A metal etch endpoint detection system

Abstract

The application discloses a metal etching endpoint detection system and particularly relates to the field of metal etching detection, comprising a spectral image acquisition module, a residual image generation module, a transient observation window management module, a transient identification module and a cross-checking judgment module. The metal etching endpoint detection system acquires spectral images through double physical sources, synchronously collects gas phase chemiluminescence and transmission light signals, avoids missing detection caused by weak single source signals, tracks long time constant slow drift and freezes baseline update in the transient observation window, maintains the integrity of the transient amplitude, prevents baseline absorption, identifies the starting moment at the sub-frame level, combines double-channel rate, amplitude and time consistency cross-checking, effectively suppresses false transients and reduces the false alarm and missing detection rates.

Description

Technical Field

[0001] This invention relates to the field of metal etching detection technology, and more specifically, to a metal etching endpoint detection system. Background Technology

[0002] With the development of visual inspection technology in micro-nano manufacturing, accurate identification of the endpoint of the etching process is particularly important. By acquiring, processing, and analyzing the spectral images and transmitted light intensity time-series signals within the process cavity, it is possible to determine whether the material layer to be processed has been completely removed, thus avoiding over-etching or under-etching. This is a key step in ensuring the quality of pattern transfer and manufacturing consistency.

[0003] Currently, the mainstream solution in the industry is single-channel endpoint detection based on emission spectrum images, which involves collecting characteristic emission spectral lines of plasma and monitoring the change of light intensity at a specific wavelength over time, and determining the endpoint based on a fixed threshold or inflection point. Meanwhile, laser interference and transmitted light intensity optical detection are also used in scenarios such as transparent media layers. The remaining thickness is inverted by counting interference fringes or changing the transmission intensity. This method does not require point-by-point alignment, has a lower cost, and can solve the basic requirements for endpoint determination under conventional graphic coverage conditions.

[0004] However, it still has some shortcomings in practical use. For example, it only relies on the spectral image data of a single physical source. The effective signal is weak and the signal-to-noise ratio is insufficient under low duty cycle and small area metal pattern. Furthermore, the signal from a single physical source cannot distinguish between metal endpoint transients and pseudo transients caused by plasma fluctuations, radio frequency disturbances, etc. The recognition algorithm based on a single source signal is difficult to balance the false alarm rate and the false detection rate. Meanwhile, the dynamic baseline tracking algorithm it employs moves with the endpoint transient itself, compressing the residual amplitude, causing weak endpoint features to be absorbed by the baseline, resulting in detection lag or missed detection. Summary of the Invention

[0005] To overcome the aforementioned deficiencies of the prior art, the present invention provides a metal etching endpoint detection system, which solves the problems mentioned in the background art through the following solutions.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A metal etching endpoint detection system, comprising: Spectral image acquisition module: used to acquire first spectral image data and second spectral image data of the detection area; the first spectral image data is the image data of the intensity of the preset characteristic emission spectral line in the gas phase chemiluminescence above the detection area changing with time, and the second spectral image data is the image data of the intensity of the probe light after passing through the detection area changing with time. The residual image generation module is used to perform slow drift component tracking on the first spectral image data and the second spectral image data respectively, and generate a first residual image sequence and a second residual image sequence; the slow drift component tracking tracks the slowly changing components in the spectral image data with a first time constant; Transient observation window management module: When any residual image sequence meets the preset candidate transient conditions, it starts a transient observation window that lasts for a preset duration, and freezes the baseline update of the slow drift component tracking corresponding to the first residual image sequence and the second residual image sequence within the transient observation window; Transient recognition module: used to identify the first transient start time based on the first residual image sequence and the second transient start time based on the second residual image sequence within the transient observation window; Cross-validation determination module: When the first residual image sequence satisfies the rate criterion and the amplitude criterion, the second residual image sequence satisfies the rate criterion, and the time difference between the start time of the first transient and the start time of the second transient falls within a preset time consistency window, it determines that a real transient event has occurred in the detection area and outputs a transient event signal.

[0007] Preferably, in the spectral image acquisition module, the wavelength of the probe light is located within the substrate transparent wavelength window of the detection area, and the substrate transparent wavelength window is 1100nm to 1700nm; In the second spectral image data, the pixel gray level at position (x,y) of a single frame of the second spectral image satisfies an exponential extinction relationship with the thickness d(x,y,t) of the metal thin film to be detected, specifically expressed as follows: ,in, Let be the pixel gray level of the second spectral image at time t; Let be the photoelectric response coefficient of the second photoelectric sensor array at position (x, y), which includes exposure time and optical collection efficiency; The intensity of the probe light incident on the upper surface of the detection area; The extinction coefficient of the thin film under test at the probe light wavelength λ2 = 1310 nm; Let t be the thickness of the film to be tested at position (x, y) at time t; The transmittance of the second narrowband filter at the probe wavelength; These are dark current noise and readout noise.

[0008] Preferably, the spectral image acquisition module presets the full width at half maximum (FWHM) of the first narrowband filter corresponding to the feature emission spectral line to be 1.5 nm; The wavelength interval between the preset characteristic emission spectrum and the radiation spectrum of other gaseous components in gas-phase chemiluminescence is not less than 3 nm.

[0009] Preferably, in the residual image generation module, the first time constant The relationship between the duration of the transient event to be detected and the following conditions is met: in, It is the shortest duration of the transient signal in the first residual image sequence.

[0010] Preferably, in the transient observation window management module, the candidate transient conditions are determined based on the average pixel grayscale value of a preset spatial region in the residual image sequence: When the absolute value of the average pixel grayscale exceeds a preset candidate amplitude threshold, and the exceeding state continues to reach a preset duration threshold, the residual image sequence is determined to satisfy the candidate transient condition. The preset candidate amplitude threshold is three times the noise standard deviation of the residual image sequence in a stable state.

[0011] Preferably, the transient recognition module searches within the transient observation window for the frame index in the first residual sequence where the absolute amplitude first exceeds a preset starting amplitude threshold, and uses this as the frame-level position at the start of the first transient. The frame-level position search formula for the first channel is: The frame index where the absolute amplitude first exceeds a preset starting amplitude threshold in the second residual sequence is searched and used as the frame-level position of the second transient start time. The frame-level position search formula for the second channel is as follows: Where, 𝑘1 and 𝑘2 are the frame indices corresponding to the start times of the first and second channels, respectively; Win is the frame sequence number within the observation window; This refers to the sampling time corresponding to the k-th frame; Linear interpolation or quadratic curve fitting is performed on the one-dimensional residual sequence within the frame-level location neighborhood to generate a first transient start time and a second transient start time that are more than 0.1 times the frame period.

