A method for predicting the lifetime of a polishing pad

Abstract

This application relates to the technical field of data processing, and in particular to a method for predicting the lifetime of a polishing pad. The method includes acquiring a sensor sequence for wafer processing and operational data including entry time and recipe identifiers; extracting local increments of the sensor sequence within the processing cycle; performing time-series accumulation based on the extracted local increments to generate a degradation trajectory; after the shielding window ends, or after the variance is no greater than a preset steady-state threshold and the shielding window is released, concatenating the degradation trajectory before suspension, the trajectory segment formed by virtual increment accumulation, and the subsequent measured trajectory obtained by subsequent local increment accumulation to form a reconstructed sequence; if no shielding window is generated, the degradation trajectory is used as the reconstructed sequence; if the current value of the reconstructed sequence is less than the previous value, the previous value is used to replace the current value, generating a monotonic sequence to predict the remaining lifetime. This application decouples and removes the coupling interference between relaxation and baseline jumps through pre-processing logic, reconstructing a monotonic degradation feature sequence.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method for predicting the lifetime of a polishing pad. Background Technology

[0002] Chemical mechanical polishing (CMP) is a core step in semiconductor manufacturing to achieve global wafer planarization, and polishing pads, as a key consumable in this process, are mainly composed of polymer materials such as polyurethane. During continuous processing, polishing pads undergo a monotonous and irreversible physical degradation process, including surface trench wear and plastic micropore collapse. Accurately monitoring the actual wear state of polishing pads and predicting their remaining lifespan is crucial for ensuring yield on mass production lines and controlling manufacturing costs. In real mass production lines, equipment typically relies on high-dimensional time-series data collected by various sensors, such as those for spindle motor torque, to continuously track the health evolution of the polishing pads.

[0003] To achieve intelligent monitoring of polishing pad wear, the industry has proposed a series of data-driven prediction schemes. For example, Chinese invention patent application CN114418885A discloses a method and system for quantitatively determining the surface wear condition of a bonded abrasive polishing pad. This technology acquires surface images and three-dimensional surface topography maps of the polishing pad at different processing stages, and preprocesses the images to extract the sharpened and filtered gradient map and its spectral data, surface structure matrix, and transformation matrix, which are then merged into multi-channel surface image samples. Simultaneously, this scheme calculates quantitative surface roughness data based on the three-dimensional surface topography map and combines it with the workpiece material removal rate to obtain and calibrate the wear condition of the polishing pad. Finally, multi-channel image samples with wear condition labels are used to train a model based on a three-dimensional convolutional neural network and a lightweight long short-term memory network architecture, thereby establishing a quantitative wear condition determination model. This approach fully leverages the excellent representational power of deep learning models for high-dimensional detail features of images.

[0004] While existing deep learning-based end-to-end prediction models excel in feature mapping, their hidden memory states are easily contaminated by reversible physical perturbations when dealing with complex real-world conditions, failing to effectively decouple high-dimensional coupling interference in the feature space. In actual semiconductor production lines, polishing equipment often experiences intermittent downtime due to shift changes or routine maintenance. When polyurethane materials lose the pressure from the process, viscoelastic stress is released, causing relaxation effects in the internal polymer chains and temporary re-opening of surface pores. This results in sensors capturing a pseudo-health signal, similar to a new polishing pad, at startup. Simultaneously, mass production lines frequently require alternating processing of wafers of different materials on the same polishing pad. This cross-material switching causes abrupt, discontinuous jumps in friction baselines, such as absolute amplitude. Existing black-box prediction models that follow Markov chain-based real-time input feedback blindly absorb these physically-affected signal fluctuations, especially in extreme failure scenarios where maintenance downtime and product switching occur simultaneously. In these cases, the sensor signal directly superimposes the signal increase caused by relaxation and the baseline decrease caused by material switching. Because the model cannot accurately identify this physical ambiguity through anomalous fluctuations in numerical values, the monotonic memory cells will be completely destroyed, leading to serious defects such as violent oscillations in the predicted remaining useful life curve, local inversion, and false alarms of premature failure. Summary of the Invention

[0005] In order to decouple and isolate the coupling interference between relaxation and baseline jump by pre-logic decoupling, and reconstruct a monotonic degradation feature sequence, this application provides a method for predicting the lifetime of a polishing pad.

[0006] This application provides a method for predicting the lifespan of a polishing pad, which adopts the following technical solution: A method for predicting the lifespan of a polishing pad, comprising: Acquire sensor sequences and operational data including entry time and recipe labeling for wafer processing; Extract the local increments of the sensor sequence within the processing cycle; Based on the extracted local increments, a time-series accumulation is performed to generate a degradation trajectory; wherein, if only the formula identifier of an adjacent wafer changes, the extraction of the local increment corresponding to the first wafer after the change of the formula identifier is suspended, and the next processed wafer is anchored as the steady-state starting point. The extraction of the local increments of subsequent processed wafers is resumed based on the steady-state starting point before the accumulation is performed. If the time interval between the arrival of only adjacent wafers is greater than a preset time threshold, a shielding window is generated, the local increment of the sensor sequence is suspended, and a virtual increment is generated based on the slope of the degradation trajectory. If the recipe identifier changes simultaneously on adjacent wafers and the time interval between wafer arrivals is greater than the preset time threshold, a shielding window is generated, the local increment of the sensor sequence is extracted and suspended, and the virtual increment is continuously generated within the shielding window based on the slope of the degradation trajectory. The variance of the sensor sequence after the recipe identifier is changed is calculated. After the shielding window ends, or after the variance is no greater than a preset steady-state threshold and the shielding window is released, the degraded trajectory before suspension, the trajectory segment formed by the accumulation of the virtual increments, and the subsequent measured trajectory obtained by the accumulation of the local increments are spliced ​​together to form a reconstruction sequence; if the shielding window is not generated, the degraded trajectory is used as the reconstruction sequence. If the current value of the reconstructed sequence is less than the previous value, then the current value is replaced with the previous value to generate a monotonic sequence to predict the remaining lifetime.

