Silicon carbide dicing control method and system based on wafer yield

Abstract

This application relates to the field of semiconductor manufacturing technology and discloses a silicon carbide dicing control method and system based on wafer yield. The method first collects dicing parameters and real-time yield data, calculates and sorts sensitivity coefficients to identify key sensitive parameters, then retrieves historical high-yield optimal parameter combinations to generate an initial scheme, compares the yield threshold to determine deviations and basic adjustment rules, then adjusts parameters according to sensitivity sorting, adaptively switches the step size based on yield response ratio, and calculates parameter correction amounts by fusing deviations and rules; simultaneously, invalid parameter combinations are stored in a taboo table, and dicing instructions are generated based on the correction amounts to execute the dicing; finally, new batch data is collected to update the sensitivity coefficients and taboo table, archives the records, and iterates continuously to form a continuous optimization closed loop. This application continuously improves yield and production efficiency through a self-learning closed loop, fundamentally solving the problems of low dicing quality and production efficiency caused by manually set parameters and lack of yield feedback.

Description

Technical Field

[0001] This application relates to the field of semiconductor manufacturing technology, and more specifically, to a silicon carbide dicing control method and system based on wafer yield. Background Technology

[0002] Silicon carbide (SiC), as a representative of third-generation semiconductor materials, possesses excellent properties such as a large bandgap, high breakdown field strength, and high thermal conductivity, and is widely used in power devices, radio frequency devices, and other fields. In the manufacturing process of SiC devices, wafer dicing is one of the key processes. Currently, mainstream SiC wafer dicing technologies include diamond wire sawing and laser cutting. However, existing dicing methods mainly rely on operator experience to set dicing parameters, lacking an intelligent parameter optimization mechanism based on wafer yield feedback. This results in a weak correlation between dicing parameters and actual wafer quality, making it impossible to dynamically adjust the dicing process based on real-time yield data, thus affecting the final wafer quality and production efficiency. Summary of the Invention

[0003] The main objective of this application is to provide a silicon carbide dicing control method and system based on wafer yield, aiming to solve the technical problems of silicon carbide wafer dicing relying on manual experience to set parameters and lacking yield feedback optimization, resulting in low dicing quality and production efficiency.

[0004] The first aspect of this application proposes a silicon carbide dicing control method based on wafer yield, comprising: Collect the cutting parameters and real-time yield data of the current batch of silicon carbide wafers, calculate the yield change range to generate and sort the sensitivity coefficients, and identify the key sensitive parameters; Based on key sensitive parameters, the optimal parameter combination of historical high-yield batches is retrieved to generate an initial parameter scheme. The deviation value is obtained by comparing with the yield threshold, and the basic adjustment rules are determined. Starting with the initial parameter scheme, the parameters are adjusted according to the sensitivity coefficient, the step size is adaptively switched based on the yield response ratio, and the parameter correction amount is calculated by integrating the deviation value and the adjustment rule. Parameter combinations that do not improve yield are stored in a taboo list to exclude invalid combinations. Control instructions are generated based on parameter correction amounts and the next batch of cutting is executed. Collect new batch yield data to update sensitivity coefficients and taboo tables, archive and adjust paths and taboo records, and return to execute initial data collection to form an iterative optimization loop.

[0005] Furthermore, the steps of collecting the cutting parameters and real-time yield data of the current batch of silicon carbide wafers, calculating the yield variation range to generate and sort sensitivity coefficients, and identifying key sensitive parameters include: Each cutting parameter and real-time yield data are collected to form paired samples. After removing abnormal samples, the valid data sequence is retained. Divide the parameter value range into multiple intervals, calculate the average yield of each interval, calculate the jump difference of the average yield of adjacent intervals, and mark the mutation-sensitive parameters. Calculate the yield gradient change rate and its variance of mutation-sensitive parameters, count the number of times the yield crosses the qualified threshold, and label parameters with high response fluctuations; Calculate the conditional entropy and entropy increase of high-response fluctuation parameters and verify their monotonicity. Calculate the local peak density and value range overlap coverage. Eliminate parameters that do not meet the requirements. Verify the uniformity of the distribution of the remaining parameters and the standard deviation of the yield. Calculate the comprehensive sensitivity coefficient by weighted fusion and sort them. Select the top N as key sensitive parameters.

[0006] Furthermore, the steps of calculating the yield gradient change rate and its variance of the mutation-sensitive parameter, counting the number of times the yield crosses the qualified threshold, and marking high-response fluctuation parameters include: Bidirectional perturbation was applied to mutation-sensitive parameters to calculate the positive and negative yield gradient change rates and mean variances, the total number of threshold crossings was counted, polar oscillation parameters and gradient stratification parameters were labeled and merged into initial screening response parameters. Calculate the stratification crossing ratio and cumulative effect coefficient, cumulative gradient and crossing increment for the initial screening response parameters, count the duration of continuous exceedance and the number of continuous crossing segments, and mark the sensitive stratification parameters. For sensitive stratified parameters, the statistical direction of the duration is calculated, the cross-correlation peak of the gradient sequence is calculated, strong and persistent low symmetry parameters are marked, and the proportion of positive crossing and the ratio of crossing asymmetry are calculated. Parameters with a positive crossing percentage exceeding the threshold, a crossing asymmetry ratio deviating from the benchmark, a continuous exceeding duration below the threshold, and a continuous crossing segment number exceeding the threshold are selected. The parameters are then weighted and sorted according to multiple indicators, and the top N parameters with high response fluctuations are output.

[0007] Furthermore, the steps of retrieving the optimal parameter combination of historical high-yield batches based on key sensitive parameters, generating an initial parameter scheme, comparing the deviation value with the yield threshold to obtain the deviation value, and determining the basic adjustment rule include: Obtain real-time yield data for the current batch, calculate the basic deviation between the yield of a single point and the preset threshold, statistically analyze the cumulative difference and average cumulative deviation of the batch, and determine the overall adjustment direction. The yield threshold is classified and the current level is marked. The threshold is split into stages and the deviation of each stage is calculated to locate the bottleneck stage and the deviation of the yield compliance rate is statistically analyzed. Search for the best historical parameter combination at the same level, calculate the level distance and segment parameter distance, weight the weighted level distance, and classify the deviation level and adjustment level. Based on a comprehensive assessment of multiple deviations, hierarchical distances, and the proportion of bottleneck segments, and combined with the deviation type to match corresponding adjustment strategies, the final basic adjustment rules are determined.

[0008] Furthermore, the steps of starting with the initial parameter scheme, adjusting parameters according to sensitivity coefficients, adaptively switching the step size based on the yield response ratio, and calculating the parameter correction amount by fusing the deviation value and adjustment rules include: Starting with the initial parameter scheme, the sensitive layers are divided and sorted according to the sensitivity coefficient, the parameter deviation is extracted, and the step size is switched according to the yield response ratio to obtain the basic sequence of parameters to be adjusted. The real-time and cumulative yield deviations are combined to obtain the comprehensive deviation. The consistency of the parameter adjustment direction is calculated, and the weighted average is used to obtain the same-direction correction base. The opposite-direction parameters are reviewed according to the proportion of the trigger direction. The initial correction coefficient is obtained by linear interpolation of the comprehensive deviation. A time series table is established to calculate the response delay. The correction coefficient is dynamically scaled according to the delay weight and the deviation trend. Adjust the step size based on the response delay, allocate the correction amount according to priority, merge the correction coefficient and the base, smoothly transition the correction amount between layers and limit the amplitude, and output the parameter correction amount.

[0009] Furthermore, the steps of merging real-time and cumulative yield deviations to obtain a comprehensive deviation, calculating the consistency of parameter adjustment directions, weighting to obtain a correction base in the same direction, and triggering direction verification for parameters in opposite directions according to their proportions include: Read the real-time and cumulative yield deviations, obtain the time-weighted comprehensive deviations by weighting them according to the time decay factor, remove outliers and segment them, mark the strong consistency and oscillation parameters, and divide the same-direction and opposite-direction parameter sets. For parameters with the same direction, outlier data is removed, the probability of consistent direction is calculated and confidence levels are assigned, the historical correlation of parameters is calculated, and the comprehensive bias is shared among highly correlated parameter groups. Determine the consistency of parameter directions within a group, mark strongly coupled same-direction groups, calculate and combine the weighted same-direction correction base and confidence same-direction correction base for each group, and statistically summarize the sensitivity of opposite and oscillating parameters. Calculate its sensitivity ratio. If it exceeds the standard, initiate a directional review. Incorporate the undetermined parameters that meet the conditions into the same-direction correction base. After recalculation, output the final base and review flag.

[0010] Furthermore, the step of storing parameter combinations that do not improve yield in a taboo table to exclude invalid combinations, generating control instructions based on parameter correction amounts, and executing the next batch cutting includes: Read parameter combinations that do not improve yield, classify them by yield reduction and store them in a taboo table, set an expiration date and create an index, and classify them by correction amount, and gradually break down large corrections. Calculate the historical response time of the correction amount, mark long response parameters and issue pre-preparation instructions, divide the parameter groups according to the degree of correlation, perform redundant expansion on the strongly correlated groups and add accompanying corrections; Integrate progressive correction, pre-preparation and accompanying correction to generate comprehensive instructions, verify their compatibility with the taboo list, eliminate invalid instructions and classify execution priorities, and record yield feedback nodes; Instructions are executed according to timing and priority. Preparations are made in advance and then progressive corrections are made. The execution status and yield feedback are verified. Abnormalities are stored in the taboo table, and normal combinations are marked as valid combinations, thus completing the cutting control.

[0011] Furthermore, the steps of calculating the historical response time of the correction amount, marking long response parameters and issuing pre-preparation instructions, dividing parameter groups according to correlation degree, redundancy expansion of strongly correlated groups, and adding accompanying corrections include: Statistical parameter corrections are analyzed based on historical response times. Parameters exceeding the average response time are marked, and advance preparation instructions containing information on correction direction, magnitude, and sensitivity are issued. The correlation degree is calculated based on parameter adjustment synchronicity, and strong correlation groups are divided. Calculate and sort the correlation weights within the group, expand bidirectionally with the median sensitivity parameter as the core, select representative parameters according to the magnitude of the correction, and construct an expanded group that balances both sensitivity and magnitude. Sign verification is performed on the extended group, the response time delay pattern is analyzed, complementary compensation parameters are introduced to form an optimization group with smooth time delay, and long response parameters are classified separately. Verify the effectiveness of redundant combinations, select matching parameters from weakly correlated groups as accompanying correction terms, limit the correction interval, generate standard control commands, and output the final optimized parameter combination.

