A method for on-line detection of defects in a silicon carbide chip manufacturing process

Abstract

This invention discloses an online defect detection method in the fabrication process of silicon carbide chips, specifically relating to the field of semiconductor defect detection technology. The method adaptively adjusts laser parameters based on the quality of real-time acquired photoluminescence signals to excite hidden defects and generate characteristic signals. Time-resolved spectral acquisition of the signals is performed, and background noise and self-absorption effects are decoupled using a nonlinear separation model to extract hidden defect features. The signal distribution and noise floor are monitored in real-time, and the discrimination threshold is dynamically adjusted to obtain candidate defect regions and their characteristics. Different detection modes are sequentially activated, and the most sensitive excitation method is selected to confirm the defect type, density, and location. This invention adaptively adjusts laser parameters based on the quality of real-time acquired photoluminescence signals, enabling stable excitation of hidden defects under dynamic conditions. By decoupling background noise and self-absorption effects, pure defect feature signals are extracted. Dynamic adjustment of the discrimination threshold and fusion of three detection modes achieve high-confidence defect identification.

Description

Technical Field

[0001] This invention relates to the field of semiconductor defect detection technology, and more specifically, to an online defect detection method in the silicon carbide chip fabrication process. Background Technology

[0002] Silicon carbide, as a third-generation semiconductor material, has broad application prospects in high-temperature, high-frequency, and high-power devices. During the fabrication of silicon carbide chips, various crystal defects are introduced through processes such as substrate growth, epitaxial deposition, ion implantation, and high-temperature annealing. Among these, hidden defects such as basal plane dislocation half-rings and shallow stacking faults may produce weak signals under conventional optical detection conditions, but they can evolve into leakage channels or cause positive voltage drift during device operation, impacting chip yield and reliability.

[0003] Currently, various technologies have been developed in the field of silicon carbide defect detection. Photoluminescence imaging can obtain information such as dislocations and stacking faults by exciting crystals to emit light, and it is used in offline sampling inspection of substrates and epitaxial wafers. Automated optical inspection is mainly used to identify macroscopic defects such as surface particles and scratches, and has the advantages of high speed and online integration. X-ray topography can accurately reveal the distribution of dislocations, but the equipment cost is high and the detection throughput is relatively limited, so it is mostly used for R&D verification or sampling inspection of key batches.

[0004] In actual production line online inspection environments, the inspection system needs to operate under conditions of high-speed continuous wafer movement, while factors such as ambient temperature and wafer surface condition fluctuate. Traditional photoluminescence detection can achieve good results under static or low-speed conditions, but during dynamic scanning, the following phenomena exist: the excitation laser parameters are fixed, making it difficult to match the differences in carrier lifetime in different regions in real time; substrate background noise and self-absorption effects in the acquired signal are coupled, interfering with weak and hidden defect signals; the defect discrimination threshold is usually based on global calibration, and its adaptability to changes in local noise substrate needs to be improved. In addition, the confidence level of a single detection mode in identifying certain hidden defects is limited, often requiring a comprehensive judgment by combining multiple analysis methods. Therefore, this invention proposes an online defect detection method in the silicon carbide chip fabrication process to solve the above problems. Summary of the Invention

[0005] To achieve the above objectives, the present invention provides the following technical solution: An online defect detection method in the silicon carbide chip fabrication process includes the following steps: Based on the quality of the real-time acquired photoluminescence signal, the wavelength, power, and pulse timing of the excitation laser are adaptively adjusted to excite hidden defects in the silicon carbide chip to generate recognizable characteristic luminescence signals, thus obtaining the original photoluminescence signal. Time-resolved spectral acquisition of the original photoluminescence signal was performed, and the hidden defect signal was decoupled from the substrate background noise and self-absorption effect using a preset nonlinear separation model to extract the characteristic signal of the hidden defect. The signal distribution and noise floor are monitored in real time during the scanning process. The defect discrimination threshold is dynamically adjusted according to the characteristic signals of hidden defects to obtain the dynamically discriminated defect candidate region and the signal characteristics of the defect candidate region. For candidate defect regions, photoluminescence detection mode, Raman spectroscopy detection mode, and nonlinear optical detection mode are sequentially activated. The most sensitive excitation mode is selected for different types of hidden defects to confirm the type, density, and location of the hidden defects.