[0012] Preferably, in the cross-validation determination module, the rate criterion is determined by the first residual image sequence, which is the first residual sequence in which the average pixel value of a preset spatial region changes over time. The second residual sequence is a second residual sequence in which the average pixel value of a preset spatial region changes over time. ; Perform adjacent frame differencing on the first residual sequence and the second residual sequence respectively: ,in, Let c be the average pixel grayscale value of the residual image of the c-th channel within the region of interest, where c=1 represents the first channel and c=2 represents the second channel. Δt represents the sampling time corresponding to the nth frame; Δt is the frame period. The amplitude criterion applies only to the first residual image sequence, and the judgment object is the maximum absolute peak amplitude of the first residual sequence within the transient observation window: Where, 𝐴1 is the maximum absolute amplitude of the first residual image sequence within the observation window; The moment the observation window is activated; Set a preset duration for the observation window.

[0013] Preferably, in the cross-validation determination module, the time consistency window value is 5ms; The preset duration of the transient observation window is three times the typical duration of the transient event to be detected; Within the transient observation window, the cross-validation module calculates the time difference frame by frame. When the first residual image sequence satisfies the rate criterion and the amplitude criterion, the second residual image sequence satisfies the rate criterion, and the time difference falls within a 5ms window, the transient event signal is immediately output without waiting for the transient observation window to end.

[0014] The technical effects and advantages of this invention are as follows: 1. This invention enables independent observation of the same metal etching endpoint event by dual physical sources through a spectral image acquisition module, effectively compensating for the weakness of the effective signal of a single physical source under low duty cycle and small area metal pattern. Even if the signal of the first channel is weak, the second channel can still independently trigger the candidate transient condition, thereby avoiding missed detection due to insufficient amplitude of the single source signal. 2. The present invention effectively suppresses the interference of background stray light and adjacent spectral components from the optical front end through the spectral image acquisition module, improves the signal-to-noise ratio of the first spectral image and the second spectral image, and provides a stable noise benchmark and threshold setting basis for candidate transient monitoring, start time identification and cross-validation judgment. 3. The present invention significantly reduces the absorption of transient signals by the dynamic baseline itself through the residual image generation module, effectively alleviating the problem of dynamic baseline tracking algorithms following the transient movement of the endpoint and compressing the residual amplitude in the prior art, thus preserving sufficient signal strength for subsequent transient identification; 4. The present invention uses a transient observation window management module to prevent the baseline from moving with the transient signal during the transient period, thereby fundamentally avoiding detection lag or missed detection caused by the absorption of weak endpoint features by the baseline, and ensuring the integrity of the true transient amplitude of the residual image sequence within the transient observation window. 5. This invention enables the precise quantification of the start time difference between two independent physical source signals through a transient identification module, providing input for the time consistency criterion of the cross-validation judgment module. This effectively distinguishes between metal endpoint transients and pseudo-transients caused by plasma fluctuations, radio frequency disturbances, etc., and avoids the inability to reliably determine the synchronization of the two signals due to insufficient time resolution. 6. The present invention effectively avoids false alarms and missed detections that may be caused by a single physical source signal or a single criterion through the cross-validation judgment module, which significantly improves the accuracy and reliability of endpoint detection. 7. This invention provides observational data for two independent physical mechanisms—gas-phase chemiluminescence and probe light transmission—through a spectral image acquisition module. Either channel can independently trigger a transient observation window, ensuring event capture capability under low duty cycle and small area patterns. The cross-validation module, based on this, uses the temporal consistency of the transient start times of the two signals, the amplitude and rate of the first channel, and the rate of the second channel for joint determination. A transient event signal is only output when the two independent physical mechanism signals are synchronously aligned in time and their change characteristics meet preset conditions. This collectively solves the core problem that a single physical source signal cannot distinguish between metal endpoint transients and pseudo-transients such as plasma fluctuations and radio frequency disturbances, achieving a balance between low false alarm rate and low false alarm rate. 8. This invention uses a long-time constant first-order recursive filtering in the residual image generation module to smoothly track the slow-drift component in a stable state, keeping the residual offset caused by baseline lag within an acceptable range; the transient observation window management module immediately freezes the baseline update after the candidate transient condition is triggered, further improving the residual amplitude retention rate during the transient period; it avoids both the baseline absorbing transient signals due to an excessively small time constant and the baseline lag offset in a stable state due to an excessively large time constant, achieving an optimal balance between residual stability in normal state and residual integrity during transient periods. Attached Figure Description

[0015] Figure 1 This is a block diagram of a metal etching endpoint detection system provided according to an embodiment of this application.

[0016] Figure 2 This is a flowchart illustrating the generation process of spectral images and residual images according to embodiments of this application.

[0017] Figure 3 This is a flowchart for identifying the transient start time according to an embodiment of this application.

[0018] Figure 4 This is a flowchart of the cross-validation determination process provided in the embodiments of this application. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0020] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to and includes any or all possible combinations of one or more of the listed items.

[0021] Hereinafter, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first," "second," and "third" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0022] As attached Figure 1 The metal etching endpoint detection system shown includes a spectral image acquisition module, a residual image generation module, a transient observation window management module, a transient recognition module, and a cross-validation judgment module.

[0023] In this embodiment, the detection area is the field of view of the metal thin film to be detected above the substrate. The detection area covered by a single frame of spectral image is 512×512 pixels. The spatial resolution is determined by the imaging magnification of the receiving optical path module and the pixel pitch of the photoelectric sensor array. In this embodiment, the pixel pitch is 25μm, the imaging magnification is 0.5 times, and the corresponding single pixel spatial resolution is 50μm. The sampling frame rate of both spectral image data is 200fps, and the exposure time is 5ms.

[0024] The spectral image acquisition module uses the light radiation signal of the preset characteristic emission spectrum line in the gas phase chemiluminescence above the detection area and the transmitted light radiation signal after the probe light passes through the detection area to perform narrowband filtering, photoelectric conversion and spatial imaging through the first receiving optical path module and the second receiving optical path module, respectively, to construct the first spectral image data and the second spectral image data, and generate two time-aligned spectral image sequences.

[0025] In one specific embodiment, the detection light source module adopts a narrow-band characteristic wavelength cold light source with an output linewidth of less than 1 nm, a center wavelength stability of better than ±0.1 nm within 2 hours of continuous operation, and an output optical power stability of better than ±0.5%, so as to ensure the amplitude stability of the second spectral image sequence in the time domain.

[0026] The first receiving optical path module includes a first focusing lens, a first narrowband filter, and a first photoelectric sensor array.

[0027] The light radiation generated by gas-phase chemiluminescence above the detection area is collected by the first focusing lens. After being filtered by the first narrow-band filter, only the preset characteristic emission spectrum lines are allowed to pass through, and background stray light is suppressed. The light radiation passing through the filter is imaged onto the first photoelectric sensor array and continuously exposed with a frame period of 5ms to output the first spectral image sequence arranged in time.

