Biaxial dfg grinder z1-z2 linkage removal amount control and grinding wheel loss optimization method

Abstract

The application relates to the technical field of precision machining, and discloses a Z1-Z2 linkage removal amount control and grinding wheel loss optimization method and system of a double-shaft DFG grinder, aiming to solve the problems of unreasonable removal amount distribution, inaccurate grinding wheel loss control and poor adaptability of wafer surface damage in the existing DFG grinding Z1-Z2. The method and system utilize wafer initial topography and material characteristics, determine Z1 rough grinding parameters and adjust the parameters through multi-objective optimization and self-adaptive control; in combination with the Z1 result, the removal amount and parameters of Z2 fine grinding are intelligently distributed by using a deep learning model, self-adaptive adjustment and abnormal Air Cut are realized; meanwhile, a grinding wheel loss prediction model is used to continuously optimize the grinding parameters. By adopting the technical scheme, the grinding efficiency and precision can be remarkably improved, the service life of the grinding wheel is prolonged, the cost is reduced, and the problems of removal amount distribution, grinding wheel loss and wafer damage are solved.

Description

Technical Field

[0001] This invention belongs to the field of precision machining technology, specifically relating to a method for controlling the removal amount and optimizing the wear of grinding wheels in a biaxial DFG grinding machine with Z1-Z2 linkage. Background Technology

[0002] With the rapid development of the semiconductor industry, integrated circuit manufacturing has placed increasingly stringent demands on wafer surface quality and processing efficiency. Dual-face grinding (DFG) machines, as key equipment, are widely used in wafer thinning and surface planarization. However, traditional grinding processes rely primarily on preset parameters and empirical adjustments, with their core principle based on the assumption of an ideal wafer state. Wafers exhibit highly complex and dynamic characteristics in actual processing: their initial surface morphology (such as damage distribution and local unevenness), material properties, and real-time conditions during the grinding process (such as wheel wear and grinding force fluctuations) constantly change with different batches and processing stages. Fixed-parameter strategies, unable to perceive and adapt to these dynamic changes in real time, are prone to low grinding efficiency, insufficient processing accuracy, or unexpectedly shortened wheel life.

[0003] Dual-axis DFG grinding machines typically employ a two-stage grinding mechanism: Z1 coarse grinding and Z2 fine grinding, to achieve progressive material removal and improved surface quality. This multi-stage grinding strategy aims to balance removal efficiency and surface quality by optimizing the removal capabilities of different grinding wheels. However, the coordinated removal volume control of the Z1 and Z2 grinding wheels and the refined optimization of wheel wear remain key technical bottlenecks limiting its performance. Faced with increasingly diverse wafer materials and more refined processing requirements, existing technologies exhibit significant limitations in terms of the intelligence and adaptability of the grinding process.

[0004] Existing technologies have significant shortcomings in controlling the removal amount of the Z1-Z2 axes in biaxial DFG grinding. For example, some solutions may extend the life of the grinding wheel by adjusting its physical parameters, but these methods are mainly designed for specific structures and fail to effectively solve the dynamic linkage matching problem of removal amounts between the Z1 and Z2 axes in biaxial DFG grinding. Especially when there is complex damage on the wafer surface, it is impossible to avoid diamond particle shedding due to excessive load on the Z2 axis, leading to abnormal wear of the grinding wheel and increasing production costs. Other technologies focus on optimizing the flow rate of the grinding fluid to assist in chip removal and cooling. Although this helps reduce grinding wheel resistance and extend its life, its core function is auxiliary and does not fundamentally solve the contradiction of improper matching of removal amounts between Z1 rough grinding and Z2 fine grinding. Especially when Z1 fails to fully remove hard spots or defects, the high-mesh Z2 grinding wheel will bear an extra burden, accelerating its wear and even causing premature failure, seriously affecting overall processing efficiency and maintenance costs. The aforementioned problems indicate that existing biaxial DFG polishing equipment and processes suffer from significant deficiencies in the precision of Z1-Z2 linkage removal control, intelligent optimization of wheel wear, and adaptive capabilities when dealing with diverse wafer materials and complex surface morphologies. Therefore, there is an urgent need for an intelligent polishing method that can achieve precise control of Z1-Z2 linkage removal and optimization of wheel wear, while adapting to dynamic changes in the wafer. Summary of the Invention

[0005] This invention discloses a method for controlling the removal amount and optimizing wheel wear in the Z1-Z2 linkage of a biaxial DFG grinding machine. The invention aims to solve problems in existing biaxial DFG grinding equipment, such as unreasonable distribution of removal amount between Z1 rough grinding and Z2 fine grinding, inaccurate control of wheel wear, and difficulty in adapting to complex wafer surface damage, thereby improving grinding efficiency, extending wheel life, and reducing production costs.

[0006] According to one aspect of the present invention, a method for controlling the removal amount and optimizing the grinding wheel wear of a biaxial DFG grinding mill with Z1-Z2 linkage is provided, comprising the following steps: Obtain initial surface morphology data and wafer material property parameters of the wafer to be ground; Based on the initial surface morphology data and wafer material characteristic parameters, the initial grinding parameters and target removal amount for the Z1 coarse grinding stage are determined. Perform Z1 rough grinding and monitor the grinding status data of the Z1 axis in real time during the Z1 rough grinding process; Based on the grinding state data and target removal amount of the Z1 coarse grinding stage, the Z1 coarse grinding parameters are adaptively adjusted. This includes triggering a local Air Cut strategy when the Z1 axis grinding force or grinding power abnormally increases beyond a preset safety threshold, until the Z1 coarse grinding is completed and the wafer surface morphology after Z1 coarse grinding is obtained. Based on the wafer surface morphology, wafer material properties and set total removal amount after Z1 rough grinding, the target removal amount and initial grinding parameters of the Z2 fine grinding stage are determined by the intelligent allocation model of grinding removal amount. Perform Z2 fine grinding and monitor the grinding status data of the Z2 axis in real time during the Z2 fine grinding process; Based on the grinding state data of the Z2 fine grinding stage and the target removal amount of Z2 fine grinding, the Z2 fine grinding parameters are adaptively adjusted. When the Z2 axis grinding force or grinding power fluctuates abnormally beyond the preset threshold, the Air Cut strategy is triggered to adjust the Z2 fine grinding depth or grinding pressure until the Z2 fine grinding is completed. Throughout the grinding process, grinding status data of Z1 and Z2 axes are continuously collected. Combined with grinding wheel model information, grinding wheel wear status and remaining life are evaluated through grinding wheel wear prediction model, and grinding parameters are adjusted based on the evaluation results.

[0007] As one embodiment of the present invention, the acquisition of the initial surface morphology data and wafer material characteristic parameters of the wafer to be ground specifically includes: The surface of the wafer to be ground is scanned non-contactly using a three-dimensional optical profilometer to obtain high-resolution three-dimensional height mapping data. The horizontal resolution of the three-dimensional height mapping data is less than or equal to one micrometer, and the vertical resolution is less than or equal to one hundred nanometers. Based on the three-dimensional height mapping data, the damage types of the wafer surface are identified and classified through image processing and feature extraction algorithms, including but not limited to frosted texture, mechanical marks, hard spots, scratches or local uneven areas, and the depth, width and distribution density of each damage area are calculated. The material properties of the wafer to be ground are obtained through a data interface. These material properties include material hardness, Young's modulus, fracture toughness, coefficient of thermal expansion, and crystal orientation.

[0008] As one embodiment of the present invention, determining the initial grinding parameters and target removal amount of the Z1 coarse grinding stage specifically includes: Based on the damage types and distribution identified in the initial surface morphology data, the initial unevenness deviation value of the entire wafer is calculated. Based on the unevenness deviation value and the wafer material characteristic parameters, combined with the preset grinding removal rate curve and grinding wheel model parameters, the initial grinding pressure, initial spindle speed, initial feed speed and target removal amount of Z1 coarse grinding stage are determined by a multi-objective optimization algorithm. The multi-objective optimization algorithm aims to maximize the defect removal rate, minimize subsurface damage, and minimize abnormal wear of the grinding wheel during the optimization process, and optimizes parameters under the constraints of satisfying the maximum allowable grinding force and the minimum material removal rate.

[0009] As one embodiment of the present invention, the step of performing Z1 rough grinding and monitoring the grinding status data of the Z1 axis in real time during the Z1 rough grinding process specifically includes: The Z1 coarse grinding is performed by a Z1 coarse grinding wheel, and the mesh size of the Z1 coarse grinding wheel ranges from 800 to 1500. The normal and tangential grinding forces of the Z1 axis are monitored in real time by a piezoelectric force sensor installed on the Z1 axis grinding spindle. The response time of the force sensor is less than one millisecond. The current and voltage of the Z1 axis spindle motor are monitored in real time by current and voltage sensors, and the actual grinding power of the Z1 axis spindle motor is calculated. The actual rotational speed of the Z1 coarse grinding wheel is monitored in real time using a speed sensor. The grinding vibration signal of the Z1 axis is monitored in real time using a vibration sensor; The actual feed depth of the Z1 coarse grinding wheel is monitored in real time by a high-precision displacement sensor, thereby calculating the amount of material removed in real time.

[0010] In one embodiment of the present invention, the adaptive adjustment of the Z1 coarse grinding parameters, based on the grinding state data and the target removal amount of the Z1 coarse grinding stage, includes triggering a local Air Cut strategy when the Z1 axis grinding force or grinding power abnormally increases beyond a preset safety threshold, until the Z1 coarse grinding is completed, and obtaining the wafer surface morphology after Z1 coarse grinding. The real-time monitored grinding status data of the Z1 axis is input to the adaptive control algorithm module, which includes a proportional-integral-derivative controller and a fuzzy logic controller. The adaptive control algorithm module dynamically adjusts the grinding pressure, spindle speed, and feed speed of the Z1 coarse grinding wheel based on the deviation between the grinding state data and the target removal amount of Z1 coarse grinding, so as to maintain a stable material removal rate and grinding force. The fuzzy logic controller defines the input variable as the grinding force deviation (e) and its rate of change (Δe), and the output variable as the grinding pressure adjustment (ΔP). The domain of e is [-50N, 50N], divided into {negative large, negative medium, zero, positive medium, positive large}, and the domain of ΔP is [-5MPa, 5MPa]. The rule is set: if e is negative large and Δe is negative large, then ΔP is positive large. The PID controller output and the fuzzy output are weighted and fused, with a weighting coefficient of 0.7 (fuzzy):0.3 (PID).

