CMP end point detection method and system and dynamic window zooming method

Abstract

The invention discloses a CMP end point detection method and system and a dynamic window scaling method.The CMP end point detection method judges an end point based on window logic and comprises the steps that a friction force signal in the grinding process is obtained, and a friction force curve is obtained according to the friction force signal; the window is dynamically adjusted to generate a dynamic window, the dynamic adjustment comprises the step of dynamically adjusting the width of the window in real time by adopting a window width dynamic contraction function, and the window width dynamic contraction function is as shown in the formula (1): W (t) = Wmax * e (-kw * df / dt); wherein W (t) is the window real-time width corresponding to a certain time point t, Wmax is the maximum window width, f in df / dt is the friction force, kw is the shrinkage sensitivity coefficient, and kw is larger than 0; and monitoring the state of the friction curve leaving the window, and if leaving from the right boundary of the window, determining that the grinding end point is reached. The width of the end point area window can be reduced, an end point signal is triggered in advance, and the over-throwing phenomenon is avoided to a great extent.

Description

Technical Field

[0001] This invention belongs to the field of semiconductor manufacturing technology, specifically relating to a CMP endpoint detection method, system, and dynamic window scaling method. Background Technology

[0002] In CMP (Chemical Mechanical Polishing) processes, endpoint detection (EPD) is crucial for controlling process defects. In the front-end CMP process, motor torque EPD (friction endpoint detection) is commonly used. It's based on the fact that the friction force varies at different film layers during CMP polishing, and the computer determines the polishing endpoint by detecting changes in this value. After collecting the friction force signal, the computer's endpoint determination algorithm often uses window logic. The size of the judgment window is manually set, and the friction force curve is analyzed based on the window's state to determine the polishing endpoint. The friction force curve's state through the window is broadly categorized into three types: window out+, window out-, and window in. The curve is a friction force curve plotted from the friction force signal collected by the computer, and the window is a manually set judgment window. The friction force curve exists in three states through the window, such as... Figure 1 As shown: when the friction force curve passes the upper boundary of the judgment window, it is in the window out+ state; when the friction force curve passes the lower boundary of the judgment window, it is in the window out- state; when the friction force curve passes the right boundary of the judgment window, it is in the window in state. Taking shallow trench isolation STI CMP as an example, when the friction force curve passes through the judgment window and the state changes from window out- to window in, the computer will determine that the polishing endpoint has been reached and stop polishing.

[0003] However, window logic has the following problems: 1. The window size is manually set, which requires a high level of experience from engineers, especially during the pre-mass production setup and testing phases. Without sufficient data support, window size settings rely entirely on experience, but the actual situation at each FAB plant, and even for each machine, is different. 2. The window size setting significantly impacts the judgment of CMP endpoint detection. Since the friction curve must pass completely through the judgment window, setting the window size too large (e.g., excessively wide window) exacerbates over-throwing. Setting the window size too small (e.g., smaller window width) may improve over-throwing, but it increases the false positive rate of CMP endpoint detection. 3. Once the window size parameter is set, it is not adjustable in the actual process. In actual CMP, even with unchanged CMP parameter settings, different wafers yield different results. However, EPD cannot adjust based on these changes; engineers often have to reset the parameters based on the results, requiring continuous adjustments and consuming significant manpower and resources. Traditional methods have drawbacks such as insufficient sensitivity of fixed windows, poor noise resistance of small windows, manual setting of verification counts, and false triggering due to sudden spike interference. Summary of the Invention

[0004] In view of all or part of the shortcomings of the prior art described above, the purpose of this invention is to provide a CMP endpoint detection method, system and dynamic window scaling method that can shrink the width of the endpoint area window, overcome the defect of insufficient sensitivity of fixed window, trigger the endpoint signal in advance, and largely avoid over-throwing phenomenon.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution: This invention provides a dynamic window scaling method for CMP endpoint detection, which dynamically adjusts the window used for CMP endpoint detection to generate a dynamic window. The dynamic adjustment includes using a window width dynamic shrinkage function to dynamically adjust the width of the window in real time. The window width dynamic shrinkage function is as follows: (1): W(t) = W max ×e^ (-kw×|df / dt|) Equation (1); where W(t) is the real-time window width corresponding to a certain time point t, W max The maximum window width is given by |df / dt|, where f is the frictional force and k is the maximum window width. w k is the shrinkage sensitivity coefficient. w >0.