[0028] The first photoelectric sensor array uses a planar imaging device, whose spectral response range covers the visible to near-ultraviolet band where the preset characteristic emission spectral lines are located.

[0029] In this embodiment, the preset characteristic emission line is selected as a chemiluminescence characteristic line at 405 nm. The center wavelength of the first narrowband filter is 405 nm, the full width at half maximum (FWHM) is 1.5 nm, the peak transmittance is 90%, and the suppression ratio of adjacent background spectra is not less than 40 dB. This preset characteristic emission line corresponds to the radiative transition of the gaseous chemical component of interest in the event to be detected. Its spectral intensity is higher than that of adjacent background spectra, and the wavelength interval between it and the radiative spectra of other gaseous components in the gaseous chemiluminescence is not less than 3 nm, thus meeting the preset spectral separation requirements. The specific chemical source of this characteristic emission line can be determined according to the specific etching reaction system. Those skilled in the art can select the corresponding gaseous chemiluminescence component as the detection target under the spectral parameters disclosed in this embodiment.

[0030] The pixel gray level at position (x, y) in a single frame of the first spectral image data can be represented by the following formula: in, Let be the pixel gray level of the first spectral image at time t; Let be the photoelectric response coefficient of the first photoelectric sensor array at position (x, y), which includes exposure time and optical collection efficiency; The spectral radiance of the gas-phase chemiluminescence at position (x, y) and time t above the detection area is given. The spectral transmittance of the first narrowband filter; The center wavelength of the first narrowband filter is 405nm. The first narrowband filter has a full width at half maximum (FWHM) of 1.5nm. These are dark current noise and readout noise.

[0031] The second receiving optical path module includes a second focusing lens, a second narrowband filter, and a second photoelectric sensor array.

[0032] The detection light source module includes a narrowband characteristic wavelength cold light source and a collimating lens group. The collimating lens group shapes the detection light into parallel monochromatic light and incident perpendicularly on the detection area. In this embodiment, the detection light wavelength is selected as 1310 nm, which is located within the transparent wavelength window of the silicon-based substrate, ranging from 1100 nm to 1700 nm. The transmittance of the silicon material at this wavelength is not less than 50%, and the metal thin film to be detected has a measurable extinction effect at 1310 nm.

[0033] The probe light passing through the detection area is collected by the second focusing lens and filtered by the second narrow-band filter, allowing only the 1310nm wavelength probe light to pass through and be imaged onto the second photoelectric sensor array. The second photoelectric sensor array uses a near-infrared band-responsive area array imaging device, continuously exposed at the same frame rate of 200fps as the first spectral image sequence, and outputs a second spectral image sequence arranged in time.

[0034] The pixel grayscale value at position (x, y) in a single frame of the second spectral image data can be represented by the following formula: ,in, Let be the pixel gray level of the second spectral image at time t; Let be the photoelectric response coefficient of the second photoelectric sensor array at position (x, y), which includes exposure time and optical collection efficiency; The intensity of the probe light incident on the upper surface of the detection area; The extinction coefficient of the thin film under test at the probe light wavelength λ2 = 1310 nm is typically 0.02 nm⁻¹ in this embodiment; Let t be the thickness of the thin film to be detected at position (x, y) at time t. In this embodiment, the typical range of the initial thickness of the thin film is 30 nm to 50 nm. The transmittance of the second narrowband filter at the probe wavelength; These are dark current noise and readout noise.

[0035] A hardware trigger pulse is generated by the same sampling clock generator, which simultaneously triggers the exposure of the first photoelectric sensor array and the second photoelectric sensor array, and records a unified timestamp in the acquired data.

[0036] In this embodiment, the sampling clock generator uses a temperature-controlled crystal oscillator with a frequency stability of ±5ppm, the rising edge jitter of the hardware trigger pulse does not exceed 0.1μs, the synchronization error of the exposure start time of the two photoelectric sensor arrays does not exceed 1μs, and the unified timestamp is obtained by counting the sampling clock with a timestamp resolution of 0.1μs.

[0037] In one specific embodiment, the metal thin film to be detected is arranged in a periodic pattern in the detection area, covering an area of ​​30% of the total detection area. The circumscribed rectangle of the thin film pattern to be detected is located in a single frame spectral image using a template matching method, and a rectangular window with a size of 400×400 pixels is selected as the region of interest, which covers the entire thin film pattern to be detected.

[0038] Template matching employs normalized cross-correlation matching. The template is taken from the region of interest (ROI) image acquired during system initialization in the full-film state, with a size of 400×400 pixels. The search range is the entire frame (512×512 pixels) with a step size of 1 pixel. When the maximum normalized cross-correlation coefficient is ≥0.85, that position is taken as the matching position and the ROI is updated. If the matching coefficient is lower than 0.85 for 5 consecutive frames, the ROI position of the previous frame remains unchanged, and a matching failure flag is output. If the actual initialization condition is not in the full-film state, the operator will re-acquire the full-film state template for calibration.

[0039] Dark current subtraction, pixel response non-uniformity correction, bad pixel removal, and inter-frame alignment are performed on the spectral image data within the region of interest.

[0040] The specific method for dark current subtraction is as follows: under no-light conditions, 100 dark frame images are continuously acquired from the two photoelectric sensor arrays respectively, the average dark current value at each pixel position is calculated, and the corresponding dark current value is subtracted from each pixel of each frame of spectral image.

[0041] The specific method for pixel response non-uniformity correction is as follows: using the gain parameters and bias parameters obtained from uniform light field illumination calibration, the pixel grayscale after dark current subtraction is linearly corrected pixel by pixel. The correction formula is: ,in, The corrected pixel grayscale; Original pixel grayscale; The average pixel grayscale of the dark frame; This is the gain correction factor; This is the bias correction coefficient.

[0042] In this embodiment, the intensity of the uniform light field used is adjusted to be on the same order of magnitude as the actual emission signal of the first channel. The calibration method for the gain correction coefficient G(x,y) and the bias correction coefficient O(x,y) is as follows: within the linear response range of the sensor, two uniform light field images with known different intensities are acquired respectively, and a linear fit is performed on each pixel position. With the light field intensity as the independent variable and the expected value of the corrected gray level as the dependent variable, G(x,y) and O(x,y) are obtained by fitting.

[0043] The specific method for defective pixel removal is as follows: use 3×3 neighborhood median filtering to detect photoelectric pixels with abnormal response. When the absolute difference between the gray value of a pixel and the median of its 3×3 neighborhood exceeds 5 times the absolute deviation MAD of the neighborhood median, the pixel is determined to be a defective pixel, and the gray value of the pixel is replaced by the neighborhood median.