[0011] When the Z1 axis grinding force or grinding power increases abnormally in a short period of time and exceeds the preset safety threshold, the adaptive control algorithm module triggers the local Air Cut strategy, which rapidly raises the Z1 coarse grinding wheel by 0.5 mm to 1 mm for a duration of 50 to 100 milliseconds to avoid wheel chipping or wafer damage, and then re-enters the grinding state with the adjusted grinding parameters. After Z1 rough grinding is completed, the wafer surface is scanned again using the three-dimensional optical profilometer to obtain the wafer surface morphology data after Z1 rough grinding, and the actual amount of material removed by Z1 rough grinding is calculated.

[0012] As one embodiment of the present invention, the determination of the target removal amount and initial grinding parameters for the Z2 fine grinding stage based on the wafer surface morphology after Z1 coarse grinding, wafer material characteristic parameters, and the set total removal amount after fine grinding, through a grinding removal amount intelligent allocation model, specifically includes: The wafer surface morphology data after Z1 rough grinding, wafer material characteristic parameters, actual wear data of Z1 rough grinding wheel, and the set total removal amount of fine grinding are used as inputs; The intelligent allocation model for grinding removal amount adopts a deep neural network. The deep neural network is trained with historical grinding data to establish a complex mapping relationship between Z1 coarse grinding result, Z2 fine grinding removal amount and grinding wheel wear. The output of the deep neural network includes the target removal amount in the Z2 fine grinding stage, as well as the initial grinding pressure, initial spindle speed, and initial feed rate of the Z2 fine grinding. The intelligent grinding removal amount allocation model comprehensively considers the severity of the remaining defects on the wafer surface after Z1 rough grinding, the bearing capacity of the Z2 fine grinding wheel, the desired final surface roughness, and the principle of minimizing wheel wear when determining the target removal amount of Z2 fine grinding, to ensure that the Z2 fine grinding wheel operates within the optimal load range.

[0013] The total removal amount of fine grinding is the total removal thickness from the initial thickness of the wafer to the target final thickness minus the actual removal amount of Z1 rough grinding.

[0014] As one embodiment of the present invention, the step of performing Z2 fine grinding and monitoring the grinding status data of the Z2 axis in real time during the Z2 fine grinding process specifically includes: The Z2 fine grinding is performed by a Z2 fine grinding wheel, and the mesh size of the Z2 fine grinding wheel ranges from three thousand to eight thousand. The normal and tangential grinding forces of the Z2 axis are monitored in real time by a piezoelectric force sensor installed on the Z2 axis grinding spindle. The response time of the force sensor is less than one millisecond. The current and voltage of the Z2 axis spindle motor are monitored in real time by current and voltage sensors, and the actual grinding power of the Z2 axis spindle motor is calculated. The actual rotational speed of the Z2 precision grinding wheel is monitored in real time using a speed sensor. The grinding vibration signal of the Z2 axis is monitored in real time using a vibration sensor; The actual feed depth of the Z2 fine grinding wheel is monitored in real time by a high-precision displacement sensor, thereby calculating the amount of material removed in real time.

[0015] As one embodiment of the present invention, the adaptive adjustment of Z2 fine grinding parameters; when the Z2 axis grinding force or grinding power fluctuates abnormally beyond a preset threshold, triggering the Air Cut strategy to adjust the Z2 fine grinding depth or grinding pressure until Z2 fine grinding is completed, specifically includes: The real-time monitored Z2 axis grinding status data is input to the adaptive control algorithm module, which includes a proportional-integral-derivative controller and a fuzzy logic controller. The adaptive control algorithm module dynamically adjusts the grinding pressure, spindle speed, and feed rate of the Z2 fine grinding wheel based on the deviation between the grinding state data and the target removal amount of Z2 fine grinding, so as to maintain a stable material removal rate and grinding force. When the Z2 axis grinding force or grinding power fluctuates abnormally within a short period of time, and the fluctuation exceeds the set grinding force threshold or grinding power threshold, the Air Cut strategy is triggered. The Air Cut strategy specifically includes: the grinding machine control system issuing a command to rapidly lift the Z2 fine grinding wheel from the wafer surface within ten milliseconds, with a lifting distance of 0.2 mm to 0.5 mm, and the lifting duration being 50 milliseconds to 100 milliseconds; During the Air Cut, the system reassesses the wafer surface condition and the causes of any abnormalities in the polishing force and polishing power. After the evaluation is completed, the system recalculates and adjusts the grinding pressure and feed speed of the Z2 fine grinding wheel based on the evaluation results and the current grinding progress, and then makes the Z2 fine grinding wheel re-contact the wafer surface with the adjusted parameters for grinding.

[0016] As one embodiment of the present invention, the grinding state data of the Z1 and Z2 axes are continuously collected throughout the grinding process, and combined with the grinding wheel model information, the grinding wheel wear state and remaining life are evaluated through a grinding wheel wear prediction model, and the grinding parameters are adjusted according to the evaluation results. Specifically, this includes: Continuously collect and store historical data such as grinding force, grinding power, spindle speed, feed rate, grinding time and actual material removal amount of Z1 axis and Z2 axis during the grinding process; The grinding wheel wear prediction model uses a temporal convolutional network or a long short-term memory network, which is trained by learning the correlation between historical grinding data and the actual wear degree of the grinding wheel; The grinding wheel wear prediction model takes real-time grinding status data and grinding wheel model information as input, and outputs the current wear amount of the grinding wheel, the degree of diamond particle shedding from the grinding wheel, the change in the flatness of the grinding wheel working surface, and the remaining service life of the grinding wheel. When the grinding wheel wear prediction model predicts that the remaining service life of any grinding wheel is lower than a preset threshold, the system will issue a warning for grinding wheel replacement or repair. Based on the output of the grinding wheel wear prediction model, the intelligent grinding removal amount allocation model and the Z1 / Z2 adaptive control algorithm will adjust the grinding parameters in real time, such as appropriately reducing the grinding pressure or feed rate, to slow down the grinding wheel wear rate and thus extend the service life of the grinding wheel.

[0017] According to another aspect of the present invention, a system for controlling the removal amount and optimizing the grinding wheel wear of a biaxial DFG grinding mill Z1-Z2 linkage is provided, comprising: The wafer initial state acquisition module is used to acquire the initial surface morphology data and wafer material property parameters of the wafer to be ground; The Z1 coarse grinding parameter determination module is used to determine the initial grinding parameters and target removal amount of the Z1 coarse grinding stage based on the initial surface morphology data and wafer material characteristic parameters. The Z1 coarse grinding execution and monitoring module is used to execute Z1 coarse grinding and monitor the grinding status data of the Z1 axis in real time during the Z1 coarse grinding process. The Z1 coarse grinding adaptive adjustment module is used to adaptively adjust the Z1 coarse grinding parameters based on the grinding state data and the target removal amount of the Z1 coarse grinding stage. The adaptive adjustment of the Z1 coarse grinding parameters includes triggering a local Air Cut strategy when the Z1 axis grinding force or grinding power abnormally increases beyond a preset safety threshold, until the Z1 coarse grinding is completed and the wafer surface morphology after Z1 coarse grinding is obtained. The intelligent grinding removal amount allocation module is used to determine the target removal amount and initial grinding parameters of the Z2 fine grinding stage based on the wafer surface morphology, wafer material characteristic parameters and the set total removal amount of fine grinding after Z1 rough grinding. The Z2 fine grinding execution and monitoring module is used to execute Z2 fine grinding and monitor the grinding status data of the Z2 axis in real time during the Z2 fine grinding process. The Z2 fine grinding adaptive adjustment module is used to adaptively adjust the Z2 fine grinding parameters based on the grinding state data of the Z2 fine grinding stage and the target removal amount of Z2 fine grinding. The Air Cut trigger and adjustment module is used to trigger the Air Cut strategy when the Z2 axis grinding force or grinding power fluctuates abnormally beyond the preset threshold, and adjust the Z2 fine grinding depth or grinding pressure until the Z2 fine grinding is completed. The grinding wheel wear prediction and optimization module is used to continuously collect grinding status data of Z1 axis and Z2 axis throughout the grinding process, and combine it with grinding wheel model information to evaluate the grinding wheel wear status and remaining life through the grinding wheel wear prediction model, and adjust the grinding parameters according to the evaluation results.

[0018] In one embodiment of the present invention, the wafer initial state acquisition module includes: A three-dimensional optical profilometer is used to perform non-contact scanning on the surface of the wafer to be ground to obtain high-resolution three-dimensional height mapping data. The horizontal resolution of the three-dimensional height mapping data is less than or equal to one micrometer, and the vertical resolution is less than or equal to one hundred nanometers. The data processing and feature extraction unit is used to identify and classify the damage types on the wafer surface based on the three-dimensional height mapping data through image processing and feature extraction algorithms, including but not limited to frosted textures, mechanical marks, hard spots, scratches or local uneven areas, and to calculate the depth, width and distribution density of each damage area. The wafer material property acquisition unit is used to acquire the material property parameters of the wafer to be ground through a data interface. The material property parameters include material hardness, Young's modulus, fracture toughness, coefficient of thermal expansion, and crystal orientation.

[0019] In one embodiment of the present invention, the Z1 rough grinding parameter determination module includes: The unevenness deviation calculation unit is used to calculate the initial unevenness deviation value of the entire wafer based on the damage type and distribution identified in the initial surface morphology data. The multi-objective optimization unit is used to determine the initial grinding pressure, initial spindle speed, initial feed rate, and target removal amount of the Z1 coarse grinding stage based on the unevenness deviation value and the wafer material characteristic parameters, combined with the preset grinding removal rate curve and grinding wheel model parameters, through a multi-objective optimization algorithm. The multi-objective optimization algorithm aims to maximize the defect removal rate, minimize subsurface damage, and minimize abnormal wear of the grinding wheel during the optimization process, and optimizes parameters under the constraints of satisfying the maximum allowable grinding force and the minimum material removal rate.

