, where the X axis is the wafer radial direction, the Y axis is the cutting feed direction, and the Z axis is the wafer thickness direction;

[0059] Specifically, the X-axis (wafer radial direction) is along the wafer radius direction and is used to characterize the radial deviation of the lateral movement trajectory of the grinding wheel. The Y-axis (cutting feed direction) is parallel to the grinding wheel spindle feed direction and is used to monitor the longitudinal position of the grinding wheel along the wafer cutting path. The Z-axis (wafer thickness direction) is perpendicular to the wafer surface and is used to quantify the change in the grinding wheel cutting depth. The positioning accuracy of the coordinate system origin O directly affects the accuracy of the trajectory deviation calculation. The geometric center of the wafer is determined by an optical positioning system (such as a CCD vision sensor). During the wafer cutting process, the actual motion trajectory of the grinding wheel must be strictly matched with the theoretical path in three-dimensional space. If the coordinate system is not defined accurately, it will lead to calculation errors in the initial trajectory deviation E0, which will affect the reliability of the subsequent edge collapse risk assessment model. The unified three-dimensional coordinate system provides a mathematical basis for the spatial association of multidimensional data. In step S1130, the grinding wheel trajectory coordinates {P i The chipping-sensitive region R in step S1400 is defined based on the spatial division of the coordinate system. Furthermore, this coordinate system aligns with the coordinate reference of the wafer edge image analysis system (e.g., edge detection algorithm) in step S1720, ensuring that the calculated profile deviation vector ΔQ (step S1730) can be directly used to determine trajectory deviation.

[0060] Step S1130: Control the grinding wheel to move along the preset cutting path without load, and collect the grinding wheel edge track coordinate sequence {P i (x i ,y i ,z i )}; where (x i ,y i ,z i ) is the coordinate of the i-th grinding wheel edge trajectory point;

[0061] Specifically, the laser ranging array consists of multiple high-precision laser displacement sensors, which are evenly distributed around the wafer cutting area in a circular layout. During the no-load motion phase, the grinding wheel moves along a preset path at a speed of v0. The laser ranging array captures the position of the grinding wheel edge (diamond abrasive area) in three-dimensional space in real time and generates a discrete trajectory coordinate sequence {P i The trajectory data from the no-load motion phase is used to calibrate the mechanical errors of the grinding wheel motion mechanism (such as screw backlash and guide rail straightness error). If loaded cutting is performed directly, the interaction between the wafer and the grinding wheel will mask the inherent errors of the equipment, resulting in the initial trajectory deviation E0 in step S1150 not being able to truly reflect the equipment's accuracy. No-load pre-calibration can be used to separate the inherent errors of the equipment from dynamic interference factors during the cutting process.

[0062] Step S1140, calculate the grinding wheel edge trajectory coordinate sequence {Pi (x i ,y i ,z i )} and the three-dimensional deviation matrix D = [δx, δy, δz] of the theoretical cutting path, where δx, δy, δz are the trajectory deviations in the X, Y, and Z axis directions respectively; i represents the i-th measured point in the grinding wheel edge trajectory, (x i ,y i ,z i ) is P i coordinates of

[0063] Specifically, the theoretical cutting path is generated by the geometric parameters (such as cutting path spacing and angle) in the wafer cutting plan and stored as a discrete coordinate sequence {Q j (x j ,y j ,z j )}. For each measured point {P i (x i ,y i ,z i )}, match the corresponding theoretical point Q through the nearest neighbor algorithm j , calculate each axial deviation: Q j represents the jth point in the theoretical cutting path, (x j ,y j ,z j ) represents Q j Coordinates of δx = x i -x j , reflecting the radial position error of the grinding wheel on the wafer; δy = y i -y j , reflecting the trajectory lag or advance in the feed direction; δz=z i -z j , reflecting the fluctuation of the grinding wheel cutting depth. The dimension of the three-dimensional deviation matrix D is the same as {P i}, for example, when the number of sampling points is 1000, D is a 1000×3 matrix.

[0064] The actual motion trajectory of the grinding wheel is affected by factors such as equipment mechanical errors and control system delays, resulting in slight deviations from the theoretical path. If these deviations are not quantified, the trajectory deviation will couple with the wafer stress distribution during loaded cutting, leading to edge chipping defects. For example, when δx is continuously positive, the grinding wheel will deviate toward the outer edge of the wafer, potentially resulting in incomplete material removal at the edge of the cut line and the formation of microcracks.

[0065] Step S1150 : generating an initial trajectory deviation E0 according to the three-dimensional deviation matrix D.

[0066] Specifically, the initial trajectory deviation, E0, is calculated using the Euclidean norm formula based on the three-dimensional deviation matrix D = [δx, δy, δz] generated in step S1140. E0 reflects the degree of trajectory deviation during the no-load phase of the grinding wheel due to factors such as mechanical clearance in the equipment and return errors in the transmission system. The vector modulus in three-dimensional space quantifies the overall trajectory deviation during no-load motion, providing a baseline reference value for real-time trajectory monitoring during the subsequent cutting process.

[0067] Taking the cutting of an 8-inch silicon wafer with a diamond grinding wheel as an example, the trajectory point coordinate sequence and deviation calculation collected during the no-load motion phase are shown in the following table:

[0068]

[0069]

[0070] Calculated based on normalized deviation E0 integrates the X / Y / Z deviations to comprehensively reflect the grinding wheel's spatial position error. For example, when the grinding wheel spindle tilts, causing an increase in δz, even if δx and δy are within the limits, E0 will still increase, triggering equipment calibration to avoid edge stress concentration caused by tilted cutting. Step S1150 establishes a quantitative benchmark for no-load trajectory accuracy by calculating the modulus of three-dimensional deviations, resolving the problem of lagging edge collapse prediction caused by the confusion between inherent equipment errors and process disturbances.

[0071] Step S1200 , correcting and compensating the wafer clamping pre-pressure F0 in the initial wafer cutting plan to obtain a corrected clamping force F1;

[0072] The core goal of step S1200 is to solve the problem of insufficient or unstable clamping force caused by micro-vibration of the clamping device during the dynamic cutting process when the initial setting value F0 of the wafer clamping pre-pressure is set. Since the mechanical structure of the wafer clamping device may produce micron-level displacement vibrations due to internal transmission clearance, external environmental disturbances (such as equipment vibration, airflow fluctuations) or wafer surface morphology differences (such as warping, uneven thickness) in an unloaded or loaded state, this micro-vibration will weaken the clamping device's fixing effect on the wafer, causing the wafer to slightly shift during the cutting process, thereby causing cutting trajectory deviation or local stress concentration, and ultimately increasing the risk of edge collapse. Step S1200 dynamically monitors the potential response signal U(t) of the clamping device, determines the clamping state in real time, and corrects the clamping force, to ensure the mechanical fixation stability of the wafer throughout the cutting process.

[0073] See also Figure 2 As shown, further, step S1200 includes:

[0074] Step S1210, in the no-load test, synchronously monitoring the potential response signal U(t) of the wafer clamping device to determine the potential response signal reference value U0;

[0075] Specifically, during the no-load test phase, the potential response signal U(t) of the contact interface between the clamping device and the wafer is collected in real time through the micro piezoelectric sensor array integrated inside the wafer clamping device. The physical essence of the potential response signal U(t) is a quantitative representation of the dynamic changes in the interface charge distribution caused by the contact friction between the clamping device and the wafer surface. When the clamping device causes uneven contact pressure distribution due to mechanical vibration or micro-slip between the wafer and the fixture, the charge distribution at the contact interface will fluctuate nonlinearly, thereby causing the time domain characteristics of U(t) to change. The piezoelectric sensor array is composed of piezoelectric ceramic sensing units distributed in a ring shape, which can cover the contact pressure distribution changes from the edge to the center area of ​​the wafer.