[0044] The specific method of inter-frame alignment is as follows: perform small translation correction on the two image sequences, estimate the subpixel translation amount by using the normalized cross-correlation of adjacent frames, and perform translation correction on the second spectral image sequence to eliminate image shift caused by vibration.

[0045] The above preprocessing operations are performed before the residual image is generated to ensure pixel spatial consistency between the two image sequences when they enter dynamic baseline tracking.

[0046] The residual image generation module, based on the first and second spectral image sequences output by the spectral image acquisition module, performs slow drift component tracking pixel-by-pixel using first-order recursive filtering to construct a first baseline image sequence and a second baseline image sequence. The first residual image sequence is generated by pixel-by-pixel difference between the first spectral image sequence and the first baseline image sequence, and the second residual image sequence is generated by pixel-by-pixel difference between the second spectral image sequence and the second baseline sequence. The residual image sequences are then subjected to spatial low-pass filtering and output to the subsequent processing module.

[0047] Slow drift component tracking is performed frame by frame. The current frame's spectral image is weighted and superimposed with the previous frame's baseline image to generate the current frame's baseline image. The weighting coefficients are determined by the first time constant and the sampling frame rate.

[0048] For the first spectral image sequence, perform a first-order recursive filter pixel by pixel, with the iterative relationship as follows: ,in, The pixel value of the first baseline image of the current frame at position (x, y); The pixel value of the first baseline image in the previous frame at position (x, y); is the pixel value of the first spectral image of the current frame at position (x, y); is the smoothing coefficient of the first channel.

[0049] For the second spectral image sequence, the same recursive structure is used to generate the second baseline image. The iterative relationship is: , where 𝛼2 is the smoothing coefficient of the second channel.

[0050] The relationship between the smoothing coefficient and the time constant is defined by the following formula: ,in, The first time constant; The sampling frame rate. In this embodiment, the sampling frame rate... =200fps, first time constant If 1.0s is selected, then =0.005.

[0051] First time constant The relationship with the typical duration of the transient event to be detected satisfies: ,in, This is the shortest duration of the transient signal in the first residual image sequence. In this embodiment, the typical duration of the transient event to be detected is 100ms, and the shortest transient duration is taken as 80ms. =1.0s satisfies =12.5, which is greater than 10 times the constraint.

[0052] The first residual image sequence is generated by pixel-by-pixel difference between the first spectral image sequence and the first baseline image sequence, and the expression is: The second residual image sequence is generated by pixel-by-pixel difference between the second spectral image sequence and the second baseline image sequence, and its expression is: ,in, and denoted as , respectively, the pixel values ​​of the first residual image and the second residual image at position (𝑥,𝑦) at time t.

[0053] The pixel values ​​of the residual image represent the fast-changing components in the current frame spectral image after removing the slow-drift components.

[0054] In this embodiment, by performing the above pixel-by-pixel interpolation on the two spectral image sequences respectively, a residual image sequence with the same spatial size (400×400 pixels) as the input spectral image in the region of interest is generated. The frame rate of the residual image sequence is the same as that of the input spectral image sequence, which is 200fps.

[0055] Spatial low-pass filtering is performed after the residual image is generated. The spatial low-pass filtering uses a 3×3 pixel neighborhood mean filter, with a filter kernel size of 3×3 and a weight of 1 / 9 for both the center pixel and its neighbors. The pixel values ​​of the residual image after filtering the c-th channel are: ,in, denoted as the residual image pixel value after spatial low-pass filtering of the c-th channel, where c=1 represents the first channel, c=2 represents the second channel, and n is the neighborhood column index.

[0056] Before performing slow drift component tracking, regions of interest are selected from the first spectral image data and the second spectral image data, respectively, and slow drift component tracking and residual image sequence generation are performed only on the pixels within the regions of interest.

[0057] In this embodiment, the region of interest (ROI) is extracted by the spectral image acquisition module, with a size of 400×400 pixels, covering the thin film pattern to be detected in the detection area. The residual image generation module performs the aforementioned slow drift component tracking, pixel-by-pixel interpolation, and spatial low-pass filtering only on this 400×400 pixel region to reduce computational load and eliminate interference from background pixels outside the detection area. For pixels outside the ROI, baseline tracking and residual calculation are not performed; they are either discarded or kept as null values.

[0058] Table 1 presents comparative data on residual compression rate and transient amplitude retention rate under different first time constants. The data in Table 1 are calculated using a first-order recursive formula under a 16-frame constant amplitude transient condition.

[0059] Table 1 In Table 1, when At 0.1s, the smoothing coefficient is too large, and the baseline rapidly follows the transient signal during the transient period, with the cumulative tracking amplitude accounting for 56%, resulting in a transient amplitude retention rate of only 44% in the residual. The transient characteristics are significantly absorbed by the baseline. At 1.0s, the baseline cumulative tracking amplitude percentage decreased to 8%, and the residual transient amplitude retention rate increased to 92%; when At 5.0s, the residual transient amplitude retention rate was further improved to 98.4%, but the excessively large time constant caused the baseline to lag in slow drift compensation, resulting in non-zero offset of the residual image in a long-term stable state.

[0060] In this embodiment, the selected =1.0s, achieving an engineering trade-off between transient amplitude retention rate and slow drift compensation hysteresis, ensuring that transient characteristics are not absorbed while avoiding residual anomalies caused by baseline hysteresis.

[0061] The transient observation window management module monitors the average pixel grayscale value of a preset spatial region in each frame of residual image based on the first residual image sequence and the second residual image sequence output by the residual image generation module, and constructs a candidate transient trigger signal. When any residual image sequence meets the preset candidate transient condition, a transient observation window for a preset duration is started, and the baseline update of the slow drift component tracking corresponding to the first residual image sequence and the second residual image sequence is frozen within the transient observation window, generating a baseline freeze control signal and an observation window start indication.

[0062] To support the backtracking determination of candidate transient conditions and the rollback activation of the observation window, the first residual image sequence and the second residual image sequence are set to a length of not less than [amount missing] before entering the transient observation window management module. A circular buffer corresponding to the frame number. After the candidate transient condition is met, the frame is retrieved from the circular buffer. Data from that point onwards is used for observation window and baseline freezing; the circular buffer also retains... The first two frames of data are used for the interpolation neighborhood of the transient recognition module. The circular buffer simultaneously stores the channel identifier, timestamp, residual pixel value, and baseline pixel value used during residual generation for the residual image; once the candidate condition is met, the data is retrieved from the buffer. The baseline image at time 1 is used as the frozen baseline.