[0020] In one embodiment of the present invention, the Z1 coarse grinding execution and monitoring module includes: Z1 coarse grinding wheel, wherein the grit size of the Z1 coarse grinding wheel ranges from 800 to 1500 grit; A piezoelectric force sensor is used to monitor the normal and tangential grinding forces of the Z1 axis in real time. The response time of the force sensor is less than one millisecond. Current and voltage sensors are used to monitor the current and voltage of the Z1 axis spindle motor in real time and to calculate the actual grinding power of the Z1 axis spindle motor. A speed sensor is used to monitor the actual speed of the Z1 coarse grinding wheel in real time. Vibration sensor, used to monitor grinding vibration signals of Z1 axis in real time; A high-precision displacement sensor is used to monitor the actual feed depth of the Z1 coarse grinding wheel in real time, thereby calculating the amount of material removed in real time.

[0021] In one embodiment of the present invention, the Z1 coarse grinding adaptive adjustment module includes: An adaptive control algorithm module is used to input real-time monitored Z1 axis grinding status data. The adaptive control algorithm module includes a proportional-integral-derivative controller and a fuzzy logic controller. The parameter dynamic adjustment unit is used to dynamically adjust the grinding pressure, spindle speed and feed speed of the Z1 coarse grinding wheel through the adaptive control algorithm module based on the deviation between the grinding state data and the Z1 coarse grinding target removal amount, so as to maintain a stable material removal rate and grinding force. The local air cut strategy execution unit is used to activate the safety threshold F when the Z1 axis grinding force or grinding power abnormally increases within a short period of time and exceeds the preset safety threshold. threshold = k × H material × V feed Where k is the safety factor (0.8-1.2), H material Wafer hardness (GPa), V feedThe feed rate is (mm / min); 'short time' is defined as the change in grinding force ΔF > 0.3 × F within 100ms. threshold The local Air Cut strategy is triggered by rapidly raising the Z1 coarse grinding wheel by 0.5 mm to 1 mm for 50 to 100 milliseconds to avoid wheel chipping or wafer damage, and then re-entering the grinding state with the adjusted grinding parameters. The Z1 rough grinding morphology acquisition unit is used to scan the wafer surface again using the three-dimensional optical profilometer after the Z1 rough grinding is completed, acquire the wafer surface morphology data after Z1 rough grinding, and calculate the actual amount of material removed by Z1 rough grinding.

[0022] In one embodiment of the present invention, the intelligent grinding removal amount allocation module includes: The data input unit is used to receive wafer surface morphology data after Z1 rough grinding, wafer material characteristic parameters, actual wear data of Z1 rough grinding wheel, and the set total removal amount of fine grinding as input; A deep neural network is trained using historical grinding data to establish a complex mapping relationship between the Z1 rough grinding result, the Z2 fine grinding removal amount, and the wear of the grinding wheel. The historical grinding data includes 20-dimensional features such as the initial wafer thickness (unit μm), hard spot density (number / cm²), actual Z1 removal amount (μm), grinding wheel mesh size, and cumulative grinding time (min), with a sample size ≥10,000. The neural network has a 3-layer fully connected structure (20 nodes in the input layer, 128 nodes / ReLU in the hidden layer, and 3 nodes in the output layer). The training uses the Adam optimizer (lr=0.001) and the mean squared error loss function. The early stopping condition is that the validation set loss does not decrease for 5 consecutive rounds.

[0023] The removal amount and parameter output unit is used to output the target removal amount of the Z2 fine grinding stage, as well as the initial grinding pressure, initial spindle speed and initial feed rate of the Z2 fine grinding. When determining the target removal amount of Z2 fine grinding, the deep neural network comprehensively considers the severity of the remaining defects on the wafer surface after Z1 rough grinding, the bearing capacity of the Z2 fine grinding wheel, the desired final surface roughness, and the principle of minimizing wheel wear, to ensure that the Z2 fine grinding wheel operates within the optimal load range.

[0024] The total removal amount of fine grinding is the total removal thickness from the initial thickness of the wafer to the target final thickness minus the actual removal amount of Z1 rough grinding.

[0025] As one embodiment of the present invention, the Z2 fine grinding execution and monitoring module includes: Z2 fine grinding wheel, wherein the grit size of the Z2 fine grinding wheel ranges from 3,000 to 8,000 grit; A piezoelectric force sensor is used to monitor the normal and tangential grinding forces of the Z2 axis in real time. The response time of the force sensor is less than one millisecond. Current and voltage sensors are used to monitor the current and voltage of the Z2 axis spindle motor in real time and to calculate the actual grinding power of the Z2 axis spindle motor. A speed sensor is used to monitor the actual speed of the Z2 fine grinding wheel in real time; Vibration sensor, used to monitor grinding vibration signals of the Z2 axis in real time; A high-precision displacement sensor is used to monitor the actual feed depth of the Z2 fine grinding wheel in real time, thereby calculating the amount of material removed in real time.

[0026] In one embodiment of the present invention, the Z2 fine grinding adaptive adjustment module and the Air Cut triggering and adjustment module include: An adaptive control algorithm module is used to input real-time monitored Z2 axis grinding status data. The adaptive control algorithm module includes a proportional-integral-derivative controller and a fuzzy logic controller. The parameter dynamic adjustment unit is used to dynamically adjust the grinding pressure, spindle speed and feed speed of the Z2 fine grinding wheel based on the deviation between the grinding state data and the Z2 fine grinding target removal amount, through the adaptive control algorithm module, so as to maintain a stable material removal rate and grinding force. An abnormal fluctuation detection unit is used to trigger the Air Cut strategy when the Z2 axis grinding force or grinding power fluctuates abnormally in a short period of time, and the fluctuation exceeds the set grinding force threshold or grinding power threshold. The Air Cut strategy execution unit is used to issue commands from the grinding machine control system to rapidly lift the Z2 fine grinding wheel from the wafer surface within ten milliseconds, with a lifting distance of 0.2 mm to 0.5 mm and a lifting duration of 50 milliseconds to 100 milliseconds. The status reassessment and parameter adjustment unit is used to reassess the wafer surface condition and the cause of abnormal grinding force and grinding power during the Air Cut. After the assessment, the system recalculates and adjusts the grinding pressure and feed speed of the Z2 fine grinding wheel according to the assessment results and the current grinding progress, and then makes the Z2 fine grinding wheel re-contact the wafer surface for grinding with the adjusted parameters.

[0027] As one embodiment of the present invention, the grinding wheel wear prediction and optimization module includes: The historical data collection and storage unit is used to continuously collect and store historical data such as grinding force, grinding power, spindle speed, feed rate, grinding time and actual material removal amount of the Z1 axis and Z2 axis during the grinding process; A grinding wheel wear prediction model is provided, which employs a temporal convolutional network or a long short-term memory network. The network is trained by learning the correlation between historical grinding data and the actual wear degree of the grinding wheel. The wear condition and life assessment unit is used to take real-time grinding condition data and grinding wheel model information as input, and output the current wear amount of the grinding wheel, the degree of diamond particle shedding of the grinding wheel, the change in the flatness of the grinding wheel working surface, and the remaining service life of the grinding wheel through the grinding wheel wear prediction model. The early warning and parameter adjustment unit is used to issue an early warning for grinding wheel replacement or refurbishment when the grinding wheel wear prediction model predicts that the remaining service life of any grinding wheel is lower than a preset threshold. Based on the output of the grinding wheel wear prediction model, the intelligent grinding removal amount allocation model and the Z1 / Z2 adaptive control algorithm will adjust the grinding parameters in real time, such as appropriately reducing the grinding pressure or feed speed, to slow down the grinding wheel wear rate and thus extend the service life of the grinding wheel.

[0028] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention utilizes high-precision three-dimensional morphology scanning of the wafer surface before Z1 rough grinding to accurately identify and quantify various damage types and unevenness, providing a reliable initial data foundation for subsequent grinding parameter optimization. This solves the problem of blind grinding caused by large differences in the initial surface of the wafer in existing technologies.

[0029] This invention proposes that the initial grinding parameters and target removal amount of the Z1 rough grinding stage are dynamically determined based on the wafer surface damage distribution and material properties through a multi-objective optimization algorithm. This ensures that Z1 rough grinding can efficiently and accurately remove major defects while minimizing subsurface damage. This strategy is significantly superior to the static or empirical parameter settings in existing technologies, avoiding under-grinding or over-grinding in Z1, and laying a uniform surface foundation for Z2 fine grinding.

[0030] This invention introduces an intelligent allocation model for the removal amount of Z1-Z2 grinding. By combining the actual morphology of Z1 after rough grinding and the wear state of the grinding wheel with a deep neural network, it intelligently allocates the target removal amount and initial grinding parameters for Z2 fine grinding. This linkage control mechanism ensures that the Z2 fine grinding wheel always operates within its optimal load range, effectively preventing diamond particle shedding and premature failure caused by Z2 bearing excessive load. It also solves the problem of abnormal wheel wear caused by improper matching of Z1 and Z2 removal amounts in existing technologies.

[0031] This invention monitors key status data such as grinding force and grinding power in real time during Z1 and Z2 grinding processes, and dynamically adjusts grinding parameters using an adaptive control algorithm. More importantly, when abnormal fluctuations in grinding force or grinding power are detected, the system can quickly trigger the Air Cut strategy, temporarily raising the grinding wheel, reassessing the grinding condition, and resuming grinding with the adjusted parameters. This effectively prevents grinding wheel damage and wafer scratches caused by sudden hard spots or local defects, improving the stability and safety of the grinding process.

[0032] This invention integrates an advanced grinding wheel wear prediction model, utilizing temporal convolutional networks or long short-term memory networks to accurately assess the wear state and remaining life of grinding wheels Z1 and Z2. This enables maintenance personnel to perform predictive maintenance based on actual wear conditions, such as timely replacement or dressing of grinding wheels, rather than relying on fixed experience cycles. By feeding the prediction results back to the grinding parameter optimization and removal amount allocation modules, the system can dynamically adjust grinding parameters to slow down grinding wheel wear, thereby maximizing grinding wheel life and significantly reducing production costs.