[0076] The potential response signal baseline value, U0, is determined as follows: During the no-load precalibration phase, the clamping device is controlled to clamp the wafer with an initial preload of F0, while maintaining the cutting platform in a stationary state (no grinding wheel motion or external excitation). The U(t) signal is collected for 10 seconds in this state. After removing the transient response data for the first 2 seconds, a sliding average is performed on the remaining 8 seconds of U(t) data, and the averaged steady-state value is taken as U0. This baseline value represents the electrical characteristic calibration value of the clamping system in an ideal stable state, providing a reference benchmark for subsequent micro-vibration detection.

[0077] Step S1220 , determining whether the clamping device is slightly vibrating based on the potential response signal U(t) and the potential response signal reference value U0 , and if so, triggering the pre-pressure compensation algorithm to generate a corrected clamping force F1 .

[0078] The method for determining whether the clamping device is micro-vibrating based on the potential response signal U(t) is: when |U(t)-U0|>μU0, it is determined that the clamping device is micro-vibrating; otherwise, it is not micro-vibrating, where μ is a preset potential fluctuation coefficient.

[0079] The method for generating the corrected clamping force F1 is as follows: according to the wafer clamping pre-pressure F0, the potential response signal U(t) and the potential response signal reference value U0, the corrected clamping force F1 is obtained by a pre-pressure compensation algorithm.

[0080] Specifically, the micro-vibration judgment logic is implemented by calculating the dynamic deviation between U(t) and U0 in real time. The micro-vibration judgment formula is: |U(t)-U0|>μU0, where μ is the preset potential fluctuation coefficient, and its value range is 0.01≤μ≤0.05 (set according to the brittleness of the wafer material and the sensitivity of the clamping device). If the above inequality holds, it indicates that the clamping device has micro-vibrations that exceed the allowable range. The traditional clamping force setting relies only on the static calibration value F0, and does not take into account the dynamic attenuation of the clamping force caused by mechanical wear, temperature changes or differences in wafer surface morphology during actual operation. For example, when the transmission screw of the clamping device has a gap due to long-term use, the actual clamping force during the closing process may be lower than F0, but static calibration cannot detect such dynamic deviations. By monitoring the degree of deviation of U(t) relative to U0, the mechanical state abnormality of the clamping device can be perceived in real time.

[0081] The pre-pressure compensation algorithm adopts a proportional-integral (PI) control strategy to dynamically adjust the clamping force according to the deviation of U(t) relative to U0. The proportional-integral (PI) control strategy contains proportional and integral terms. The proportional term is used to quickly respond to the instantaneous fluctuation of U(t), and the integral term is used to eliminate steady-state errors. Traditional compensation methods only adjust the clamping force based on historical statistics and cannot respond to transient disturbances in real time. For example, when cutting highly brittle materials (such as gallium arsenide), the contact impact between the grinding wheel and the wafer may cause high-frequency micro-vibration of the clamping device. Such transient disturbances need to be quickly suppressed by the proportional term, while long-term drifts (such as sensor zero point offset caused by temperature) need to be corrected by the integral term.

[0082] Step S1220 addresses the issue of dynamic attenuation of the clamping force due to mechanical wear, environmental disturbances, or wafer morphology variations by monitoring the clamping device's potential response signal in real time. The PI compensation algorithm employed suppresses both transient disturbances and steady-state drift, ensuring the wafer remains stably clamped during the cutting process. This, combined with the initial parameter setting in step S1100, the cutting force monitoring in step S1700, and the compensation verification in step S3000, forms a closed-loop control loop, significantly reducing the risk of edge collapse caused by unstable clamping and improving the cutting yield.

[0083] Step S1300: Obtain historical cutting data and extract key parameters when edge chipping occurs in the historical cutting data. The key parameter sequence when edge chipping occurs includes grinding wheel wear W(t), wafer edge stress distribution S(x, y) and edge chipping location coordinates (x b ,y b );

[0084] Specifically, the technical purpose of step S1300 is to establish a correlation model between edge chipping defects and multi-dimensional process parameters through a systematic analysis of historical cutting data, thereby providing data support for the subsequent delineation of dynamic sensitive areas. In the wafer cutting process, the generation of edge chipping defects has significant nonlinear characteristics, and its cause involves the coupling of multiple factors such as dynamic wear of the grinding wheel, anisotropy of the wafer material, and fluctuations in the clamping force. Traditional methods rely solely on instantaneous data from a single sensor and are unable to capture the potential correlation between historical defect patterns and the current process status.

[0085] Grinding wheel wear is defined as the cumulative loss of material at the edge of the grinding wheel during the cutting process, and its quantitative expression is: Where D0 is the initial diameter of the grinding wheel, and D(t) is the actual diameter at time t. A laser ranging sensor array mounted on the grinding wheel spindle collects real-time grinding wheel outer diameter data. This parameter characterizes the impact of grinding wheel sharpness degradation on cutting force. When W(t) exceeds 1%, cutting efficiency at the wheel edge decreases significantly, which can lead to stress concentration at the wafer edge.

[0086] The wafer edge stress distribution S(x,y) is acquired by attaching a thin film piezoelectric sensor array to the wafer edge area. During the cutting process, each sensor outputs the local stress value in real time to form the stress distribution matrix S(x,y). b ,y b ) Scan the edge of the cut wafer using a high-resolution optical microscope (resolution 0.1μm) and use image processing algorithms to extract the geometric center coordinates of the edge chipping defect. The specific algorithm includes:

[0087] Edge contour extraction: Use the Canny operator to detect the wafer edge contour and generate a binary image;

[0088] Defect area segmentation: Based on the morphological dilation operation, the chipped edge area is separated from the normal edge;

[0089] Centroid calculation: locate the centroid of the edge collapse area after segmentation and obtain (x b ,y b For example, after a certain cut, the pixel coordinates of the chipping area are detected to be (1200, 800). Through pixel-to-physical coordinate conversion, the actual coordinates are obtained as (12.0mm, 8.0mm). This parameter is used to construct a spatial distribution model of the chipping-sensitive area.

[0090] This step solves the problem of single dimension and lack of spatiotemporal correlation in traditional methods of edge chipping defect analysis through multi-dimensional analysis of historical data. The grinding wheel wear W(t) is directly related to the edge fragility function in step S1500, providing a basis for dynamic weight allocation; the grinding wheel wear W(t) and the wafer edge stress distribution S(x,y) are directly related to the edge fragility function in step S1500; the coordinates of the edge chipping location (x b ,y b ) provides a spatial benchmark for the sensitive area delineation in step S1400.

[0091] Step S1400 , defining an edge chipping sensitive region R based on key parameters when an edge chipping defect occurs;

[0092] See also Figure 3 As shown, further, step S1400 includes:

[0093] Step S1410: The edge collapse position coordinates (x b ,y b ) as the center, construct a square edge collapse sensitive area R with a side length of 2a; where a is the side length parameter;

[0094] Step S1420: The vertex coordinates of the square edge chipping sensitive area R are (x b -a,y b -a)、(x b -a,y b +a)、(x b +a,yb+a)、(x b +a,yb-a).