[0006] In a preferred embodiment, adaptively adjusting the wavelength, power, and pulse timing of the excitation laser specifically includes the following steps: The photoluminescence signal at the current scanning position is acquired in real time, and the signal-to-noise ratio of the signal is calculated by subtracting the common logarithm of the noise root mean square value from the common logarithm of the common logarithm of the signal peak intensity. At the same time, the defect response intensity is calculated by subtracting the background signal intensity of the adjacent defect-free area from the signal intensity at the current scanning position, and then dividing by the background signal intensity. The adjacent defect-free area refers to the area around the current position where the signal intensity is stable and there are no abnormal changes. When the signal-to-noise ratio (SNR) is lower than the first preset SNR threshold, or the defect response intensity is lower than the first preset response intensity threshold, the power of the excitation laser is increased step by step in a fixed increment of 5% of the current power value. After each increase, the SNR and defect response intensity are recalculated until the SNR reaches the second preset SNR threshold or the power reaches the preset maximum safe power. If the SNR still does not meet the standard after power adjustment, the wavelength of the excitation laser is switched to a preset backup wavelength of 266 nanometers. In pulse timing adjustment, the transient decay curve of the photoluminescence signal is acquired and fitted to obtain the carrier lifetime. The fitting method is as follows: with time as the independent variable and signal intensity as the dependent variable, the time value corresponding to the centroid of the area under the decay curve is calculated; then the laser pulse width is set to half of the carrier lifetime and the repetition frequency is set to one-third of the carrier lifetime, so that the excitation pulse matches the carrier lifetime of the hidden defect. The adjusted wavelength, power, and pulse timing parameters are used as the initial excitation conditions for the next scan position.

[0007] In a preferred embodiment, exciting hidden defects in a silicon carbide chip to generate distinguishable characteristic light emission signals to obtain the original photoluminescence signal specifically includes the following steps: The surface of a silicon carbide chip is irradiated with adaptively adjusted laser parameters to excite electronic transitions in hidden defect regions. The photoluminescence signal intensity was recorded as a function of time after the laser pulse stopped. The luminescence intensity and attenuation characteristic parameters of the hidden defects were calculated. The luminescence intensity was calculated by taking the maximum intensity value on the curve. The attenuation characteristic parameter was the average attenuation time, which was calculated by integrating the product of time and signal intensity over a period of three times the time required for the signal to attenuate to one-eth of its peak value, starting from the moment the laser pulse stopped, and then dividing by the integral value of the signal intensity over the same time period. For a plane dislocation half-ring, the first average attenuation time was calculated; for a shallow stacking fault, the second average attenuation time was calculated. In the defect-free region of the silicon carbide chip, the reference decay time of the substrate band edge emission is obtained using the same laser parameters and calculation methods. The first average decay time is compared with the reference decay time: if the two are equal or the difference is less than the preset allowable deviation range, it is initially determined that the signal cannot be distinguished from the substrate background, the position is marked as the area to be re-examined and re-acquisition is triggered, and adaptive adjustment is re-executed after re-acquisition is triggered; if the two are not equal and the difference is greater than or equal to the preset allowable deviation range, it is initially confirmed that the hidden defect is a base plane dislocation half-ring. The second average decay time is compared with the reference decay time: if the two are equal or the difference is less than the preset allowable deviation range, it is initially determined that the signal cannot be distinguished from the substrate background, the position is marked as the area to be re-examined and re-acquisition is triggered, and adaptive adjustment is re-executed after re-acquisition is triggered; if the two are not equal and the difference is greater than or equal to the preset allowable deviation range, it is initially confirmed that the hidden defect is a shallow stacking fault. When the same location simultaneously meets the preliminary confirmation conditions for both a plane dislocation half-loop and a shallow stacking fault, the defect is initially identified as a plane dislocation half-loop according to the preset priority rules. Since a plane dislocation half-loop contains stacking faults in its structure and poses a greater threat to device reliability, it is not counted as a shallow stacking fault again. The photon counts located in the characteristic emission band of the hidden defect are extracted using a preset bandpass filter. The characteristic emission band is from 500 nanometers to 600 nanometers. The photon counts are then divided by the acquisition time to obtain the emission intensity. The calculated luminescence intensity and the first or second average decay time corresponding to the preliminary judgment result are correlated with the current spatial coordinates to generate the original photoluminescence signal containing the luminescence intensity and decay characteristics of the hidden defects.

[0008] In a preferred embodiment, time-resolved spectral acquisition of the original photoluminescence signal specifically includes the following steps: At the same time the excitation laser pulse stops, the recorded continuous time change curve is sampled at a preset fixed time interval. The total sampling time is set to the time required for the signal intensity to decay to one percent of its peak value. The sampled discrete signal intensity values ​​are arranged in chronological order to form the original time series data. The original time series data is divided into a predetermined number of continuous and non-overlapping time windows, with each time window containing multiple sampling points; The arithmetic mean of the intensity values ​​of all sampling points within each time window is calculated to obtain the average signal intensity of that time window. The average signal intensities of all time windows are then arranged in chronological order to generate discrete time-resolved spectral data.