[0063] The monitoring object for candidate transient conditions is the average pixel grayscale value of a preset spatial region in the residual image sequence. The preset spatial region is all pixels of the region of interest. In this embodiment, the region of interest size is 400×400 pixels, and the total number of pixels is... =160,000.

[0064] For each frame of the residual image, the spatial average is calculated to obtain a one-dimensional residual mean sequence: ,in, Let be the average pixel grayscale value of the residual image within the region of interest at time t; where is the pixel value of the residual image after spatial low-pass filtering; ROI is the set of pixels in the region of interest. The above formula applies to both the first and second channels, respectively using . and Substitute into the calculation of the corresponding channel .

[0065] when Exceeding the preset candidate amplitude threshold And this exceeded state continues to reach a preset duration threshold. When the residual image sequence satisfies the candidate transient condition, i.e., CandTrig=1, if Make , Otherwise, CandTrig=0.

[0066] in, This serves as a candidate transient trigger flag. This refers to the sampling time corresponding to the k-th frame; The starting frame index for the first occurrence of the state. ; The threshold value for candidate amplitude; Duration threshold The corresponding frame number, in this embodiment =30ms =6.

[0067] Candidate amplitude threshold Take the noise standard deviation of the residual image sequence in a steady state. Three times the duration threshold Take the typical duration of the transient event to be detected 0.3 times.

[0068] In this embodiment, Take the first channel =1.2LSB was used as a uniform noise benchmark, which was calculated by continuously acquiring 1000 frames of steady-state residual images.

[0069] Right now =3.6 LSB.

[0070] When a certain frame When the amplitude exceeds 3.6 LSB for the first time, the trigger counter is set to 1; for each subsequent frame that exceeds 3.6 LSB, the counter is incremented by 1; if a frame with an amplitude not exceeding 3.6 LSB occurs before the counter reaches 6, the counter is reset to zero, and the system waits for the next exceedance event. The candidate transient condition is determined to be met when the counter reaches 6.

[0071] Candidate transient conditions are applied to the first residual image sequence and the second residual image sequence, respectively. When either residual image sequence satisfies the candidate transient condition, the transient observation window is activated.

[0072] Under conditions of low duty cycle and small area metal patterning, the effective signal of the first residual image sequence may be weak, and the transient amplitude may be lower than the candidate amplitude threshold of 3.6 LSB of the first channel. At this time, the first channel does not trigger the candidate condition. However, the probe light transmission signal in the second residual image sequence generates a step transient when the thin film is cleared, with an amplitude of up to 15.6 LSB, far exceeding the 3.6 LSB threshold, and continues to reach the preset frame number. Therefore, the second channel independently triggers the transient observation window.

[0073] When any residual image sequence meets the candidate transient condition, a transient observation window is activated for a preset duration. The preset duration of the transient observation window is... Take the typical duration of the transient event to be detected Three times that in this embodiment =300ms, corresponding to a 60-frame time interval at a frame rate of 200fps. The transient observation window is defined temporally as the detection moment when the candidate transient condition is first met. From, until The time interval.

[0074] The spatial scope of the transient observation window covers all pixels in the region of interest. When a candidate transient condition is triggered by a local spatial sub-region, the spatial scope of the transient observation window can also be the local sub-region and its neighborhood. In this embodiment, candidate transient monitoring is based on the average pixel grayscale value of the entire region of interest.

[0075] At the same time as the transient observation window is started, the transient observation window management module generates a baseline freeze control signal Frozen(t)=1 and outputs it to the residual image generation module. At the same time, it generates an observation window start indication WinStart(t)=1 and outputs it to the transient recognition module.

[0076] During the duration of the transient observation window, i.e. Within, the first baseline image corresponding to the first residual image sequence. Second baseline image corresponding to the second residual image sequence Each pixel value remains unchanged from the baseline value at the start of the transient observation window. The iterative calculation of slow drift component tracking is paused, but the spectral image sequence continues to be acquired and stored.

[0077] The frozen baseline image remains as follows: ,in, and These are the pixel values ​​of the first and second baseline images at position (x, y) at the start of the observation window, respectively.

[0078] During baseline freezing, residual calculation continues frame-by-frame, but the baseline is no longer updated. The residual image pixel values ​​are then: ,in, and The pixel values ​​of the first and second spectral images of the current frame.

[0079] Table 2 presents comparative data on the retention rate of residual amplitude during transients in the baseline frozen state and the unfrozen state. The data in Table 2 were obtained experimentally from standard etched samples at 200 fps.

[0080] Table 2 If no real transient event is determined to have occurred by the end of the transient observation window, the baselines corresponding to the first residual image sequence and the second residual image sequence are updated from the baseline value at the time of freezing and continue to perform first-order recursive filtering tracking; if a real transient event is determined to have occurred by the end of the transient observation window, the baseline is kept frozen until an external reset signal is received.

[0081] In this embodiment, without determining a real transient event, the baseline recovery update is performed on the baseline value since the freeze time. and The first-order recursive filtering continues, and the smoothing coefficient φ = 0.005 in the residual image generation module is used for the recursive filtering iteration relationship.

[0082] If a true transient event has been identified, Frozen(t) remains unchanged at 1 until an external reset signal arrives and Frozen(t) becomes 0. The baseline continues to be updated from the baseline value at the time of freezing, or the external controller decides whether to reinitialize the baseline.

[0083] In an alternative implementation, the candidate transient condition can be a pixel-level counting condition: when the number of pixels in the residual image whose amplitude exceeds the candidate threshold reaches a preset proportion, the candidate transient condition is determined to be met.

[0084] In another alternative implementation, baseline recovery after the observation window ends can employ a jump recovery rule: if no transient is detected during the observation window, the baseline value is directly updated to a smoothed estimate of the current frame's spectral image after the observation window ends, and tracking restarts; if a transient is detected, the baseline remains frozen until an external reset. The jump recovery rule can eliminate baseline shifts accumulated during the observation window more quickly, but it introduces a brief baseline jump at the recovery time.

[0085] All of the above alternatives can be replaced within the overall framework of this embodiment.

[0086] The transient recognition module, based on the first residual image sequence and the second residual image sequence within the transient observation window initiated by the transient observation window management module, constructs the first residual sequence and the second residual sequence over time for the pixel average value of a preset spatial region, respectively. It searches for the frame index that first exceeds the preset starting amplitude threshold as the frame-level position by means of a preset starting amplitude threshold, and performs linear interpolation or quadratic curve fitting on the one-dimensional residual sequence in the neighborhood of the frame-level position to generate the first transient start time and the second transient start time with sub-frame-level time precision.