[0006] This invention also provides a CMP endpoint detection method, which determines the endpoint based on window logic, including: The friction force signal during the grinding process is acquired, and the friction force curve of the layer to be ground is obtained based on the friction force signal. The window used for CMP endpoint detection is dynamically adjusted to generate a dynamic window. The dynamic adjustment includes dynamically adjusting the width of the window in real time using a window width dynamic shrinkage function. The window width dynamic shrinkage function is as follows (1): W(t) = W max ×e^ (-kw×|df / dt|) Equation (1); Where W(t) is the real-time width of the window at a certain time point t, W max The maximum window width is given by |df / dt|, where f is the frictional force and k is the maximum window width. w k is the shrinkage sensitivity coefficient. w >0; Monitor the state of the friction force curve leaving the window. If the friction force curve leaves the right boundary of the window, it is determined that the grinding end point has been reached.

[0007] |df / dt| represents the absolute value of the slope at a given time point t on the friction curve. This invention drives exponential width contraction through slope, dynamically reducing the window width in the endpoint region (even down to 30% of the conventional width, i.e., 0.3W), overcoming the insufficient sensitivity of fixed windows. This invention can automatically adjust the window width dynamically without manual setting, triggering the CMP endpoint signal in advance and largely avoiding CMP over-throwing. Existing window width adjustments typically involve calculating and saving a fixed value, but the friction curves obtained during CMP are often different, making fixed window widths unsuitable. This invention provides a dynamic window that can adjust the window width in real time based on the slope, resulting in more precise adjustments and applicability to multiple processes, thus solving the sensitivity deficiency caused by fixed windows in traditional methods.

[0008] The dynamic window width shrinkage function, i.e., equation (1), uses |df / dt| calculated in real time via the five-point central difference method. When |df / dt| is less than or equal to 5, the window width is adjusted. |df / dt| being less than or equal to 5 is not strictly defined; rather, a slope absolute value less than 5 usually indicates a smaller slope, closer to the grinding endpoint. In practical applications, the slope range triggering window width shrinkage can be adjusted as needed.

[0009] When the friction force signal has noise fluctuations, the dynamic adjustment also includes using a window height noise adaptive function to dynamically adjust the height of the window in real time. The window height noise adaptive function is as follows (3): H(t) = H θ +k h ×σ(t), Equation (3); Where H(t) is the real-time height of the window at a certain time point t, H θ Let k be the base window height, σ(t) be the sliding standard deviation of the friction force signal within the calculation window, and k be the base window height. h k is the noise tolerance coefficient. h >0. By dynamically adjusting the window height, the window height automatically expands by, for example, 2-3 times under noise conditions, overcoming the defect of poor noise resistance of small windows.

[0010] Before obtaining the friction force curve, if the friction force signal exhibits noise fluctuations, anti-interference processing is also performed on the friction force signal. This anti-interference processing employs a real-time spike filter for preliminary noise reduction. After the initial noise processing, dynamic window height adjustment can be enabled simultaneously.

[0011] The mathematical model output of the real-time spike filter is defined as shown in equation (2), which eliminates impulse noise greater than 3σ based on four-point median filtering: Equation (2); Among them, f t f is the sampled value of the friction force signal at the current moment. t-1 f is the sampled value from the previous time step. t-2 The sampled values ​​are from the first two time steps, f t+1 The sampled value at the next time step, σ short Median is the median function used to calculate the standard deviation of the window. Real-time processing via four-point median filtering eliminates impulse noise greater than 3σ, preventing false triggering due to sudden spike interference.