[0007] Optionally, the step of continuously generating the virtual increment based on the slope of the degradation trajectory within the shielding window and calculating the variance of the sensor sequence after changing the formula identifier includes: Extract the average slope of the degradation trajectory corresponding to a preset number of wafers before suspension; The virtual increment is continuously generated by linearly projecting the virtual projection within the shielding window according to the average slope. A data sliding window is constructed synchronously, and the sensor sequence after the formula identifier is changed is imported piece by piece and steady-state feature values ​​are extracted. The sliding variance of the steady-state feature values ​​within the data sliding window is calculated as the variance. If the variance corresponding to a consecutive preset number of wafers is not greater than the preset steady-state threshold, then the linear projection is stopped.

[0008] Optionally, the step of concatenating the degraded trajectory before suspension, the trajectory segment formed by the accumulation of virtual increments, and the subsequent measured trajectory obtained by the accumulation of local increments into a reconstructed sequence includes: In response to stopping the linear projection, the end value obtained by combining the degenerate trajectory before suspension with the sum of all the virtual increments is extracted as the stitching start point; After removing the shielding window, process the corresponding local increment of the wafer according to the wafer processing sequence; Using the splicing starting point as a reference, the local increments are successively accumulated to obtain the subsequent measured trajectory; The degraded trajectory generated before suspension, the trajectory segment formed by accumulating all the virtual increments generated within the shielding window, and the subsequent measured trajectory are spliced ​​together according to the time sequence to form the reconstruction sequence.

[0009] Optionally, the step of suspending the extraction of the local increment corresponding to the first wafer after the change of the formula identifier, anchoring the next processed wafer as the steady-state starting point, and using the steady-state starting point as a reference to restore the extraction of the local increment of subsequent processed wafers before performing the accumulation includes: Determine the cross-material node where the formula identifier change occurs in adjacent wafers; The local increment corresponding to the first processed wafer after the cross-material node is suspended and extracted; Anchor the next processed wafer of the first processed wafer as the steady-state starting point; Based on the steady-state starting point, the local increment of the wafer for subsequent processing is recovered; Using the end value of the degradation trajectory accumulated before the cross-material node as a reference, the recovered and extracted local increments are successively accumulated onto the reference along the wafer processing sequence to generate the degradation trajectory that continues across materials.

[0010] Optionally, the step of replacing the current value with the previous value if the current value of the reconstructed sequence is less than the previous value, and generating a monotonic sequence to predict the remaining lifetime, includes: Along the wafer processing timeline, extract the previous value and the current value that are adjacent to each other in the reconstructed sequence; Calculate the difference between the current value and the previous value. If the difference is less than zero, then determine that the current value is less than the previous value. Replace the current value with the previous value, and use the replaced current value as the previous value for the next comparison; The extraction and replacement are performed sequentially along the time sequence until the difference between all adjacent items in the reconstructed sequence is not less than zero, and the monotonic sequence is output to predict the remaining lifetime.

[0011] Optionally, generating a monotonic sequence to predict remaining lifetime includes: Set the preset failure threshold for the polishing pad; Extract a predetermined number of consecutive sequence terms from the end of the monotonic sequence, and calculate the decay approximation slope of the sequence terms; The difference between the preset failure threshold and the value of the last term of the monotonic sequence is obtained as the available decay margin. When the decay approximation slope is greater than zero, the available decay margin is linearly extrapolated based on the decay approximation slope to estimate the remaining processing cycle required to reach the preset failure threshold, and this cycle is used as the remaining lifetime of the polishing pad; when the decay approximation slope is not greater than zero, the remaining lifetime of the most recent effective prediction is used.

[0012] Optionally, if the time interval between the arrival of only adjacent wafers is greater than a preset time threshold, a shielding window is generated to suspend the extraction of the local increment from the sensor sequence, and a virtual increment is generated based on the slope of the degradation trajectory, including: The time interval between machine entry and exit that is greater than the preset time threshold is taken as the downtime, and the number of wafers to be shielded is determined based on the downtime. The shielding window is generated with the number of wafers shielded as the span, and the local increment of the sensor sequence is extracted within the shielding window; Extract the average slope of the degradation trajectory corresponding to a preset number of wafers before suspension, and perform linear projection of the average slope within the shielding window to generate the virtual increment.

[0013] Optionally, the extraction of local increments in the sensor sequence within the processing cycle includes: The processing cycle corresponding to a single wafer is defined as an independent time window, and sensor data at the beginning and end of the preset time period within the independent time window are removed. The sensor sequences after rejection are smoothed using a moving average filter; Extract the steady-state feature values ​​of the smoothed sequence; Calculate the difference between the steady-state characteristic value corresponding to the current wafer and the steady-state characteristic value corresponding to the previous processed wafer, and align the polarity of the difference in conjunction with the evolution trend of the sensor sequence. Use the aligned difference as the local increment corresponding to the current wafer.

[0014] Optional, also includes: During the import of the data sliding window, the extraction of the local increments from the sensor sequence is suspended; The sliding variance is continuously updated based on the steady-state feature values ​​extracted after changing the formula identifier; In response to stopping the linear projection, the wafer corresponding to the sensor sequence that was last imported into the data sliding window is anchored as the steady-state starting point; Based on the steady-state starting point, the local increment of the subsequently processed wafer is recovered.

[0015] Optionally, after generating the monotonic sequence to predict the remaining lifetime, the method further includes: The predicted remaining lifespan is obtained and compared in real time with a preset intervention threshold; In response to the remaining lifetime not being greater than the preset intervention threshold, an interception command is issued; The interception command blocks subsequent wafer processing operations and triggers a polishing pad replacement warning.