[0033] In summary, this invention constructs a comprehensive and efficient biaxial DFG grinding solution by precisely sensing the initial state of the wafer, intelligently allocating the Z1-Z2 linkage removal amount, real-time adaptive control of the grinding process, and accurate prediction and optimization of grinding wheel wear. This solution can significantly improve processing efficiency, ensure processing accuracy, and greatly extend grinding wheel life, thereby reducing the overall cost of semiconductor wafer manufacturing. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the overall technical solution architecture of the method for controlling the removal amount and optimizing the grinding wheel wear of the Z1-Z2 linkage of the biaxial DFG grinding machine proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the intelligent allocation model for grinding removal in this invention; Figure 3 This is a logical flow diagram of the Z1-Z2 linkage grinding method of the biaxial DFG grinding machine of the present invention; Figure 4 This is a schematic diagram illustrating the core principle framework of the adaptive control of the grinding process and the Air Cut anomaly handling strategy in this invention; Figure 5 This is a schematic diagram of the core principle framework of the grinding wheel wear prediction and optimization feedback mechanism in this invention. Detailed Implementation

[0035] This invention aims to provide a method for controlling the Z1-Z2 linkage removal amount and optimizing wheel wear in a biaxial differential grinding machine. This addresses problems in existing biaxial grinding equipment, such as improper matching of removal amounts between Z1 rough grinding and Z2 fine grinding, abnormal wheel wear caused by differences in workpiece surface flatness, and the lack of an effective air-cut adjustment strategy. These problems directly affect the efficiency, accuracy, and cost control of semiconductor wafer processing. Traditional grinding methods often use static preset parameters, which are difficult to adapt to the complex and varied damage types and material properties of wafer surfaces. This results in insufficient removal by the Z1 grinding wheel when processing hard spots or deep defects, causing the high-mesh Z2 grinding wheel to bear additional load and accelerate wear. Simultaneously, there is a lack of effective identification and response mechanisms for localized air-cut conditions, leading to low grinding efficiency or unnecessary wheel wear. This invention introduces an intelligent allocation model, adaptive control strategy, and grinding wheel wear optimization mechanism to achieve refined and intelligent management of the grinding process, thereby maximizing the extension of grinding wheel life and improving processing efficiency and stability.

[0036] This embodiment discloses a method for controlling the amount of material removed by the Z1-Z2 linkage and optimizing the wear of the grinding wheel in a biaxial differential grinding machine.

[0037] S101: Obtain the initial state data of the workpiece and the initial state data of the grinding wheel before grinding.

[0038] This step is fundamental to the intelligentization of the entire grinding process, aiming to comprehensively understand the initial physical properties of the wafer to be processed, the distribution of surface defects, and the geometric parameters and performance of the grinding wheel used. Accurate initial data is a prerequisite for the subsequent intelligent allocation model to make accurate decisions.

[0039] S1011: Obtain initial physical and defect data of the wafer to be ground.

[0040] This sub-step uses a data acquisition system integrated into the front end of the grinding equipment to perform non-contact scanning and measurement on each wafer to be processed. The acquired data covers multiple key dimensions, including the initial thickness of the wafer, overall flatness, surface roughness, and the spatial distribution and severity of various defects (such as the depth of the frosting marks, mechanical marks, scratches, etc.).

[0041] Specifically, the initial thickness of the wafer can be sampled at multiple preset points using a high-precision capacitive thickness gauge or laser thickness gauge to form a wafer thickness distribution map. Typical thickness measurement accuracy can reach the micrometer or even sub-micrometer level. Overall flatness is calculated by fitting and analyzing multi-point thickness measurement data to determine indicators such as total thickness variation (total thickness variation) or local thickness variation (local thickness variation). Surface roughness is obtained by scanning with a non-contact optical profilometer or atomic force microscope to acquire parameters such as average roughness (arithmetic mean deviation) and root mean square roughness (root mean square deviation). These instruments typically employ laser interferometry, white light interferometry, or probe scanning principles, providing surface morphology information at the nanometer or even angstrom level.

[0042] For defects on wafer surfaces, such as frosted textures and mechanical marks, depth and lateral dimensions are key parameters. These defects can be identified and quantified using high-resolution machine vision systems combined with image processing algorithms. Vision systems are typically equipped with multi-source illumination to capture reflected images of the wafer surface from different angles. Through algorithms such as contrast enhancement, edge detection, and morphological manipulation, the geometric features of the defects are extracted. Defect depth can be measured using optical sectioning or confocal microscopy. All acquired data, such as the coordinates and values ​​of thickness measurement points, statistical parameters of roughness sampling areas, and the type, location, size, and depth of defects, are stored in a structured data format (e.g., multidimensional arrays or database records) and appended with timestamps and unique wafer identifiers for traceability and subsequent processing. Typical ranges of these data vary depending on the wafer material (e.g., silicon, silicon carbide, gallium nitride) and processing stage; for example, initial thicknesses range from hundreds of micrometers to several millimeters, surface roughness from tens of nanometers to several micrometers, and maximum defect depths can reach several micrometers. During the data acquisition process, the system performs real-time data verification. For example, it cross-compares data consistency using redundant sensors and marks or removes outliers that exceed the preset physical range to ensure data reliability.

[0043] S1012: Obtain the initial geometry and performance data of the currently configured grinding wheel.

[0044] This sub-step aims to obtain detailed information about the grinding wheels mounted on the Z1 and Z2 axes of the grinding machine. This information is crucial for assessing their grinding capabilities, predicting wear, and optimizing grinding strategies. The data obtained includes the type, grit size, diameter, initial thickness, bond type, abrasive material, manufacturer parameters, and historical dressing records of the grinding wheels.

[0045] The grinding wheel type (e.g., metal-bonded diamond wheel, resin-bonded diamond wheel) and mesh size (abrasive grain size, e.g., Z1 is 800 mesh, Z2 is 2000 mesh) are entered by the operator or read from the grinding wheel management system. The initial diameter and thickness of the grinding wheel are measured using high-precision laser diameter gauges and thickness gauges. These measuring devices are typically installed in dedicated stations inside or outside the grinding machine and can perform measurements automatically with micron-level accuracy. Physicochemical parameters such as bond type and abrasive material are usually obtained from datasheets provided by the manufacturer and stored in the grinding wheel database.

[0046] Historical dressing records include the cumulative number of dressings, the amount removed during each dressing, and the post-dressing wheel morphology data. This data reflects the current health and performance degradation trend of the grinding wheel. Post-dressing wheel morphology data can be obtained using a 3D scanner to assess whether the dressing effect met expectations and whether irregular wear exists. All grinding wheel data, including its unique identifier, model, service life threshold, and recommended dressing interval, is integrated into a centralized grinding wheel management database. Data verification mechanisms include checking whether input parameters conform to manufacturer specifications, whether measured values ​​fluctuate within a reasonable range, and comparing trends with historical data. For example, if the measured grinding wheel thickness is much smaller than the initial value and there is no corresponding dressing record, there may be data entry errors or abnormal wear.

[0047] S1013: Preprocess and standardize the acquired initial data.

[0048] This sub-step aims to clean, standardize, fill missing values, handle outliers, and scale features of the heterogeneous raw data obtained in S1011 and S1012, so that it can be adapted to the input requirements of the subsequent intelligent model and improve the training efficiency and prediction accuracy of the model.

[0049] Data cleaning includes removing duplicate records and correcting measurement errors. Format standardization involves converting data from different sensors or data sources into a unified numerical or coded format. For example, wafer defect types (such as scratches and mechanical marks) need to be encoded into numerical categories. For missing values, interpolation methods (e.g., linear interpolation, cubic spline interpolation) or methods based on statistical models (e.g., mean imputation, regression prediction) can be used to fill in the missing values. Outlier handling identifies, corrects, or removes measurement points that significantly deviate from the data distribution, for example, through the three-standard-deviation rule or algorithms based on isolated forests. These outliers may be caused by sensor malfunctions or environmental interference.

[0050] Feature scaling unifies features with different dimensions and numerical ranges (e.g., wafer thickness at the millimeter level, surface roughness at the nanometer level) to a similar scale, preventing certain features from dominating the model due to excessively large values. Common methods include min-max normalization (scaling the data to a range of zero to one) or Z-score standardization (converting the data into a distribution with a mean of zero and a standard deviation of one). The standardized data is then used as input to the subsequent intelligent grinding removal allocation model in the form of a multi-dimensional feature vector. For example, the initial state of a wafer might be represented as a feature vector containing one hundred or more dimensions, including thickness, roughness, density of different defects, and average depth. The initial state of the grinding wheel is similarly encoded as another feature vector, including mesh size, abrasive hardness, and bond type. These preprocessed and standardized feature vectors are securely transmitted to the intelligent grinding removal allocation module.

[0051] S102: Based on the acquired initial state data, determine the initial allocation parameters of the Z1-Z2 linkage removal amount through the intelligent allocation model of grinding removal amount.

[0052] This step is one of the core components of the method of this invention. It aims to intelligently determine the amount of material to be removed in the rough grinding (Z1) and fine grinding (Z2) stages, as well as the corresponding grinding machine operating parameters, based on the initial state of the wafer and grinding wheel using an advanced predictive model. This intelligent allocation mechanism breaks through the limitations of traditional experience-based settings or fixed-ratio allocations, achieving precise matching of grinding tasks.

[0053] S1021: Construct an intelligent allocation model for grinding removal amount.

[0054] The intelligent grinding removal amount allocation model is a predictive model based on machine learning or deep learning architecture. Its core function is to output the optimal Z1 and Z2 removal amount ratio and corresponding grinding parameters based on the input wafer initial state feature vector and grinding wheel initial state feature vector. This model is trained offline using a large amount of historical grinding data, which includes the actual grinding removal amount, grinding parameters, and final processing results (such as surface quality and grinding wheel wear) under different wafer types, defect conditions, and grinding wheel configurations.