[0012] It also includes a delayed triggering mechanism, which selects the state of N(t) consecutive windows as windowin to trigger the endpoint, N(t) = ceil(W max / W(t)), Equation (4); ceil is the rounding function. The window in state represents the state in which the friction curve leaves the right boundary of the window. After adjusting the window width of the endpoint area, the width is smaller than the conventional window size, and it is even more necessary to avoid misjudgment in CMP endpoint detection. Set the number of times the friction curve passes through the same state of the judgment window. For example, when the friction curve passes through the judgment window state three times in a row, it is determined that this is the polishing endpoint and the polishing stop signal is triggered. If the window size and the number of times the friction curve passes through the window state are manually set, the improvement effect on CMP endpoint detection and defects is not significant. This invention uses a delayed triggering mechanism. The smaller the window width of the endpoint area, the more consecutive window in states are required. By automatically adjusting the number of window in states through a function, the probability of misjudgment in CMP endpoint detection can be greatly reduced. By intelligently inversely adjusting the number of consecutive windows N(t), the number of verifications for small windows can be automatically increased without the need for manual setting of the number of verifications.

[0013] This invention also provides an endpoint detection system for a CMP machine, which determines the endpoint based on window logic, including: The curve acquisition module is used to acquire the friction force signal during the grinding process and obtain the friction force curve of the layer to be ground based on the friction force signal. The window width dynamic adjustment module is used to dynamically adjust the window width according to the slope corresponding to the time point selected on the endpoint region of the friction curve. The larger the absolute value of the slope, the smaller the window width. The window height dynamic adjustment module is used to dynamically adjust the height of the window based on the sliding standard deviation of the friction force signal within the calculation window. The larger the sliding standard deviation, the larger the window height. An anti-interference module is used to perform preliminary noise processing on the friction force signal in advance; The endpoint determination module is used to determine the grinding endpoint based on the state of the friction curve leaving the window.

[0014] The present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the CMP endpoint detection method described above.

[0015] The present invention also provides a storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned CMP endpoint detection method. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the specific embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 The window logic determines the state of the friction curve passing through the window when the CMP endpoint is reached; Figure 2 This is a flowchart illustrating a CMP endpoint detection method provided by the present invention; Figure 3 This is an algorithm logic diagram of a CMP endpoint detection method provided in Example 2; Figure 4 This is a friction force curve exemplified in Example 2; Figure 5 This is a situation where the existing window misjudges when there is noise fluctuation. Figure 6 This is a schematic diagram of the height and width changes of the dynamic window in Example 2 during noise fluctuations and endpoint triggering. Detailed Implementation

[0018] The technical solutions in specific embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] It should be noted that, in order to describe the technical solution more specifically, the steps described in the following embodiments do not strictly correspond one-to-one with the steps described in the invention content section. The following embodiments take chemical mechanical polishing (CMP) of shallow trench isolation (STI) as an example, but it is not intended to be limiting and can be used for polishing any semiconductor structure.

[0020] Example 1 A dynamic window scaling method for CMP endpoint detection involves dynamically adjusting the window used for CMP endpoint detection to generate a dynamic window. The dynamic adjustment includes dynamically adjusting the width of the window in real time using a window width dynamic shrinkage function, which is given by equation (1): W(t) = W max ×e^ (-kw×|df / dt|) Equation (1); where W(t) is the real-time window width corresponding to a certain time point t, W max The maximum window width is given by |df / dt|, where f is the frictional force and k is the maximum window width. wk is the shrinkage sensitivity coefficient. w >0.

[0021] Example 2 A CMP endpoint detection method, based on window logic to determine the endpoint, see [link to relevant documentation]. Figures 1 to 3 ,include: The friction force signal during the grinding process is acquired. The computer inputs the friction force signal in real time. When the friction force signal has noise fluctuations, anti-interference processing is performed on the friction force signal. The anti-interference processing uses a real-time spike filter for preliminary noise processing. The mathematical model of the real-time spike filter is introduced for anti-interference processing. The output of the mathematical model of the real-time spike filter is defined as shown in the following equation (2). Based on the four-point median filter, impulse noise greater than 3σ is eliminated: Equation (2); Among them, f t f is the sampled value of the friction force signal at the current moment. t-1 f is the sampled value from the previous time step. t-2 The sampled values ​​are from the first two time steps, f t+1 The sampled value at the next time step, σ short The standard deviation of the calculation window is 0.1 seconds, and Median is the median function. In this embodiment, the minimum unit of time data for the machine's endpoint is 0.1 seconds.