[0016] In summary, this application includes the following beneficial technical effects: 1. This application employs a pre-emptive operational event determination mechanism to simultaneously monitor the time and material domains at the data entry point, accurately identifying two types of interference events: viscoelastic relaxation and frictional baseline jumps. For coupled scenarios where both types of events are triggered simultaneously, a dual-thread collaborative logic of continuous front-end virtual incremental generation and back-end sliding variance steady-state verification is adopted. Under the dual blind zone of physical ambiguity in the signal, the coupling interference of relaxation and baseline jumps can be effectively separated without relying on complex mathematical decoupling models. This fundamentally solves the core technical problem of existing technologies, which suffer from the destruction of monotonic memory cells, violent oscillations in prediction curves, and premature failure false alarms due to the inability to accurately identify physical ambiguity.

[0017] 2. This application transforms incomparable absolute signal amplitudes under multiple operating conditions into relative increments characterizing irreversible wear of the polishing pad by linking local incremental extraction with cross-material steady-state anchoring logic. Combined with the shielding window suspension and virtual incremental linear projection mechanism based on the slope of the degradation trajectory, it effectively blocks the pseudo-health signal caused by relaxation effect and the pollution of degradation trajectory by abnormal fluctuations in the physical break-in period during the initial stage of cross-material processing. It fundamentally eliminates the impact of the abrupt transition of the friction baseline caused by wafer processing of different materials on the degradation feature sequence, and realizes smooth trajectory reconstruction of cross-material continuity.

[0018] 3. This application constructs a reconstructed sequence by splicing the degradation trajectory before suspension, the trajectory segment composed of virtual incremental accumulation, and the subsequent measured trajectory, and applies a monotonicity constraint to the reconstructed sequence to generate a strictly non-decreasing monotonic sequence. This ensures that the final sequence used for prediction fully conforms to the monotonic and irreversible physical decay nature of the wear of the polishing pad surface grooves and the collapse of plastic micropores. This completely avoids the local inversion and abnormal oscillation problems that are prone to occur in the remaining lifetime prediction curve in the prior art, thereby significantly improving the stability and reliability of remaining lifetime prediction in the complex scenario of frequent operating condition switching in semiconductor mass production lines. Attached Figure Description

[0019] Figure 1 This is a logic flowchart of the lifetime prediction method according to an embodiment of this application; Figure 2 This is a schematic diagram illustrating the trajectory evolution of cross-material connection processing in an embodiment of this application; Figure 3 This is a schematic diagram of the shielding window setting and virtual incremental projection under the relaxation effect in an embodiment of this application; Figure 4 This is a logic diagram of dual-thread collaborative processing when two types of interference events are triggered simultaneously in an embodiment of this application. Detailed Implementation

[0020] The following combination Figures 1-4 This application will be described in further detail.

[0021] like Figure 1 As shown in the embodiments, this application discloses a method for predicting the lifetime of polishing pads, applied to the chemical mechanical polishing process in semiconductor manufacturing. This method addresses the core deficiencies of existing technologies by constructing a dual-logic decoupling and virtual degradation trajectory reconstruction method based on operating condition events. Through pre-processing operating condition event determination, interference decoupling, trajectory reconstruction, and monotonicity constraints, physical artifacts in sensor signals are removed, generating a feature sequence that conforms to the physical laws of irreversible degradation of polishing pads. This enables stable and accurate prediction of the remaining lifetime of polishing pads under complex operating conditions.

[0022] When implementing this method, the sensor sequence of the polishing equipment throughout the entire wafer processing flow, as well as the corresponding operating condition data for each wafer, are collected synchronously. The sensor sequence covers several types of continuous time-series data directly related to polishing pad wear, such as spindle motor torque, polishing head pressure, acoustic emission signals, and polishing chamber temperature. The sampling frequency is adapted to the signal type. Slow-changing process signals such as spindle torque, polishing head pressure, and polishing chamber temperature are sampled at frequencies from 10Hz to 100Hz, while high-frequency transient signals such as acoustic emission are sampled at frequencies from 100kHz to 1MHz. This approach not only fully captures the characteristic changes corresponding to polishing pad wear but also avoids generating excessive redundant data.

[0023] For each sensor sequence acquired, subsequent full-process steps are executed independently, and the polishing pad life is ultimately determined based on the prediction results of each sequence. Simultaneously, the system accesses the operating condition data issued by the Manufacturing Execution System (MES) via the SECS / GEM protocol, a common protocol for semiconductor equipment. The operating condition data for each wafer includes a unique arrival time and recipe identifier. The recipe identifier is directly linked to the wafer material category; wafers of the same material use a unified recipe identifier, while wafers of different materials use mutually distinguishable recipe identifiers.

[0024] After completing the basic data acquisition, using the arrival time of each wafer as the core benchmark, the corresponding operating condition data for each wafer is bound one-to-one with the sensor sequences collected within the same time period. After binding, the time coverage of the operating condition data and sensor sequences for each wafer is verified to ensure that the complete processing cycle of a single wafer from arrival to departure corresponds to a uniquely matched sensor sequence and operating condition data, preventing data crosstalk between adjacent wafers. This establishes a unified timing benchmark for subsequent wafer-by-wafer data processing and event determination.

[0025] Based on multi-source data with completed time-series binding, a dual-path parallel logic monitoring logic is constructed, namely time domain monitoring and material domain monitoring. Both monitoring paths use a single wafer as the smallest execution unit and run the judgment logic synchronously. Unlike the existing technology that processes all sensor data indiscriminately, this approach completes the pre-identification of interference sources from the data entry point.

[0026] Before execution, two types of core thresholds are preset. One type is the preset time threshold. The threshold setting is determined based on the Shore A hardness of the polyurethane substrate of the polishing pad. The higher the Shore A hardness, the longer the viscoelastic relaxation recovery time of the polyurethane substrate, and the higher the corresponding preset time threshold setting. For polyurethane polishing pads with a Shore hardness of 50HA to 80HA, the preset time threshold is set to 10 minutes by default. When the Shore hardness is lower than 60HA, it can be adjusted to 5-8 minutes, and when it is higher than 70HA, it can be adjusted to 12-15 minutes.