[0012] Furthermore, the steps of collecting new batch yield data to update the sensitivity coefficient and taboo table, archiving and adjusting paths and taboo records, and returning to execute the initial data collection to form an iterative optimization loop include: Collect new batch yield data, compare it with historical yield standards, filter parameter combinations that do not improve yield, calculate time-weighted sensitivity coefficient, mark time-sensitive parameters and degradation parameters, enter invalid combinations into the taboo table and adjust the recording weights; The parameters are adjusted to complete the path, the timeliness value is calculated segment by segment, the path unit is divided and its success rate is predicted, and a comprehensive path sequence is formed. The adjusted paths and taboo records are classified and organized, differentially encoded and compressed, path primitives are extracted, and duplicate and invalid records are removed. Verify the consistency between the archived records and the taboo list, determine the return strategy, recalibrate the parameters if the criteria are not met, and start iterative optimization if the criteria are met, forming a four-driven iterative optimization loop.

[0013] A second aspect of this application also proposes a silicon carbide dicing control system based on wafer yield, comprising: The acquisition and identification module is used to acquire the cutting parameters and real-time yield data of the current batch of silicon carbide wafers, calculate the yield change range to generate and sort the sensitivity coefficients, and identify key sensitive parameters. The retrieval and generation module is used to retrieve the optimal parameter combination of historical high-yield batches based on key sensitive parameters, generate an initial parameter scheme, compare the deviation value with the yield threshold, and determine the basic adjustment rules. The adaptive calculation module is used to start from the initial parameter scheme, adjust the parameters according to the sensitivity coefficient, adaptively switch the step size according to the yield response ratio, and calculate the parameter correction amount by fusing the deviation value and the adjustment rules. The taboo execution module is used to store parameter combinations that do not improve yield into the taboo table to exclude invalid combinations, generate control instructions based on parameter correction amount, and execute the next batch cutting. The closed-loop update module is used to collect new batch yield data, update sensitivity coefficients and taboo tables, archive adjustment paths and taboo records, and return to execute initial data collection to form an iterative optimization loop.

[0014] The first aspect of this plan brings the following benefits: This application uses sensitivity coefficients to accurately locate key parameters, significantly improving the targeting of parameter control; it quickly generates initial solutions based on historical high-yield data, reducing trial-and-error costs; adaptive step size balances adjustment efficiency and cutting stability, effectively improving yield and production efficiency; taboo tables avoid invalid parameter combinations, preventing repeated and inefficient debugging; and it continuously optimizes the process through closed-loop iteration, forming a self-learning mechanism to steadily improve cutting quality and output efficiency, fundamentally solving the problems of low quality and efficiency caused by manual parameter setting and optimization without feedback. Attached Figure Description

[0015] Figure 1 This is a schematic flowchart of a silicon carbide dicing control method based on wafer yield according to an embodiment of this application; Figure 2 This is a schematic diagram of the structure of a silicon carbide dicing control system based on wafer yield according to an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a computer device according to an embodiment of this application; The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0017] Those skilled in the art will understand that, unless explicitly stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in the specification of this application means the presence of features, integers, steps, operations, elements, modules, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, modules, components, and / or groups thereof. It should be understood that when an element is “connected” or “coupled” to another element, it may be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein may include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any modules and all combinations of one or more associated listed items.

[0018] Those skilled in the art will understand that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0019] Reference Figure 1 This application provides a silicon carbide dicing control method based on wafer yield, comprising: S1: Collect the cutting parameters and real-time yield data of the current batch of silicon carbide wafers, calculate the yield change range to generate and sort the sensitivity coefficients, and identify the key sensitive parameters; S2: Based on key sensitive parameters, retrieve the optimal parameter combination of historical high-yield batches, generate an initial parameter scheme, compare it with the yield threshold to obtain the deviation value, and determine the basic adjustment rules; S3: Starting from the initial parameter scheme, adjust the parameters according to the sensitivity coefficient, adaptively switch the step size based on the yield response ratio, and calculate the parameter correction amount by integrating the deviation value and the adjustment rule. S4: Store parameter combinations that do not improve yield in the taboo table to exclude invalid combinations, generate control instructions based on parameter correction amount and execute the next batch cutting; S5: Collect new batch yield data, update sensitivity coefficients and taboo tables, archive adjusted paths and taboo records, and return to execute initial data collection to form an iterative optimization loop.

[0020] In step S1, the data preparation process begins by collecting 12 key cutting parameters, including cutting speed, tension, feed rate, laser power, and cutting frequency, throughout the entire cutting process of the current batch of 6-inch silicon carbide wafers. Simultaneously, real-time yield data for each wafer, such as edge chipping, cracking, thickness deviation, and warpage, are collected, resulting in 8000 parameter-yield paired samples. Abnormal fluctuation samples are removed using the 3σ principle, retaining 7650 valid data sets. The value range of each parameter is divided into 10 intervals with equal frequency, and the average yield of each interval is calculated, along with the jump difference between adjacent intervals. Three parameters sensitive to sudden changes—laser power, cutting speed, and line tension—are identified. These parameters are subjected to ±5% bidirectional perturbation to calculate the yield gradient change rate, variance, and pass threshold crossing frequency, identifying parameters with high response fluctuations. Further calculations are made of conditional entropy, yield entropy increase, local peak density, and value domain overlap coverage to eliminate redundant parameters and verify data uniformity and yield standard deviation characteristics. A comprehensive sensitivity coefficient is calculated based on local peak density and gradient variance, and the top two coefficients are selected as key sensitive parameters for laser power and cutting speed. Using a self-developed process analysis platform, the yield change amplitude is calculated based on single-parameter unit perturbation, generating sensitivity coefficients and automatically sorting them to accurately identify and lock key sensitive parameters, providing a priority basis for subsequent parameter adjustments. This step, through multi-dimensional data collection and quantitative sensitivity screening, accurately identifies the core parameters affecting yield, avoiding blind parameter tuning and improving the efficiency and targeting of subsequent optimization.

[0021] In step S2, based on the two key sensitive parameters of laser power and cutting speed locked in S1, the process files of high-yield batches with a historical yield of ≥98% for 6-inch silicon carbide wafers of the same specification and under the same cutting equipment are retrieved. The optimal parameter combination of 18.5W laser power and 12mm / s cutting speed is extracted, and the compensation amount is allocated according to the sensitivity ratio to generate the inherited and corrected initial parameter scheme. Simultaneously, the real-time yield data of the current batch is compared hierarchically with the preset qualified threshold of 95%, excellent threshold of 97%, and outstanding threshold of 99%. The basic deviation of the yield of a single wafer from the threshold, the overall cumulative deviation of the batch, and the median deviation are calculated. Combining the dynamic deviation trends of the three sub-batches (before, during, and after), the data in the oscillation range is removed to obtain the memory-weighted average cumulative deviation, thus clarifying the overall parameters. The initial parameters need to be adjusted towards "slightly increasing power and reducing speed." The cutting process is then broken down into three segments: feeding, steady cutting, and finishing. The percentage of yield meeting the target and the deviation value for each segment are calculated. The steady cutting segment is identified as the yield bottleneck, with a yield of only 93.2%, below the target threshold. Subsequently, the weighted hierarchical distance between the current parameters and the historical best parameters is calculated, categorized into medium deviation levels and three adjustment ranges. Combining the deviation type, bottleneck segment percentage, and hierarchical distance, a "segment breakthrough leap" adjustment strategy is matched, forming the basic adjustment rules for prioritizing steady cutting segment optimization and coordinating the entire process. Simultaneously, a fixed mapping relationship is established between yield deviation, batch trend, threshold level, and adjustment strategy to ensure that subsequent parameter corrections are based on evidence, providing complete rule support for S3 adaptive correction calculations. This step relies on historical best experience to quickly generate an initial solution, accurately locating deviations and bottlenecks, making subsequent adjustments clearer and more efficient.

[0022] In step S3, starting with the initial parameter scheme generated in S2, the two key parameters of laser power and cutting speed are prioritized based on the sensitivity coefficient calculated in S1, with priority given to adjusting the laser power parameter, which has a higher impact on yield. During parameter adjustment, the yield response ratio is introduced as the core basis for step size switching: for laser power, due to its fast response to yield, it is a highly sensitive and fast-response parameter, and a large step size of 0.8W is used to quickly approach the optimal range; for cutting speed, its yield response is relatively smooth, so a small step size of 0.2mm / s is used for fine-tuning and stabilization, avoiding process oscillations or wafer damage due to large adjustments. Subsequently, the memory-weighted average cumulative deviation obtained in S2 is integrated with the basic adjustment rules to establish a calculation logic of "weighted correction of parameters in the same direction and damping adjustment of parameters in opposite directions": for laser power, since the adjustment direction is consistent with the deviation correction direction, a weight coefficient of 1.2 is assigned for positive compensation; for cutting speed, a damping coefficient of 0.7 is used for micro-adjustment. Simultaneously, taking into full account the delay characteristics of the cutting process, a dynamic compensation coefficient is introduced for parameters with long response cycles, scaling the correction amount in real time to ensure continuous and compliant correction results. Ultimately, a unique and precise parameter correction amount is calculated, outputting a target parameter combination of 19.3W laser power and 11.6mm / s cutting speed. This provides accurate and scientific data support for the execution of S4 commands, achieving a leap from qualitative adjustment to quantitative calculation. This step achieves precision and dynamism in parameter adjustment, balancing optimization efficiency and process stability, and avoiding blind parameter tuning.

[0023] In step S4, firstly, parameter combinations that were tried during the S3 adjustment process but failed to improve yield, or even caused a decrease in yield, are collected and stored in a taboo table according to the magnitude of yield reduction. A valid duration and retrieval index are set to prevent subsequent re-entry into invalid parameter ranges. Simultaneously, the final correction values ​​of the laser power (19.3W) and cutting speed (11.6mm / s) output from S3 are read and categorized as medium-amplitude corrections. No gradual splitting is required; the complete correction value is executed directly. Based on parameter response characteristics, historical average response times are calculated, and laser power is marked as a long-response parameter. A pre-preparation command is issued three control cycles in advance, including the target correction direction, estimated amplitude, and sensitivity coefficient, allowing the cutting equipment to complete power preheating and optical path calibration in advance, reducing execution lag. Based on the historical synchronous adjustment frequency of parameters, the correlation is calculated, and laser power and cutting speed are identified as a strongly correlated parameter group. Bidirectional redundancy expansion is performed with median sensitivity as the core, and line tension is added as an accompanying correction parameter. The correction value is 35% of the original value, ensuring coordinated parameter changes and stable cutting conditions. Subsequently, the progressive correction amount, pre-preparation instructions, and accompanying correction items are integrated to generate standardized comprehensive control instructions. Each instruction is verified for compatibility with the taboo list, and instructions containing invalid combinations are removed. These instructions are then prioritized and yield feedback nodes are recorded. During execution, a pre-preparation instruction is issued first. After the equipment completes its pre-response, the main parameter correction and accompanying correction are executed sequentially according to priority. The execution status and real-time yield feedback are monitored throughout the process. If a correction timeout, abnormal magnitude, or no yield improvement occurs, the current combination is immediately marked and added to the taboo list. If all executions are normal and the yield increases synchronously, it is marked as a valid combination, completing the batch cutting control loop and providing reliable execution results for S5 iteration updates. This step, through taboo avoidance and predictive execution, avoids ineffective trial and error, improves instruction stability and cutting safety, and ensures continuous yield improvement.