[0022] The Median function calculates the median by averaging the two middle values ​​after sorting the four points, i.e., median(a,b,c,d)=(x(2)+x(3)) / 2, where x(1)≤x(2)≤x(3)≤x(4). If the absolute value of the difference between the current sample value and the previous sample value is greater than three times the standard deviation σ of the 0.1-second calculation window... short At this point, the median is used to calculate the average value, reducing the impact of noise fluctuations on endpoint detection. This model has two key characteristics: first, filtering is triggered when the instantaneous change in the friction force signal exceeds 3σ, reducing endpoint misjudgment caused by signal fluctuations; second, it has a dynamic response, with normal data remaining undistorted and outliers replaced by the median. Therefore, the friction force signal is filtered to eliminate significant noise signals. Real-time processing using four-point median filtering eliminates impulse noise greater than 3σ, preventing false triggering due to sudden spike interference.

[0023] After anti-interference processing, the friction force curve of the layer to be polished is obtained based on the friction force signal. This does not refer to the complete friction force curve, which is obtained only after polishing is complete. The shape and function of the friction force curve are given here for ease of description; in reality, the friction force curve is obtained in real time based on the polishing process and the signal acquisition sequence. See [link to documentation]. Figure 4 The obtained friction curve takes a simple piecewise function as an example. The horizontal axis is time in seconds, and the vertical axis is the change in friction force. The function relationship within 0-100 s is: f(t)=-0.05t^2+5t. There are noise fluctuations between 89 and 94 s. After that, the friction force tends to stabilize and the change in friction force tends to be 0.

[0024] Set the EPD endpoint detection window, for example, with a width of 10N and a height of 10N. At 50-60 seconds, trigger the window out state (the friction curve enters from the left boundary of the window and exits from the bottom boundary), at which point the endpoint detection system starts. Between 89-94 seconds, due to significant noise fluctuations, typical window detection systems are prone to misjudgment or failure due to multiple window state changes (see [link to documentation]). Figure 5 The dynamic window algorithm engine is activated, which incorporates two aspects: slope monitoring and noise assessment. On the one hand, the width of the window changes with the slope of the friction curve; on the other hand, its height changes with the noise level of the friction curve.

[0025] The window used for CMP endpoint detection is dynamically adjusted to generate a dynamic window. The dynamic adjustment includes real-time dynamic adjustment of the window width using a window width dynamic shrinkage function and real-time dynamic adjustment of the window height using a window height noise adaptive function. In other embodiments, only one of the width dynamic adjustment or height dynamic adjustment may be performed according to the actual situation. This invention is not strictly limited.

[0026] (a) Noise assessment: When the friction force signal has noise fluctuations (there are noise fluctuations in this embodiment), the height of the window is dynamically adjusted in real time using a window height noise adaptive function, which is as follows (3): H(t) = H θ +k h ×σ(t), Equation (3); Where H(t) is the real-time height of the window at a certain time point t.

[0027] H θ The base window height is set to 10N (signal amplitude, for example only) by default in this embodiment.

[0028] σ(t) represents the sliding standard deviation of the friction force signal within the calculation window, which is set to 1 s by default (for example only). Specifically, a fixed-length window is set on the data sequence, the standard deviation of the data within this window is calculated, and then the window is allowed to slide one data point at a time, repeating the calculation to obtain a series of local standard deviations that vary over time.

[0029] k h k is the noise tolerance coefficient. h >0, in this embodiment the default value is 4.0 (for example only).