[0027] Another type is the preset steady-state threshold. Since steady-state determination uses a sliding window containing data from multiple wafers, the preset steady-state threshold needs to be reasonably set based on the historical variance fluctuation benchmark within the sliding window span of the polishing pad under stable processing conditions. Stable processing conditions refer to the working condition when more than 10 wafers of the same material are processed smoothly. The higher the process precision requirement, the more stringent the corresponding steady-state threshold setting. The default steady-state threshold for conventional wafer planarization processes is 1.1 to 1.5 times the mean variance of steady-state characteristic values ​​within a historical sliding window of the same span. This completely avoids steady-state misjudgment caused by confusing the microscopic variance within a single wafer with the macroscopic variance between multiple wafers, providing a scientific basis for judging the smooth convergence of signals under new working conditions.

[0028] After completing the preliminary judgment of the working condition event, the local increments of the corresponding sensor sequence are extracted one by one along the wafer processing timeline. First, the wafer entry time and exit time, which have been aligned and bound by the timeline, are used as the starting and ending boundaries to define the complete processing cycle of a single wafer as an independent time window. Each independent time window corresponds to the entire processing flow of only one wafer, so that the subsequent increment extraction can accurately correspond to the processing of a single wafer.

[0029] Next, the sensor sequence bound to the independent time window is retrieved. Sensor data at the beginning and end of the preset time period within the independent time window are first removed. The preset time period at the beginning and end is set to 5% of the single wafer processing cycle, with a minimum of 2s and a maximum of 5s. This setting can fully cover the transient fluctuation stages of wafer cutting into the polishing pad, such as speed-up and pressure building, and speed-down and pressure-relief, without excessively losing effective data from the stable processing segment.

[0030] After removing the first and last data, the sensor sequence was smoothed by a moving average filter. The window length of the moving average filter was adaptively set to the number of sampling points corresponding to 0.1s at the corresponding sensor sampling frequency. This not only smoothed out the white noise and random fluctuations in the sampling process, but also avoided excessively smoothing out the signal evolution trend caused by the continuous wear of the polishing pad.

[0031] After smoothing, the mean value of the middle section or the overall mean value of the smoothed sequence is extracted as the steady-state characteristic value. Then, the difference between the steady-state characteristic value corresponding to the current wafer and the steady-state characteristic value corresponding to the previous processed wafer is calculated. This difference is then polarized by considering the physical trend of wear evolution in the current sensor sequence. The aligned difference is used as the local increment corresponding to the current single wafer. The reason for performing polarity alignment is to perfectly accommodate positively correlated signals such as polishing cavity temperature, which increases with wear work, and negatively correlated signals such as spindle torque, which decreases with surface aging, uniformly converting them into positive values ​​characterizing the polishing pad consumption trend.

[0032] Meanwhile, when the processing is subjected to minor physical disturbances that cause reverse abnormal fluctuations, the increment after polarity alignment can accurately present a negative value, thus truly preserving these negative perturbation characteristics. This provides real underlying data support for the subsequent precise removal of artifacts through monotonicity constraints. The local increment extracted here objectively characterizes the macroscopic baseline drift of the polishing pad during the alternation of processing adjacent wafers. This is completely different from the existing technology that directly compares the start and end signal drop within a single wafer, thus avoiding the interference of workpiece planarization fluctuations on the consumable degradation characteristics from a physical perspective.

[0033] After extracting the local increments of a single wafer, the effective local increments extracted one wafer at a time are continuously accumulated along a defined wafer processing sequence. The accumulated time sequence is the degradation trajectory of the polishing pad. The temporal trend of this degradation trajectory corresponds perfectly to the physical degradation process of groove wear and plastic micropore collapse on the polishing pad surface. This transforms the discrete wear characteristics of a single wafer into a continuous full-lifecycle degradation time sequence curve, completely reconstructing the entire wear evolution of the polishing pad from brand new to failure.

[0034] If the previous material domain monitoring only triggered the friction baseline transition event, it will first lock the cross-material node corresponding to the formula identifier change. This node is the timing boundary between two adjacent wafers with different formula identifiers. During the initial single-wafer processing of cross-material switching, the polishing pad surface has a short physical break-in period to adapt to the new friction baseline, and the sensor signal is very susceptible to abnormal fluctuations caused by the break-in period.

[0035] At this stage, if the calculated incremental results are simply discarded, the volatile steady-state eigenvalues ​​of this wafer will still be used as subtractions in the underlying difference formula for the calculation of the next normal wafer. This will inevitably inject huge pseudo-incremental noise in the reverse, causing a chain reaction of underlying mathematical errors and completely contaminating the subsequent degradation trajectory. Therefore, it is necessary to directly suspend the extraction of the local increment corresponding to the first processed wafer after the cross-material node and anchor the next processed wafer of this first processed wafer as the steady-state starting point, because at this time the polishing pad has basically passed the short physical break-in period.

[0036] Then, using this steady-state starting point as a benchmark, the local increment of the subsequent processed wafer is recovered and extracted. That is, the wafer corresponding to the steady-state starting point is used as the first comparison benchmark, and the logic of calculating the difference between the steady-state characteristic value of the current processed wafer and the previous processed wafer is recovered, so as to obtain the pure local increment one by one.

[0037] Finally, using the end value of the degradation trajectory accumulated before the cross-material node as a reference, the recovered local increments are successively accumulated onto this reference value along the wafer processing timeline to generate a cross-material continuous degradation trajectory. In this way, there is no need to perform complex normalization calibration on the signal baselines of different materials, and the cascading pollution of the differential formula by abnormal fluctuations during the cross-material break-in period is completely cut off from the mathematical level. This perfectly realizes the smooth continuation of degradation trajectories under different processing conditions and avoids the problems of trajectory abrupt changes and wear characteristic distortion caused by cross-material processing in existing technologies.