[0055] The model can employ various architectures. For example, a time-series prediction model based on Long Short-Term Memory (LSTM) networks can be used to process historical grinding process data, or a convolutional neural network (CNN) can be used to process image features of the wafer surface morphology. In more complex scenarios, a Transformer model incorporating an attention mechanism can be used to better capture the complex nonlinear relationships between different input features. The model's input layer receives a normalized wafer initial state feature vector (e.g., containing features such as wafer thickness, flatness, defect density, and roughness) and a grinding wheel initial state feature vector (e.g., containing features such as grinding wheel mesh size, diameter, bond type, and wear state) from step S1013. The model's hidden layers learn the mapping relationship between input features and output parameters through multiple nonlinear transformations (e.g., fully connected layers, activation functions, and batch normalization layers). The model's output layer directly predicts a series of initial grinding parameters, such as Z1 removal amount, Z2 removal amount, Z1 feed rate, Z2 feed rate, Z1 spindle speed, and Z2 spindle speed. The training objective of the model is to minimize the error between the predicted parameters and the actual optimal parameters, while considering multiple optimization objectives such as processing efficiency, surface quality, and grinding wheel life.

[0056] The intelligent grinding removal amount allocation model employs a deep neural network. Its specific architecture includes: an input layer receiving the standardized feature vector, followed by three fully connected hidden layers, each containing 256, 128, and 64 neurons respectively, using the ReLU activation function. Batch normalization and Dropout (with a dropout rate of 0.2) are used between hidden layers to prevent overfitting. The output layer outputs continuous values ​​such as Z1 removal amount and Z2 removal amount using a linear activation function based on the properties of the grinding parameters. The model is trained using tens of thousands of sets of historical grinding batch data, covering different wafer materials, initial damage types, grinding wheel wear levels, and final processing quality. The training process uses the Adam optimizer with mean squared error (MSE) as the loss function, and the learning rate is initially set to 0.001 with a learning rate decay strategy. The model's validation set is used for hyperparameter tuning and early stopping.

[0057] S1022: Calculate the initial removal ratio of Z1 and Z2 based on model predictions.

[0058] In this sub-step, the intelligent allocation model for grinding removal amount receives the preprocessed and normalized wafer initial state feature vector F provided by S1013. wafer_init and the eigenvector F of the initial state of the grinding wheel wheel_init As input, the model uses an internal, complex nonlinear mapping function M. alloc The optimal Z1 removal amount R for the current grinding task is predicted. Z1 and Z2 removal amount R Z2This process can be summarized by the following mathematical expression: Among them, R Z1 and R Z2 P represents the thickness of material to be removed in the Z1 rough grinding and Z2 fine grinding stages, respectively. other It is a vector that contains key parameters other than the amount of material removed during the grinding process, such as the feed rates of Z1 and Z2, spindle speed, grinding pressure, etc. alloc The function is an intelligent allocation model built by S1021. It dynamically allocates removal tasks by comprehensively considering factors such as the initial defect level of the wafer (e.g., deeper roughness requires more Z1 removal), material hardness, target final surface roughness (e.g., higher surface quality requires more Z2 removal), and the current wear state of the grinding wheel. This dynamic allocation mechanism aims to ensure that the Z1 stage can efficiently and fully remove most irregularities and defects, reduce the pressure on the Z2 stage, and avoid over-grinding in Z1, thereby optimizing the overall grinding effect and extending the grinding wheel life. For example, for wafers with deep mechanical marks, the model will allocate a relatively large R... Z1 This ensures that defects are thoroughly removed during the rough grinding stage and prevented from propagating to the fine grinding stage; however, for wafers with good initial flatness but extremely high requirements for final surface roughness, the model may assign a relatively small R0. Z1 and larger R Z2 This is to ensure a fine grinding effect.

[0059] S1023: Determine the initial operating parameters of the grinder.

[0060] After determining the initial removal ratio of Z1 and Z2, this sub-step further determines the specific operating parameters of the biaxial DFG grinder at startup. These parameters include, but are not limited to, Z1 spindle speed, Z2 spindle speed, Z1 feed rate, Z2 feed rate, coolant flow rate, and the oscillation frequency and amplitude of the grinding disc.

[0061] In S1022, the intelligent grinding removal amount allocation model not only predicts the removal amount but also outputs these initial operating parameters. For example, based on the hardness of the wafer material (e.g., silicon carbide wafers typically require lower feed rates and higher rotation speeds to prevent edge chipping) and the grit of the grinding wheel, the model recommends an optimal combination of Z1 and Z2 spindle speeds. Higher rotation speeds generally improve removal efficiency but may also increase grinding heat and vibration, thus requiring consideration in conjunction with coolant flow rate and feed rate. The feed rate setting directly affects the grinding force and removal efficiency; the model optimizes the initial feed rate based on the target removal amount and material properties. For example, the Z1 coarse grinding stage typically uses a relatively high feed rate to quickly remove material, while the Z2 fine grinding stage uses a lower feed rate to achieve a finer surface finish.

[0062] The coolant flow rate is set to effectively remove grinding heat, prevent wafer overheating damage, and promptly remove grinding debris to reduce grinding wheel clogging. The model comprehensively considers grinding power and material heat generation characteristics to recommend the initial coolant flow rate. All these initial parameters are loaded into the grinding machine's motion controller and process controller as instructions for the first grinding cycle. Before loading the parameters, the system performs parameter validity checks, such as checking whether the rotational speed exceeds the motor's maximum limit and whether the feed rate is lower than the minimum stable feed rate, to avoid equipment damage or processing errors.

[0063] S103: Based on the initial allocation parameters, start the biaxial DFG grinding machine to perform Z1-Z2 linkage grinding, and monitor the status data during the grinding process in real time.

[0064] This step describes the actual execution and data acquisition of the grinding process, and is a crucial link connecting intelligent decision-making and physical execution. It ensures that the grinder operates according to the optimized parameters determined in S102, while continuously acquiring multi-dimensional real-time data during the grinding process, providing a data foundation for subsequent adaptive control and anomaly handling.

[0065] S1031: Initialize the grinder to run to the set parameters.

[0066] Upon receiving the initial operating parameters determined by S1023, including Z1 spindle speed, Z2 spindle speed, Z1 feed rate, Z2 feed rate, and coolant flow rate, the central control unit of the dual-axis DFG grinding machine sends instructions to each actuator. First, the workpiece clamping module precisely positions and fixes the wafer to be ground onto the grinding disc. Then, the cooling and chip removal system activates, supplying grinding fluid to the grinding area at a preset flow rate to ensure adequate cooling and timely removal of grinding debris during the grinding process. The Z1 and Z2 spindle motors accelerate until they reach the set speeds. The Z-axis servo system controls the Z1 and Z2 grinding wheels to gradually descend according to their respective initial feed rates until stable contact is established with the wafer surface and material removal begins. The entire initialization process is typically completed within a few seconds, accompanied by real-time monitoring of the status of each actuator, such as motor current, speed feedback, and valve opening, to ensure all parameters are accurately set and operating stably. Any initialization failure or parameter deviation will trigger an alarm and pause the grinding task.

[0067] S1032: Start the grinding process and continuously collect multi-dimensional real-time data.

[0068] Once the grinding machine reaches a stable operating state, various sensors installed on the grinding equipment begin to collect multi-dimensional real-time data in parallel at high sampling frequencies (e.g., 100 Hz to 1 kHz). This data is the core basis for adaptive control of the grinding process and optimization of grinding wheel wear.

[0069] The types of data collected include: 1. Grinding Force and Torque Data: Measured in real time by triaxial force and torque sensors mounted on the spindles of grinding wheels Z1 and Z2. These data directly reflect the intensity of the interaction between the grinding wheel and the wafer and the removal efficiency. For example, excessive grinding force may lead to wafer damage or rapid wheel wear, while insufficient grinding force may indicate inadequate removal efficiency or the presence of air cuts.

[0070] 2. Wafer Thickness Data: Wafer thickness is monitored online in real time during the grinding process using a non-contact laser thickness gauge or eddy current sensor. Measurement points are typically distributed at different radial locations on the wafer to capture changes in the overall thickness distribution. Measurement accuracy is usually below micrometers, with a fast response time, capable of capturing transient trends in thickness changes.

[0071] 3. Temperature Data: The surface temperature of the area where the grinding wheel contacts the wafer is measured using an infrared temperature sensor or thermocouple array. Localized overheating is a significant cause of wafer damage and grinding wheel binder aging; therefore, accurate temperature monitoring is crucial for preventing thermal damage.

[0072] 4. Vibration and Acoustic Data: Mechanical vibrations and generated sound waves during the grinding process are monitored using accelerometers and acoustic sensors (e.g., microphones or acoustic emission sensors). Abnormal vibration patterns or high-frequency acoustic emission signals may indicate wheel clogging, abrasive grain shedding, or wafer abnormalities.

[0073] 5. Spindle motor current and voltage data: These reflect the actual load and grinding power of the spindle motor, and can indirectly assess the grinding status and the contact between the grinding wheel and the workpiece.

[0074] All collected real-time data, along with its precise timestamps, are transmitted via a high-speed data bus to the data acquisition and monitoring system for preliminary processing.

[0075] S1033: Perform preliminary processing and caching of real-time acquired data.

[0076] Following data acquisition, this sub-step performs preliminary preprocessing and temporary storage of the massive amounts of raw real-time data to facilitate efficient analysis by subsequent modules. Preliminary processing includes: 1. Time synchronization: Since different sensors may have different sampling frequencies and start-up sequences, it is necessary to timestamp all data to ensure that all measurements correspond to the same physical time, so as to facilitate subsequent multimodal data fusion analysis.

[0077] 2. Filtering and Noise Reduction: Various electromagnetic interferences and mechanical vibrations exist in the grinding environment, which may cause a large amount of noise in the sensor signal. By applying digital filters (e.g., median filtering, Kalman filtering, wavelet transform), high-frequency noise is removed, the effective signal is retained, and the signal-to-noise ratio of the data is improved.

[0078] 3. Unit Conversion and Formatting: Convert the raw sensor output (e.g., analog voltage signal) into a value with standard physical units (e.g., Newton, micrometer, degree Celsius) and format it into a uniform data structure, such as a fixed-length time series vector or frame.

[0079] 4. Real-time caching: Processed data is temporarily stored in a high-speed cache (e.g., a random access memory array or a solid-state drive array) for real-time access and analysis by the grinding process adaptive control module and the Air Cut anomaly handling module. This caching mechanism ensures data stream continuity and low latency, meeting the requirements of real-time control.