[0030] The greater the noise, the larger σ(t), the higher the window, and the more difficult it is to trigger boundary penetration. By dynamically adjusting the height driven by noise, the window height automatically expands by, for example, 2-3 times under noise conditions, overcoming the poor noise immunity of small windows. After dynamic height adjustment of the detection window, when the noise fluctuation of the friction force curve (signal) is large, the window becomes higher. At this time, the friction force curve enters from the left boundary of the window and exits from the lower boundary. After height adjustment, it passes through the noise signal normally, without misjudgment or failure (see...). Figure 6 In this embodiment, the window height adapts to noise fluctuations detected by the window detection system. The noise fluctuations are detected by the signal collector, meaning that the fluctuations before and after each data point collection are too large. At this time, the aforementioned real-time spike filter is triggered. After the filter processes the data, a friction curve is generated. However, the filter only initially filters the noise. Therefore, the height adjuster (window height noise adaptive function) is activated at the same time to enhance the anti-interference capability of the endpoint detection.

[0031] The standard deviation of σ(t) calculated at this time is 5.43N, and the calculation method is as follows: f(t) = -0.05t^2 + 5t; f(89)=48.95N, f(90)=45N, f(91)=50N, f(92)=40N, f(93)=55N, f(94)=40N; According to the standard deviation calculation formula, σ(t) = 5.43N.

[0032] H(t) = 10N + 5.43N × 4 = 31.72N. At this point, the window height is three times the base window height, and the window will still detect the endpoint signal normally.

[0033] (ii) Slope monitoring: When the CMP grinding endpoint is approaching at approximately 97 seconds, the friction stabilizes and the slope (absolute value) gradually decreases. At this point, the width shrinker is triggered, and the window width is dynamically adjusted in real time using a window width dynamic shrinking function. The width adjuster is mainly used to reduce the trigger time of endpoint detection and reduce over-polishing.

[0034] The dynamic shrinkage function of the window width is as follows (1): W(t) = W max ×e^ (-kw×|df / dt|) Equation (1); Where W(t) is the real-time width of the window at a certain time point t.

[0035] W max This is the maximum window width, which defaults to 10 seconds (for example only).

[0036] In |df / dt|, f represents the frictional force, and |df / dt| is the absolute value of the slope of the frictional force curve at a specific time point t. When the polishing endpoint is reached, the frictional force curve gradually flattens out, approaching 0. When the slope is between -5 and +5, i.e., when |df / dt| is less than or equal to 5, width contraction is triggered, adjusting the width of the window. This adjustment can be performed multiple times in real-time until the window width reaches its minimum.

[0037] In this embodiment, |df / dt| is calculated in real time using the five-point central difference method. The slope of the endpoint region in this embodiment is -4.62, and the calculation method is as follows: f(t) = -0.05t^2 + 5t; f(96)=19.2N, f(96.1)=18.7N, f(96.2)=18.3N, f(96.3)=17.8N, f(96.4)=17.4N; |df / dt|=(- f(96.4)+8 f(96.3)-8 f(96.1)+ f(96)) / (12×0.1)=-4.62.

[0038] k w k is the shrinkage sensitivity coefficient. w >0, in this embodiment the default value is 0.05 / N (for example only). This coefficient is only a preliminary setting value, and this value will change depending on the specific situation.

[0039] At this point, W(t) = 10s × e^(-t / t) (-0.05×4.62) =7.9s, at which point the window width is 7.9s. Window detection has a lag; if the window width were 10s, the signal collector would need to collect 10s of signal before generating the window. Compared to the conventional 10s window width, this embodiment shrinks the width of the detection window before the endpoint, triggering the endpoint signal 2.1 seconds earlier than a typical endpoint detection window (see...). Figure 6 In this model, as the friction slope |df / dt| increases, the window exponent decreases (in fact, in some specific cases it can be as small as 0.3W). maxThis improves the over-throwing phenomenon of CMP.

[0040] The system monitors the state of the friction curve as it leaves the window. "Window out+" (Out+) indicates it has touched the upper boundary, "window out-" (Out-) indicates it has touched the lower boundary, and "window in" (IN) indicates it has left from the right boundary. If the friction curve leaves from the right boundary of the window, the grinding endpoint is determined to have been reached.