[0009] In a preferred embodiment, a preset nonlinear separation model is used to decouple the hidden defect signal from the substrate background noise and self-absorption effect, and the characteristic signal of the hidden defect is extracted. Specifically, this includes the following steps: Discrete-time resolved spectral data is used as input, and a preset nonlinear separation model is invoked. This model uses an independent component analysis algorithm and is trained in advance using mixed signal samples containing known hidden defect signals, substrate background noise and self-absorption effects to obtain three independent signal component basis functions. The input discrete time-resolved spectral data is decomposed into three signal components, where the first component corresponds to the substrate background noise, the second component corresponds to the signal distortion caused by the self-absorption effect, and the third component corresponds to the characteristic signal of the hidden defect. The amplitude of the third component in each time window is calculated, and the amplitude sequence is correlated with the spatial coordinates acquired simultaneously to obtain the temporal and spatial distribution of the hidden defect feature signal. Subtracting the first and second components from the discrete-time resolved spectral data yields the decoupled hidden defect feature signal.

[0010] In a preferred embodiment, real-time monitoring of signal distribution and noise floor during the scanning process specifically includes the following steps: Within the current scan line or scan frame, continuously acquire the original photoluminescence signal intensity values ​​corresponding to each spatial location to form a one-dimensional or two-dimensional signal distribution map; In the signal distribution map, the entire map is traversed with a preset sliding window size. The minimum signal strength is calculated in each window, and the minimum value is used as the local noise basis estimate for that window. Arrange the local noise base estimates of all windows according to their spatial location to form a noise base distribution map; The noise floor distribution map is updated in real time. For each new raw signal strength value collected, the noise floor of the window in which it is located is recalculated, and old data that has exceeded the preset historical time period is removed.

[0011] In a preferred embodiment, the defect discrimination threshold is dynamically adjusted based on the characteristic signal of the hidden defect to obtain the dynamically discriminated defect candidate region and the signal characteristics of the defect candidate region. This specifically includes the following steps: Multiply the value at the corresponding position in the noise floor distribution map by a preset multiplier coefficient, and then multiply by a preset mapping coefficient to obtain the benchmark value of the dynamic defect discrimination threshold applicable to the hidden defect feature signal. The mapping coefficient is pre-calibrated based on the statistical relationship of signal strength before and after decoupling. The intensity of the decoupled hidden defect feature signal is compared with the benchmark value of the dynamic defect discrimination threshold: if the feature signal intensity is greater than or equal to the benchmark value, the spatial location is marked as a defect candidate point; if it is less than the benchmark value, it is marked as a non-defect point. Connectivity analysis is performed on adjacent defect candidate points to merge spatially connected candidate points into a single defect candidate region, and the boundary coordinates of this region are recorded. Calculate the arithmetic mean and maximum value of the characteristic signal intensity of all candidate points within each defect candidate region, and use them as the signal feature output of that defect candidate region.

[0012] In a preferred embodiment, for the defect candidate region, multiple detection modes are sequentially activated from photoluminescence detection mode, Raman spectroscopy detection mode, and nonlinear optical detection mode, specifically including the following steps: The candidate defect regions are sorted according to their spatial coordinates to form a queue to be detected; For each candidate defect region in the queue, perform the following operations in sequence: first, enable the photoluminescence detection mode and collect the photoluminescence spectrum of the region; then, enable the Raman spectroscopy detection mode and collect the Raman shift spectrum of the region; finally, enable the nonlinear optical detection mode and collect the second harmonic or two-photon excitation signal of the region. For each candidate defect region, data from all three detection modes must be collected.

[0013] In a preferred embodiment, the most sensitive excitation method is selected for different types of hidden defects to confirm the type, density, and location of the hidden defects. This specifically includes the following steps: Photoluminescence spectrum, Raman shift spectrum, and nonlinear optical signal are compared with a pre-defined standard feature database of various hidden defects. The matching degree is calculated as follows: for each detection mode, the acquired signal feature curve is subtracted point by point from the standard feature curve of the corresponding defect type in the database. The absolute values ​​of the differences at each point are summed and divided by the total number of points of the standard feature curve to obtain the average absolute deviation. Then, the ratio of the average absolute deviation to the pre-defined maximum allowable deviation is subtracted from one to obtain the matching degree between the signal and the defect type under the detection mode. The matching degree value is between zero and one. For each defect candidate region, the defect type corresponding to the detection mode with the highest matching degree is selected as the final defect type of the region, and the sensitive activation mode corresponding to the type is recorded as the recommended parameter for subsequent review; if the final defect type is inconsistent with the previous judgment result, the current final confirmation result shall prevail. The density of the same defect type is obtained by counting the number of the same defect type in all defect candidate regions and dividing it by the total area of ​​the batch of wafers. The spatial coordinate boundaries, final defect type, and density data of each defect candidate region are summarized, and a list of confirmed defect information is output.