[0038] like Figure 2 As shown in the figure, the local incremental processing logic is intuitively illustrated when a change in the formulation identifier of an adjacent wafer (triggering a cross-material node) occurs. The scattered points in the figure clearly indicate the location of "suspending the first wafer" and the subsequent action of anchoring the next wafer as the "steady-state starting point of the new material." This illustration clearly reveals that this method does not require complex benchmark normalization calibration. It simply suspends the abnormal fluctuation data of the first wafer after the material transition and recalculates the local increment based on the next wafer at the steady-state starting point. This physically severs the cascading contamination of the degradation trajectory caused by the initial break-in anomalies during the transition between old and new materials, achieving a smooth continuation of local increments across operating conditions.

[0039] If a viscoelastic relaxation event is triggered by time-domain monitoring, the time interval between incoming processes exceeding a preset time threshold will first be used as the downtime. Then, based on the downtime and a preset mapping rule, the number of wafers to be shielded will be determined. The preset mapping rule mentioned here is uniformly... ,in To take the smaller function, The rounding up symbol, This refers to the number of wafers used for shielding, measured in wafers. The downtime is expressed in minutes. This is the empirical coefficient for positive correlation, with a value ranging from 0.2 to 1. The basic compensation coefficient ranges from 2 to 5. The maximum number of wafers to be shielded is preset to match the relaxation saturation limit of the polymer polyurethane material. The value is adjusted synchronously with the Shore A hardness of the polyurethane substrate of the polishing pad. When the Shore A hardness is between 50HA and 60HA... For Shore hardness, the range is 0.2 to 0.4; for 60HA to 70HA, the range is 0.4 to 0.7; and for 70HA to 80HA, the range is 0.7 to 1. The higher the Shore hardness, the longer the relaxation recovery period of the polyurethane material. A larger value allows for more precise matching of the relaxation characteristics of polishing pads made of different materials, while also... Effectively constrains the infinite divergence of the shielding span under extreme operating conditions.

[0040] Next, a shielding window is generated with the calculated number of wafers shielded as the span. Within the span of the shielding window, the local incremental extraction and accumulation operations of the corresponding sensor sequence of the subsequent wafers are suspended one by one. The sensor data is still collected and stored normally, only blocking the calculation link that leads to the degradation trajectory, thus avoiding the false health signal pollution degradation trajectory caused by the relaxation effect from the root.

[0041] While generating the shielding window, the average slope of the degradation trajectory corresponding to 3 to 5 consecutive wafers of the same material and under stable processing conditions before suspension is extracted. The average slope is obtained by performing a univariate linear fitting on the selected consecutive wafer sequence points with the cumulative number of processed wafers as the x-axis and the degradation trajectory value as the y-axis. This range of numbers can not only fully cover the degradation characteristics of the stable processing state before shutdown, but also avoid the influence of random fluctuations in single wafer data on the accuracy of the slope, ensuring that the calculated average slope completely matches the actual wear trend of the polishing pad.

[0042] Then, based on the calculated average slope, a linear projection is performed within the generated shielding window to generate virtual increments that match the wafer processing timing, one by one. The dimensions of the virtual increments are completely consistent with the previously extracted local increments. These virtual increments are used to fill the increment gaps within the shielding window to ensure the timing continuity and physical rationality of the degradation trajectory under shutdown conditions.

[0043] like Figure 3 As shown in the figure, the data-targeted intervention strategy for the viscoelastic relaxation effect of the polishing pad caused by shutdown is illustrated. The figure clearly defines the stage triggered by "shutdown" and plots the "shielding window" interval adaptively generated based on the shutdown duration. Simultaneously, dashed lines visually demonstrate the virtual incremental projection within the shielding window based on the slope of the degradation trajectory before shutdown. This figure proves that the proposed method can accurately identify and isolate the relaxation fallback illusion (pseudo-health signal) caused by shutdown. By directly filling the characteristic gap of this stage through linear projection of the historical degradation trend, it fundamentally blocks the contamination of the monotonic physical wear trajectory by the reversible relaxation effect, effectively ensuring the temporal continuity and physical rationality of the degradation trajectory under abnormal operating conditions.

[0044] If both time-domain and material-domain monitoring simultaneously determine that a viscoelastic relaxation event and a friction baseline jump event are triggered, a shielding window is generated to suspend the extraction of local increments from the sensor sequence. The termination time of this shielding window is dynamically determined by the subsequent steady-state determination result. This approach is necessary because when both types of events are triggered simultaneously, the polishing pad is both in the nonlinear recovery period of material relaxation and facing a precipitous jump in the friction baseline caused by the processing of the new material. The sensor signal exhibits two completely opposite fluctuation trends, resulting in strong physical ambiguity. A fixed-length shielding window either cannot completely cover the interference period or excessively shields the area, affecting processing efficiency. Current technology is completely unable to decouple this coupling interference. Subsequently, a dual-threaded processing flow is simultaneously initiated, involving continuous generation of virtual increments in the foreground and steady-state verification in the background. This silently completes the steady-state determination of the new operating condition without interrupting the decay trajectory.

[0045] In the front-end processing flow, the average slope of the degradation trajectory corresponding to 3 to 5 consecutive stable processed wafers of the same material before suspension is used. Linear projection is continuously performed within an extended shielding window according to this average slope. A corresponding virtual increment is generated synchronously after each wafer is processed, until a command to stop linear projection is received. Throughout this process, the trajectory is always continued based on the stable degradation trend before shutdown, completely unaffected by fluctuation signals under new operating conditions, ensuring the consistency of the degradation trajectory throughout the entire cycle.