[0041] After the friction curve passes through the dynamic window, the computer system determines its window state. When its state is "window in" (the friction curve enters from the left boundary of the window and leaves from the right boundary), a hysteresis verification mechanism is triggered. The hysteresis verification mechanism is as follows: the hysteresis triggering mechanism selects the state of N(t) consecutive windows as "window in" as the trigger endpoint, where N(t) = ceil(W max / W(t)), Equation (4); ceil is the floor function; the window in state represents the state where the friction curve leaves the right boundary of the window. If the window state is window in, the counter = counter + 1, otherwise, the counter = 0. If the counter ≥ N(t), the endpoint signal is triggered. The smaller the dynamic window width, the more consecutive counts are required to improve the accuracy of endpoint detection.

[0042] In this embodiment, the verification is triggered based on the window width at the last moment. The triggering condition is that N(t) consecutive dynamic windows are in the IN state, where N(t) = ceil(10 / 7.9) = 2. This means that two IN states are required, and data is collected once every 0.1 seconds. After collecting the second data, the confirmation of the two IN states is completed. Therefore, compared to the case of a 10-second window width without verification, this embodiment can trigger the endpoint signal 2 seconds earlier. By intelligently inversely adjusting the number of consecutive windows N(t), the verification count for smaller windows is automatically increased without the need for manual setting of the verification count.

[0043] Example 3 A system for detecting the endpoint of a CMP machine is provided, which determines the endpoint based on window logic, including: The curve acquisition module is used to acquire the friction force signal during the grinding process and obtain the friction force curve of the layer to be ground based on the friction force signal. The window width dynamic adjustment module is used to dynamically adjust the window width according to the slope corresponding to the time point selected on the endpoint region of the friction curve. The larger the absolute value of the slope, the smaller the window width. In this embodiment, the window width dynamic adjustment module may, but is not limited to, use W(t)=W mentioned in Embodiment 2. max ×e^ (-kw×|df / dt|)The function dynamically adjusts the window width (see Example 2 for details on how to adjust it, which will not be repeated here).

[0044] The window height dynamic adjustment module is used to dynamically adjust the height of the window based on the sliding standard deviation of the friction force signal within the calculation window. The larger the sliding standard deviation, the larger the window height. In this embodiment, the window height dynamic adjustment module may, but is not limited to, use H(t)=H mentioned in Embodiment 2. θ +k h The ×σ(t) function is dynamically adjusted (see Example 2 for details on how to adjust it, which will not be repeated here).

[0045] An anti-interference module is used to perform preliminary noise processing on the friction force signal. In this embodiment, the anti-interference module uses a real-time spike filter based on four-point median filtering to eliminate impulse noise greater than 3σ (see Embodiment 2 for details on how to remove noise, which will not be repeated here). The endpoint determination module is used to determine the grinding endpoint based on the state of the friction curve leaving the window. In this embodiment, the state of the friction curve leaving the window is monitored: window out+ indicates touching the upper boundary, window out- indicates touching the lower boundary, and window in indicates leaving from the right boundary. If the friction curve leaves from the right boundary of the window, the grinding endpoint is determined to have been reached. Alternatively, a hysteresis verification mechanism can be used: N(t) = ceil(W max The / W(t) function is used for verification to improve the accuracy of endpoint detection (see Example 2 for details on how to verify, which will not be repeated here).

[0046] Example 4 A computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a CMP endpoint detection method as described in Embodiment 2.

[0047] Example 5 A storage medium is provided on which a computer program is stored, which, when executed by a processor, implements a CMP endpoint detection method as described in Embodiment 2.

[0048] The algorithm logic based on friction endpoint detection is based on window logic, which is an existing algorithm logic in this field (not described in more detail here). This invention improves the algorithm logic of window logic by detecting the slope of the friction curve, thereby improving the process conditions of CMP.

[0049] The above description of the embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. It should be noted that those skilled in the art can make several improvements and modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the scope of protection of the claims of the present invention.