[0024] In step S5, firstly, the yield data of silicon carbide wafers after executing the S4 correction instruction in this batch is collected. Real-time yield, edge breakage ratio, thickness uniformity, and other indicators are compared with historical yield standards. Parameter combinations that did not improve yield in this adjustment are screened out. The time-weighted sensitivity coefficients of laser power and cutting speed are recalculated according to time decay weights. Time-sensitive parameters with ranking fluctuations exceeding three places and decay parameters with decreased yield contribution are marked. Invalid combinations are updated and entered into the taboo table, and the timeliness weight of the records is adjusted. Subsequently, the parameter adjustment path from S1 to S4 is completely organized according to the execution sequence, divided into three stages: initial inheritance, adaptive step size adjustment, and instruction execution. The ratio of yield improvement to execution time is calculated segment by segment to obtain the timeliness value. The path is broken down into independent units, and combined with historical records to predict unit success rates, forming a comprehensive path sequence sorted by timeliness and success rate. Next, differential encoding compression is performed on the adjustment path and taboo records, retaining only the parameter changes of adjacent steps. High-frequency effective adjustment segments are extracted as path primitives, and duplicate and non-gain archived data are removed to reduce storage usage and improve retrieval efficiency. After archiving, the consistency between the archived data and the taboo list is verified. Four indicators are comprehensively evaluated: the number of time-sensitive parameters, the prediction accuracy of the success rate, the path compression rate, and the network evolution efficiency. Currently, there are two time-sensitive parameters and the prediction accuracy is above 85%, meeting the stable optimization conditions. Therefore, the process directly returns to step S3 for further optimization based on timeliness value and success rate prediction. If the indicators do not meet the standards, the process returns to step S1 to recalibrate sensitivity and rebuild the model. Finally, the complete adjusted path, sensitivity ranking changes, taboo records, and timeliness value distribution are uniformly archived into the process database, forming an experience knowledge base that can be directly inherited by subsequent batches. This achieves a four-pronged optimization loop driven by intergenerational accumulation and continuous iteration. This step realizes data closed-loop iteration and experience accumulation, allowing cutting control to continuously self-optimize and steadily improve yield and production efficiency in the long term.

[0025] In one embodiment, the steps of collecting the cutting parameters and real-time yield data of the current batch of silicon carbide wafers, calculating the yield variation to generate and sort sensitivity coefficients, and identifying key sensitive parameters include: S10: Collect each cutting parameter and real-time yield data to form paired samples, and retain the valid data sequence after removing abnormal samples; S11: Divide the parameter value range into multiple intervals, calculate the average yield of each interval, calculate the jump difference of the average yield of adjacent intervals, and mark the mutation-sensitive parameters. S12: Calculate the yield gradient change rate and its variance of the mutation-sensitive parameters, count the number of times the yield crosses the qualified threshold, and mark the high response fluctuation parameters; S13: Calculate the conditional entropy and entropy increase of the high-response fluctuation parameters and verify their monotonicity. Calculate the local peak density and value range overlap coverage. Eliminate parameters that do not meet the requirements. Verify the uniformity of the distribution of the remaining parameters and the standard deviation of the yield. Calculate the comprehensive sensitivity coefficient by weighted fusion and sort them. Select the top N as key sensitive parameters.

[0026] In this embodiment, the first step is to collect cutting parameters and yield data to form valid paired samples. Taking a batch of 6-inch silicon carbide wafer laser cutting as an example, six cutting parameters are collected: laser power, cutting speed, line tension, feed rate, frequency, and spot diameter. Simultaneously, the chipping amount, crack rate, thickness deviation, warpage, and yield judgment results of each wafer are collected. A total of 7200 sets of parameter-yield paired samples are formed according to the one-to-one correspondence of each wafer. The 3σ principle is used to remove abnormal data with excessive differences. 312 abnormal samples caused by equipment jumps and detection interference are removed, and 6888 sets of valid data sequences are retained to ensure that the parameters and yield data are true and reliable and without significant deviation. Secondly, parameter intervals were divided and mutation-sensitive parameters were marked. The value ranges of the six parameters were divided into 10 intervals with equal frequency, that is, the parameter value ranges were divided into M intervals with equal frequency or equal width. The average yield was calculated for each interval, and the jump difference between the average yields of adjacent intervals was calculated. The jump amplitude threshold was set to 8%. When the yield difference between adjacent intervals exceeded this threshold, it was marked as a mutation. The calculated interval jump differences of the laser power and cutting speed parameters reached 11.2% and 9.7% respectively, which were significantly higher than other parameters. Therefore, these two parameters were marked as mutation-sensitive parameters.

[0027] Next, the yield gradient change rate, variance, and threshold crossing count are calculated, and high-response fluctuation parameters are marked. The laser power and cutting speed are perturbed by ±5%, that is, the parameters sensitive to sudden changes are perturbed by a fixed step size or an adaptive step size. The yield gradient change rate and its variance are calculated point by point. At the same time, the number of times the yield crosses the 95% qualified threshold is counted. The gradient variance of the laser power is 0.032 and the number of crossings reaches 17. The gradient variance of the cutting speed is 0.027 and the number of crossings reaches 14. Both meet the high response judgment conditions, so they are marked as high-response fluctuation parameters. That is, when the gradient variance exceeds the preset variance threshold and the number of crossings exceeds the preset number threshold, it is marked as a high-response fluctuation parameter. Next, conditional entropy, entropy increase, local peak density, and value range overlap coverage are calculated and screened for verification. Conditional entropy and yield entropy increase are calculated for two high-response fluctuation parameters, verifying that entropy increase shows a monotonic trend with parameter changes. Then, the local peak density of the yield standard deviation within the parameter's neighborhood and the value range overlap coverage between parameters are calculated, eliminating invalid dimensions with low peak density or those strongly correlated with redundant parameters. The uniformity of the remaining parameter data distribution is verified, and the yield standard deviation of the middle sample is less than 20% of the entire sample, meeting the stability screening criteria. Finally, the local peak density is used for... By weighting and fusing with gradient variance, a comprehensive sensitivity coefficient of 0.91 for laser power and 0.76 for cutting speed were calculated. These coefficients were then sorted from highest to lowest, and the top two were selected. Laser power and cutting speed were ultimately identified as the key sensitive parameters for this batch. Specifically, the comprehensive sensitivity coefficient was calculated by weighting local peak density and gradient variance, and then sorted in descending order. The top N coefficients were selected, where N is determined based on the total number of parameters and process complexity, ranging from 2 to 5. These were identified as key sensitive parameters, providing a clear priority basis for subsequent historical parameter matching and adaptive adjustments. This step, through multi-dimensional quantitative screening, accurately identifies key sensitive parameters, improves the targeting of control, avoids ineffective parameter tuning, and enhances overall optimization efficiency.

[0028] In one embodiment, the steps of calculating the yield gradient change rate and its variance of the mutation-sensitive parameter, counting the number of times the yield crosses the pass threshold, and marking high-response fluctuation parameters include: S120: Perform bidirectional perturbation on mutation-sensitive parameters, calculate the positive and negative yield gradient change rate and mean variance, count the total number of threshold crossings, mark polar oscillation parameters and gradient stratification clear parameters, and merge them into initial screening response parameters. S121: Calculate the stratification crossing ratio and cumulative effect coefficient, cumulative gradient and crossing increment for the initial screening response parameters, count the duration of continuous exceeding the standard and the number of continuous crossing segments, and mark the sensitive stratification parameters. S122: Statistically calculate the duration of the direction of sensitive hierarchical parameters, calculate the peak value of the cross-correlation of gradient sequences, mark the strong and persistent low symmetry parameters, and calculate the proportion of positive crossing and the ratio of crossing asymmetry. S123: Filter parameters whose positive crossing percentage exceeds the threshold, whose crossing asymmetry ratio deviates from the benchmark, whose continuous exceeding duration is less than the threshold, and whose continuous crossing segment number exceeds the threshold. Rank the parameters by weighted fusion of multiple indicators and extract the top N high-response fluctuation parameters for output.

[0029] In this embodiment, the mutation-sensitive parameters are first bidirectionally perturbed and merged to obtain the initial screening response parameters. Taking the two mutation-sensitive parameters, laser power and cutting speed, which have been determined above, as the objects, bidirectional perturbations with fixed step sizes of +0.5% in the positive direction and -0.5% in the negative direction are carried out respectively. The yield gradient change rate is calculated point by point, and the mean and variance are obtained. At the same time, the total number of times the yield crosses the 95% qualified threshold is counted across the entire range. After calculation, the positive and negative gradient directions of laser power are opposite, and it is determined to be a polar oscillation parameter. The gradient change rate of cutting speed shows a low-medium-high pyramid distribution, and it is determined to be a parameter with clear gradient stratification. The two types of parameters are merged to obtain laser power and cutting speed as the initial screening response parameters. Secondly, the stratification crossing ratio and cumulative effect coefficient were calculated, and sensitive stratification parameters were marked. The stratification crossing ratio was obtained by calculating the ratio of yield crossing density at low and high levels for the two initial screening response parameters. At the same time, the cumulative effect coefficient was obtained by calculating the ratio of single-step instantaneous gradient to cumulative multi-step gradient. The gradient increment and threshold crossing increment were accumulated segment by segment, and the duration of continuous gradient exceedance and the maximum number of consecutive threshold crossing segments were statistically analyzed. The laser power stratification crossing ratio was 3.6 and the cumulative effect coefficient was 1.42. The cutting speed stratification crossing ratio was 0.28 and the cumulative effect coefficient was 0.73. Both of them deviated significantly from the baseline of 1 and met the sensitive stratification judgment conditions. Therefore, they were marked as sensitive stratification parameters.

[0030] Next, the directional duration was statistically analyzed and strong-duration low-symmetry parameters were marked. The directional duration was obtained by statistically analyzing the maximum number of consecutive steps of the same-direction gradient for sensitive layer parameters. The cross-correlation peak value of the positive and negative perturbation gradient sequences was calculated. The directional duration of laser power accounted for 72% and the cross-correlation peak value was 0.41. The directional duration of cutting velocity accounted for 68% and the cross-correlation peak value was 0.37, both of which met the criteria for strong-duration low-symmetry. At the same time, the positive crossing ratio and the crossing asymmetry ratio were calculated. The positive crossing ratio of laser power was 76% and the crossing asymmetry ratio was 2.4. The positive crossing ratio of cutting velocity was 73% and the crossing asymmetry ratio was 2.1.