[0046] In the background processing flow, a data sliding window is constructed synchronously. The length of the data sliding window is set to 3 wafers, and the sliding step is 1 wafer. This window length ensures the stability of variance calculation and keeps the delay of steady-state determination within an acceptable range for mass production lines, without excessively affecting subsequent incremental extraction. Then, the sensor sequences collected after the formula change are imported wafer by wafer. As soon as a new wafer's data is imported, the dataset in the window is updated synchronously. The steady-state characteristic values ​​of the corresponding sensor sequences for each wafer in the data sliding window are extracted, and the variance of the steady-state characteristic values ​​corresponding to multiple wafers in the window is calculated and used as the sliding variance for determination.

[0047] During the data import process in the data sliding window, the local incremental extraction operation is continuously suspended to avoid invalid incremental pollution decay trajectories extracted when the baseline is not stable. At the same time, based on the steady-state characteristic values ​​continuously extracted after the formula identifier is changed, the sliding variance value is updated synchronously to track in real time whether the macroscopic absolute friction baseline has converged and stabilized under the new operating conditions.

[0048] The real-time calculated sliding variance is continuously compared with a preset steady-state threshold. If the sliding variance of two to three consecutive wafers is not greater than the preset steady-state threshold, the steady-state misjudgment caused by accidental fluctuations in single-batch data is avoided. This ensures that the shielding is only removed after the polishing pad has truly passed the material relaxation period and fully adapted to the friction baseline of the new material. Once the judgment condition is met, the logic determines that the polishing pad has entered a stable processing state, issues a command to stop linear projection, and officially removes the shielding window.

[0049] In response to the command to stop linear projection, the wafer corresponding to the sensor sequence of the last imported data sliding window is anchored as the steady-state starting wafer. Starting from the next wafer after the steady-state starting wafer, the local incremental extraction operation resumes. Then, the end value obtained by combining the accumulated degradation trajectory generated before suspension with the accumulated virtual increments within the shielding window is used as the stitching starting point. The local increments corresponding to the wafers processed along the wafer processing time sequence after the shielding window is removed are obtained. Using the stitching starting point as the reference, the subsequently extracted local increments are accumulated upwards to obtain the subsequent measured trajectory. The trajectory segment formed by the accumulated degradation trajectory generated before suspension and the accumulated virtual increments within the shielding window is then completely stitched together with the subsequent measured trajectory in chronological order to form a reconstruction sequence that is not only chronologically continuous but also completely preserves the absolute cumulative wear amount throughout the entire life cycle. If no shielding window is generated during the entire wafer processing process, the previously accumulated degradation trajectory is directly used as the reconstruction sequence to complete the reconstruction of the degradation characteristic sequence under all operating conditions.

[0050] like Figure 4 As shown in the figure, the dual-thread collaborative mechanism of simultaneous operation of foreground degradation trajectory continuation and background sliding variance monitoring is illustrated. The upper curve represents the system's ability to continuously project virtual increments linearly at the original slope within the extended shielding window, unaffected by lower-level signal interference. The lower curve tracks the sliding variance change of the new material's steady-state eigenvalues. When the sliding variance curve continuously falls below the preset steady-state threshold, a variance convergence signal is precisely triggered to release the shielding window, completing the smooth splicing of the foreground virtual trajectory and the subsequent measured trajectory. This figure demonstrates that when facing the physical ambiguity blind zone caused by the superposition of relaxation effects and baseline jumps, this invention does not rely on complex black-box models for forced fitting, but only on logical decoupling and steady-state verification to perfectly isolate coupling interference, achieving smooth reconstruction and seamless integration of extreme operating condition data.

[0051] After obtaining the reconstructed sequence, its monotonicity is enforced to generate a monotonic sequence that conforms to the physical law of irreversible decay of the polishing pad. Along the wafer processing timeline, adjacent previous and current sequence values ​​within the reconstructed sequence are extracted one by one. The difference between the current and previous sequence values ​​is then calculated. If the difference is less than zero, the current sequence value is considered less than the previous sequence value, indicating a negative perturbation that was not completely intercepted by the preceding steps. The current sequence value is then replaced with the previous sequence value, and the replaced current sequence value is used as the previous sequence value for the next comparison. This extraction and replacement operation is repeated along the timeline until the difference between all adjacent terms in the reconstructed sequence is not less than zero, outputting a strictly non-decreasing monotonic sequence. This process perfectly matches the monotonic irreversible physical decay nature caused by the wear of surface grooves and the collapse of plastic micropores on the polishing pad, completely avoiding the local inversion and abnormal oscillation problems that easily occur in prediction curves in existing technologies. It ensures the physical rationality of the subsequent lifetime prediction input from the data layer.

[0052] After generating a monotonic sequence, the remaining lifetime of the polishing pad can be predicted based on this sequence. First, a preset failure threshold for the polishing pad is set. Since this solution generates the degradation trajectory by extracting and successively accumulating tiny local pure increments, representing an absolute increase in wear value as processing continues, the preset failure threshold must be set based on the maximum cumulative wear increment limit calibrated from the historical full-lifecycle failure data of the same model of polishing pad. This threshold setting combines the physical wear characteristics of the polyurethane polishing pad with the yield requirements of the wafer planarization process. In a brand-new state, the surface trenches of the polishing pad are intact and the micropore structure is full. As processing continues, the physical wear accumulation increases synchronously. When this cumulative characteristic value reaches the preset failure threshold, it means that the trench wear and micropore collapse of the polishing pad have reached a critical state, making it impossible to stably guarantee the global planarization effect of the wafer, and the process yield will decline significantly.

[0053] Then, the sequence terms of 5 to 10 consecutive wafers from the end of the monotonic sequence are extracted. This range of numbers can fully capture the stable degradation trend of the polishing pad in the current stage, while also avoiding the impact of random fluctuations of a small number of sequence terms on the slope calculation. Next, using the wafer processing time (cumulative number of processed wafers) as the independent variable and the corresponding value of the monotonic sequence as the dependent variable, a univariate linear fit is performed on the extracted sequence terms to obtain the degradation approximation slope of the sequence terms. Then, the difference between the preset failure threshold and the value of the last term of the monotonic sequence is obtained as the usable degradation margin.