[0095] Specifically, step S1400 converts the spatial distribution characteristics of historical edge collapse events into a sensitive area model that can be monitored in real time, thereby achieving dynamic focusing of risk areas during the cutting process. By delineating spatially sensitive areas, the problem of evenly distributed monitoring resources and inability to focus on high-risk areas in traditional methods is solved. The method for determining the edge length parameter a is: a = k·σ x' , where σ x' is the standard deviation of the x' coordinate of the historical chipping position, and k is the coverage factor (usually k = 3, covering 99.7% of the normal distribution data). For example, the statistics of a certain type of wafer show σ x' =1.2mm, then a=3.6mm, and the side length of the sensitive area 2a=7.2mm. b ,y b ) is converted into the mechanical coordinate system of the actual cutting platform. Assuming that the geometric center of the wafer is O(0,0), and the X / Y axis direction of the cutting platform is consistent with the wafer coordinate system, the vertex coordinates of the sensitive area are: (x b-a,y b -a)、(x b -a,y b +a)、(x b +a,y b +a)、(x b +a,y b -a).

[0096] In traditional wafer dicing processes, edge chipping defect monitoring typically uses a global uniform sampling strategy, which makes it impossible to dynamically adjust monitoring focus based on historical defect distribution and real-time process status. This leads to two core problems:

[0097] Waste of monitoring resources: Excessive data collection in low-risk areas, while insufficient sampling in high-risk areas leads to missed defects.

[0098] Lack of spatial correlation: The spatial distribution characteristics of historical chipping are not combined with the real-time cutting trajectory, making it impossible to predict the chain chipping caused by grinding wheel path deviation.

[0099] Step S1400 uses kernel density estimation to generate a defect distribution heat map based on the historical edge collapse coordinates extracted in step S1300, and identifies the spatial density peak area; with the density peak point as the center, the side length 2a of the sensitive area is determined to form a dynamic monitoring boundary; combined with the three-dimensional coordinate system in step S1120 and the grinding wheel trajectory data in step S1130, the theoretical sensitive area is mapped to the actual cutting platform to eliminate the influence of clamping errors.

[0100] Step S1500 , defining an edge fragility function F(x,y) based on key parameters when edge chipping defects occur;

[0101] The edge fragility function F(x,y) is defined based on the key parameters when the edge collapse defect occurs, including:

[0102] Based on the grinding wheel wear W(t) and the wafer edge stress distribution S(x,y), the edge fragility function F(x,y) is defined.

[0103] Specifically, this step aims to establish a dynamic vulnerability evaluation system for the wafer edge area. By quantifying the coupling effect of grinding wheel wear and wafer stress, it solves the problem that single sensor data cannot reflect the dynamic interaction between material properties and equipment status. Grinding wheel wear can lead to cutting edge blunting, increase cutting resistance and induce stress concentration at the wafer edge. The stress distribution S(x,y) is strongly correlated with the wafer material properties (such as the anisotropy of silicon single crystals) and the cutting path geometric parameters (such as cutting depth and grinding wheel penetration angle). The edge vulnerability function F(x,y) is defined by linearly superimposing the grinding wheel wear W(t) and the wafer stress S(x,y).

[0104] Step S1600: Based on the edge vulnerability function, a high-risk area spatial mapping model is established to obtain a high-risk area spatial mapping matrix M(x, y);

[0105] See also Figure 4 As shown, further, step S1600 includes:

[0106] Step S1610 , generating a two-dimensional fragility heat map of the wafer according to the edge fragility function;

[0107] The method for generating a two-dimensional fragility heat map of a wafer includes: dividing the wafer surface into n2 grid units, and the coordinates of the center point of each grid unit are (x j' ,y j' ), calculate the edge fragility value F(x j' ,y j' ), according to the edge fragility value F(x j' ,y j' ), generate a two-dimensional fragility heat map of the wafer; where (x j' ,y j' ) is the coordinate of the center point of the j'th grid cell, F(x j' ,y j' ) is the edge vulnerability value of the j'th grid cell, 1≤j'≤n2.

[0108] Step S1620: Divide the wafer into three risk zones according to the two-dimensional fragility heat map of the wafer; the three risk zones include a red zone, an orange zone, and a green zone;

[0109] Red area: F(x j' ,y j' )>F th +σ F , where σ F is the historical standard deviation, F th is the vulnerability threshold;

[0110] Orange area: F th ≤F(x j' ,y j' )≤F th +σ F ;

[0111] Green area: F(x j' ,y j' ) <F th .

[0112] Step S1630 : spatially superimpose the red area and the edge chipping sensitive area R to generate a spatial mapping matrix M(x, y).

[0113] Specifically, traditional methods rely on single sensor data (such as cutting force or image) to statically delineate risk areas. These methods fail to dynamically couple multidimensional factors such as grinding wheel wear, material stress distribution, and trajectory deviation, leading to delayed risk prediction. Historical statistical compensation strategies fail to distinguish between differences in mechanical properties across different regions, resulting in crude adjustments to cutting parameters (such as global speed reduction) and an inability to accurately suppress localized edge chipping. The coupled effect of grinding wheel wear and wafer microdisplacement can trigger sudden edge chipping, but existing models lack a risk spatial weighting mechanism, making it difficult to trigger timely graded warnings. Step S1610 aims to transform the edge fragility function F(x,y) defined in step S1500 into a visual spatial distribution model to address the existing issues of fuzzy positioning of edge chipping risk areas and the inability to accurately quantify the risk level at each location. The wafer surface is discretized into square regions of equal area for spatial quantitative analysis. The grid side length is determined based on the wafer size and the characteristic size of the edge chipping defect, ensuring that the defect distribution covers at least two to three grids to avoid missed detections. During the cutting process, grinding wheel wear leads to a nonlinear increase in cutting force, while stress concentration at the wafer edge reduces the material's crack resistance. By rasterizing the vulnerability map, the edge collapse sensitive area R defined in step S1400 is combined with the dynamically changing device state parameters (W(t) and S(x,y)) to achieve spatially refined positioning of the risk area.

[0114] Step S1620 solves the problem that the risk level classification in the prior art is rigid and cannot reflect the statistical distribution characteristics. th and the historical standard deviation σ F , wafers are divided into three levels: red (high risk), orange (medium risk), and green (low risk). The specific judgment rules are:

[0115] Red area: F(x j' ,y j' )>F th +σ F ;

[0116] Orange area: F th ≤F(x j' ,y j' )≤F th +σ F ;

[0117] Green area: F(x j' ,y j' ) <F th .

[0118] F th Determined based on the 95% quantile of historical edge collapse data. F It is the standard deviation of F(x,y) in historical data, reflecting the range of vulnerability fluctuation. FUsed to dynamically adjust the threshold bandwidth. When the wafer material batch is changed (such as switching from silicon to gallium arsenide), σ F The threshold value will be recalculated based on the initial cutting data of the new material to ensure its adaptability. F Instead of a fixed threshold, the risk area division can be adapted to different process conditions. The three-level division is linked with the three-level monitoring mode in step S1710 to concentrate high-frequency monitoring resources in the red area, reducing the overall data volume and lowering the real-time computing load.

[0119] Step S1630 aims to integrate the static sensitive area (R in step S1400) and the dynamic risk area (the red area in step S1620) to solve the problem of missed detection caused by a single area definition. The generation rule of the spatial mapping matrix M(x,y) is:

[0120] If the grid belongs to both the red area and the edge collapse sensitive area R, then M(x j' ,y j' )=1.5.

[0121] If the grid belongs only to the red area or R, then M(x j' ,y j' )=1.2.

[0122] Other areas M(x j' ,y j' )=1.0.

[0123] The gradient weights of the matrix M(x,y) (1.5>1.2>1.0) guide the compensation system to prioritize high-risk grids. For example, when resources are limited, active cooling is only implemented in areas where M≥1.2, improving cooling efficiency.