[0087] Within the transient observation window, the average pixel value of a preset spatial region in the first residual image sequence over time constitutes the first residual sequence. The preset spatial area is the spatial range of the transient observation window, i.e., the region of interest is 400×400 pixels, and the total number of pixels is... =160,000. The formula for calculating the first residual sequence is: ,in, Let be the average pixel grayscale value of the first residual image within the region of interest at time t; is the pixel value at time t after spatial low-pass filtering in the first residual image sequence; ROI is the set of pixels in the region of interest.

[0088] Similarly, a second residual sequence is constructed from the second residual image sequence. The calculation formula is: ,in, is the pixel value at time t in the second residual image sequence after spatial low-pass filtering.

[0089] The preset initial amplitude threshold is set independently for the first residual image sequence and the second residual image sequence, respectively. In this embodiment, the noise standard deviation of the first residual image sequence in a steady state is... =1.2LSB, the first channel has a preset starting amplitude threshold. Take 3 =3.6 LSB; Noise standard deviation of the second residual image sequence in steady state. 0.8 LSB, preset starting amplitude threshold for the second channel A value of 5.0 LSB, which is 6.25 times higher than the noise level of the second channel, is used to adapt to the edge features of step transients in the second residual image sequence, so as to avoid the start time being shifted forward due to noise triggering the step front in advance.

[0090] Within the transient observation window, frame-level position searches are performed on the first residual sequence and the second residual sequence, respectively: Within the transient observation window, the frame index in the first residual sequence where the absolute amplitude first exceeds a preset starting amplitude threshold is searched and used as the frame-level position of the first transient start time; the frame index in the second residual sequence where the absolute amplitude first exceeds the preset starting amplitude threshold is searched and used as the frame-level position of the second transient start time. The frame-level position search formula for the first channel is: The frame-level position search formula for the second channel is: Where, 𝑘1 and 𝑘2 are the frame indices corresponding to the start times of the first and second channels, respectively; Win is the frame sequence number within the observation window; This refers to the sampling time corresponding to the k-th frame; =3.6 LSB, =5.0 LSB.

[0091] In one specific embodiment, let the starting time of the first satisfaction of the candidate transient condition be... The corresponding absolute time is 100ms, and the frame period Δt = 5ms. Therefore, frame 18 is 90ms, frame 19 is 95ms, frame 20 is 100ms, frame 21 is 105ms, and frame 22 is 110ms. The values ​​of the first residual sequence within the observation window for frame indices 18 to 22 are shown in Table 3. In the table, frame 20 corresponds to... Frames 18 and 19 are The first two frames, frames 21 and 22 are The last two frames.

[0092] Table 3 In Table 3, the amplitudes of the first residual sequence at frame indices 18 and 19 are 2.1 LSB and 2.8 LSB, respectively, neither exceeding the preset starting amplitude threshold of 3.6 LSB for the first channel. The amplitude at frame index 20 is 4.2 LSB, exceeding the 3.6 LSB threshold for the first time. Therefore, frame index 20 is determined as the frame-level position of the first transient start time, corresponding to a time of 100 ms. If subframe-level refinement is not performed, the first transient start time is directly taken as 100 ms at frame index 20, with a recognition time resolution of 5 ms frame period.

[0093] Within the frame-level location neighborhood, linear interpolation is performed on the first residual sequence to obtain the first transient start time with sub-frame-level temporal precision. Frame index 19 (amplitude 2.8 LSB) and frame index 20 (amplitude 4.2 LSB) are selected as interpolation endpoints, and linear interpolation is used to calculate the moment when the residual amplitude rises from below the threshold to above the threshold. The linear interpolation formula is: , among which, 𝑡 19 =95ms, =3.6 LSB, =2.8 LSB, =4.2 LSB, Δt = 5 ms; calculated, t1 = 97.857 ms. Since frame index 19 corresponds to 95 ms and frame index 20 corresponds to 100 ms, the residual curve between the above two frame points has a monotonically increasing linear change characteristic. The intersection time of 97.857 ms obtained by linear interpolation can be used as the subframe level estimate of the start time of the first transient.

[0094] Among the five sampling points in frame indices 18 to 22, the two endpoints used for interpolation are the closest frame points before and after the threshold. The interpolation error is no more than 0.5 times the frame period of 5ms, that is, no more than 2.5ms. In this embodiment, the residual change is approximately linear, and the interpolation accuracy is within 0.1 times the frame period, that is, the time resolution is better than 0.5ms.

[0095] Furthermore, the frame-level position search and sub-frame-level refinement process for the second channel is similar to that of the first channel, but the threshold and residual variation characteristics are different. The values ​​of the second residual sequence within the observation window for frame indices 18 to 22 are shown in Table 4. The table shows the correspondence between frame indices and time points. The frame assignments are the same as in Table 3.

[0096] Table 4 In Table 4, the amplitudes of the second residual sequence at frame indices 18 and 19 are 1.8 LSB and 2.5 LSB, respectively, neither of which exceeds the preset starting amplitude threshold of 5.0 LSB for the second channel; the amplitude at frame index 20 is 8.0 LSB, exceeding the 5.0 LSB threshold for the first time. Therefore, frame index 20 is determined as the frame-level position of the start time of the second transient, corresponding to a time of 100ms.

[0097] Subframe-level refinement was performed on the second residual sequence. Frame index 19 (amplitude 2.5 LSB) and frame index 20 (amplitude 8.0 LSB) were selected as interpolation endpoints, and linear interpolation was used to calculate the start time of the second transient. The linear interpolation formula is as follows: , among which, 𝑡 19 =95ms, Thr_start2=5.0LSB, 𝑟2,ROI(𝑡 19 = 2.5 LSB, 𝑟2, ROI(𝑡 20 =8.0 LSB, Δt = 5 ms. Calculated, t2 = 97.273 ms. The subframe-level identification result at the start of the second transient is 97.273 ms.

[0098] When the residual curve exhibits nonlinear changes within the frame-level neighborhood, a quadratic curve fitting method can be used to determine the intersection point of the fitted curve and a preset initial amplitude threshold. The quadratic curve fitting uses three sampling points—the frame-level location and its preceding and following frames—to perform a second-order polynomial fitting. The fitting function is: Where 𝑟(𝑡) is the one-dimensional residual amplitude after fitting; 𝑡 is time; 𝑎, 𝑏, and 𝑐 are fitting coefficients, determined by the mean residual of three adjacent frames and the corresponding time. After fitting, the coefficients satisfying the following conditions are calculated. The smallest real root is taken as the transient start time. Both the first and second transient start times are identified based on the first and second residual image sequences.

[0099] In an alternative implementation, the transient start time can be identified using the moment of maximum rate of change: the frame index with the largest absolute rate of change is found on the first-order difference curve of the residual sequence, and that point is used as the start time. Those skilled in the art can choose between the threshold search method described in this embodiment and the maximum rate of change method, depending on the actual detection scenario.