[0014] The technical effects and advantages of this invention are as follows: This invention adaptively adjusts the wavelength, power, and pulse timing of the excitation laser based on the quality of the real-time acquired photoluminescence signal. This enables stable excitation of distinguishable characteristic luminescence signals from hidden defects such as basal plane dislocation semi-rings and shallow stacking faults in silicon carbide chips under dynamic conditions such as high-speed scanning and temperature fluctuations, avoiding missed detections caused by dynamic response lag or signal nonlinear distortion. By performing time-resolved spectral acquisition on the original photoluminescence signal and using a preset nonlinear separation model to decouple the hidden defect signal from substrate background noise and self-absorption effects, pure hidden defect characteristic signals can be extracted from the mixed signal, eliminating the influence of background interference on the accuracy of discrimination. By dynamically adjusting the defect discrimination threshold in real-time by monitoring the signal distribution and noise substrate, and by combining photoluminescence, Raman spectroscopy, and nonlinear optics detection modes sequentially to confirm the defect type, density, and location, the invention can adaptively maintain detection sensitivity under local noise substrate changes and improve the confidence of hidden defect identification through multimodal fusion, ultimately outputting a complete list of defect information. Attached Figure Description

[0015] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is a schematic diagram of an online defect detection method in the silicon carbide chip fabrication process according to the present invention. Detailed Implementation

[0016] 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 of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Reference Figure 1 The following examples were obtained: Example 1: An online defect detection method in the silicon carbide chip fabrication process, comprising the following steps: Based on the quality of the real-time acquired photoluminescence signal, the wavelength, power, and pulse timing of the excitation laser are adaptively adjusted to excite hidden defects in the silicon carbide chip to generate recognizable characteristic luminescence signals, thus obtaining the original photoluminescence signal. Time-resolved spectral acquisition of the original photoluminescence signal was performed, and the hidden defect signal was decoupled from the substrate background noise and self-absorption effect using a preset nonlinear separation model to extract the characteristic signal of the hidden defect. The signal distribution and noise floor are monitored in real time during the scanning process. The defect discrimination threshold is dynamically adjusted according to the characteristic signals of hidden defects to obtain the dynamically discriminated defect candidate region and the signal characteristics of the defect candidate region. For candidate defect regions, photoluminescence detection mode, Raman spectroscopy detection mode, and nonlinear optical detection mode are sequentially activated. The most sensitive excitation mode is selected for different types of hidden defects to confirm the type, density, and location of the hidden defects.

[0018] Adaptive adjustment of the wavelength, power, and pulse timing of the excitation laser includes the following steps: The photoluminescence signal at the current scanning position is acquired in real time, and the signal-to-noise ratio (SNR) is calculated as follows: 20 times the common logarithm of the peak signal intensity minus 20 times the common logarithm of the noise root mean square (RMS). Simultaneously, the defect response intensity is calculated by subtracting the background signal intensity of the adjacent defect-free region from the signal intensity at the current scanning position, and then dividing by the background signal intensity. The adjacent defect-free region refers to an area where the signal intensity is stable and there are no abnormal abrupt changes around the current position. This step is to determine if the current excitation conditions are good enough. If the SNR is too low, the weak, hidden defect signal will be drowned out by noise. The defect response intensity is determined by how much stronger the signal is at the candidate defect position compared to the surrounding normal area. If the difference is not significant, it indicates that the excitation conditions have not successfully excited the defect.

[0019] For example, scanning a six-inch silicon carbide epitaxial wafer, the calculated signal-to-noise ratio at the current point is only 10 dB, while the normal requirement is over 20 dB. Assuming the root mean square noise value is 5 counts and the peak signal value is 50 counts, multiplying by the commonly used logarithm (20 x 1.7 - 20 x 0.7) yields exactly 10 dB. Simultaneously, observing the defect response intensity, the signal intensity at the current point is 60, while the background in the surrounding stable area is 55, resulting in a calculated response intensity of only 0.09, far below the preset 1.0. This indicates insufficient laser power or an incorrect wavelength, requiring adjustment.