[0124] Step S1600 quantifies grinding wheel wear, stress distribution, and trajectory deviation into spatial weights, resolving the issue of isolated parameters in traditional models. Regionally differentiated processing based on M(x,y) transforms the compensation strategy from a one-size-fits-all approach to a regionally tailored approach, improving yield. Through matrix weighting and dynamic threshold adjustment, sudden edge collapse is effectively suppressed, reducing the rate of missed early warnings.

[0125] Step S1700: Based on the high-risk area spatial mapping model, a three-level monitoring mode is adopted to collect multi-dimensional cutting process data in real time to determine the abnormal trajectory deviation state; the multi-dimensional cutting process data includes cutting force data and wafer edge images;

[0126] Furthermore, step S1700 includes:

[0127] Step S1710, obtain the real-time cutting position, activate the three-level monitoring mode according to the real-time cutting position, and collect cutting force data; the cutting force data includes the X-axis component Fx , Y-axis component F y and the Z-axis component F z ;

[0128] The three-level monitoring mode is:

[0129] If the real-time cutting position is within the red area, the first monitoring mode is activated;

[0130] If the real-time cutting position is within the orange area, the second monitoring mode is activated;

[0131] If the real-time cutting position is within the green area, the third monitoring mode is maintained.

[0132] Step S1720 , obtaining an image of the wafer edge during the cutting process, and extracting a real-time edge contour coordinate sequence using an edge detection algorithm;

[0133] Step S1730 , calculating the deviation vector ΔQ between the real-time edge contour coordinate sequence and the theoretical edge contour coordinate sequence;

[0134] Step S1740: Calculate the angle between the real-time cutting path and the theoretical cutting path

[0135] Step S1750, according to the angle and the deviation vector ΔQ to determine the abnormal trajectory deviation state.

[0136] When n1 consecutive sampling When |ΔQ| increases, it is determined to be an abnormal trajectory deviation state; is the acute angle threshold, and |ΔQ| is the modulus of the deviation vector ΔQ.

[0137] Specifically, when the real-time cutting position enters the high-risk red zone, the high-frequency monitoring mode, i.e., the first monitoring mode, is activated, the cutting force data sampling frequency is increased to 1kHz, and the X, Y, and Z axis cutting force components (F x ,F y ,F z In the medium-risk orange area, the medium-frequency monitoring mode (500Hz), i.e. the second monitoring mode, is adopted; in the low-risk green area, the baseline monitoring mode (100Hz), i.e. the third monitoring mode, is maintained. x It is the cutting resistance of the grinding wheel in the radial direction of the wafer (X axis), reflecting the lateral contact stress between the grinding wheel and the edge of the wafer. y It is the thrust resistance of the grinding wheel along the cutting feed direction (Y axis), which is positively correlated with the wear of the grinding wheel. zIt is the vertical pressure of the grinding wheel in the thickness direction of the wafer (Z axis), and its abnormal fluctuation may cause micro cracks on the surface of the wafer. The three-axis force signal is acquired in real time by the six-axis force sensor integrated in the grinding wheel spindle, and is denoised and normalized by the signal conditioning circuit. Traditional fixed-frequency sampling is prone to miss transient abnormal signals (such as instantaneous slippage of the grinding wheel) in high-risk areas, while high-frequency sampling in low-risk areas leads to data redundancy. This step dynamically adapts the sampling frequency through a three-level mode to balance data accuracy and computing resource consumption. Analyze a certain axial force (such as F x ) cannot distinguish whether the trajectory deviation is due to grinding wheel wear or insufficient clamping force. The synchronous acquisition of the three-axis force components provides a multi-dimensional correlation basis for subsequent deviation judgment. x ,F y ,F z ) can be used for collaborative analysis, for example, when F x Abnormally elevated and F y When it decreases, it can be determined as the lateral deviation of the grinding wheel (X-axis trajectory deviation); if F z If the value of the scalar is increased simultaneously, it may be accompanied by wafer warpage (step S2210, determining if the cut is too deep). This collaborative analysis improves the accuracy of determining the cause of trajectory deviation. Step S1710 dynamically switches the monitoring mode based on the spatial mapping matrix M(x,y), focusing data collection on high-risk areas.

[0138] A high-resolution linear array CCD camera is positioned along the wafer cutting path to capture grayscale images of the wafer edge area. The original image is Gaussian filtered to remove noise, and the Canny operator is used to extract the wafer edge contour, generating a binary image. The edge pixel points are converted into a coordinate sequence using a chain code tracking algorithm (Freeman chain code) to obtain a real-time edge contour coordinate sequence. The theoretical edge contour coordinate sequence is extracted from the wafer design file. For each real-time contour point, the nearest neighbor point is searched in the theoretical contour sequence, the deviation component is calculated, and the deviation components of all points are counted to generate an overall deviation vector. Traditional methods use the maximum deviation value as the judgment basis, which can easily misjudge local material defects as trajectory deviations. Step S1730 reflects the overall deviation trend by averaging the deviation vector. The tens of thousands of contour points are reduced to a two-dimensional vector ΔQ to reduce the complexity of subsequent calculations. If a region has a local protrusion due to wafer impurities, but the deviations in other regions are normal, the modulus of ΔQ remains within a safe range, avoiding false triggering of warnings.

[0139] In step S1740, the grinding wheel position P at the current time t is obtained according to the grinding wheel spindle encoder data. t (x t ,y t ) and the position P at the previous moment t-Δt t-Δt (x t-Δt ,y t-Δt), calculate the grinding wheel motion direction vector Extracting theoretical grinding wheel motion direction vector from initial cutting plan where v 0x and v 0y The X and Y axis cutting velocity components defined in step S1110. Angle calculation Traditional methods only monitor displacement deviations and are unable to identify the angle between the grinding wheel's motion direction and the theoretical path (such as tangential offset caused by wheel side slip). Traditional methods only monitor displacement deviations and are unable to identify the angle between the grinding wheel's motion direction and the theoretical path (such as tangential offset caused by wheel side slip). When the grinding wheel slides sideways due to vibration, even if the displacement deviation ΔQ is within the limit, the angle φ can promptly indicate directional anomalies.

[0140] In step S1750, when n1 (e.g., 5) consecutive samplings are performed, When |ΔQ| increases, the trajectory is considered to be in an abnormal deviation state. A single instantaneous anomaly may be caused by noise, and the continuous determination mechanism can effectively reduce the false alarm rate. A monotonically increasing |ΔQ| indicates a worsening deviation trend, requiring early warning to avoid irreversible damage.

[0141] Step S1700, through a three-level monitoring model, multidimensional data fusion, and dynamic decision logic, addresses the inaccuracy and response lag issues inherent in traditional trajectory deviation detection methods. This, in synergy with other steps in this embodiment (such as risk area delineation and compensation parameter generation), forms a closed-loop control system, improving the accuracy of edge chipping defect prediction and compensation, meeting the process requirements for high-yield wafer dicing.

[0142] Step S1800 , constructing a wafer edge chipping risk assessment model to provide wafer edge chipping warning.