[0031] Finally, high-response fluctuation parameters are output through multi-index weighted screening. Thresholds are set for the positive crossing percentage (70%), crossing asymmetry ratio (benchmark 1.5), continuous exceeding duration (5 steps), and continuous crossing segment count (3 segments). Both laser power and cutting speed must meet all conditions: positive crossing percentage exceeding the threshold, crossing asymmetry ratio significantly deviating from the benchmark, continuous exceeding duration less than 5 steps, and continuous crossing segment count greater than 3 segments. Weighted fusion sorting is performed based on the degree of deviation of layered crossing ratio, cumulative effect coefficient deviation, directional persistence percentage, and continuous crossing segment count. The laser power comprehensive score is 0.89, and the cutting speed comprehensive score is 0.75. The top two scores are selected, and the final output laser power and cutting speed are the high-response fluctuation parameters. This provides accurate intermediate results for subsequent sensitivity calculations and key parameter determination, maintaining complete consistency with the 6-inch silicon carbide wafer laser cutting scenario described earlier, forming a complete closed-loop logic. This step, through multi-level quantitative screening and multi-dimensional index verification, accurately identifies high-response fluctuation parameters, improving the reliability of sensitivity calculations and the accuracy of control.

[0032] In one embodiment, the steps of retrieving the optimal parameter combination of historical high-yield batches based on key sensitive parameters, generating an initial parameter scheme, comparing the deviation value with the yield threshold to determine the basic adjustment rule include: S20: Obtain the real-time yield data of the current batch, calculate the basic deviation between the yield of a single point and the preset threshold, calculate the cumulative difference and average cumulative deviation of the batch, and determine the overall adjustment direction; S21: Classify the yield threshold and mark the current level, break the threshold by stage and calculate the deviation of each segment, locate the bottleneck segment, and count the deviation of the yield compliance rate. S22: Retrieve the best historical parameter combination at the same level, calculate the level distance and segment parameter distance, weight the weighted level distance, and classify the deviation level and adjustment level; S23: Based on a comprehensive judgment of multiple deviations, hierarchical distances, and the proportion of bottleneck segments, and combined with the deviation type to match the corresponding adjustment strategy, the final basic adjustment rules are determined.

[0033] In this embodiment, the first step is to acquire the current batch yield data and calculate the deviation to determine the overall adjustment direction. Following the 6-inch silicon carbide wafer laser cutting scenario described earlier, laser power and cutting speed are used as key sensitive parameters. Real-time yield data for each wafer in the current batch is read, and after removing outliers, an effective yield sequence is formed. 95% is used as the acceptable yield threshold. The basic deviation between the yield at a single point and the threshold is calculated point by point. The cumulative difference and average cumulative deviation for the entire batch are statistically analyzed, and the median yield value is extracted and subtracted from the threshold to obtain the median deviation. The calculation shows that the current batch's average yield is 93.2%, the basic deviation is -1.8%, and both the cumulative deviation and the median deviation are negative. Furthermore, multiple wafers show the same deviation direction. Therefore, the overall adjustment direction is determined to be to positively improve the yield, and the parameters need to be corrected in the direction of increasing the yield. Secondly, the yield thresholds were graded and bottlenecks were identified. The yield was divided into three levels: qualified (≥95%), good (≥97%), and excellent (≥99%). The current batch yield was only 93.2%, which is below the qualified level. The cutting process was then divided into three stages: feeding stage, steady cutting stage, and finishing stage. The deviation between the yield of each stage and the corresponding threshold was calculated. The yield of the steady cutting stage was only 91.5%, which was significantly lower than the other two stages. The steady cutting stage was identified as the bottleneck for the current yield. At the same time, the deviation between the batch yield compliance rate and the target rate was -6.8%.

[0034] Next, the best historical parameters are retrieved and the weighted level distance is calculated. The process files of the same specifications, equipment and materials with the highest yield (≥98%) batches are retrieved, and the best parameter combination is extracted: laser power 18.5W and cutting speed 12mm / s. The level distance and segment parameter distance between the current parameters and the best historical parameters are calculated respectively. The weighted level distance is obtained by combining the sensitivity coefficient, which is determined to be a medium deviation level and divided into three adjustment range levels. Finally, a comprehensive assessment and matching adjustment strategy is implemented. This involves integrating multiple indicators such as basic deviation, cumulative deviation, compliance rate deviation, weighted level distance, and bottleneck segment percentage. Considering that the current batch falls into the category of "below acceptable level, bottleneck segment concentrated, and moderate deviation," a **breakthrough transition** adjustment strategy is matched. This strategy focuses on the stable cutting segment as the core breakthrough target, prioritizing the correction of key sensitive parameters while ensuring overall process stability. This forms a complete set of basic adjustment rules, including adjustment direction, priority, magnitude, and execution stage. This provides a clear basis for subsequent adaptive step-size adjustments and parameter correction calculations, maintaining consistency with the key parameters, batch scenarios, and data definitions mentioned earlier, forming a coherent closed-loop implementation logic. This step, through multi-dimensional deviation analysis and the inheritance of historical experience, accurately identifies bottlenecks and clarifies strategies, making parameter adjustments more efficient and better suited to actual process requirements.

[0035] In one embodiment, the steps of acquiring the real-time yield data of the current batch, calculating the basic deviation between the yield of a single point and a preset threshold, statistically analyzing the cumulative difference and average cumulative deviation of the batch, and determining the overall adjustment direction include: S201: Read the yield data of single point in the current batch, remove outliers to obtain the effective sequence, divide the sub-batch by time, calculate the dynamic threshold by combining with historical batches and obtain the dynamic deviation of each sub-batch. S202: Calculate the baseline deviation and median deviation, mark the time sequence coherence segments, and count the cumulative number of nodes and the average interval steps within each sub-batch coherence segment. S203: Compare the cumulative difference trend of sub-batch, remove oscillating data, perform high and low frequency decomposition on monotonic trend deviation, calculate the mean using a sliding window and sum them up to obtain the total cumulative difference. S204: Eliminate high-fluctuation intervals, accumulate trend term deviations by time weight, calculate the memory weighted average cumulative deviation, and determine the overall adjustment direction of parameters by considering the consistency of the signs of multiple deviations.

[0036] In this embodiment, the first step is to read the yield data of the current batch at a single point and calculate the dynamic deviation. Continuing with the 6-inch silicon carbide wafer laser cutting scenario described earlier, a total of 6888 sets of valid single-point yield data for the current batch are read. The 3σ principle is used to eliminate extreme yield values ​​caused by detection interference or instantaneous equipment malfunctions, resulting in a regular and effective yield sequence. The batch is then evenly divided into three sub-batches—previous, middle, and later—according to the cutting time sequence, with an equal sample size for each batch. At the same time, the historical yield threshold compliance status of the previous three batches is retrieved and calculated to obtain a dynamic qualified threshold of 95.2%. The difference between the single-point yield of each sub-batch and the dynamic threshold is calculated to obtain the dynamic deviation of the previous, middle, and later sub-batches. Secondly, the basic deviation and median deviation are calculated and the temporal continuity segments are marked. Based on the dynamic threshold, the difference between the yield of a single point and the threshold is calculated to obtain the basic deviation. The median yield of the whole batch is extracted and the difference is made with the threshold to obtain the median deviation. The segments with three or more consecutive pieces showing the same deviation along the cutting time are marked as temporal continuity segments. The cumulative number of nodes of continuity segments in each sub-batch and the average interval steps of adjacent continuity segments are counted. Among them, the continuity segments of the sub-batch have the highest proportion, the shortest interval, and the strongest deviation continuity.

[0037] Next, the trend was compared and the total cumulative difference was calculated. The cumulative difference trends of the three sub-batches were compared one by one. The first sub-batches fluctuated slightly, the second sub-batches tended to be stable, and the middle batch showed a clear monotonically decreasing trend. Data from the first sub-batches with obvious oscillation characteristics were removed. The monotonically changing trend deviation of the middle batch was decomposed into high and low frequency components to separate the trend component that changed steadily and the fluctuation component that changed instantaneously. The mean was calculated segment by segment using 5 sliding windows and accumulated to obtain the total cumulative difference after removing the fluctuation interference. Finally, a weighted calculation is performed to determine the overall adjustment direction. High-fluctuation intervals with fluctuations exceeding 50% of the trend term are eliminated. The remaining pure trend term deviations are weighted and accumulated according to time weights, with the weight of recent cutting segments set to twice that of longer-term segments to reduce historical data interference, resulting in a memory-weighted average cumulative deviation. Since the basic deviation, median deviation, and memory-weighted average cumulative deviation all have the same negative sign, it indicates that the current overall yield is consistently below the threshold. Therefore, the overall parameter adjustment direction is determined to be positive, aiming to improve the yield. Priority is given to gain-based corrections to the two key sensitive parameters: laser power and cutting speed. The entire process maintains complete consistency with the sample size, key parameters, and process scenarios described earlier, ensuring data coherence and logical closure, providing the most reliable directional basis for subsequent deviation calculations and strategy matching. This step, through time-series decomposition, dynamic thresholds, and weighted statistics, accurately determines the adjustment direction, improves the reliability of deviation judgment, and avoids misadjustments and haphazard adjustments.

[0038] In one embodiment, the steps of starting with an initial parameter scheme, adjusting parameters according to sensitivity coefficients, adaptively switching the step size based on the yield response ratio, and calculating the parameter correction amount by fusing the deviation value and adjustment rules include: S30: Starting from the initial parameter scheme, divide and sort the sensitive layers according to the sensitivity coefficient, extract the parameter deviation, switch the step size according to the yield response ratio, and obtain the basic sequence of parameters to be adjusted; S31: Combine real-time and cumulative yield deviations to obtain a comprehensive deviation, calculate the consistency of parameter adjustment direction, weight the same-direction correction base, and trigger the review of opposite-direction parameters according to their proportion. S32: Obtain the initial correction coefficient by linear interpolation of the comprehensive deviation, establish a time series table to calculate the response delay, and dynamically scale the correction coefficient according to the delay weight and deviation trend; S33: Adjust the step size according to the response delay, allocate the correction amount according to priority, merge the correction coefficient and the base, smoothly transition the inter-layer correction amount and limit the amplitude, and output the parameter correction amount.

[0039] In this embodiment, the first step is to obtain the basic sequence of parameters to be adjusted. Starting with the initial parameter scheme of 18.5W laser power and 12mm / s cutting speed for a 6-inch silicon carbide wafer generated in S2, the parameters are divided into high-sensitivity layers and medium-sensitivity layers according to the sensitivity coefficient determined in S1. The laser power sensitivity of 0.91 is assigned to the high-sensitivity layer, and the cutting speed sensitivity of 0.76 is assigned to the medium-sensitivity layer, and they are arranged in descending order of sensitivity. The deviation between the current parameters and the historical best parameters is extracted, and the yield response ratio is calculated in real time. When the yield response ratio is greater than 1.2, a large step size is used, and when it is less than 0.8, a small step size is switched. This yields the basic sequence of parameters to be adjusted, which includes parameter priority, deviation, and adaptive step size, ensuring complete consistency with the key parameters and data caliber mentioned above. Secondly, the deviations are merged and the same-direction correction base is calculated. The current batch real-time yield deviation of -1.8% and cumulative yield deviation of -2.3% are read and merged according to time decay weight to obtain the time-weighted comprehensive deviation of -2.1%. After removing outliers, the deviations are segmented according to time sequence. The consistency between the adjustment direction of laser power and cutting speed and the sign of the comprehensive deviation is calculated one by one. Both parameters are in the same direction. The same-direction correction base is obtained by weighting according to the sensitivity coefficient. The sensitivity ratio of the opposite parameter is 0, and there is no need to trigger the direction verification, which ensures that the parameter adjustment direction is uniform and reliable.