[0020] When the signal-to-noise ratio (SNR) is lower than the first preset SNR threshold, or the defect response intensity is lower than the first preset response intensity threshold, the excitation laser power is increased incrementally in steps of 5% of the current power value. After each increase, the SNR and defect response intensity are recalculated until the SNR reaches the second preset SNR threshold, or the power reaches the preset maximum safe power. If the SNR still does not meet the standard after power adjustment, the excitation laser wavelength is switched to a preset backup wavelength of 266 nanometers. This step provides a specific adjustment strategy: adjust the power first, as it is the easiest to change. A 5% step size will not cause laser overshoot or damage to the wafer. The calculation is repeated after each adjustment to see if there is any improvement. If the SNR is still insufficient even after adjusting the power to the maximum safe power, it indicates that the absorption efficiency of this wavelength itself is poor, and the backup wavelength is switched.

[0021] For example, the initial laser power is 50 milliwatts, with a signal-to-noise ratio (SNR) of 10 dB. The first preset SNR threshold is 15 dB, which is insufficient. Therefore, the power is increased by 5% each time. The first time, it's increased to 52.5 milliwatts, and the SNR becomes 11 dB; the second time, it's increased to 55.1 milliwatts, and it becomes 12.5 dB; ... When it reaches 65 milliwatts, the SNR reaches 16 dB, exceeding the second preset threshold of 15 dB, so it stops. If the maximum safe power of 80 milliwatts still only results in 14 dB, then the wavelength is switched from the commonly used 355 nanometers to 266 nanometers, and the process restarts.

[0022] In pulse timing adjustment, the transient decay curve of the photoluminescence signal is acquired and fitted to obtain the carrier lifetime. The fitting method is as follows: with time as the independent variable and signal intensity as the dependent variable, the time value corresponding to the centroid of the area under the decay curve is calculated. Then, the laser pulse width is set to half of the carrier lifetime, and the repetition frequency is set to one-third of the carrier lifetime to match the excitation pulse with the carrier lifetime of the hidden defect. After the power and wavelength are adjusted, the pulse timing also needs to be adjusted. Different defects have different carrier lifetimes. If the pulse width and repetition frequency do not match the lifetime, either the excitation will be insufficient or the signal will overlap. Calculating the centroid of the area under the decay curve is to find the average time point of signal decay, and this value is the carrier lifetime. Setting the pulse width to half of the lifetime ensures that the pulse ends before most carriers recombine after excitation, avoiding signal interference from the pulse itself. Setting the repetition frequency to one-third of the lifetime means that the interval between two pulses is three times the lifetime, so that the signal from the previous excitation is basically decayed before the next one is emitted, without crosstalk.

[0023] For example, by measuring the transient decay curve of a base-plane dislocation half-ring, the calculated carrier lifetime corresponding to the barycenter is 100 nanoseconds. The laser pulse width is set to 50 nanoseconds, and the repetition frequency is set to 1 / 300 MHz, or approximately 3.3 MHz. Each pulse lasts 50 nanoseconds, and then there is a 300-nanosecond wait before the next pulse is emitted, allowing sufficient time for the signal to decay completely.

[0024] The adjusted wavelength, power, and pulse timing parameters are used as the initial excitation conditions for the next scan position. This step transfers the optimized parameters from the current point to the next scan point. Because the surface characteristics of silicon carbide wafers change slowly—for example, the edges warp more than the center, or the impurity concentration varies at different locations—using the optimal parameters from the previous point as the starting point for the next point can significantly reduce the number of adjustments and increase the scan speed. If the conditions at the next point change significantly, the system will automatically adjust again, thus always maintaining the optimal excitation state.

[0025] To excite hidden defects in a silicon carbide chip to generate a distinguishable characteristic light-emitting signal, thus obtaining the original photoluminescence signal, the following steps are involved: The surface of the silicon carbide chip is irradiated with adaptively adjusted laser parameters to excite electron transitions in the hidden defect region. After the preceding series of adjustments to power, wavelength, and pulse timing, the laser has found its optimal operating state for the current scanning position. Irradiating the chip surface with these parameters ensures that the laser energy is precisely absorbed by the hidden defect region. Silicon carbide is a wide bandgap semiconductor; under normal conditions, electrons are bound in the valence band and cannot move. After the laser strikes the chip, electrons absorb photon energy and transition to the conduction band, forming electron-hole pairs between the conduction and valence bands. These carriers recombine over time, releasing photons, which constitute the photoluminescence signal. Hidden defects are called hidden defects because they do not produce a distinguishable light-emitting signal under normal excitation conditions. Therefore, the adaptively adjusted parameters must be used to address this, such as increasing the power, changing the wavelength to a band with higher absorption efficiency, and ensuring the pulse width matches the carrier lifetime.