[0054] To avoid system crashes caused by the monotonicity-forced smoothing in the preceding steps, which could result in a horizontal constant sequence and a slope approaching zero, a linear extrapolation is performed on the available decay margin based on the calculated decay approximation slope to estimate the remaining processing cycles required to reach the preset failure threshold. If the decay approximation slope is not greater than zero, the extrapolation is suspended, and the most recent effective predicted slope or remaining lifetime is automatically used. Finally, the remaining processing cycles are used as the remaining lifetime of the polishing pad. For multiple independently processed sensor sequences, the corresponding remaining lifetimes are predicted separately, and the minimum value is taken as the final remaining lifetime of the polishing pad.

[0055] After predicting the remaining lifetime, the lifetime warning and processing intervention logic are executed simultaneously. The predicted remaining lifetime is obtained and compared in real time with a preset intervention threshold. The preset intervention threshold is set within a reasonable range greater than the number of wafers processed in a single batch, such as 30 to 50 wafer processing cycles. This threshold setting is deeply adapted to the continuous production rhythm of semiconductor mass production lines. Considering that the number of wafers processed in a single batch (lot) on a mass production line is typically 25, setting an intervention threshold greater than this batch number allows sufficient time for polishing pad preparation, replacement, and equipment debugging, ensuring the smooth completion of the current batch without forced interruption of production. This completely avoids interrupting continuous production of the entire batch due to temporary consumable replacement. If the remaining lifetime is not greater than the preset intervention threshold, an interception command is issued, and the processing of subsequent batches of wafers awaiting entry is blocked based on the interception command, simultaneously triggering a polishing pad replacement warning.

[0056] In summary, this method, through the setting of dual-path parallel logic monitoring, achieves accurate identification and hierarchical processing of two types of interference events: viscoelastic relaxation and frictional baseline jumps. This provides an event-level triggering basis for subsequent interference decoupling, avoiding the disturbance contamination problem caused by the indiscriminate absorption of the entire signal in existing technologies. Through the linkage logic of local incremental extraction and cross-material steady-state anchoring, the incomparable absolute signal amplitudes under multiple operating conditions are transformed into relative increments characterizing the irreversible wear of the polishing pad, fundamentally eliminating the impact of the abrupt transitions in the frictional baseline caused by wafer processing of different materials on the decay trajectory. Furthermore, by using a mechanism of shielding window suspension and virtual incremental linear projection, the relaxation... The pseudo-health signal caused by the hesitant effect contaminates the degradation trajectory, while ensuring the temporal continuity and physical rationality of the degradation trajectory during downtime. For coupled scenarios where two types of interference events are triggered simultaneously, a dual-thread collaborative logic of continuous front-end virtual incremental generation and back-end sliding variance steady-state verification achieves smooth reconstruction of the degradation trajectory even in the dual blind zone of physical ambiguity in the signal, eliminating interference removal without relying on complex mathematical decoupling models. Finally, through monotonicity-forced constraints, it ensures that the sequence used for lifetime prediction fully conforms to the physical laws of irreversible polishing pad degradation, completely eliminating oscillations and local inversions in the remaining lifetime prediction curve. This method significantly reduces the feature decoupling burden in the lifetime prediction process, and can still output stable and reliable remaining lifetime prediction results even in complex scenarios with frequent operating condition switching in semiconductor mass production lines.

[0057] The above are all preferred embodiments of this application, and are not intended to limit the scope of protection of this application. Therefore, all equivalent changes made in accordance with the structure, shape and principle of this application should be covered within the scope of protection of this application.

Claims

1. A method for predicting the lifespan of a polishing pad, characterized in that, include: Acquire sensor sequences and operational data including entry time and recipe labeling for wafer processing; Extract the local increments of the sensor sequence within the processing cycle; Based on the extracted local increments, a time-series accumulation is performed to generate a degradation trajectory; wherein, if only the formula identifier of an adjacent wafer changes, the extraction of the local increment corresponding to the first wafer after the change of the formula identifier is suspended, and the next processed wafer is anchored as the steady-state starting point. The extraction of the local increments of subsequent processed wafers is resumed based on the steady-state starting point before the accumulation is performed. If the time interval between the arrival of only adjacent wafers is greater than a preset time threshold, a shielding window is generated, the local increment of the sensor sequence is suspended, and a virtual increment is generated based on the slope of the degradation trajectory. If the recipe identifier changes simultaneously on adjacent wafers and the time interval between wafer arrivals is greater than the preset time threshold, a shielding window is generated, the local increment of the sensor sequence is extracted and suspended, and the virtual increment is continuously generated within the shielding window based on the slope of the degradation trajectory. The variance of the sensor sequence after the recipe identifier is changed is calculated. After the shielding window ends, or after the variance is no greater than a preset steady-state threshold and the shielding window is released, the degraded trajectory before suspension, the trajectory segment formed by the accumulation of the virtual increments, and the subsequent measured trajectory obtained by the accumulation of the local increments are spliced ​​together to form a reconstruction sequence; if the shielding window is not generated, the degraded trajectory is used as the reconstruction sequence. If the current value of the reconstructed sequence is less than the previous value, then the current value is replaced with the previous value to generate a monotonic sequence to predict the remaining lifetime.