[0040] Next, the initial correction coefficient is calculated and dynamically scaled. The time-weighted comprehensive deviation is assigned to the low deviation range, and the initial correction coefficient of 1.25 is obtained by linear interpolation between the upper and lower limits of the range. A parameter adjustment and yield response time table is established according to the cutting sequence. The laser power response delay is 2 steps and the cutting speed response delay is 1 step. The correction coefficient is dynamically scaled according to the rule that the longer the delay, the lower the weight. At the same time, the coefficient is slightly strengthened in combination with the continuous downward trend of the deviation to eliminate the control deviation caused by the response lag. Finally, the correction amount is allocated and the final result is output. The step size is further optimized based on the response delay; if the delay exceeds two steps, the step size is halved. 70% of the correction amount is allocated to the laser power and 30% to the cutting speed according to sensitivity priority. The dynamic correction coefficient is fused with the same-direction correction base to smoothly transition the correction amount from the high-sensitivity layer to the medium-sensitivity layer, and a ±5% amplitude limit constraint is set to avoid overshoot damage to the wafer. The final calculation yields a laser power correction of 19.3W and a cutting speed correction of 11.6mm / s, forming a continuous, stable, and compliant parameter correction amount. This provides accurate data for S4 instruction execution, maintaining a closed-loop consistency with the previously described scenario, parameters, and deviation logic throughout the process. This step, through adaptive step size, deviation fusion, and delay compensation, achieves precise and controllable parameter correction, improving yield, efficiency, and process stability.

[0041] In one embodiment, the steps of merging real-time and cumulative yield deviations to obtain a comprehensive deviation, calculating the consistency of parameter adjustment directions, weighting to obtain a corrective base in the same direction, and triggering directional verification for parameters in opposite directions according to their proportions include: S310: Read the real-time and cumulative yield deviations, obtain the time-weighted comprehensive deviations by weighting the time decay factor, remove outliers and segment them, mark strong consistency and oscillation parameters, and divide the same-direction and opposite-direction parameter sets. S311: Remove outliers from parameters with the same direction, calculate the probability of consistent direction and classify confidence levels, calculate the historical correlation of parameters, and share the comprehensive bias among highly correlated parameter groups. S312: Determine the consistency of parameter directions within a group, mark strongly coupled same-direction groups, calculate the same-direction correction base and confidence same-direction correction base of each group separately and combine them, and statistically summarize the sensitivity of opposite and oscillating parameters. S313: Calculate its sensitivity ratio. If it exceeds the standard, start the directional review. Incorporate the undetermined parameters that meet the conditions into the same-direction correction base, recalculate, and output the final base and review mark.

[0042] In this embodiment, the deviation is first read and a time-weighted comprehensive deviation and parameter set is constructed. Continuing with the 6-inch silicon carbide wafer laser cutting scenario described earlier, the determined real-time yield deviation of -1.8% and cumulative yield deviation of -2.3% are read and weighted by the time decay factor. The weight of recent batches is set to twice that of distant batches, resulting in a time-weighted comprehensive deviation of -2.1%. Abnormal fluctuation data is eliminated using the three-standard-deviation criterion. The deviation is evenly divided into five consecutive segments according to the cutting sequence. The parameter adjustment direction and deviation sign are compared segment by segment. The laser power and cutting speed both show a continuous three-segment same-direction characteristic and are marked as strongly consistent parameters. There are no oscillating parameters with more than two direction switching. The two parameters are directly included in the same-direction parameter set. There are no opposite-direction parameters, ensuring complete consistency with the deviation values ​​and key sensitive parameters described earlier. Secondly, confidence screening and correlation calculation are performed on the same-direction parameters. Extreme abnormal data of laser power and cutting speed are removed from the same-direction set. The consistency probability of the adjustment direction and the sign of the time-weighted comprehensive deviation are calculated. If the consistency probability of both parameters exceeds 90%, they are marked as high-confidence same-direction parameters. At the same time, the correlation between parameters is calculated based on historical adjustment records. The correlation between laser power and cutting speed reaches 78%, which exceeds the 70% threshold. It is judged as a high-correlation parameter group. The time-weighted comprehensive deviation is shared by the whole group, avoiding the deviation tearing caused by independent calculation of a single parameter.

[0043] Next, strongly coupled groups are marked and the merged correction base is calculated. Groups with a 100% same-direction ratio within highly correlated parameter groups are marked as strongly coupled same-direction groups. The same-direction correction base is calculated by weighting the sum of the group's sensitivity, and simultaneously, a confidence-based same-direction correction base is calculated by weighting the sensitivity coefficient and confidence level. The two bases are merged to obtain the initial same-direction correction base. The sum of the sensitivity of opposite-direction and oscillating parameters is calculated; currently, there are no opposite-direction or oscillating parameters, so this value is 0. Finally, the sensitivity ratio is calculated and the final base is output. The ratio of the sensitivity of opposite-direction and oscillating parameters to the total sensitivity of all sensitive parameters is calculated; the result is 0, far below the 15% verification threshold, so no direction verification is needed, and there are no pending parameters to be included. The initial same-direction correction base is directly normalized and calibrated, and the final same-direction correction base and the no-verification marker are output. This provides a stable and reliable numerical basis for subsequent correction coefficient calculation, step size allocation, and parameter limiting. The data, scenarios, and logic throughout the process are highly consistent with the preceding text, forming a complete closed loop. This step improves the reliability of the correction base by using weighted bias, confidence judgment, and coupling analysis, avoids directional conflicts, and ensures that parameter adjustments are stable and effective.

[0044] In one embodiment, the steps of reading the real-time and cumulative yield deviations, weighting the deviations by a time decay factor to obtain the time-weighted comprehensive deviation, removing outliers and segmenting the deviations, marking strong consistency and oscillation parameters, and dividing the parameter sets into in-direction and out-of-direction categories include: S3101: Read the real-time and cumulative yield deviations, obtain the time-weighted comprehensive deviations by weighting the time decay factor, remove abnormal data, divide the deviation levels and segment them according to the time sequence. S3102: Obtain the sliding composite deviation by merging through a sliding window, determine the overall reference symbol, and calculate the parameter symbol frequency and direction persistence index; S3103: Obtain multi-scale comprehensive bias according to long and short periods, remove conflicting data, calculate the bias sign distribution entropy and parameter clustering distance, and divide the near and far center layers; S3104: Label stable unidirectional, stable anisodirectional, and two types of oscillation parameters according to their sign, persistence, and entropy value, and classify them into their respective sets. After verifying mutual exclusivity, output the classification and labeling results.

[0045] In this embodiment, the first step is to read and calculate the time-weighted comprehensive deviation. Continuing with the 6-inch silicon carbide wafer laser cutting scenario described earlier, the determined real-time yield deviation of -1.8% and cumulative yield deviation of -2.3% are read and weighted according to the time decay factor. The weight of recent batches is set to twice that of distant batches, resulting in a time-weighted comprehensive deviation of -2.1%. The three-standard-deviation criterion is used to eliminate abnormal data caused by detection interference and instantaneous equipment fluctuations. The comprehensive deviation is divided into three levels: high, medium, and low, based on its numerical value. The current -2.1% is classified into the medium deviation level. The effective deviation data is evenly divided into five consecutive segments according to the cutting sequence to ensure that the data volume of each segment is balanced, the timing is complete, and it is completely consistent with the deviation values ​​and key sensitive parameters described earlier. Secondly, the overall reference sign and persistence index are determined by a sliding window. The deviation data are linearly weighted and merged using a sliding window with a window size equal to the size of the most recent five batches to obtain the sliding comprehensive deviation. The median sign of all valid deviation values ​​is extracted as negative and used as the overall reference sign. The frequency of sign occurrence and direction persistence index of laser power and cutting speed in the five time segments are calculated respectively. The direction persistence index is the proportion of consecutive batches in the same direction. Both parameters remain negative in the five time segments, and the direction persistence index reaches 100%, showing a highly stable same-direction characteristic.

[0046] Next, multi-scale fusion and clustering stratification were performed. Time-weighted merging was performed according to three batches for short periods and ten batches for long periods to obtain the comprehensive deviation of short-scale and long-scale data. Both deviations were negative, and there was no data with scale conflict, so all of them were retained. Using the overall reference symbol as the cluster center, the Euclidean distance between the laser power, cutting speed and the cluster center was calculated. The distances of both parameters were extremely small, so they were uniformly classified into the near-center layer, and there were no far-center layer parameters. At the same time, the distribution entropy of each deviation symbol was calculated. The values ​​were all lower than the preset threshold, indicating that the symbol distribution was highly concentrated and without chaotic fluctuations. Finally, parameter labeling and set division are performed. Based on sign consistency, directional persistence, and distribution entropy, laser power and cutting speed are labeled as stable, unidirectional parameters, with no stable, anisotropic, scale-consistent, or clustered oscillation parameters. The two stable unidirectional parameters are grouped into the unidirectional parameter set, with no anisotropic or oscillating parameters. The sets are rigorously verified for mutual exclusivity and no overlap. The final output includes a classification set and a multi-dimensional labeling report, providing accurate and reliable classification criteria for subsequent confidence calculations, correlation analysis, and correction cardinality calculations. The entire process maintains a high degree of consistency with the preceding text in terms of scenario, numerical values, and logic, forming a complete closed loop. This step, through multi-scale and multi-dimensional deviation analysis, accurately classifies parameter types, improves the reliability of directional judgment, and lays a stable foundation for correction calculations.

[0047] In one embodiment, the step of storing parameter combinations that do not improve yield in a taboo table to exclude invalid combinations, generating control instructions based on parameter correction amounts, and executing the next batch cutting includes: S40: Read parameter combinations that do not improve yield, store them in a taboo table according to yield reduction level, set validity period and create index, and at the same time, classify them according to the size of correction, and gradually break down large corrections. S41: Calculate the historical response time of the correction amount, mark long response parameters and issue pre-preparation instructions, divide parameter groups according to correlation degree, perform redundant expansion on strongly correlated groups and add accompanying corrections; S42: Integrate progressive correction, pre-preparation and accompanying correction to generate comprehensive instructions, verify their compatibility with the taboo list, eliminate invalid instructions and divide execution priorities, and record yield feedback nodes; S43: Execute instructions according to timing and priority, prepare in advance and then make progressive corrections, verify the execution status and yield feedback, store abnormalities in the taboo table, and mark normal ones as valid combinations to complete the cutting control.