[0143] See also Figure 5 As shown, further, step S1800 includes:

[0144] Step S1810 , measuring the distance variation rate between the grinding wheel and the wafer edge in the red and orange areas, and constructing a regional enhancement item of the wafer edge chipping risk assessment model based on the distance variation rate;

[0145] Step S1820 , constructing a trajectory reference item of a wafer edge chipping risk assessment model based on the initial trajectory deviation E0;

[0146] Step S1830 , constructing a regional vulnerability term using the spatial mapping matrix M(x, y) and the edge vulnerability function F(x, y);

[0147] Step S1840 , constructing a profile deviation term from the deviation vector ΔQ;

[0148] Step S1850 , constructing a wafer edge chipping risk assessment model based on the region enhancement item, the trajectory reference item, the region vulnerability item, and the profile deviation item;

[0149] Step S1860, obtaining a wafer edge chipping risk prediction value according to a wafer edge chipping risk assessment model;

[0150] Step S1870 , when the wafer edge chipping risk prediction value exceeds the wafer edge chipping risk threshold, or when it is determined that the trajectory is in an abnormal deviation state, a wafer edge chipping warning is triggered.

[0151] Specifically, step S1800 addresses the following technical issues: the existing technology relies on single sensor data for edge chipping defect prediction and cannot respond to dynamic cutting parameter changes in real time. This is achieved by establishing a multi-dimensional coupled wafer edge chipping risk assessment model.

[0152] The separation of historical data and real-time dynamic parameters: Traditional methods are based only on static historical statistics and cannot integrate real-time variables such as grinding wheel wear and wafer positioning offset during the cutting process;

[0153] Coarse-grained regional risk assessment: Risk levels are not divided according to the material properties and stress distribution differences in different regions of the wafer, resulting in insufficient warning sensitivity;

[0154] Lack of quantification of nonlinear risk factors: Dynamic factors such as grinding wheel trajectory deviation and cutting force mutation have not formed a coupled analysis model with material fragility, resulting in a lag in risk threshold determination.

[0155] In step S1810, a laser interferometer is used to monitor the distance D(t) between the edge of the grinding wheel and the wafer cutting path in real time. The sampling frequency is synchronized with the three-level monitoring mode of step S1700 to calculate the macro change rate. Where Δt is the sampling interval (Δt=1ms in the red area and Δt=2ms in the orange area). Normalized to a dimensionless parameter and multiplied by the regional weight coefficient η (η = 0.3 in the red area and η = 0.15 in the orange area) to obtain the regional enhancement term The micro-distance variation rate reflects the instantaneous rate of change of the distance between the grinding wheel and the wafer edge. A positive value indicates an increase in the distance (grinding wheel retreats), and a negative value indicates a decrease in the distance (grinding wheel intrudes). In the red / orange area (such as the 5mm range from the wafer edge), the wafer material is more brittle and the grinding wheel wear is aggravated. The micro-distance variation rate can capture the track jitter caused by grinding wheel vibration or thermal deformation. For example, when the grinding wheel has radial runout due to wear, Periodic negative fluctuations trigger an increase in the weight of regional enhancements, prioritizing the model's response to dynamic anomalies in high-risk areas. Through high-frequency sampling and regional weighting, the contribution of micro-variation rates to edge collapse risk is specifically increased in red and orange areas. This reduces risk identification and response time compared to traditional, uniform monitoring of the entire area.

[0156] In step S1820, the initial trajectory deviation E0 represents the impact of the inherent mechanical error of the equipment (such as guide rail straightness and spindle radial runout) on the trajectory accuracy, and the installation error is eliminated by no-load pre-calibration. E0 is divided by the maximum trajectory deviation allowed by the equipment. max , and obtain the dimensionless reference term. After the equipment is cold started or shut down for a long time, the thermal deformation of the mechanical structure may cause the initial trajectory deviation to increase. By incorporating E0 into the model, the contribution of the equipment inherent error and the dynamic offset during the cutting process can be distinguished, avoiding the misjudgment of the equipment inherent error as a real-time risk. The reference term separates the static error of the equipment from the dynamic risk, allowing the model to focus on real-time controllable factors (such as cutting speed and clamping force). Cooperating with the no-load pre-calibration in step S1100, ensure that E0 is updated before each cutting to adapt to changes in equipment status (such as trajectory deviation correction after grinding wheel replacement).

[0157] Based on the spatial overlay results from step S1630, M = 1.5 for red areas, M = 1.2 for orange areas, and M = 1.0 for green areas. Multiplying M(x, y) by F(x, y) yields the regional vulnerability term. This regional vulnerability term integrates historical data (edge ​​collapse location statistics) with real-time parameters (stress and wear), enabling the model to adaptively adjust risk weights in the spatial dimension. Combined with the high-risk area mapping from step S1600, this ensures that the model applies enhanced monitoring only to red / orange areas, reducing computational resource consumption.

[0158] The method for constructing the contour deviation term from the deviation vector ΔQ in step S1840 is: the deviation vector ΔQ of the real-time edge contour coordinate sequence and the theoretical contour is [ΔQ x ,ΔQ y ,ΔQ z ], where ΔQ x is the deviation vector in the X-axis direction, ΔQ y is the deviation vector in the Y-axis direction, ΔQ z is the deviation vector in the Z-axis direction; calculate the modulus |ΔQ| of the deviation vector ΔQ and normalize it to |ΔQ| / |ΔQ| max as the profile deviation term, where |ΔQ| max The maximum allowable profile deviation is quantified by the modulus of the 3D deviation vector to prevent a single axial deviation from masking the global risk.

[0159] In step S1850, the regional enhancement item, trajectory reference item, regional fragility item and contour deviation item are weighted to construct a wafer chipping risk assessment model; through sub-item coupling, the model responds to equipment static error, material fragility, dynamic contour offset and micro-distance mutation at the same time, and the chipping prediction accuracy is improved compared with the single parameter model. In step S1870, the warning condition adopts a dual judgment mechanism of "risk value exceeding the standard" and "trajectory abnormality". φ is the angle between the real-time cutting path and the theoretical path, φ th It is an acute angle threshold (such as 5°). When the angle continues to exceed the limit and the deviation vector increases, it indicates that the cutting trajectory has a nonlinear deviation that cannot be ignored. Even if the risk value does not exceed the standard, an early warning is required. The dual early warning mechanism solves the problem of early warning lag in the existing technology. For example, in a certain cutting, the risk value R is close to the threshold but has not exceeded the standard, but the abnormal trajectory deviation is judged to be true. At this time, the early warning is triggered and the cutting parameters are adjusted to avoid the early warning delay caused by the judgment of a single indicator. This step is combined with the trajectory abnormality judgment in step S1750 to form a composite early warning system of "state assessment + trend prediction", which improves the timeliness and reliability of the early warning.

[0160] Step S1800 integrates data streams from no-load calibration, clamping force correction, area mapping, and three-level monitoring to construct a multi-source information fusion chipping risk assessment model. By coupling multi-dimensional parameters and dynamically assigning weights, this model achieves accurate prediction and real-time warning of wafer chipping risk, resolving the issues of traditional methods that rely on a single data source and experience delayed response times.

[0161] Step S2000: After the wafer edge collapse warning is triggered, a compensation plan is formulated for the initial wafer cutting plan; the compensation plan includes cutting speed adjustment, trajectory micro-displacement compensation, and cutting angle compensation;

[0162] Step S2000 dynamically adjusts the cutting speed, trajectory and cutting angle based on the risk prediction value to address the defect that the compensation strategy relies on historical statistical data and cannot respond to nonlinear risks.

[0163] Furthermore, step S2000 includes:

[0164] Step S2100: adjusting the first cutting speed v0 in the initial wafer cutting plan based on the wafer chipping risk prediction value; applying micro-displacement compensation to the grinding wheel in the X-axis and Y-axis directions based on the deviation vector ΔQ;

[0165] The direction of the deviation vector ΔQ includes the X-axis direction deviation vector ΔQ x , Y-axis deviation vector ΔQ y and the Z-axis deviation vector ΔQ z , apply micro displacement compensation Δx to the grinding wheel in the X-axis direction, and Δx is determined by ΔQ xCalculation results: Apply micro displacement compensation Δy to the grinding wheel in the Y-axis direction, where Δy is calculated from ΔQ y Calculated.