[0100] The cross-validation judgment module constructs a cross-validation judgment framework composed of rate criterion, amplitude criterion, and time consistency criterion based on the first residual image sequence and the second residual image sequence within the transient observation window, as well as the first transient start time and the second transient start time output by the transient recognition module, and generates a transient event signal used to determine whether a real transient event has occurred in the detection area.

[0101] Specifically, the cross-validation judgment module updates the cumulative maximum value required for each criterion frame by frame within the transient observation window, including the maximum absolute rate of the first channel, the maximum absolute amplitude of the first channel, and the maximum absolute rate of the second channel, and checks whether the time difference can be calculated in each frame.

[0102] When the first residual image sequence satisfies the rate criterion and the amplitude criterion, the second residual image sequence satisfies the rate criterion, and the time difference falls within the time consistency window, the transient event signal EventSignal=1 is immediately output without waiting for the transient observation window to end; if all conditions are not met when the transient observation window ends, EventSignal=0 is output at the end of the observation window, indicating that no real transient event has occurred in the detection area.

[0103] The rate criterion is used to determine whether the rate of change of transients in the residual image sequence reaches a preset rate threshold. The rate criterion is used to determine the first residual image sequence, which is the first residual sequence in which the average pixel value of a preset spatial region changes over time. The second residual sequence is a second residual sequence in which the average pixel value of a preset spatial region changes over time. .

[0104] Perform frame-to-frame difference analysis on the first and second residual sequences respectively to obtain the change per unit time. The calculation formula is as follows: ,in, Let c be the average pixel grayscale value of the residual image of the c-th channel within the region of interest, where c=1 represents the first channel and c=2 represents the second channel. Δt is the sampling time corresponding to the nth frame; Δt is the frame period, and in this embodiment, Δt = 5ms.

[0105] The rate criterion requires that the maximum absolute rate during the transient period exceeds a preset rate threshold: ; ,in, The first channel rate threshold, This is the rate threshold for the second channel.

[0106] In this embodiment, 1000 frames of residual images are continuously acquired under steady-state conditions. The 99th percentile of the absolute value of the difference between adjacent frames in the one-dimensional residual sequence after ROI averaging is calculated. The first channel is 0.10 LSB / ms, and the second channel is 0.15 LSB / ms. The rate threshold is taken as 3 times this percentile, i.e. =0.30 LSB / ms =0.45 LSB / ms.

[0107] In one specific embodiment, the one-dimensional residual sequence data of the first and second channels within the transient observation window at frame indices 18 to 22. The first channel's... The values ​​are 2.1 LSB, 2.8 LSB, 4.2 LSB, 6.5 LSB, and 9.8 LSB respectively; the second channel's... The values ​​are 1.8 LSB, 2.5 LSB, 8.0 LSB, 15.2 LSB, and 20.1 LSB, respectively.

[0108] The difference between adjacent frames is calculated as follows: the differences between adjacent frames in the first channel are 0.7 LSB, 1.4 LSB, 2.3 LSB, and 3.3 LSB, respectively, corresponding to changes per unit time of 0.14 LSB / ms, 0.28 LSB / ms, 0.46 LSB / ms, and 0.66 LSB / ms, respectively, and the maximum absolute rate. =0.66 LSB / ms, which is greater than the first channel rate threshold of 0.30 LSB / ms, therefore the first residual image sequence meets the rate criterion; the adjacent frame differences of the second channel are 0.7 LSB, 5.5 LSB, 7.2 LSB, and 4.9 LSB, respectively, corresponding to unit time changes of 0.14 LSB / ms, 1.10 LSB / ms, 1.44 LSB / ms, and 0.98 LSB / ms, respectively, with the maximum absolute rate being... =1.44 LSB / ms, which is greater than the second channel rate threshold of 0.45 LSB / ms. Therefore, the second residual image sequence satisfies the rate criterion.

[0109] The amplitude criterion is used to determine whether the amplitude of transients in the first residual image sequence reaches a preset amplitude threshold.

[0110] The amplitude criterion applies only to the first residual image sequence. The criterion is based on the maximum absolute peak amplitude of the first residual sequence within the transient observation window. The calculation formula is as follows: Where, 𝐴1 is the maximum absolute amplitude of the first residual image sequence within the observation window; The moment the observation window is activated; In this embodiment, a preset duration is set for the observation window. =300ms.

[0111] The amplitude criterion requires that the maximum absolute amplitude exceeds a preset amplitude threshold: .

[0112] In this embodiment, the noise standard deviation of the first residual image sequence in a steady state =1.2LSB, set =6.0 LSB, or 5 times the noise standard deviation. This is also determined based on the temporal standard deviation of the one-dimensional residual sequence after averaging 1000 frames of ROI under steady-state conditions, and is taken as 5 times. In the above embodiment, the maximum absolute amplitude of the one-dimensional residual sequence in the first channel within the observation window reaches 23.5 LSB, which is greater than... =6.0LSB, then the first residual image sequence satisfies the amplitude criterion.

[0113] The time consistency window is used to compare the time difference between the start time of the first transient and the start time of the second transient.

[0114] The time difference is calculated by taking the absolute value of the difference between the start time of the first transient and the start time of the second transient: Where, 𝑡1=97.857ms is the start time of the first transient; 𝑡2=97.273ms is the start time of the second transient; This represents the time difference between the start times of the two transients. In this embodiment, =0.584ms.

[0115] Preset time consistency window The value is set to 5.0ms, which is less than the preset duration of the transient observation window, 300ms.

[0116] The time consistency window is determined based on the following: the physical response time difference between the two signals is approximately 0.3 ms for gas-phase chemiluminescence and approximately 0.1 ms for photodetector, with a sum of approximately 0.4 ms; the subframe interpolation error is no greater than 0.5 ms; and the combined error does not exceed 0.9 ms. This embodiment selects... =5.0ms, which is greater than the synthesis error and consistent with the frame period, thus avoiding time difference windowing caused by normal sampling jitter.

[0117] The time difference judgment condition is: ≤ In this embodiment, 0.584ms ≤ 5.0ms, which satisfies the time consistency criterion.

[0118] When the time difference between the start time of the first transient and the start time of the second transient falls within the preset time consistency window, it indicates that the two independent physical mechanism signals are synchronized in time, supporting the determination of a real transient event; when the time difference is outside the window, the two signals are from different sources and are determined to be pseudo-transients.

[0119] When the first residual image sequence satisfies the rate criterion and the amplitude criterion, the second residual image sequence satisfies the rate criterion, and the time difference between the start time of the first transient and the start time of the second transient falls within a preset time consistency window, it is determined that a real transient event has occurred in the detection area, and a transient event signal is output; otherwise, it is determined that no real transient event has occurred in the detection area, and no transient event signal is output.