[0048] In this embodiment, the first step is to read invalid parameter combinations and construct a taboo table. Simultaneously, the parameter correction amounts are categorized and progressively split. Following the 6-inch silicon carbide wafer laser cutting scenario described earlier, parameter combinations that failed to improve yield or even caused yield reduction during historical debugging are read. These are categorized into three levels based on the yield reduction: slight reduction, significant reduction, and severe reduction. All invalid combinations are categorized and stored in the taboo table, with an 8-hour validity period and a fast retrieval index to avoid repeated trial and error. Simultaneously, the final correction amounts for the laser power (19.3W) and cutting speed (11.6mm / s) output by S3 are read. Based on the magnitude, this is determined to be a medium-amplitude correction, not meeting the criteria for large-amplitude splitting. Therefore, the complete correction amount is used directly without progressive splitting, maintaining consistency with the key parameters and correction results described earlier throughout the process. Secondly, the response time is calculated and pre-preparation and redundancy expansion are completed. The historical average response time of laser power and cutting speed is statistically analyzed. If the laser power response time exceeds twice the average, it is marked as a long response parameter. Pre-preparation instructions containing target direction, estimated amplitude and sensitivity coefficient are issued three control cycles in advance, so that the equipment can complete the optical path preheating and power calibration in advance. The correlation is calculated by adjusting the synchronization frequency according to the parameter history. The correlation between laser power and cutting speed exceeds 70%, and it is divided into a strongly correlated parameter group. Bidirectional redundancy expansion is carried out with median sensitivity as the core. Linear tension is selected as the accompanying correction parameter, and the correction amount is 35% of the original value to ensure that the parameters change in a coordinated manner and the cutting state is stable and shock-free.

[0049] Next, comprehensive instructions are generated and taboo checks and priority assignments are completed. Progressive corrections, pre-preparation instructions, and accompanying corrections are integrated into standardized comprehensive control instructions. Each instruction is matched and verified against the taboo list, automatically eliminating instructions containing invalid parameter combinations to ensure instruction safety and feasibility. Execution priorities are assigned according to the principles of "pre-preparation priority, main parameter priority, and high sensitivity priority," and the yield feedback node corresponding to each instruction is recorded synchronously, clarifying the data collection timing and judgment criteria. Finally, instructions are executed sequentially and closed-loop verification is performed, strictly following the timing and priority. Pre-preparation instructions are issued first, waiting for the equipment's pre-response to complete, then main parameter corrections and accompanying corrections are executed. The execution status, parameter availability, and real-time yield feedback are monitored throughout the process. If correction timeouts, magnitude deviations, or no yield improvement occur, the current parameter combination is immediately marked and added to the taboo list. If all executions are normal and the yield significantly increases, it is marked as a valid combination, officially completing the batch cutting control. This provides real and reliable execution results and data support for the S5 closed-loop iteration. The overall scenario, parameters, and logic are completely consistent with the previous sections, forming a coherent closed-loop implementation process. This step avoids ineffective trial and error and improves instruction stability and cutting control security through taboo avoidance, predictive execution, and redundant coordination.

[0050] In one embodiment, the steps of calculating the historical response time of the correction amount, marking long response parameters and issuing pre-preparation instructions, dividing parameter groups according to correlation, redundancy expansion of strongly correlated groups, and adding accompanying corrections include: S410: Statistical parameter correction amount historical response time, mark parameters that exceed the average response time, issue advance preparation instructions containing correction direction, magnitude and sensitivity information, calculate correlation degree based on parameter adjustment synchronicity and divide into strong correlation groups; S411: Calculate and sort the correlation weights within the group, expand bidirectionally with the median sensitivity parameter as the core, select representative parameters according to the magnitude of the correction, and construct an expanded group that balances both sensitivity and magnitude. S412: Perform sign verification on the extended group, analyze the response time delay pattern, introduce complementary compensation parameters, form a time delay smoothing optimization group, and classify long response parameters separately; S413: Verify the effectiveness of redundant combinations, select matching parameters from weakly correlated groups as accompanying correction terms, limit the correction interval, generate standard control commands, and output the final optimized parameter combination.

[0051] In this embodiment, the response time is first statistically analyzed and long-response parameters are marked. Simultaneously, a pre-preparation command is issued and strongly correlated groups are defined. Using the 6-inch silicon carbide wafer laser cutting scenario described earlier, the historical correction response times of the two key sensitive parameters, laser power and cutting speed, are statistically analyzed. The average response time is calculated to be 1.2 seconds, with the single-shot response time for laser power reaching 2.7 seconds, significantly exceeding the average, and thus marked as a long-response parameter. A pre-preparation command is issued three control cycles in advance, explicitly including the positive correction direction, amplitude range, and sensitivity coefficient information, enabling the cutting equipment to complete optical path calibration and power preheating in advance, eliminating lag effects. Then, the correlation degree is calculated based on the historical adjustment synchronization of the parameters. The synchronization rate of laser power and cutting speed adjustments exceeds 70%, classifying them as strongly correlated parameter groups. The remaining weakly correlated parameters are categorized separately to ensure complete consistency with the parameters, response characteristics, and data caliber described earlier. Secondly, the correlation weights are calculated and a dual-balanced expansion group is constructed. The correlation weights of the parameters in the strong correlation group are calculated and sorted by sensitivity. The cutting speed with moderate sensitivity is used as the core for high and low bidirectional expansion. At the same time, the correction amount is divided into three levels: small, medium and large. The corresponding representative parameters are selected respectively to take into account the distribution of sensitivity and the rationality of the amplitude. The expansion group with dual balance of sensitivity and amplitude is constructed to avoid parameter concentration or imbalance.

[0052] Next, sign verification and time-delay smoothing optimization are performed. The correction direction of all parameters in the extended group is checked to ensure that they are all consistent with the direction of yield improvement. The time delay law of parameter response is analyzed. The laser power time delay is relatively long and the cutting speed time delay is relatively short. Line tension is introduced as a complementary compensation parameter to balance the timing difference between fast and slow parameters, forming a time-delay smoothing optimization group. Laser power is separately classified as a long response parameter and implemented with special predictive control. Finally, redundant combinations are verified and accompanying correction terms are selected to generate the final instruction. The validity of the redundant extended combinations is verified, and invalid terms with no gain or high fluctuations are eliminated. The line tension with the highest matching degree with the main parameter is selected from the weakly correlated parameter group as the accompanying correction term, and a safe correction range of ±5% is limited to prevent over-adjustment from damaging the wafer. The main parameter correction, accompanying correction, and pre-preparation instruction are integrated into a standard control instruction, outputting the final optimized parameter combination of laser power 19.3W, cutting speed 11.6mm / s, and accompanying fine-tuning of line tension. This provides a complete, stable, and safe control basis for subsequent instruction execution, taboo verification, and batch cutting. The entire process, including scenarios, values, and logic, is highly consistent with the previous steps, forming a tight closed loop. This step improves instruction response speed and operational stability through time delay prediction, correlation extension, and accompanying correction, effectively ensuring the cutting yield.

[0053] In one embodiment, the steps of calculating and sorting the correlation weights within the group, performing bidirectional expansion with the median sensitivity parameter as the core, and selecting representative parameters according to the magnitude of the correction to construct an expanded group that balances both sensitivity and magnitude include: S4110: Calculate and sort the correlation weights of the parameters of the strongly correlated group, filter the core layer according to the threshold, and determine the dynamic threshold range based on the mean of the core sensitivity. S4111: Based on this interval, perform bidirectional expansion of high and low sensitivity, group and classify according to adjustment frequency and correction magnitude, and select representative parameters for each gradient; S4112: Merge high and low frequency groups, calculate amplitude statistics and select representative parameters for each level, construct amplitude horizontal coverage and amplitude radiation axis, verify the extended group and remove abnormal parameters; S4113: Cross-overlay the sensitivity extension and amplitude coverage matrix, remove duplicates and fill gaps to complete multi-dimensional redundant extension, and combine the output matrices to form a data processing closed loop.

[0054] In this embodiment, the correlation weights of parameters within the strongly correlated group are first calculated and sorted, and the core layer is selected. Using the 6-inch silicon carbide wafer laser cutting scenario described earlier, the determined strongly correlated parameter groups of laser power and cutting speed are taken as the processing objects. The correlation between the two parameters and the overall yield fluctuation is calculated, resulting in a laser power correlation weight of 0.89 and a cutting speed correlation weight of 0.75. These parameters are then sorted from high to low weight. A weight threshold of 0.7 is set. Both parameters meet the requirements and are included in the core layer. Using the average sensitivity of the core layer parameters (0.82) as a benchmark, a dynamic threshold range of 0.74–0.90 is determined by a 10% fluctuation, providing a unified benchmark for subsequent bidirectional expansion. The entire process remains completely consistent with the key sensitive parameters, sensitivity values, and process scenarios described earlier. Secondly, based on the dynamic threshold range, bidirectional expansion of high and low sensitivity is carried out, and grouping and grading are completed according to the adjustment frequency and correction magnitude. High-sensitivity alternative process parameters are expanded above the dynamic threshold range, and low-sensitivity auxiliary process parameters are expanded below the range. At the same time, the adjustment frequency and correction magnitude of each parameter in historical processing are statistically analyzed, and all parameters are divided into three groups: high frequency with small amplitude, medium frequency with medium amplitude, and low frequency with large amplitude. Parameters with distinct characteristics and strong representativeness are selected step by step to avoid expansion that is too sparse or dense.

[0055] Next, the data within the groups are merged and an amplitude coverage system is constructed. Abnormal parameters are removed, and the high-frequency and mid-frequency groups are combined into a regular adjustment group, while the low-frequency group is retained as an extreme compensation group. The mean, variance, and range of the correction values ​​for each group are calculated. Within each gradient, the parameter closest to the statistical center is selected as the representative of that level. An amplitude horizontal coverage band is constructed with the correction value as the horizontal axis and an amplitude radiation axis is constructed with the sensitivity influence degree as the vertical axis, forming a complete two-dimensional coverage structure. The preliminary expansion group is then numerically verified, and abnormal parameters that deviate too far from the radiation axis or conflict with the core parameters are removed to ensure that the expansion structure is regular and reasonable.

[0056] Finally, multi-dimensional cross-over and redundancy optimization are performed to form a closed-loop data structure. The bidirectional sensitivity expansion results are cross-over superimposed with the amplitude coverage matrix, matched one by one according to the numerical intervals. Blank intervals appearing after superposition are filled, and duplicate parameter combinations are deduplicated, completing multi-dimensional redundancy expansion of sensitivity, amplitude, and frequency. The final output is a standardized parameter expansion matrix combination, providing structurally balanced and comprehensive data support for subsequent time-delay smoothing optimization, accompanying correction selection, and control command generation. All data calibers, parameter objects, and process logic are consistent with the previous steps, forming a complete and rigorous implementation closed loop. This step, through multi-dimensional balanced expansion and matrix superposition, makes the parameter combination more reasonable, improving control stability and yield optimization effects.