[0166] Specifically, step S2100 aims to compensate for wafer chipping risk in real time by dynamically adjusting the cutting speed and grinding wheel position, addressing the existing compensation strategy's reliance on fixed parameters and inability to respond to real-time risk changes. This step combines wafer chipping risk assessment results with real-time trajectory deviation data to achieve adaptive adjustment of cutting parameters. Based on the wafer chipping risk prediction value, the initial cutting speed v0 is corrected in real time. The specific method is: when the risk prediction value indicates an increased chipping risk, the cutting speed is proportionally reduced to reduce frictional heat accumulation and edge stress concentration during contact between the grinding wheel and the wafer. This adjustment mechanism avoids the limitations of fixed speed settings in complex working conditions. For example, when grinding wheel wear causes abnormal cutting forces, reducing the speed can effectively control local thermal stresses within the material's tolerance range. Based on the deviation vector ΔQ, micro-displacement compensation is applied to the grinding wheel in the X and Y axes. A specific implementation method is: when the real-time edge profile coordinates deviate from the theoretical profile, the grinding wheel position is adjusted in the opposite direction of the deviation vector, with the adjustment amount positively correlated with the deviation magnitude. For example, if the real-time edge profile deviates by 0.1mm from the wafer center in the X-axis direction, the grinding wheel is controlled to move a corresponding distance in the positive X-axis direction to correct for trajectory deviations caused by micro-drifts in wafer positioning or equipment vibration. This compensation process references the historical trajectory deviation patterns recorded in the three-dimensional deviation matrix D to ensure that the compensation direction and amplitude conform to the inherent error characteristics of the equipment, avoiding blind adjustments that may cause new trajectory anomalies. Speed ​​is reduced to minimize thermal stress, while micro-displacements are applied to correct mechanical deviations, forming a dual-dimensional "thermal-mechanical" control system that effectively addresses complex factors such as grinding wheel wear and micro-drifts in the wafer.

[0167] Cutting speed adjustment is directly linked to risk prediction values, enabling process parameters to dynamically adjust in response to real-time risk. For example, when the risk assessment model in step S1800 detects increased grinding wheel wear and stress concentration within the red zone, the cutting speed is automatically reduced, minimizing frictional heat generation per unit time and thus lowering the probability of edge material cracking due to excessive thermal stress. This mechanism, combined with the regional enhancement item (the rate of change in the micro-distance between the grinding wheel and the wafer edge) in step S1800, forms a closed-loop control system of "risk identification-parameter adjustment," resolving the problem of compensation lagging behind risk changes in traditional methods. Micro-displacement compensation based on real-time edge profile deviation can effectively address nonlinear perturbations such as micro-drifts in wafer positioning. For example, if the wafer position shifts due to slight vibration of the vacuum chuck during the cutting process, the edge image detection in step S1720 and the deviation calculation in step S1730 can capture this shift in real time. Micro-displacement compensation in step S2100 can then rapidly adjust the grinding wheel position to ensure that the deviation between the actual cutting path and the theoretical path remains within a safe range. This operation works in conjunction with the three-level monitoring mode of step S1710 to obtain high-precision deviation data through high-frequency monitoring in high-risk areas, thereby improving the accuracy of compensation. The combination of cutting speed adjustment and trajectory compensation covers two key dimensions: thermal stress control and mechanical positioning error correction. For example, when cutting brittle materials (such as gallium arsenide), excessive speed can easily cause thermal edge collapse, and trajectory deviation can easily cause mechanical stress concentration. The coordinated adjustment of the two can reduce both risk factors at the same time, significantly improving the comprehensive suppression effect of edge collapse defects compared to single parameter adjustment.

[0168] Step S2200: Based on the Z-axis component F of the cutting force data collected in real time z Dynamically correct the grinding wheel cutting angle θ0 in the initial wafer cutting plan;

[0169] Furthermore, step S2200 includes:

[0170] Step S2210, according to F z Determine whether the grinding wheel cuts in too deeply. If so, adjust the grinding wheel cutting angle θ0. Otherwise, there is no need to adjust θ0.

[0171] When F z >F zth When the grinding wheel cuts too deep, F zth is the vertical force threshold;

[0172] Step S2220, according to θ0, F z and F zth Calculate the adjusted grinding wheel cutting angle θ1.

[0173] Specifically, step S2200 monitors the vertical component F of the cutting force in real time. zThe cutting angle θ0 of the grinding wheel is adjusted dynamically to solve the problem of edge stress concentration caused by improper cutting depth, ensuring that the contact mechanical state between the grinding wheel and the wafer matches the material's bearing capacity. z With the preset vertical force threshold F zth , to judge whether the grinding wheel cuts too deep. zth Based on the F when no edge collapse occurs in the historical cutting data z The statistical maximum value is determined, which represents the critical value of the vertical cutting force that the wafer material can withstand under the current process conditions. z More than F zth If the cutting depth is too deep, the contact stress of the grinding wheel will exceed the safe range of the material, and there is a risk of edge cracking. Otherwise, the cutting depth is considered reasonable and no adjustment is required. If the cutting depth is too deep, the contact area between the grinding wheel and the wafer can be reduced by reducing the cutting angle θ0, thereby reducing the vertical cutting force. The specific adjustment method is: based on the initial cutting angle θ0, according to F z More than F zth The angle is reduced according to the preset rules. For example, if F z F zth 1.5 times, the cut-in angle is reduced by a certain amount, which is determined by historical test data to ensure that the adjusted F z Fall back to a safe range. The adjusted entry angle θ1 must meet the engineering safety range of the angle between the grinding wheel axis and the tangent direction of the wafer surface to avoid excessive angle adjustment that may lead to reduced cutting efficiency or increased vibration.

[0174] The core function of the grinding wheel cutting angle θ0 is to control the cutting contact area. If the angle is too small, the contact area will increase, resulting in F z Increase and local stress concentration; if the angle is too large, vibration may be caused by uneven cutting force. z Monitoring and dynamic adjustment of θ1 can adapt to the mechanical properties of different wafer materials (such as silicon, gallium arsenide) and thickness. For example, when cutting thin wafers, the material has low shear strength and the allowed F zth The lower the threshold, the more sensitive the angle adjustment strategy is needed to ensure that the contact area is dynamically matched with the material properties. z Real-time feedback adjustment of the cutting angle directly solves the stress concentration problem caused by improper cutting depth. For example, when cutting a 500μm thick silicon wafer, if the grinding wheel wear causes a sudden increase in cutting force (F z >F zth ), the system automatically reduces the cutting angle and the contact area, so that F zRapidly descend to a safe range to avoid edge cracking due to excessive stress. The cut-in angle adjustment complements the risk assessment model of step S1800. The risk model identifies high stress areas through the edge fragility function, while the cut-in angle adjustment regulates the local contact mechanical state in real time. For example, in the red risk area, even if the edge collapse warning has not been triggered, if F z If the pressure is continuously above the threshold, the contact stress can be actively reduced through angle adjustment to prevent the risk from accumulating to a critical state, forming a dual protection mechanism of "regional risk warning-local mechanical regulation". Together with the cutting speed adjustment of step S2100 and the trajectory monitoring of step S1700, a control closed loop is formed to suppress the risk of edge collapse from the two aspects of motion parameters (speed, trajectory) and contact parameters (cut-in angle), solving the one-sidedness of traditional single parameter compensation. Through the above technical means, step S2200 realizes the precise control of the contact mechanical state during the cutting process, effectively reduces the edge collapse defects caused by improper cutting depth, and improves the stability and yield of the wafer cutting process under complex working conditions.