[0026] The photoluminescence signal intensity was recorded as a function of time after the laser pulse stopped. The luminescence intensity and attenuation characteristic parameters of the hidden defects were calculated. The luminescence intensity was calculated by taking the maximum intensity value on the curve. The attenuation characteristic parameter was the average attenuation time, which was calculated by integrating the product of time and signal intensity over a period of three times the time required for the signal to attenuate to one-eth of its peak value, starting from the moment the laser pulse stopped, and then dividing by the integral value of the signal intensity over the same time period. For a plane dislocation half-ring, the first average attenuation time was calculated; for a shallow stacking fault, the second average attenuation time was calculated. Once the laser pulse stops, the excitation process ends, and what remains is the slow recombination of charge carriers. The signal intensity recorded at this point over time is the decay curve. This curve contains two key pieces of information: first, the height of the highest point, which represents the luminous intensity and how much light signal was generated in the defect region; and second, the rate of decay, which represents the carrier lifetime, which varies depending on the defect.

[0027] The significance of using the integral formula to calculate the average decay time is that the numerator is the integral of time multiplied by the signal intensity, and the denominator is the integral of the signal intensity. The calculated value is the "center of gravity" moment of the entire decay process. In practice, we measure a wafer. A suspicious location is scanned, parameters are adjusted, a laser is applied, and the decay curve is recorded. If the calculated average decay time is 85 nanoseconds, it is recorded as the first average decay time, initially suspected to be a plane dislocation half-ring; if it is 120 nanoseconds, it is recorded as the second average decay time, initially suspected to be a shallow stacking fault. The different decay times of the two types of defects are due to their different energy level structures and non-radiative recombination rates; this difference is the basis for distinguishing them.

[0028] In a defect-free region of a silicon carbide chip, a reference decay time for substrate bandgap emission is obtained using the same laser parameters and calculation methods. The substrate itself also emits light, called bandgap emission, and this signal is always present. If the decay time of the substrate's own emission is unknown, it's impossible to determine whether the previously calculated average decay time is contributed by defects or the substrate. Therefore, it's necessary to measure the area in the defect-free region—that is, a defect-free location on the wafer—using the exact same laser parameters and methods to obtain a reference value. For example, select a location on an epitaxial wafer, confirm with a high-powered microscope that there are no dislocations or stacking faults, then use the same laser parameters to measure the area, record the decay curve, and calculate the average decay time, assuming it's 100 nanoseconds. This 100 nanoseconds is the reference decay time, and the average decay times calculated for all subsequent questionable locations must be compared.

[0029] The first average decay time is compared with the reference decay time: if the two are equal or the difference is less than the preset allowable deviation range, it is initially determined that the signal cannot be distinguished from the substrate background, the position is marked as the area to be re-examined and re-acquisition is triggered, and adaptive adjustment is re-executed after re-acquisition is triggered; if the two are not equal and the difference is greater than or equal to the preset allowable deviation range, it is initially confirmed that the hidden defect is a base plane dislocation half-ring. The second average decay time is compared with the reference decay time: if the two are equal or the difference is less than the preset allowable deviation range, it is initially determined that the signal cannot be distinguished from the substrate background, the position is marked as the area to be re-examined and re-acquisition is triggered, and adaptive adjustment is re-executed after re-acquisition is triggered; if the two are not equal and the difference is greater than or equal to the preset allowable deviation range, it is initially confirmed that the hidden defect is a shallow stacking fault. The purpose of the comparison is to determine whether there are hidden defects at this location. If the calculated average decay time and the reference decay time in the defect-free region are almost the same, it means that the signal at this location is the background signal of the substrate itself and there are no defect characteristics. It is then marked as needing further investigation, possibly due to a false alarm or insufficient excitation conditions. Therefore, a re-acquisition is triggered, returning to the adaptive adjustment step for another attempt. If the calculated average decay time and the reference value are significantly different, and the difference exceeds the preset allowable deviation range, it indicates that the carrier recombination behavior at this location is different from that of a normal substrate. This preliminarily confirms the presence of a defect, and a preliminary determination of its type is made based on whether it is the first or second average decay time. Regarding the preset allowable deviation range, there are clear standards in the existing technology. The international standard IEC63068-3:2020 (or equivalently GB / T43493.3-2023) specifies a non-destructive method for detecting defects in 4H-SiC homoepitaxial wafers using photoluminescence technology, focusing on two typical defects: basal plane dislocations and stacking faults, and providing their characteristic emission wavelengths and morphological descriptions. The standard requires periodic calibration of system parameters to ensure that excitation light intensity fluctuations are less than ±5%. It also requires the establishment of a reference wafer library containing photoluminescence image features of typical defects. Based on this, the preset allowable deviation range is typically set between 5% and 10% of the reference decay time. For example, if the reference decay time is 100 nanoseconds, a 5% allowable deviation range would be ±5 nanoseconds. If the calculated average decay time at a suspicious location is between 95 and 105 nanoseconds, it is considered indistinguishable from the substrate background; if it is less than 95 nanoseconds or greater than 105 nanoseconds, a defect is confirmed. This value is based on the fact that, under the same process conditions and within the same batch of wafers, the decay time fluctuation in defect-free areas typically does not exceed 3% to 5%. Setting the allowable deviation range above this range ensures the authenticity of the distinguishing signal while preventing frequent false alarms due to excessively small fluctuations.