[0057] In one embodiment, the steps of collecting new batch yield data to update the sensitivity coefficient and taboo table, archiving and adjusting paths and taboo records, and returning to perform initial data collection to form an iterative optimization loop include: S50: Collect new batch yield data, compare with historical yield standards, filter parameter combinations that do not improve yield, calculate time-weighted sensitivity coefficient, mark time-sensitive parameters and degradation parameters, enter invalid combinations into the taboo table and adjust the recording weight; S51: Organize parameters to adjust the complete path, calculate the timeliness value in segments, divide the path into units and predict their success rate to form a comprehensive path sequence. S52: Classify and organize the adjusted paths and taboo records, perform differential coding and compression, extract path primitives, and remove duplicate and invalid records; S53: Verify the consistency between the archived records and the taboo list, determine the return strategy, if the standard is not met, recalibrate the parameters, if the standard is met, start iterative optimization, forming a four-driven iterative optimization loop.

[0058] In this embodiment, the first step is to collect new batch yield data and update relevant parameters. Following the 6-inch silicon carbide wafer laser cutting scenario described earlier, based on the control commands executed by S4, the cutting yield data for the new batch (of the same specifications as described earlier) is collected. Specifically, this includes the yield value, edge chipping, thickness deviation, and other data for each wafer, forming a complete yield dataset. Then, compared with historical yield standards (the previously set 95% pass threshold), parameter combinations that do not improve yield (such as a combination of 18W laser power and 3mm / s cutting speed, resulting in a yield of only 82%, failing to meet the pass standard) are identified and marked as invalid parameter combinations. Next, the time-weighted sensitivity coefficient is calculated. Based on the parameter weights determined in S1, and combined with the new batch yield data, the sensitivity coefficients of laser power and cutting speed are recalculated. The laser power sensitivity coefficient is updated to 0.89, and the cutting speed is 0.76. The time-sensitive parameter (laser power) and the stable parameter (cutting speed) are marked. At the same time, invalid parameter combinations (such as laser power 17W and cutting speed 4mm / s) are entered into the taboo table, and the weights of the taboo table are adjusted to ensure consistency with the parameter caliber and taboo table rules mentioned above.

[0059] Secondly, the parameter adjustment paths and taboo records are organized. The complete process of this parameter adjustment (from the initial parameters of 18W laser power and 12mm / s cutting speed to 19.3W laser power and corresponding cutting speed) is archived step by step. The time value of each adjustment stage is calculated segment by segment (e.g., the time value of the initial adjustment stage is 0.8, and the time value of the stabilization stage is 1.2). Path units are divided, and effective adjustment paths and invalid paths are marked to form a complete path sequence. Then, differential encoding is performed on the path sequence to extract effective adjustment segments (e.g., the path from laser power adjustment from 18W to 19.3W). Duplicate and invalid paths are eliminated, and the effective path units are organized and archived. At the same time, the correlation between the path and the yield improvement is verified to ensure that the path record corresponds one-to-one with the yield change.

[0060] Finally, the consistency of the archived data is verified and the iterative loop is initiated. The consistency of the adjustment path, taboo records, and yield data is checked. After confirming there are no conflicts, the valid path, updated sensitivity coefficients, and taboo tables are archived synchronously. Since all indicators meet the standards, there is no need to recalibrate the parameters; iterative optimization is directly initiated, returning to the initial data acquisition steps, forming a complete closed loop of "acquisition-adjustment-feedback-update." This ensures that all data (such as laser power parameters, yield values, and taboo records) remains consistent with the preceding data, with no data gaps and a logically coherent closed loop. This step achieves iterative optimization closed loop by updating parameter sensitivity, archived valid paths, and taboo records in real time, ensuring the continuity and reliability of parameter adjustments and improving yield optimization efficiency.

[0061] In one embodiment, the steps of adjusting the parameters to complete the path, calculating the timeliness value segment by segment, dividing the path into units and predicting their success rate to form a comprehensive path sequence include: S510: Organize and adjust the path according to the time sequence and divide it into segments; record key operations such as parameter adjustment and step size switching; divide the path into units and mark the execution time; divide the stage according to the success rate mutation point. S511: Retrieve historical records, assign time decay weights, calculate historical success rates using weighted averages, and calculate three-dimensional feature values ​​based on sensitivity, response speed, and stability. S512: Match the unit with the historical path pattern, determine the baseline success rate according to the matching type, establish a feature value-success rate mapping table and remove low success rate units. S513: Combine contextual similarity to correct prediction results, calculate and sort the comprehensive value based on timeliness and success rate, extract effective segments, and form a comprehensive path sequence driven by four factors.

[0062] In this embodiment, the adjustment path is first organized and segmented according to time sequence. Taking the 6-inch silicon carbide wafer cutting process as a scenario, combined with the control instructions executed in S4 above (laser power adjusted from 18W to 19.3W, and cutting speed adjusted accordingly), the entire process of this parameter adjustment is organized in time sequence. Starting from the initial parameter scheme (laser power 18W, cutting speed 12mm / s), each key node of parameter adjustment is recorded in sequence: initial parameters → parameter range division → gradient adjustment → final parameters (19.3W laser power, corresponding cutting speed). The process is segmented according to time sequence, with each segment corresponding to an operation node. The parameter value, adjustment time, and operation type (such as parameter adjustment, step size switching) of each node are recorded. The execution time of each path unit is marked to ensure complete correspondence with the parameters and steps above, with no data gaps.

[0063] Next comes the core calculation and analysis. Historical adjustment records from S1-S5 and the high-yield parameter scheme from S2 are retrieved. Combined with the parameter adjustment process in S3, different weights are assigned to different stages of the adjustment operation based on time decay (0.8 for recent adjustments and 0.2 for long-term adjustments). The success rate of historical adjustments is calculated (e.g., 92% success rate for laser power adjustment and 88% success rate for cutting speed adjustment). Simultaneously, the response speed of parameter adjustments is statistically analyzed (laser power response delay is 2 steps, cutting speed response delay is 1 step). Then comes path unit division and matching. The current adjustment path is divided into three units: initial parameter setting unit (corresponding to historical parameter retrieval in S2), parameter adjustment unit (corresponding to core calculation in S3), and result verification unit (corresponding to execution in S4). Each unit is labeled with the corresponding parameter value and adjustment time, forming a traceable path sequence. Simultaneously, adjustment path patterns from similar historical scenarios are retrieved and compared with the current path to calculate contextual similarity (e.g., consistency in parameter adjustment direction and magnitude). Historical path segments that highly match the current adjustment are selected, and their effective adjustment logic is extracted and added to the current path.

[0064] Finally, a comprehensive sorting and filtering process is performed. All path units are sorted according to their comprehensive value (success rate × timeliness weight), invalid adjustment segments (such as parameter fine-tuning without gain) are eliminated, and valid adjustment steps are retained to form a complete comprehensive path sequence. This ensures that each path unit corresponds to subsequent parameter correction and instruction execution steps, echoing the control instruction generation in S4 and the closed-loop iteration in S5. The data throughout the process remains completely consistent with the parameters (laser power, cutting speed) and scenario (6-inch silicon carbide cutting) mentioned earlier, with no data conflicts, providing complete path support for subsequent iterative optimization. This step, through standardized path organization, precise matching of historical patterns, and weighted calculation of comprehensive value, ensures the completeness and effectiveness of path records, eliminates invalid operation segments, provides a reliable path basis for subsequent iterative optimization, and facilitates continuous optimization of parameter adjustments. This step, through time-series organization, historical matching, and value sorting, accurately filters valid adjustment paths, eliminates invalid operations, ensures the continuity and reliability of subsequent iterative optimization, and improves process adjustment efficiency.

[0065] refer to Figure 2 A silicon carbide dicing control system based on wafer yield includes: The acquisition and identification module 100 is used to acquire the cutting parameters and real-time yield data of the current batch of silicon carbide wafers, calculate the yield change range to generate and sort the sensitivity coefficients, and identify key sensitive parameters. The retrieval and generation module 200 is used to retrieve the optimal parameter combination of historical high-yield batches based on key sensitive parameters, generate an initial parameter scheme, compare the deviation value with the yield threshold, and determine the basic adjustment rules. The adaptive calculation module 300 is used to adjust parameters by sorting them according to sensitivity coefficients, starting from the initial parameter scheme, adaptively switching the step size based on the yield response ratio, and calculating the parameter correction amount by fusing the deviation value and adjustment rules. The taboo execution module 400 is used to store parameter combinations that do not improve yield into the taboo table to exclude invalid combinations, generate control instructions based on parameter correction amount, and execute the next batch cutting. The closed-loop update module 500 is used to collect new batch yield data, update sensitivity coefficients and taboo tables, archive adjustment paths and taboo records, and return to execute initial data collection to form an iterative optimization loop.

[0066] Reference Figure 3 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 3 As shown, this computer device includes a processor, memory, network interface, and database connected via a bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores operations, computer programs, and the database. The internal memory provides an environment for the operation and execution of the computer programs stored in the non-volatile storage media. The database stores data such as silicon carbide cutting control methods based on wafer yield. The network interface is used for communication with external terminals via a network connection. When executed by a processor, this computer program implements a silicon carbide dicing control method based on wafer yield, including the following steps: acquiring dicing parameters and real-time yield data of the current batch of silicon carbide wafers, calculating the yield change range to generate and sort sensitivity coefficients, and identifying key sensitive parameters; retrieving the optimal parameter combination of historical high-yield batches based on the key sensitive parameters, generating an initial parameter scheme, comparing the deviation value with the yield threshold to obtain the deviation value, and determining the basic adjustment rules; starting from the initial parameter scheme, adjusting parameters according to the sensitivity coefficients, adaptively switching the step size based on the yield response ratio, and calculating the parameter correction amount by fusing the deviation value and the adjustment rules; storing parameter combinations that do not improve yield in a taboo table to exclude invalid combinations, generating control instructions based on the parameter correction amount, and executing the dicing of the next batch; acquiring new batch yield data to update the sensitivity coefficients and taboo table, archiving adjustment paths and taboo records, and returning to execute the initial data acquisition to form an iterative optimization loop.

[0067] One embodiment of this application also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements a silicon carbide dicing control method based on wafer yield, including the following steps: collecting dicing parameters and real-time yield data of the current batch of silicon carbide wafers; calculating the yield change amplitude to generate and sort sensitivity coefficients; identifying key sensitive parameters; retrieving the optimal parameter combination of historical high-yield batches based on the key sensitive parameters; generating an initial parameter scheme; comparing the deviation value with the yield threshold to obtain the deviation value and determining the basic adjustment rules; starting from the initial parameter scheme, adjusting parameters according to the sensitivity coefficients; adaptively switching the step size based on the yield response ratio; fusing the deviation value and the adjustment rules to calculate the parameter correction amount; storing parameter combinations that do not improve yield in a taboo table to exclude invalid combinations; generating control instructions based on the parameter correction amount and executing the dicing of the next batch; collecting new batch yield data to update the sensitivity coefficients and taboo table; archiving the adjustment path and taboo records; and returning to execute the initial data collection to form an iterative optimization loop.