[0175] Step S3000 , after compensating the initial wafer cutting plan according to the compensation plan, a compensated wafer cutting plan is obtained, wafer cutting is performed according to the compensated wafer cutting plan, and the compensation effect is verified.

[0176] Step S3000 ensures the effectiveness of compensation measures through joint judgment of multiple indicators, avoids the vicious cycle of "adjustment-failure-missing judgment", and accumulates historical data to provide an optimization basis for subsequent cutting.

[0177] Furthermore, step S3000 includes:

[0178] Step S3100: Continuously monitor the cutting force fluctuation amplitude ΔF and the profile deviation stability σ ΔQ and the standard deviation of the potential response σU;

[0179] Step S3200: When ΔF<ΔF std And σ ΔQ <σ Qth And σ U <σ Uth When , the compensation is determined to be effective; otherwise, the secondary compensation is triggered, where ΔF std is the preset cutting force fluctuation standard value, σ Qth is the preset contour deviation stability threshold, σ Uth is the preset potential response standard deviation threshold.

[0180] Specifically, step S3000 aims to verify the effectiveness of the compensation solution in real time, ensuring that the risk of wafer chipping is effectively controlled after adjusting cutting parameters. This data feedback forms a closed-loop process optimization process, addressing the existing issues of a lack of quantitative evaluation of compensation effectiveness and the resulting inability to continuously optimize. Step S3000 determines the effectiveness of compensation measures through the collaborative analysis of multi-dimensional monitoring indicators.

[0181] The six-dimensional force sensor on the grinding wheel spindle collects the cutting force components in the X, Y, and Z axes in real time, and calculates the fluctuation range of each component per unit time. ΔF reflects the stability of the mechanical load during the cutting process. Excessive fluctuation indicates abnormal contact between the grinding wheel and the wafer, which may cause edge stress concentration. Based on the real-time edge profile deviation vector ΔQ calculated in step S1730, its standard deviation within a certain time window is calculated. σ ΔQ The deviation degree and fluctuation trend of the actual wafer edge profile from the theoretical profile are quantified. The smaller the value, the higher the trajectory control accuracy and the lower the risk of edge cracking. The potential response signal is monitored in real time and its standard deviation is calculated. U Reflects the vibration stability of the wafer clamping system. The larger the value, the more significant the micro-vibration of the clamping device, which may cause micro-displacement of the wafer and thus cause edge collapse. When ΔF is less than the preset cutting force fluctuation standard value ΔF std , σ ΔQ Less than the contour deviation stability threshold σ Qth , σU is less than the potential response standard deviation threshold σ Uth When the compensation is determined to be effective, it indicates that the current cutting parameter adjustment has restored the system to a stable state and the risk of edge collapse is within a controllable range. If any indicator fails to meet the standard, secondary compensation is triggered: steps S2100 (cutting speed and trajectory compensation) and S2200 (cutting angle adjustment) are repeated, and the compensation parameters (such as the adjusted cutting speed and cutting angle) and monitoring results are stored in the historical database to provide data accumulation for subsequent process optimization.

[0182] ΔF is directly related to the cutting angle adjustment effect of step S2200. If the cutting is too deep, the vertical cutting force F z Abnormal, ΔF will increase significantly. By comparing ΔF with ΔF std , it can be judged whether the angle adjustment effectively reduces the contact stress. σ ΔQ Reflects the accuracy of the micro-displacement compensation of the trajectory in step S2100. When the real-time cutting path deviates due to micro-offset of wafer positioning or equipment vibration, σ ΔQ Increased by σ Qth By comparison, it can be verified whether the compensation amount is sufficient to correct the trajectory error. U Verify the reliability of the clamping force correction in step S1200. If the clamping device vibrates and the potential response signal fluctuates, σ U Exceed σUth , indicating that further adjustment of the clamping force is needed to ensure stable wafer fixation. The combined monitoring of these three factors eliminates the reliance on a single indicator for compensation verification, significantly reducing the risk of missed detections.

[0183] In existing technologies, the effectiveness of compensation strategies lacks systematic verification methods and often relies on manual judgment, resulting in repeated occurrence of edge chipping defects. Step S3000 transforms the abstract concept of "compensation effectiveness" into a quantifiable judgment condition through real-time monitoring of three core indicators: ΔF measures the stability of mechanical load to avoid stress concentration caused by improper cutting depth or grinding wheel wear; σ ΔQ Quantify the trajectory control accuracy to solve the problem of missed trajectory deviation in traditional methods; U Monitor the reliability of the clamping system to eliminate the risk of edge collapse caused by micro-displacement of the wafer from the source. The three combined judgments cover the main causes of edge collapse defects and form a three-dimensional effect evaluation system. Through the secondary compensation mechanism and historical data storage, step S3000 establishes an "execution-verification-optimization" feedback chain: when the compensation is invalid (such as σ ΔQ If the target is not met), the system automatically traces the three-dimensional deviation matrix D recorded in step S1140, analyzes the inherent error mode of the equipment, and adjusts the micro-displacement compensation strategy in a targeted manner; the compensation parameters and effect data accumulated in the historical database provide a pre-optimization solution for subsequent wafer cutting of the same type, and gradually improve the system's adaptability to complex working conditions such as material property fluctuations and equipment aging.

[0184] Example 2

[0185] This embodiment provides a prediction and compensation system for edge chipping defects during wafer cutting based on embodiment 1. Figure 6 Shown, including:

[0186] Trajectory Abnormal Deviation Status Judgment Module: This module is used to obtain the initial wafer cutting plan and historical cutting data, extract the key parameters when edge chipping defects occur in the historical cutting data, define the edge fragility function F(x,y) based on the key parameters when edge chipping defects occur, and establish a high-risk area spatial mapping model based on the edge fragility function F(x,y); obtain the real-time cutting position, and based on the high-risk area spatial mapping model, adopt a three-level monitoring mode to collect multi-dimensional cutting process data in real time to determine the trajectory abnormal deviation status;

[0187] Wafer edge chipping warning module: used to build a wafer edge chipping risk assessment model and provide wafer edge chipping warning based on the wafer edge chipping risk assessment model and the results of abnormal trajectory deviation status determination;

[0188] Wafer cutting plan compensation module: After the wafer edge collapse warning is triggered, a compensation plan is formulated for the initial wafer cutting plan; after compensating the initial wafer cutting plan according to the compensation plan, a compensated wafer cutting plan is obtained;

[0189] Compensation effect verification module: According to the compensated wafer cutting plan, wafer cutting is carried out to verify the compensation effect.

[0190] In the abnormal trajectory deviation state determination module, defining the edge fragility function F(x,y) includes: defining the edge fragility function F(x,y) based on the grinding wheel wear W(t) and the wafer edge stress distribution S(x,y).

[0191] In the trajectory abnormal deviation state determination module, the determination of the trajectory abnormal deviation state includes:

[0192] Step S1710, obtain the real-time cutting position, activate the three-level monitoring mode according to the real-time cutting position, and collect cutting force data; the cutting force data includes the X-axis component F x , Y-axis component F y and the Z-axis component F z ;

[0193] Step S1720 , obtaining an image of the wafer edge during the cutting process, and extracting a real-time edge contour coordinate sequence using an edge detection algorithm;

[0194] Step S1730 , calculating the deviation vector ΔQ between the real-time edge contour coordinate sequence and the theoretical edge contour coordinate sequence;

[0195] Step S1740: Calculate the angle between the real-time cutting path and the theoretical cutting path

[0196] Step S1750, according to the angle and the deviation vector ΔQ to determine the abnormal trajectory deviation state.