Claims

1. A dynamic window scaling method for CMP endpoint detection, characterized in that, The window used for CMP endpoint detection is dynamically adjusted to generate a dynamic window. The dynamic adjustment includes dynamically adjusting the width of the window in real time using a window width dynamic shrinkage function, which is as follows: W(t) = W max ×e^ (-kw×|df / dt|) Equation (1); where W(t) is the real-time window width corresponding to a certain time point t, W max The maximum window width is given by |df / dt|, where f is the frictional force and k is the maximum window width. w k is the shrinkage sensitivity coefficient. w >0.

2. A CMP endpoint detection method, characterized in that, Determining the endpoint based on window logic includes: The friction force signal during the grinding process is acquired, and the friction force curve of the layer to be ground is obtained based on the friction force signal. The window used for CMP endpoint detection is dynamically adjusted to generate a dynamic window. The dynamic adjustment includes dynamically adjusting the width of the window in real time using a window width dynamic shrinkage function. The window width dynamic shrinkage function is as follows (1): W(t) = W max ×e^ (-kw×|df / dt|) Equation (1); Where W(t) is the real-time width of the window at a certain time point t, W max The maximum window width is given by |df / dt|, where f is the frictional force and k is the maximum window width. w k is the shrinkage sensitivity coefficient. w >0; Monitor the state of the friction force curve leaving the window. If the friction force curve leaves the right boundary of the window, it is determined that the grinding end point has been reached.

3. The CMP endpoint detection method according to claim 2, characterized in that, In the dynamic shrinkage function of the window width, i.e., formula (1), the |df / dt| is calculated in real time by the five-point central difference method; when the |df / dt| is less than or equal to 5, the width of the window is adjusted.

4. The CMP endpoint detection method according to claim 2, characterized in that, When the friction force signal has noise fluctuations, the dynamic adjustment also includes using a window height noise adaptive function to dynamically adjust the height of the window in real time. The window height noise adaptive function is as follows (3): H(t) = H θ +k h ×σ(t), Equation (3); Where H(t) is the real-time height of the window at a certain time point t, H θ Let k be the base window height, σ(t) be the sliding standard deviation of the friction force signal within the calculation window, and k be the base window height. h k is the noise tolerance coefficient. h >0.

5. A CMP endpoint detection method according to claim 2, characterized in that, Before obtaining the friction force curve, if there is noise fluctuation in the friction force signal, the friction force signal is further subjected to anti-interference processing, which uses a real-time spike filter for preliminary noise processing.

6. The CMP endpoint detection method according to claim 5, characterized in that, The mathematical model output of the real-time spike filter is defined as shown in equation (2), which eliminates impulse noise greater than 3σ based on four-point median filtering: Equation (2); Among them, f t f is the sampled value of the friction force signal at the current moment. t-1 f is the sampled value from the previous time step. t-2 The sampled values ​​are from the first two time steps, f t+1 The sampled value at the next time step, σ short Median is the median function used to calculate the standard deviation of a window.

7. A CMP endpoint detection method according to claim 2, characterized in that, It also includes a delayed triggering mechanism, which selects the state of N(t) consecutive windows as window in to trigger the endpoint, N(t) = ceil(W max / W(t)), Equation (4); ceil is the up-rounding function; window in state represents the state in which the friction force curve leaves the right boundary of the window.

8. An endpoint detection system for a CMP machine, characterized in that, Determining the endpoint based on window logic includes: The curve acquisition module is used to acquire the friction force signal during the grinding process and obtain the friction force curve of the layer to be ground based on the friction force signal. The window width dynamic adjustment module is used to dynamically adjust the window width according to the slope corresponding to the time point selected on the endpoint region of the friction curve. The larger the absolute value of the slope, the smaller the window width. The window height dynamic adjustment module is used to dynamically adjust the height of the window based on the sliding standard deviation of the friction force signal within the calculation window. The larger the sliding standard deviation, the larger the window height. An anti-interference module is used to perform preliminary noise processing on the friction force signal in advance; The endpoint determination module is used to determine the grinding endpoint based on the state of the friction curve leaving the window.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements a CMP endpoint detection method according to any one of claims 2-7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a CMP endpoint detection method as described in any one of claims 2-7.

Images