[0068] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media provided in this application and used in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-speed SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0069] The above description is only a preferred embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural changes made based on the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A silicon carbide dicing control method based on wafer yield, characterized in that, include: Collect the cutting parameters and real-time yield data of the current batch of silicon carbide wafers, calculate the yield change range to generate and sort the sensitivity coefficients, and identify the key sensitive parameters; Based on key sensitive parameters, the optimal parameter combination of historical high-yield batches is retrieved to generate an initial parameter scheme. The deviation value is obtained by comparing with the yield threshold, and the basic adjustment rules are determined. Starting with the initial parameter scheme, the parameters are adjusted according to the sensitivity coefficient, the step size is adaptively switched based on the yield response ratio, and the parameter correction amount is calculated by integrating the deviation value and the adjustment rule. Parameter combinations that do not improve yield are stored in a taboo list to exclude invalid combinations. Control instructions are generated based on parameter correction amounts and the next batch of cutting is executed. Collect new batch yield data to update sensitivity coefficients and taboo tables, archive and adjust paths and taboo records, and return to execute initial data collection to form an iterative optimization loop.

2. The silicon carbide dicing control method based on wafer yield according to claim 1, characterized in that, The steps of collecting the cutting parameters and real-time yield data of the current batch of silicon carbide wafers, calculating the yield variation to generate and sort sensitivity coefficients, and identifying key sensitive parameters include: Each cutting parameter and real-time yield data are collected to form paired samples. After removing abnormal samples, the valid data sequence is retained. Divide the parameter value range into multiple intervals, calculate the average yield of each interval, calculate the jump difference of the average yield of adjacent intervals, and mark the mutation-sensitive parameters. Calculate the yield gradient change rate and its variance of mutation-sensitive parameters, count the number of times the yield crosses the qualified threshold, and label parameters with high response fluctuations; Calculate the conditional entropy and entropy increase of high-response fluctuation parameters and verify their monotonicity. Calculate the local peak density and value range overlap coverage. Eliminate parameters that do not meet the requirements. Verify the uniformity of the distribution of the remaining parameters and the standard deviation of the yield. Calculate the comprehensive sensitivity coefficient by weighted fusion and sort them. Select the top N as key sensitive parameters.

3. The silicon carbide dicing control method based on wafer yield according to claim 2, characterized in that, The steps of calculating the yield gradient change rate and its variance of the mutation-sensitive parameters, counting the number of times the yield crosses the qualified threshold, and marking high-response fluctuation parameters include: Bidirectional perturbation was applied to mutation-sensitive parameters to calculate the positive and negative yield gradient change rates and mean variances, the total number of threshold crossings was counted, polar oscillation parameters and gradient stratification parameters were labeled and merged into initial screening response parameters. Calculate the stratification crossing ratio and cumulative effect coefficient, cumulative gradient and crossing increment for the initial screening response parameters, count the duration of continuous exceedance and the number of continuous crossing segments, and mark the sensitive stratification parameters. For sensitive stratified parameters, the statistical direction of the duration is calculated, the cross-correlation peak of the gradient sequence is calculated, strong and persistent low symmetry parameters are marked, and the proportion of positive crossing and the ratio of crossing asymmetry are calculated. Parameters with a positive crossing percentage exceeding the threshold, a crossing asymmetry ratio deviating from the benchmark, a continuous exceeding duration below the threshold, and a continuous crossing segment number exceeding the threshold are selected. The parameters are then weighted and sorted according to multiple indicators, and the top N parameters with high response fluctuations are output.

4. The silicon carbide dicing control method based on wafer yield according to claim 1, characterized in that, The steps of retrieving the optimal parameter combination of historical high-yield batches based on key sensitive parameters, generating an initial parameter scheme, comparing the deviation value with the yield threshold to determine the basic adjustment rule include: Obtain real-time yield data for the current batch, calculate the basic deviation between the yield of a single point and the preset threshold, statistically analyze the cumulative difference and average cumulative deviation of the batch, and determine the overall adjustment direction. The yield threshold is classified and the current level is marked. The threshold is split into stages and the deviation of each stage is calculated to locate the bottleneck stage and the deviation of the yield compliance rate is statistically analyzed. Search for the best historical parameter combination at the same level, calculate the level distance and segment parameter distance, weight the weighted level distance, and classify the deviation level and adjustment level. Based on a comprehensive assessment of multiple deviations, hierarchical distances, and the proportion of bottleneck segments, and combined with the deviation type to match corresponding adjustment strategies, the final basic adjustment rules are determined.

5. The silicon carbide dicing control method based on wafer yield according to claim 1, characterized in that, The steps of starting with the initial parameter scheme, adjusting parameters according to sensitivity coefficients, adaptively switching the step size based on the yield response ratio, and calculating the parameter correction amount by fusing the deviation value and adjustment rules include: Starting with the initial parameter scheme, the sensitive layers are divided and sorted according to the sensitivity coefficient, the parameter deviation is extracted, and the step size is switched according to the yield response ratio to obtain the basic sequence of parameters to be adjusted. The real-time and cumulative yield deviations are combined to obtain the comprehensive deviation. The consistency of the parameter adjustment direction is calculated, and the weighted average is used to obtain the same-direction correction base. The opposite-direction parameters are reviewed according to the proportion of the trigger direction. The initial correction coefficient is obtained by linear interpolation of the comprehensive deviation. A time series table is established to calculate the response delay. The correction coefficient is dynamically scaled according to the delay weight and the deviation trend. Adjust the step size based on the response delay, allocate the correction amount according to priority, merge the correction coefficient and the base, smoothly transition the correction amount between layers and limit the amplitude, and output the parameter correction amount.

6. The silicon carbide dicing control method based on wafer yield according to claim 5, characterized in that, The steps of merging real-time and cumulative yield deviations to obtain a comprehensive deviation, calculating the consistency of parameter adjustment directions, weighting to obtain a correction base in the same direction, and triggering direction verification for parameters in opposite directions according to their proportions include: Read the real-time and cumulative yield deviations, obtain the time-weighted comprehensive deviations by weighting them according to the time decay factor, remove outliers and segment them, mark the strong consistency and oscillation parameters, and divide the same-direction and opposite-direction parameter sets. For parameters with the same direction, outlier data is removed, the probability of consistent direction is calculated and confidence levels are assigned, the historical correlation of parameters is calculated, and the comprehensive bias is shared among highly correlated parameter groups. Determine the consistency of parameter directions within a group, mark strongly coupled same-direction groups, calculate and combine the weighted same-direction correction base and confidence same-direction correction base for each group, and statistically summarize the sensitivity of opposite and oscillating parameters. Calculate its sensitivity ratio. If it exceeds the standard, initiate a directional review. Incorporate the undetermined parameters that meet the conditions into the same-direction correction base. After recalculation, output the final base and review flag.

7. The silicon carbide dicing control method based on wafer yield according to claim 1, characterized in that, The steps of storing parameter combinations that do not improve yield in a taboo table to exclude invalid combinations, generating control instructions based on parameter correction amounts, and executing the next batch cutting include: Read parameter combinations that do not improve yield, classify them by yield reduction and store them in a taboo table, set an expiration date and create an index, and classify them by correction amount, and gradually break down large corrections. Calculate the historical response time of the correction amount, mark long response parameters and issue pre-preparation instructions, divide the parameter groups according to the degree of correlation, perform redundant expansion on the strongly correlated groups and add accompanying corrections; Integrate progressive correction, pre-preparation and accompanying correction to generate comprehensive instructions, verify their compatibility with the taboo list, eliminate invalid instructions and classify execution priorities, and record yield feedback nodes; Instructions are executed according to timing and priority. Preparations are made in advance and then progressive corrections are made. The execution status and yield feedback are verified. Abnormalities are stored in the taboo table, and normal combinations are marked as valid combinations, thus completing the cutting control.

8. The silicon carbide dicing control method based on wafer yield according to claim 7, characterized in that, The steps of calculating the historical response time of the correction amount, marking long response parameters and issuing pre-preparation instructions, dividing parameter groups according to correlation degree, redundancy expansion of strongly correlated groups and adding accompanying corrections include: Statistical parameter corrections are analyzed based on historical response times. Parameters exceeding the average response time are marked, and advance preparation instructions containing information on correction direction, magnitude, and sensitivity are issued. The correlation degree is calculated based on parameter adjustment synchronicity, and strong correlation groups are divided. Calculate and sort the correlation weights within the group, expand bidirectionally with the median sensitivity parameter as the core, select representative parameters according to the magnitude of the correction, and construct an expanded group that balances both sensitivity and magnitude. Sign verification is performed on the extended group, the response time delay pattern is analyzed, complementary compensation parameters are introduced to form an optimization group with smooth time delay, and long response parameters are classified separately. Verify the effectiveness of redundant combinations, select matching parameters from weakly correlated groups as accompanying correction terms, limit the correction interval, generate standard control commands, and output the final optimized parameter combination.

9. The silicon carbide dicing control method based on wafer yield according to claim 1, characterized in that, The steps of collecting new batch yield data, updating the sensitivity coefficient and taboo table, archiving and adjusting paths and taboo records, and returning to execute the initial data collection to form an iterative optimization loop include: Collect new batch yield data, compare it with historical yield standards, filter parameter combinations that do not improve yield, calculate time-weighted sensitivity coefficient, mark time-sensitive parameters and degradation parameters, enter invalid combinations into the taboo table and adjust the recording weights; The parameters are adjusted to complete the path, the timeliness value is calculated segment by segment, the path unit is divided and its success rate is predicted, and a comprehensive path sequence is formed. The adjusted paths and taboo records are classified and organized, differentially encoded and compressed, path primitives are extracted, and duplicate and invalid records are removed. Verify the consistency between the archived records and the taboo list, determine the return strategy, recalibrate the parameters if the criteria are not met, and start iterative optimization if the criteria are met, forming a four-driven iterative optimization loop.

10. A silicon carbide cutting control system based on wafer yield, characterized in that, include: The acquisition and identification module is used to acquire the cutting parameters and real-time yield data of the current batch of silicon carbide wafers, calculate the yield change range to generate and sort the sensitivity coefficients, and identify key sensitive parameters. The retrieval and generation module is used to retrieve the optimal parameter combination of historical high-yield batches based on key sensitive parameters, generate an initial parameter scheme, compare the deviation value with the yield threshold, and determine the basic adjustment rules. The adaptive calculation module is used to start from the initial parameter scheme, adjust the parameters according to the sensitivity coefficient, adaptively switch the step size according to the yield response ratio, and calculate the parameter correction amount by fusing the deviation value and the adjustment rules. The taboo execution module is used to store parameter combinations that do not improve yield into the taboo table to exclude invalid combinations, generate control instructions based on parameter correction amount, and execute the next batch cutting. The closed-loop update module is used to collect new batch yield data, update sensitivity coefficients and taboo tables, archive adjustment paths and taboo records, and return to execute initial data collection to form an iterative optimization loop.

Images