[0080] S104: Based on real-time monitored grinding process status data, the grinding parameters are dynamically adjusted through adaptive control of the grinding process and Air Cut anomaly handling strategy.

[0081] This step is crucial for achieving precise control and robust operation of the grinding process in this invention. By analyzing the real-time data collected by S103, the system can dynamically adjust grinding parameters to cope with transient changes in the wafer and grinding wheel, and promptly handle potential air cuts or abnormal situations.

[0082] S1041: Real-time assessment of grinding status and target deviation.

[0083] This sub-step continuously compares the real-time grinding data processed in S1033 with the target parameters set in S102, calculating the deviation between the current grinding state and the preset optimization target. Key evaluation indicators include: 1. Grinding force deviation: The difference between the Z1 and Z2 grinding forces measured in real time and the target grinding force. The target grinding force is set in S102 based on the characteristics of the grinding wheel, the wafer material, and the required removal efficiency.

[0084] 2. Removal Rate Deviation: The difference between the current material removal rate, calculated from real-time thickness measurement data, and the target removal rate. For example, if the removal rate is lower than expected, it may indicate wheel passivation or insufficient feed.

[0085] 3. Thickness uniformity deviation: The deviation between the real-time thickness difference at different measurement points on the wafer and the target thickness uniformity index.

[0086] 4. Temperature deviation: The deviation between the real-time temperature of the grinding area and the upper limit of the safe temperature or the target stable temperature.

[0087] 5. Vibration spectrum analysis: The real-time vibration data is analyzed by means of Fourier transform and other methods to identify the presence of abnormal frequency components and compare them with the baseline vibration mode.

[0088] These deviation values ​​serve as inputs to the control system, and are weighted and integrated through a feedback control algorithm to form a comprehensive grinding status evaluation signal. The evaluation frequency is matched with the data sampling frequency to ensure timely response to rapid changes in the grinding process.

[0089] S1042: Determine if there is an Air Cut condition or grinding abnormality.

[0090] Based on continuous evaluation of the grinding status, this sub-step uses preset thresholds and anomaly detection algorithms to determine whether an air cut (partial no-load) has occurred or other grinding anomalies exist.

[0091] Air Cut Detection: The system determines that an air cut has occurred when the real-time grinding force of the Z1 or Z2 grinding wheel remains consistently below the preset air cut force threshold. For example, the target grinding force of the Z1 grinding wheel is typically between several Newtons and tens of Newtons. If the real-time force remains consistently below 0.5 Newtons or 1 Newton, an air cut may occur. This usually happens when local high points on the wafer surface have been removed, or when wafer warping causes some areas to lose contact with the grinding wheel. Additionally, a significant decrease in the spindle motor current can also serve as an auxiliary criterion.

[0092] Judgmentation of grinding abnormalities: 1. Force / torque abrupt change: The grinding force or torque rises or falls abnormally rapidly within a very short time (e.g., tens of milliseconds), which may indicate that the wafer clamping is loose, the grinding wheel is partially broken, or foreign objects have entered the grinding area.

[0093] 2. Abnormal temperature: If the temperature in the grinding area continues to exceed the safe limit or there is an abnormal temperature surge, it indicates that there may be a risk of overheating or thermal damage.

[0094] 3. Abnormal vibration: The root mean square value of the vibration signal continues to rise, or new high-energy frequency peaks appear in the spectrum, which may be related to uneven wear of the grinding wheel, spindle imbalance or wafer breakage.

[0095] 4. Sudden drop / sudden increase in removal rate: The removal rate is significantly different from the expected rate, which may indicate wheel clogging (sudden drop) or a sudden change in wafer material (sudden increase).

[0096] The judgment logic is typically implemented using a rule-based expert system combined with statistical methods (e.g., moving average, exponential smoothing) and machine learning models (e.g., support vector machines for classifying abnormal patterns). Once an air cut or grinding abnormality is detected, the system immediately triggers the corresponding processing mechanism.

[0097] S1043: Perform dynamic adaptive adjustment of grinding parameters.

[0098] When S1041 detects a deviation between the grinding state and the target, or when S1042 determines the Air Cut condition, this sub-step will adjust the grinding parameters in real time and dynamically according to the preset control strategy so that the grinding process returns to the optimal state.

[0099] Grinding parameter adjustment mechanism: 1. Feed Rate Adjustment: This is the most commonly used adjustment method. When the grinding force is lower than the target value (or the removal rate is lower than expected), the system increases the corresponding Z-axis feed rate to improve removal efficiency. Conversely, when the grinding force is too high or the removal rate is too fast, the feed rate is reduced to protect the wafer and grinding wheel. The real-time feed rate V of the Z-axis... f (t) Requires adjustment based on real-time grinding force F grind (t) and the target grinding force F target The deviation is adaptively adjusted. Assuming a positive correlation between grinding force and feed rate, the adjustment amount ΔV of the feed rate is then... f It can be represented as: Among them, G p This is the grinding force feedback gain coefficient, used to adjust the sensitivity of the response. When the real-time grinding force is lower than the target force, the feed rate increases; conversely, it decreases. This strategy helps maintain the stability of the grinding process and removal efficiency. This adjustment, based on proportional control of the grinding force feedback, enables a fast response. More complex control algorithms, such as model predictive control (MMCC) or reinforcement learning-based strategies, can consider multi-objective optimization (efficiency, quality, loss) and predict future states for adjustments.

[0100] 2. Spindle speed adjustment: In some situations, such as when the grinding force is consistently low and there are many hard spots on the wafer surface, the spindle speed can be appropriately increased to increase the cutting speed of the grinding wheel, thereby improving the removal capacity. However, speed adjustment should be done with caution to avoid introducing excessive heat or vibration.

[0101] 3. Grinding pressure adjustment: The grinding pressure is adjusted by changing the downward pressure on the Z-axis. This is similar to the feed rate adjustment, but it may be achieved through a different control loop.

[0102] Air Cut Adjustment Strategy: When an air cut is detected, the system executes a specific compensation strategy. For example, the feed rate of the Z1 or Z2 grinding wheel can be temporarily increased to quickly re-establish effective contact with the wafer surface. Simultaneously, the system may record the area and duration of the air cut for subsequent analysis of the wafer's initial flatness or the wear of the grinding wheel. For air cuts in the Z1 rough grinding stage, a larger feed rate increase is typically allowed to remove voids as quickly as possible; while the air cut process in the Z2 fine grinding stage is gentler to avoid impacting the fine surface.

[0103] S1044: Emergency strategies for handling Air Cut or abnormal situations.

[0104] When S1042 detects a serious grinding anomaly, this sub-step will immediately trigger an emergency handling mechanism to prevent equipment damage, wafer scrapping, or operator injury. Emergency strategies include: 1. Pause or stop grinding: When a grinding force overload, spindle motor stall, temperature runaway or strong abnormal vibration is detected, the system will immediately perform a grinding pause or emergency stop operation, raise the grinding wheel to a safe position, and stop all grinding operations.

[0105] 2. Alarm and Log Recording: Issues audible and visual alarms to the operator and displays detailed anomaly information on the control interface. Simultaneously, it records the time and type of the anomaly, all relevant sensor data, and system status parameters in the event log for post-incident analysis and troubleshooting.

[0106] 3. Diagnostic and Recovery Recommendations: In more advanced systems, pre-defined fault trees or expert knowledge bases can be used to perform preliminary diagnoses of the causes of anomalies and provide operators with possible recovery steps or maintenance recommendations. For example, if a grinding wheel blockage is detected, the system may recommend performing a dressing cycle.

[0107] 4. Parameter rollback and safety reset: The system may roll back some key parameters to a safe state before the anomaly occurred and perform a safety reset on the relevant actuators to ensure a safe restart after the fault is cleared.

[0108] For non-urgent but persistent air cut situations, if the problem persists after the system performs parameter adjustments in S1043, it may be recorded as a potential problem and the operator will be reminded to check the flatness of the wafer fixture or grinding pad after the grinding cycle ends.

[0109] S105: Continuously monitor grinding wheel wear data and evaluate and optimize grinding strategies through grinding wheel wear prediction and optimization feedback mechanisms.

[0110] This step is crucial for maximizing grinding wheel life and reducing production costs in this invention. By monitoring and predicting the wear state of the grinding wheel in real time, and iteratively optimizing the grinding strategy based on the prediction results, the service life of the grinding wheel can be effectively extended, the number of dressing operations reduced, and stable grinding performance maintained.

[0111] S1051: Real-time monitoring and cumulative grinding wheel wear related indicators.

[0112] Throughout the grinding process, the system continuously collects and accumulates multiple indicators directly or indirectly related to grinding wheel wear. These indicators include: 1. Grinding wheel thickness variation: The remaining thickness of the grinding wheel is measured periodically or continuously using a high-precision laser thickness gauge or a contact displacement sensor. Each measurement is recorded and compared with the initial thickness of the grinding wheel and historical dressing data to calculate the actual wear. The accuracy can reach the micrometer level.

[0113] 2. Grinding wheel morphology changes: During non-grinding periods or specific maintenance windows, surface morphology data of the grinding wheel is acquired using a 3D optical scanner to assess the sharpness, uniformity, and wear of the abrasive grains and the bond. For example, analyzing the distribution of abrasive grain protrusion height can determine whether the grinding wheel is dulled.

[0114] 3. Grinding Power Accumulation: Instantaneous grinding power is calculated based on real-time current and voltage data of the spindle motor, and the accumulated grinding power consumption is obtained by time integration. Grinding power consumption is closely related to the amount of material removed and the wear of the grinding wheel.

[0115] 4. Acoustic Emission Signal Characteristics: Acoustic emission sensors can capture weak high-frequency sound waves generated when abrasive grains break, detach, or the grinding wheel surface becomes clogged. By analyzing the amplitude, count rate, energy, and other characteristics of the acoustic emission signal, the microscopic wear state and degree of clogging of the grinding wheel can be assessed in real time.