[0120] The final decision logic can be represented by the following formula: EventSignal=(Rate1∧Amp1∧Rate2∧TimeConsist), where Rate1 is the Boolean value of the first residual image sequence satisfying the rate criterion; Amp1 is the Boolean value of the first residual image sequence satisfying the amplitude criterion; Rate2 is the Boolean value of the second residual image sequence satisfying the rate criterion; TimeConsist is the Boolean value of the time difference falling within the time consistency window; ∧ represents the logical AND operation.

[0121] When all four conditions above are true, EventSignal=1, and a transient event signal is output; when any one condition is false, EventSignal=0, and no transient event signal is output.

[0122] In this embodiment, the first residual image sequence satisfies the rate criterion (Rate1=1) and the amplitude criterion (Amp1=1), and the second residual image sequence satisfies the rate criterion (Rate2=1). The time difference of 0.584ms falls within the 5.0ms window (TimeConsist=1), so EventSignal=1, and it is determined that a real transient event has occurred in the detection area.

[0123] The transient event signal is represented by at least one of the following: a switching level transition, a digitally encoded event message, or an error code.

[0124] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other. In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A metal etching endpoint detection system, characterized in that, include: Spectral image acquisition module: used to acquire the first spectral image data and the second spectral image data of the detection area; The first spectral image data is the image data of the intensity of the preset characteristic emission spectral line in the gas phase chemiluminescence above the detection area changing with time, and the second spectral image data is the image data of the intensity of the probe light after passing through the detection area changing with time. The residual image generation module is used to perform slow drift component tracking on the first spectral image data and the second spectral image data respectively, and generate a first residual image sequence and a second residual image sequence. The slow drift component tracking tracks the slowly changing components in the spectral image data using a first time constant. Transient observation window management module: When any residual image sequence meets the preset candidate transient conditions, it starts a transient observation window that lasts for a preset duration, and freezes the baseline update of the slow drift component tracking corresponding to the first residual image sequence and the second residual image sequence within the transient observation window; Transient recognition module: used to identify the first transient start time based on the first residual image sequence and the second transient start time based on the second residual image sequence within the transient observation window; Cross-validation determination module: When the first residual image sequence satisfies the rate criterion and the amplitude criterion, the second residual image sequence satisfies the rate criterion, and the time difference between the start time of the first transient and the start time of the second transient falls within a preset time consistency window, it determines that a real transient event has occurred in the detection area and outputs a transient event signal.

2. The metal etching endpoint detection system according to claim 1, characterized in that: The spectral image acquisition module detects light whose wavelength is within the substrate transparent wavelength window of the detection area, wherein the substrate transparent wavelength window is 1100nm to 1700nm. In the second spectral image data, the pixel gray level at position (x,y) of a single frame of the second spectral image satisfies an exponential extinction relationship with the thickness d(x,y,t) of the metal thin film to be detected, specifically expressed as follows: ， in, Let be the pixel gray level of the second spectral image at time t; Let be the photoelectric response coefficient of the second photoelectric sensor array at position (x, y), which includes exposure time and optical collection efficiency; The intensity of the probe light incident on the upper surface of the detection area; The extinction coefficient of the thin film under test at the probe wavelength λ = 1310 nm; Let t be the thickness of the film to be detected at position (x, y) at time t; The transmittance of the second narrowband filter at the probe wavelength; These are dark current noise and readout noise.

3. The metal etching endpoint detection system according to claim 1, characterized in that: The spectral image acquisition module presets the full width at half maximum (FWHM) of the first narrowband filter corresponding to the feature emission spectral line to be 1.5 nm. The wavelength interval between the preset characteristic emission spectrum and the radiation spectrum of other gaseous components in gas-phase chemiluminescence is not less than 3 nm.

4. The metal etching endpoint detection system according to claim 1, characterized in that: The residual image generation module, the first time constant The relationship between the duration of the transient event to be detected and the following conditions is met: in, It is the shortest duration of the transient signal in the first residual image sequence.

5. The metal etching endpoint detection system according to claim 1, characterized in that: The transient observation window management module determines the candidate transient conditions based on the average pixel grayscale value of a preset spatial region in the residual image sequence. When the absolute value of the average pixel grayscale exceeds the preset candidate amplitude threshold and continues to reach the preset duration threshold, the residual image sequence is determined to meet the candidate transient condition. The preset candidate amplitude threshold is three times the noise standard deviation of the residual image sequence in a stable state.

6. The metal etching endpoint detection system according to claim 1, characterized in that: The transient identification module searches within the transient observation window for the frame index in the first residual sequence where the absolute amplitude first exceeds a preset starting amplitude threshold, and uses this as the frame-level position at the start of the first transient. The frame-level position search formula for the first channel is as follows: ; The frame index where the absolute amplitude first exceeds a preset starting amplitude threshold in the second residual sequence is searched and used as the frame-level position of the second transient start time. The frame-level position search formula for the second channel is as follows: ; Wherein, 𝑘1 and 𝑘2 are the frame indices corresponding to the start times of the first and second channels, respectively; Win is the frame sequence number within the observation window; This refers to the sampling time corresponding to the k-th frame; Linear interpolation or quadratic curve fitting is performed on the one-dimensional residual sequence within the frame-level location neighborhood to generate a first transient start time and a second transient start time that are more than 0.1 times the frame period.

7. The metal etching endpoint detection system according to claim 1, characterized in that: The cross-validation determination module determines the rate criterion based on the first residual image sequence, which is defined as the first residual sequence in which the average pixel value of a preset spatial region changes over time. The second residual sequence is a second residual sequence in which the average pixel value of a preset spatial region changes over time. ; Perform adjacent frame differencing on the first residual sequence and the second residual sequence respectively: ,in, Let c be the average pixel grayscale value of the residual image of the c-th channel within the region of interest, where c=1 represents the first channel and c=2 represents the second channel. Δt represents the sampling time corresponding to the nth frame; Δt is the frame period. The amplitude criterion applies only to the first residual image sequence, and the judgment object is the maximum absolute peak amplitude of the first residual sequence within the transient observation window: Where, 𝐴1 is the maximum absolute amplitude of the first residual image sequence within the observation window; The moment the observation window is activated; Set a preset duration for the observation window.

8. The metal etching endpoint detection system according to claim 1, characterized in that: The cross-validation determination module has a time consistency window value of 5ms. The preset duration of the transient observation window is three times the typical duration of the transient event to be detected; Within the transient observation window, the cross-validation module calculates the time difference frame by frame. When the first residual image sequence satisfies the rate criterion and the amplitude criterion, the second residual image sequence satisfies the rate criterion, and the time difference falls within a 5ms window, the transient event signal is immediately output without waiting for the transient observation window to end.

Images