[0197] When n1 consecutive sampling When |ΔQ| increases, it is determined to be an abnormal trajectory deviation state; is the acute angle threshold, and |ΔQ| is the modulus of the deviation vector ΔQ.

[0198] In the wafer edge chipping warning module, the wafer edge chipping risk assessment model is constructed, and the wafer edge chipping warning is performed based on the wafer edge chipping risk assessment model and the result of the abnormal trajectory deviation state determination, including:

[0199] Step S1810 , measuring the distance variation rate between the grinding wheel and the wafer edge in the red and orange areas, and constructing a regional enhancement item of the wafer edge chipping risk assessment model based on the distance variation rate;

[0200] Step S1820 , constructing a trajectory reference item of a wafer edge chipping risk assessment model based on the initial trajectory deviation E0;

[0201] Step S1830 , constructing a regional vulnerability term using the spatial mapping matrix M(x, y) and the edge vulnerability function F(x, y);

[0202] Step S1840 , constructing a profile deviation term from the deviation vector ΔQ;

[0203] Step S1850 , constructing a wafer edge chipping risk assessment model based on the region enhancement item, the trajectory reference item, the region vulnerability item, and the profile deviation item;

[0204] Step S1860, obtaining a wafer edge chipping risk prediction value according to a wafer edge chipping risk assessment model;

[0205] Step S1870 , when the wafer edge chipping risk prediction value exceeds the wafer edge chipping risk threshold, or when it is determined that the trajectory is in an abnormal deviation state, a wafer edge chipping warning is triggered.

[0206] The methods and systems of the present application may be implemented in many ways. For example, the methods and systems of the present application may be implemented using software, hardware, firmware, or any combination of software, hardware, and firmware. The above order of steps used in the method is for illustration only, and the steps of the method of the present application are not limited to the order specifically described above unless otherwise specified.

[0207] In addition, the parts of the above technical solutions provided in the embodiments of the present application that are consistent with the implementation principles of the corresponding technical solutions in the prior art are not described in detail to avoid excessive redundancy.

[0208] The above-described specific embodiments further illustrate the objectives, technical solutions, and beneficial effects of the present invention. It should be understood that the above description is merely a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. A method for predicting and compensating edge chipping defects during wafer dicing, characterized in that: The method comprises: Obtain the initial wafer cutting plan and historical cutting data, extract the key parameters when edge chipping defects occur in the historical cutting data, define the edge fragility function F(x,y) based on the key parameters when edge chipping defects occur, and establish a high-risk area spatial mapping model based on the edge fragility function F(x,y); obtain the real-time cutting position, and based on the high-risk area spatial mapping model, adopt a three-level monitoring mode to collect multi-dimensional cutting process data in real time and determine the abnormal trajectory deviation status; construct a wafer edge chipping risk assessment model, and issue a wafer edge chipping warning based on the wafer edge chipping risk assessment model and the results of the abnormal trajectory deviation status determination; After the wafer edge collapse warning is triggered, a compensation plan is formulated for the initial wafer cutting plan; after compensating the initial wafer cutting plan according to the compensation plan, a compensated wafer cutting plan is obtained; according to the compensated wafer cutting plan, wafer cutting is carried out to verify the compensation effect.

2. The method for predicting and compensating edge chipping defects during wafer dicing according to claim 1, wherein: The wafer cutting initial plan includes at least a wafer clamping pre-pressure F0; the key parameters when the edge chipping defect occurs in the historical cutting data include at least a grinding wheel wear W(t) and a wafer edge stress distribution S(x, y).

3. The method for predicting and compensating edge chipping defects during wafer dicing according to claim 2, wherein: The method further comprises: In the no-load test, the potential response signal U(t) of the wafer clamping device is synchronously monitored to determine the potential response signal reference value U0; Whether the clamping device is slightly vibrating is determined based on the potential response signal U(t) and the potential response signal reference value U0. If it is slightly vibrating, the pre-pressure compensation algorithm is triggered to generate a corrected clamping force F1.

4. The method for predicting and compensating edge chipping defects during wafer dicing according to claim 3, wherein: The method for generating the corrected clamping force F1 is as follows: according to the wafer clamping pre-pressure F0, the potential response signal U(t) and the potential response signal reference value U0, the corrected clamping force F1 is obtained by a pre-pressure compensation algorithm.

5. The method for predicting and compensating edge chipping defects during wafer dicing according to claim 2, wherein: The method for defining the edge fragility function F(x,y) is as follows: linearly superimposing the grinding wheel wear W(t) and the wafer stress S(x,y) to define the edge fragility function F(x,y).

6. The method for predicting and compensating edge chipping defects during wafer dicing according to claim 1, wherein: The establishment of a high-risk area spatial mapping model based on the edge vulnerability function F(x,y) includes: Generate a two-dimensional fragility heat map of the wafer based on the edge fragility function; According to the two-dimensional vulnerability heat map of the wafer, the wafer is divided into three levels of risk areas; the three levels of risk areas of the wafer include red area, orange area and green area.

7. The method for predicting and compensating edge chipping defects during wafer dicing according to claim 6, wherein: The method for generating a two-dimensional fragility heat map of a wafer includes: The wafer surface is divided into n2 grid cells, and the coordinates of the center point of each grid cell are (x j' ,y j' ), calculate the edge fragility value F(x j' ,y j' ), according to the edge fragility value F(x j' ,y j' ), generate a two-dimensional fragility heat map of the wafer; where (x j' ,y j' ) is the coordinate of the center point of the j'th grid cell, F(x j' ,y j' ) is the edge vulnerability value of the j'th grid cell, 1≤j'≤n2.

8. The method for predicting and compensating edge chipping defects during wafer dicing according to claim 6, wherein: The three-level monitoring mode is: If the real-time cutting position is within the red area, the first monitoring mode is activated; If the real-time cutting position is within the orange area, the second monitoring mode is activated; If the real-time cutting position is within the green area, the third monitoring mode is maintained; The first monitoring mode is a high-frequency monitoring mode, the second monitoring mode is a medium-frequency monitoring mode, and the third monitoring mode is a low-frequency monitoring mode.

9. The method for predicting and compensating edge chipping defects during wafer dicing according to claim 1, wherein: The multi-dimensional cutting process data includes cutting force data and wafer edge images; The method for determining the abnormal trajectory deviation state is as follows: Perform edge detection on wafer edge images during the cutting process and extract real-time edge contour coordinate sequences; Calculate the deviation vector ΔQ between the real-time edge profile coordinate sequence and the theoretical edge profile coordinate sequence; Calculate the angle between the real-time cutting path and the theoretical cutting path According to the angle and the deviation vector ΔQ to determine the abnormal trajectory deviation state.

10. The method for predicting and compensating edge chipping defects during wafer dicing according to claim 9, wherein: According to the angle and the deviation vector ΔQ, the determination of the abnormal trajectory deviation state includes: When n1 consecutive sampling When |ΔQ| increases, it is determined to be an abnormal trajectory deviation state; is the acute angle threshold, |ΔQ| is the modulus of the deviation vector ΔQ; The method for performing wafer edge collapse warning is as follows: obtaining a wafer edge collapse risk prediction value according to a wafer edge collapse risk assessment model; triggering a wafer edge collapse warning when the wafer edge collapse risk prediction value exceeds a wafer edge collapse risk threshold, or when it is determined that the trajectory is in an abnormal deviation state.

Images