[0116] 5. Vibration spectrum changes: Uneven wear or clogging of the grinding wheel can lead to specific vibration modes during grinding. By continuously monitoring the output of the vibration sensor and performing spectrum analysis, changes in vibration characteristics related to the condition of the grinding wheel can be identified.

[0117] All of this data is timestamped and stored in the grinding wheel management database, forming a complete lifecycle profile of the grinding wheel and providing a rich data foundation for subsequent wear prediction.

[0118] S1052: Assess the current health status of the grinding wheel using a grinding wheel wear prediction model.

[0119] This sub-step uses the grinding wheel wear-related indicators accumulated in S1051 as input, and through a pre-trained grinding wheel wear prediction model, evaluates the current health status of the grinding wheel in real time and predicts its remaining effective life or the next dressing window.

[0120] Grinding wheel wear prediction models can be based on various techniques, such as support vector regression machines, recurrent neural networks (RNNs), or deep learning models. The training data for the model comes from historical grinding wheel usage data, including the initial parameters of the grinding wheel, cumulative grinding power consumption, measured wear depth, performance degradation (such as decreased removal efficiency and deterioration of surface quality), and dressing timing.

[0121] The real-time wear depth W(t) of the grinding wheel can be obtained from the accumulated grinding power E. grind and the wear coefficient K of the grinding wheel wear A prediction is made. Considering the initial state and material properties of the grinding wheel, the wear depth prediction model can be expressed as: Among them, W initial This represents the initial effective wear depth of the grinding wheel. For a moment Instantaneous grinding power, K wear This represents the wear coefficient related to the grinding wheel material and bond type. This model allows for real-time assessment of the remaining life of the grinding wheel and optimization of dressing timing.

[0122] This formula provides a basic linear cumulative wear model. In practical applications, the model becomes more complex, for example, by taking multiple features such as grinding force, temperature, and vibration as inputs to predict the overall health index of the grinding wheel. The model output includes the current percentage of wear on the grinding wheel, the predicted number of wafers that can be ground, or the predicted time point for the next dressing. For example, when the wear on the grinding wheel reaches 70%, the system will mark it as "needs attention," and when it reaches 90%, it will mark it as "needs dressing."

[0123] S1053: Generate grinding strategy optimization suggestions.

[0124] Based on the grinding wheel wear prediction results of S1052, this sub-step generates specific grinding strategy optimization suggestions to balance grinding wheel life, processing efficiency and product quality.

[0125] Optimization suggestions may include: 1. Dressing Timing Recommendation: When the grinding wheel wear prediction model indicates that the grinding wheel performance has decreased to a certain threshold, the system recommends performing a dressing operation. For example, if grinding wheel passivation leads to increased grinding force and decreased removal efficiency, the system will recommend a dressing operation to restore the grinding wheel's cutting ability.

[0126] 2. Z1-Z2 Removal Ratio Adjustment: If the prediction indicates that the Z2 grinding wheel is wearing too quickly, it may mean that the initial defects or unevenness of the wafer were not sufficiently removed during the Z1 rough grinding stage. The system will suggest appropriately increasing the removal amount of Z1 in subsequent grinding tasks to reduce the load on Z2 and thus extend the life of the Z2 grinding wheel. Conversely, if the Z1 grinding wheel is wearing significantly, it may be recommended to adjust the Z1 removal amount.

[0127] 3. Fine-tuning of grinding parameters: Based on the real-time wear condition of the grinding wheel, it is recommended to fine-tune parameters such as feed rate, rotation speed, and coolant flow rate of Z1 and Z2. For example, for a slightly passivated grinding wheel, the feed rate or rotation speed can be appropriately increased to maintain the removal efficiency, but it must be within the safe tolerance range of the grinding wheel and the wafer.

[0128] 4. Grinding wheel replacement recommendation: When the wear of the grinding wheel reaches the end of its design life or irreversible serious damage occurs, the system will recommend replacing the grinding wheel with a new one.

[0129] These optimization suggestions are presented in the form of structured instructions, which can be confirmed and executed by the operator, or fed back directly to the intelligent allocation model for grinding removal in a fully automated system for iterative updates.

[0130] S1054: Feedback optimization suggestions to the intelligent allocation model for grinding removal amount.

[0131] This sub-step is crucial for the closed-loop optimization of the grinding strategy. The optimization suggestions generated in S1053, especially the suggestions on adjusting the Z1-Z2 removal ratio and fine-tuning the grinding parameters, will be fed back to the intelligent grinding removal allocation model as new training samples or update rules (S1021).

[0132] Through this feedback mechanism, the intelligent allocation model can continuously learn and adapt to the actual wear conditions and performance changes of the grinding wheels, thus providing more accurate and forward-looking parameters when allocating grinding tasks to new wafers in the future. For example, if the system discovers through historical data analysis that the wear pattern of grinding wheel Z1 changes under specific conditions, then in subsequent grinding tasks, the intelligent allocation model will consider this new grinding wheel wear characteristic when calculating the removal amounts of Z1 and Z2, dynamically adjusting its allocation strategy to balance removal efficiency and grinding wheel life. This iterative optimization process enables the entire grinding system to possess self-learning and adaptive capabilities, continuously improving grinding efficiency, extending grinding wheel life, and reducing operating costs.

[0133] Dual-axis DFG grinding machine system

[0134] The method described in this invention can be implemented using a biaxial DFG polishing machine system. This system is a highly integrated automated device designed to support refined and intelligent wafer polishing processes.

[0135] 1. Workpiece Clamping and Conveying Module: This module is responsible for the precise loading, positioning, clamping, and unloading of wafers after grinding. Wafers are securely fixed to the grinding disc using vacuum chucks or mechanical clamps, ensuring stability and positioning accuracy during high-speed grinding. This module also includes a wafer transfer robotic arm, responsible for transferring wafers from the feed hopper to the grinding disc, and then transferring them out after grinding.

[0136] 2. Z1 Rough Grinding Unit: This unit is equipped with a Z1 grinding wheel (typically a high-grit diamond wheel, such as 800 grit) for rough grinding of wafers. This unit features an independent, high-precision Z-axis servo control system, capable of achieving micron-level vertical feed accuracy. The Z1 spindle is driven by a high-performance servo motor, allowing for a wide range of speed adjustments to accommodate different grinding materials and removal efficiency requirements. A triaxial force sensor and torque sensor are integrated on the spindle for real-time monitoring of grinding force and torque. Coolant nozzles and chip removal ports are also integrated into this unit to ensure effective cooling of the grinding area and efficient chip removal.

[0137] 3. Z2 Fine Grinding Unit: This unit is configured in parallel or series with the Z1 coarse grinding unit, equipped with Z2 grinding wheels (typically diamond wheels with smaller or higher grit, such as 2000 grit) for fine grinding of wafers. Similar to the Z1 unit, the Z2 unit also features an independent high-precision Z-axis servo control system and a high-performance spindle motor. Its design prioritizes fine grinding and surface quality, therefore, it typically has higher requirements for vibration suppression and spindle runout control. Similarly, force sensors, torque sensors, coolant supply, and chip removal systems are all standard configurations.

[0138] 4. Data Acquisition and Monitoring System: This is a distributed sensor network and a centralized data processor. It includes various high-precision sensors, such as non-contact laser thickness gauges, infrared temperature sensors, accelerometers, acoustic emission sensors, and electrical sensors for monitoring motor current and voltage. These sensors continuously acquire multi-dimensional data during the grinding process at a high sampling frequency. The data acquisition system is responsible for preprocessing, time-synchronizing, and formatting the raw data, and then transmitting it to the central control unit.

[0139] 5. Central Control Unit: As the brain of the entire grinding machine, the central control unit is a high-performance industrial computer equipped with a real-time operating system. It integrates the following key software modules:

[0140] The intelligent allocation module for grinding removal amount runs the intelligent allocation model described in S102 and calculates the initial removal amount and operating parameters of Z1 and Z2 based on the initial wafer and grinding wheel data.

[0141] The adaptive control module for the grinding process implements the real-time grinding status assessment and dynamic parameter adjustment logic in S104. Based on sensor feedback data, it performs closed-loop control of the Z-axis feed rate, spindle speed, and other parameters.

[0142] Air Cut Anomaly Handling Module: Responsible for Air Cut condition judgment and emergency response strategies in S104.

[0143] Grinding wheel wear prediction and optimization feedback module: Runs the grinding wheel wear prediction model described in S105, evaluates the health status of the grinding wheel, and generates optimization suggestions to be fed back to the intelligent allocation module.

[0144] Motion control module: Directly controls actuators such as the Z-axis servo motor, spindle motor, and coolant pump to achieve precise motion control and flow regulation.

[0145] 6. Cooling and Chip Removal System: This system includes a coolant pump, filter, flow controller, nozzle array, and waste liquid collection device. It is responsible for providing a stable flow of polishing fluid, effectively removing polishing heat, and flushing away polishing chips in the polishing area to prevent them from accumulating on the grinding wheel or wafer surface.

[0146] 7. Human-Machine Interface: Provides an intuitive graphical user interface, through which operators can input task information, view the real-time status of the grinding process, monitor the health of the grinding wheel, receive abnormal alarms and optimization suggestions, and perform manual intervention or parameter adjustments.

[0147] The system tightly couples the various functional modules through a central control unit, realizing closed-loop management from data acquisition, intelligent decision-making, physical execution to process feedback, ensuring that the dual-axis DFG grinding machine has high precision, high efficiency and high reliability in wafer processing.