[0030] When the same location simultaneously meets the preliminary confirmation conditions for a plane dislocation half-loop and a shallow stacking fault, the defect is initially identified as a plane dislocation half-loop according to a preset priority rule. This is because a plane dislocation half-loop structurally contains a stacking fault and poses a greater threat to device reliability, so it is not counted again as a shallow stacking fault. This situation does indeed occur in actual testing. A plane dislocation half-loop itself contains a stacking fault; two Shockley incomplete dislocations are sandwiched between a surface defect, which is the stacking fault. Therefore, when detecting a plane dislocation half-loop, its signal characteristics contain both dislocation characteristics (corresponding to the first average decay time) and stacking fault characteristics (corresponding to the second average decay time), both conditions being met simultaneously. In this case, it cannot be counted as both a dislocation and a stacking fault, as this would result in double counting. According to the priority rule, it is first judged as a base plane dislocation half-loop, because this kind of thing poses a much greater threat to the reliability of the device than a shallow stacking fault alone. When the bipolar current is turned on, the base plane dislocation half-loop will evolve into a through stacking fault, causing the forward voltage of the body diode to drift continuously, eventually causing the entire chip to fail.

[0031] In actual scanning, the following situation may be encountered: the first average decay time calculated for a location is 85 nanoseconds, which exceeds the allowable deviation range from the reference value of 100 nanoseconds. At the same time, the second average decay time is 120 nanoseconds, which is also significantly different from the reference value. If both conditions are met, then according to the priority rule, only one base plane dislocation half-loop will be counted in the end, and shallow stacking faults will not be counted separately.

[0032] A pre-set bandpass filter is used to extract photon counts within the characteristic emission band of the hidden defect, which is between 500 and 600 nanometers. The photon count is then divided by the acquisition time to obtain the emission intensity. While the decay time dimension has been the primary focus, emission intensity is also crucial. Besides defect emission, photoluminescence signals contain stray light and environmental interference, necessitating a bandpass filter to filter out light in specific bands. The 500-600 nanometer band is a region where the characteristic emission of hidden defects is concentrated. Dividing the extracted photon count by the acquisition time yields the number of photons per unit time, which is the emission intensity. This intensity value will be recorded as a characteristic parameter of the defect for screening candidate defect regions.

[0033] The international standard BSIEC 63068-3:2020 specifies the characteristic detection wavelengths for different defects. Planar dislocations are typically observed using long-pass filters with wavelengths greater than 650 nm for bright line contrast. Stacking faults exhibit dark contrast at wavelengths greater than 650 nm, transitioning to bright contrast at wavelengths of 420 nm. Here, we select 500 to 600 nm as the characteristic emission band for hidden defects, representing a compromise of a wide wavelength range. The aim is to capture as many emission signals from both types of hidden defects as possible in a single acquisition, avoiding signal loss due to a narrow wavelength band.

[0034] The calculated luminescence intensity and the corresponding first or second average decay time based on the preliminary judgment are correlated with the current spatial coordinates to generate a raw photoluminescence signal containing the luminescence intensity and decay characteristics of the hidden defects. At this step, the data for each scanning position is complete: the spatial coordinates indicate the location on the wafer, the luminescence intensity indicates the signal strength, and the first average decay time is included for a plane dislocation semi-ring, while the second average decay time is included for a shallow stacking fault. This information is packaged together to form a tagged data record, which is called the raw photoluminescence signal.