2. The method for predicting the lifespan of a polishing pad according to claim 1, characterized in that, The process of continuously generating the virtual increment within the shielding window based on the slope of the degradation trajectory, and calculating the variance of the sensor sequence after changing the formula identifier, includes: Extract the average slope of the degradation trajectory corresponding to a preset number of wafers before suspension; The virtual increment is continuously generated by linearly projecting the virtual projection within the shielding window according to the average slope. A data sliding window is constructed synchronously, and the sensor sequence after the formula identifier is changed is imported piece by piece and steady-state feature values ​​are extracted. The sliding variance of the steady-state feature values ​​within the data sliding window is calculated as the variance. If the variance corresponding to a consecutive preset number of wafers is not greater than the preset steady-state threshold, then the linear projection is stopped.

3. The method for predicting the lifespan of a polishing pad according to claim 2, characterized in that, The step of concatenating the degraded trajectory before suspension, the trajectory segment formed by the accumulation of virtual increments, and the subsequent measured trajectory obtained by the accumulation of local increments into a reconstructed sequence includes: In response to stopping the linear projection, the end value obtained by combining the degenerate trajectory before suspension with the sum of all the virtual increments is extracted as the stitching start point; After removing the shielding window, process the corresponding local increment of the wafer according to the wafer processing sequence; Using the splicing starting point as a reference, the local increments are successively accumulated to obtain the subsequent measured trajectory; The degraded trajectory generated before suspension, the trajectory segment formed by accumulating all the virtual increments generated within the shielding window, and the subsequent measured trajectory are spliced ​​together according to the time sequence to form the reconstruction sequence.

4. The method for predicting the lifespan of a polishing pad according to claim 1, characterized in that, The process of suspending the extraction of the local increment corresponding to the first wafer after the change of the formula identifier, anchoring the next processed wafer as the steady-state starting point, and using the steady-state starting point as a reference to recover the extraction of the local increment of subsequent processed wafers before performing the accumulation includes: Determine the cross-material node where the formula identifier change occurs in adjacent wafers; The local increment corresponding to the first processed wafer after the cross-material node is suspended and extracted; Anchor the next processed wafer of the first processed wafer as the steady-state starting point; Based on the steady-state starting point, the local increment of the wafer for subsequent processing is recovered; Using the end value of the degradation trajectory accumulated before the cross-material node as a reference, the recovered and extracted local increments are successively accumulated onto the reference along the wafer processing sequence to generate the degradation trajectory that continues across materials.

5. The method for predicting the lifespan of a polishing pad according to claim 1, characterized in that, The step of replacing the current value with the previous value if the current value of the reconstructed sequence is less than the previous value, and generating a monotonic sequence to predict the remaining lifetime, includes: Along the wafer processing timeline, extract the previous value and the current value that are adjacent to each other in the reconstructed sequence; Calculate the difference between the current value and the previous value. If the difference is less than zero, then determine that the current value is less than the previous value. Replace the current value with the previous value, and use the replaced current value as the previous value for the next comparison; The extraction and replacement are performed sequentially along the time sequence until the difference between all adjacent items in the reconstructed sequence is not less than zero, and the monotonic sequence is output to predict the remaining lifetime.

6. The method for predicting the lifespan of a polishing pad according to claim 5, characterized in that, The generation of monotonic sequences to predict remaining lifetime includes: Set the preset failure threshold for the polishing pad; Extract a predetermined number of consecutive sequence terms from the end of the monotonic sequence, and calculate the decay approximation slope of the sequence terms; The difference between the preset failure threshold and the value of the last term of the monotonic sequence is obtained as the available decay margin. When the decay approximation slope is greater than zero, the available decay margin is linearly extrapolated based on the decay approximation slope to estimate the remaining processing cycle required to reach the preset failure threshold, and this cycle is used as the remaining lifetime of the polishing pad; when the decay approximation slope is not greater than zero, the remaining lifetime of the most recent effective prediction is used.

7. The method for predicting the lifespan of a polishing pad according to claim 1, characterized in that, If the time interval between the arrival of only adjacent wafers is greater than a preset time threshold, a shielding window is generated, suspending the extraction of the local increment from the sensor sequence, and generating a virtual increment based on the slope of the degradation trajectory, including: The time interval between machine entry and exit that is greater than the preset time threshold is taken as the downtime, and the number of wafers to be shielded is determined based on the downtime. The shielding window is generated with the number of wafers shielded as the span, and the local increment of the sensor sequence is extracted within the shielding window; Extract the average slope of the degradation trajectory corresponding to a preset number of wafers before suspension, and perform linear projection of the average slope within the shielding window to generate the virtual increment.

8. The method for predicting the lifespan of a polishing pad according to claim 1, characterized in that, The extraction of local increments in the sensor sequence within the processing cycle includes: The processing cycle corresponding to a single wafer is defined as an independent time window, and sensor data at the beginning and end of the preset time period within the independent time window are removed. The sensor sequences after rejection are smoothed using a moving average filter; Extract the steady-state feature values ​​of the smoothed sequence; Calculate the difference between the steady-state characteristic value corresponding to the current wafer and the steady-state characteristic value corresponding to the previous processed wafer, and align the polarity of the difference in conjunction with the evolution trend of the sensor sequence. Use the aligned difference as the local increment corresponding to the current wafer.

9. The method for predicting the lifespan of a polishing pad according to claim 2, characterized in that, Also includes: During the import of the data sliding window, the extraction of the local increments from the sensor sequence is suspended; The sliding variance is continuously updated based on the steady-state feature values ​​extracted after changing the formula identifier; In response to stopping the linear projection, the wafer corresponding to the sensor sequence that was last imported into the data sliding window is anchored as the steady-state starting point; Based on the steady-state starting point, the local increment of the subsequently processed wafer is recovered.

10. The method for predicting the lifespan of a polishing pad according to claim 1, characterized in that, Following the generation of the monotonic sequence to predict the remaining lifetime, the method further includes: The predicted remaining lifespan is obtained and compared in real time with a preset intervention threshold; In response to the remaining lifetime not being greater than the preset intervention threshold, an interception command is issued; The interception command blocks subsequent wafer processing operations and triggers a polishing pad replacement warning.

Images