Claims

1. A method for controlling the removal amount and optimizing wheel wear in a biaxial DFG grinding mill with Z1-Z2 linkage, characterized in that, include: Obtain initial surface morphology data and wafer material property parameters of the wafer to be ground; Based on the initial surface morphology data and wafer material characteristic parameters, the initial grinding parameters and target removal amount of the Z1 coarse grinding stage are determined by a multi-objective optimization algorithm. The multi-objective optimization algorithm aims to maximize the defect removal rate, minimize subsurface damage, and minimize abnormal wear of the grinding wheel, and optimizes the parameters under the constraints of satisfying the maximum allowable grinding force and the minimum material removal rate. Z1 rough grinding is performed, and the grinding status of the Z1 axis is monitored in real time during the Z1 rough grinding process. The intelligent allocation model for grinding removal amount receives the preprocessed and standardized wafer initial state feature vector F. wafer_init and the eigenvector F of the initial state of the grinding wheel wheel_init As input, the model uses an internal, complex nonlinear mapping function M. alloc The optimal Z1 removal amount R for the current grinding task is predicted. Z1 and Z2 removal amount R Z2 This process can be summarized by the following mathematical expression: Among them, R Z1 and R Z2 P represents the thickness of material to be removed in the Z1 rough grinding and Z2 fine grinding stages, respectively. other It is a vector; Based on the grinding status data of the Z1 coarse grinding stage and the target removal amount of Z1 coarse grinding, the Z1 coarse grinding parameters are adaptively adjusted. This includes when the Z1 axis grinding force or grinding power abnormally increases and exceeds the preset safety threshold, or when the Z1 axis grinding force or grinding power abnormally increases and exceeds the preset safety threshold in a short period of time, the Z1 coarse grinding wheel is rapidly raised by 0.5 mm to 1 mm for a duration of 50 ms to 100 ms to avoid triggering a local Air Cut strategy due to wheel edge chipping or wafer damage. Then, the grinding state is re-entered with the adjusted grinding parameters until the Z1 coarse grinding is completed and the wafer surface morphology after Z1 coarse grinding is obtained. Based on the wafer surface morphology after Z1 coarse grinding, the wafer material characteristic parameters, and the set total removal amount of fine grinding, the target removal amount and initial grinding parameters of the Z2 fine grinding stage are determined by the intelligent allocation model of grinding removal amount. Perform Z2 fine grinding and monitor the grinding status data of the Z2 axis in real time during the Z2 fine grinding process; Based on the grinding state data of the Z2 fine grinding stage and the target removal amount of the Z2 fine grinding, the Z2 fine grinding parameters are adaptively adjusted. When the Z2 axis grinding force or grinding power fluctuates abnormally in a short period of time, and the fluctuation exceeds the set grinding force threshold or grinding power threshold, the Air Cut strategy is triggered to adjust the Z2 fine grinding depth or grinding pressure until the Z2 fine grinding is completed. Throughout the grinding process, grinding status data for the Z1 and Z2 axes are continuously collected. Combined with grinding wheel model information, the wear status and remaining life of the grinding wheel are assessed using a grinding wheel wear prediction model. This model employs a temporal convolutional network or a long short-term memory network, predicting wear by learning the correlation between historical grinding data and actual grinding wheel wear. Grinding parameters are adjusted based on the assessment results. The real-time wear depth W(t) of the grinding wheel can be measured by the accumulated grinding power consumption E. grind The wear coefficient K of the grinding wheel wear To make predictions, considering the initial state of the grinding wheel and its material properties, the wear depth prediction model can be expressed as follows: Among them, W initial This represents the initial effective wear depth of the grinding wheel. For a moment Instantaneous grinding power, K wear This is the wear coefficient related to the grinding wheel material and the type of bond.

2. The method for controlling the removal amount and optimizing wheel wear of a biaxial DFG grinding mill Z1-Z2 linkage according to claim 1, characterized in that, The process of obtaining the initial surface morphology data and wafer material property parameters of the wafer to be ground includes: The surface of the wafer to be ground is scanned non-contactly using a three-dimensional optical profilometer to obtain high-resolution three-dimensional height mapping data. The horizontal resolution of the three-dimensional height mapping data is less than or equal to one micrometer, and the vertical resolution is less than or equal to one hundred nanometers. Based on the three-dimensional height mapping data, the damage types on the wafer surface are identified and classified through image processing and feature extraction algorithms. The damage types include frosted texture, mechanical marks, hard spots, scratches or local uneven areas, and the depth, width and distribution density of each damage area are calculated. The material properties of the wafer to be ground are obtained through a data interface. These material properties include material hardness, Young's modulus, fracture toughness, coefficient of thermal expansion, and crystal orientation.

3. The method for controlling the removal amount and optimizing wheel wear in the Z1-Z2 linkage of a biaxial DFG grinding mill according to claim 1, characterized in that, The process of performing Z1 rough grinding and monitoring the grinding status data of the Z1 axis in real time during the Z1 rough grinding includes: The Z1 coarse grinding is performed by a Z1 coarse grinding wheel, and the mesh size of the Z1 coarse grinding wheel ranges from 800 to 1500. The normal and tangential grinding forces of the Z1 axis are monitored in real time by a piezoelectric force sensor installed on the Z1 axis grinding spindle. The response time of the force sensor is less than one millisecond. The current and voltage of the Z1 axis spindle motor are monitored in real time by current and voltage sensors, and the actual grinding power of the Z1 axis spindle motor is calculated. The actual rotational speed of the Z1 coarse grinding wheel is monitored in real time using a speed sensor. The grinding vibration signal of the Z1 axis is monitored in real time using a vibration sensor; The actual feed depth of the Z1 coarse grinding wheel is monitored in real time by a high-precision displacement sensor, thereby calculating the amount of material removed in real time.

4. The method for controlling the removal amount and optimizing wheel wear of a biaxial DFG grinding mill Z1-Z2 linkage according to claim 1, characterized in that, The process of performing Z2 fine grinding and monitoring the grinding status data of the Z2 axis in real time during the Z2 fine grinding includes: The Z2 fine grinding is performed by a Z2 fine grinding wheel, and the mesh size of the Z2 fine grinding wheel ranges from three thousand to eight thousand. The normal and tangential grinding forces of the Z2 axis are monitored in real time by a piezoelectric force sensor installed on the Z2 axis grinding spindle. The response time of the force sensor is less than one millisecond. The current and voltage of the Z2 axis spindle motor are monitored in real time by current and voltage sensors, and the actual grinding power of the Z2 axis spindle motor is calculated. The actual rotational speed of the Z2 precision grinding wheel is monitored in real time using a speed sensor. The grinding vibration signal of the Z2 axis is monitored in real time using a vibration sensor; The actual feed depth of the Z2 fine grinding wheel is monitored in real time by a high-precision displacement sensor, thereby calculating the amount of material removed in real time.

5. The method for controlling the removal amount and optimizing wheel wear of a biaxial DFG grinding mill Z1-Z2 linkage according to claim 1, characterized in that, The Z2 grinding parameters are adaptively adjusted based on the grinding state data of the Z2 grinding stage and the target removal amount of the Z2 grinding. When the Z2 axis grinding force or grinding power fluctuates abnormally within a short period of time, and the fluctuation exceeds the set grinding force threshold or grinding power threshold, the Air Cut strategy is triggered to adjust the Z2 grinding depth or grinding pressure until the Z2 grinding is completed, including: The real-time monitored Z2 axis grinding status data is input to the adaptive control algorithm module, which includes a proportional-integral-derivative controller and a fuzzy logic controller. The adaptive control algorithm module dynamically adjusts the grinding pressure, spindle speed, and feed rate of the Z2 fine grinding wheel based on the deviation between the grinding state data and the target removal amount of Z2 fine grinding, so as to maintain a stable material removal rate and grinding force. When the Z2 axis grinding force or grinding power fluctuates abnormally within a short period of time, and the fluctuation exceeds the set grinding force threshold or grinding power threshold, the Air Cut strategy is triggered. The Air Cut strategy specifically includes: the grinding machine control system issuing a command to rapidly lift the Z2 fine grinding wheel from the wafer surface within ten milliseconds, with a lifting distance of 0.2 mm to 0.5 mm, and the lifting duration being 50 milliseconds to 100 milliseconds; During the Air Cut, the system reassesses the wafer surface condition and the causes of any abnormalities in the polishing force and polishing power. After the evaluation is completed, the system recalculates and adjusts the grinding pressure and feed speed of the Z2 fine grinding wheel based on the evaluation results and the current grinding progress, and then makes the Z2 fine grinding wheel re-contact the wafer surface with the adjusted parameters for grinding.

6. A system for controlling the amount of material removed and optimizing wheel wear in a biaxial DFG grinding mill with Z1-Z2 linkage, comprising the method of any one of claims 1-5, characterized in that, include: The wafer initial state acquisition module is used to acquire the initial surface morphology data and wafer material property parameters of the wafer to be ground; The Z1 coarse grinding parameter determination module is used to determine the initial grinding parameters and target removal amount of the Z1 coarse grinding stage based on the initial surface morphology data and wafer material characteristic parameters. The Z1 coarse grinding execution and monitoring module is used to execute Z1 coarse grinding and monitor the grinding status data of the Z1 axis in real time during the Z1 coarse grinding process. The Z1 coarse grinding adaptive adjustment module is used to adaptively adjust the Z1 coarse grinding parameters according to the grinding state data of the Z1 coarse grinding stage and the target removal amount of Z1 coarse grinding, until the Z1 coarse grinding is completed and the wafer surface morphology after Z1 coarse grinding is obtained. This includes triggering a local air cut strategy when the Z1 axis grinding force or grinding power abnormally increases beyond a preset safety threshold. The intelligent grinding removal amount allocation module is used to determine the target removal amount and initial grinding parameters of the Z2 fine grinding stage based on the wafer surface morphology after Z1 coarse grinding, the wafer material characteristic parameters, and the set total removal amount of fine grinding, through the intelligent grinding removal amount allocation model. The Z2 fine grinding execution and monitoring module is used to execute Z2 fine grinding and monitor the grinding status data of the Z2 axis in real time during the Z2 fine grinding process. The Z2 fine grinding adaptive adjustment module is used to adaptively adjust the Z2 fine grinding parameters based on the grinding state data of the Z2 fine grinding stage and the target removal amount of the Z2 fine grinding. The Z2 Air Cut triggering and adjustment module is used to trigger the Air Cut strategy and adjust the Z2 fine grinding depth or grinding pressure when the Z2 axis grinding force or grinding power fluctuates abnormally in a short period of time and the fluctuation exceeds the set grinding force threshold or grinding power threshold, until the Z2 fine grinding is completed. The grinding wheel wear prediction and optimization module is used to continuously collect grinding status data of Z1 axis and Z2 axis throughout the grinding process, and combine it with grinding wheel model information to evaluate the grinding wheel wear status and remaining life through the grinding wheel wear prediction model, and adjust the grinding parameters according to the evaluation results.

Images