[0035] The time-resolved spectral acquisition of the original photoluminescence signal includes the following steps: Simultaneously with the cessation of the laser pulse, the recorded continuous-time variation curve is sampled at preset fixed time intervals. For example, the fixed time interval is set to ten nanoseconds based on the main carrier lifetime range of silicon carbide material, and the total sampling time is set to the time required for the signal intensity to decay to one percent of its peak value. The continuous analog signal is converted into discrete digital points because computers can only process discrete data. The carrier lifetime caused by common basal plane dislocations and stacking faults in silicon carbide is generally between tens and hundreds of nanoseconds. A sampling interval of ten nanoseconds can ensure that enough points are sampled in each lifetime cycle without the data volume being too large. For example, a signal with a lifetime of one hundred nanoseconds can be sampled at ten points, which can basically restore the decay shape. The total sampling time is taken until the peak value decays to one percent, which is usually around one to two microseconds. This way, the slow decay of shallow stacking fault signals is not lost, and time is not wasted waiting for it to completely return to zero. When actually measuring an epitaxial wafer, the timing starts the moment the laser pulse stops, and an intensity value is recorded every ten nanoseconds until only one percent of the peak value remains.

[0036] The sampled discrete signal intensity values ​​are arranged in chronological order to form the original time series data. The sampled points are arranged in a column according to time sequence. The first point is the intensity at zero nanosecond, the second point is the intensity at ten nanosecond, and so on. This sequence is the original time series data, which retains all the information about the signal's changes over time, but it is not yet a spectrum and will be processed later.

[0037] The original time series data is divided into a predetermined number of continuous and non-overlapping time windows, each containing multiple sampling points. Since a single sampling point often contains random noise, analyzing a single point directly would be very unstable. Therefore, several consecutive sampling points are bundled into a window, for example, five sampling points (fifty nanoseconds) per window. The windows do not overlap, so each window contains five intensity values. The number of windows is predetermined, for example, the entire decay curve is divided into twenty windows.

[0038] The arithmetic mean of the intensity values ​​of all sampling points within each time window is calculated to obtain the average signal intensity of that time window. The average signal intensities of all time windows are then arranged in chronological order to generate discrete time-resolved spectral data. The sum of the intensities of several sampling points in each window is divided by the number of points to obtain an average value, which represents the average luminous intensity of that time period. Twenty windows yield twenty average values, which, when arranged in chronological order, constitute discrete time-resolved spectral data. This data is smoother than the original sampling sequence, removes high-frequency noise, and reduces the time resolution from ten nanoseconds to the window width (e.g., fifty nanoseconds). This resolution is sufficient to distinguish the attenuation characteristics of different hidden defects, as the attenuation time difference between basal plane dislocation half-rings and shallow stacking faults is typically greater than tens of nanoseconds.

[0039] The hidden defect signal is decoupled from the substrate background noise and self-absorption effect using a pre-defined nonlinear separation model, and the characteristic signal of the hidden defect is extracted. The specific steps include: Discrete-time resolved spectral data is used as input, and a preset nonlinear separation model is invoked. This model uses an independent component analysis algorithm and is trained in advance using mixed signal samples containing known hidden defect signals, substrate background noise and self-absorption effects to obtain three independent signal component basis functions. The training process of the nonlinear separation model is as follows: Three types of pure source signal samples are prepared. The first type is taken from a silicon carbide epitaxial wafer with a thickness of less than 10 micrometers and no defects. The attenuation curve of the substrate background noise is obtained using weak excitation conditions and denoted as vector N. The second type is taken from a thin sheet sample containing high-density basal plane dislocation semi-rings. The attenuation curve of the pure hidden defect signal is extracted through spatial filtering and averaging processing and denoted as vector D. The third type is taken from a silicon carbide substrate with a thickness of more than 300 micrometers. The signal distortion curve caused by self-absorption effect is measured using the differential double pulse method and denoted as vector A. These three types of source signals are mixed in a random ratio to generate one thousand sets of mixed signal samples. The mixing formula for each set of samples is M=aD+bN+c*A, where a, b, and c are random coefficients uniformly distributed between 0 and 1. Gaussian noise with a signal-to-noise ratio of 20 decibels is added after mixing to simulate the actual detection environment. Each set of mixed signals is used as observation data, and a fast independent component analysis algorithm is used for iterative calculation. Each iteration estimates the unmixing matrix by maximizing non-Gaussianity, until the matrix change between two adjacent iterations is less than one-thousandth. The three independent component basis functions obtained after training are denoted as B_defect, B_noise, and B_abs, respectively. Each basis function is a real number vector of length L (L equals the number of time windows), for example, B_defect=[0.12,0.35,0.78,0.92,0.61,0.23,0.08], representing the standard waveform of the hidden defect signal over a continuous time window.

[0040] The input discrete-time resolved spectral data is decomposed into three signal components. The first component corresponds to the substrate background noise, the second component corresponds to the signal distortion caused by self-absorption, and the third component corresponds to the characteristic signal of the hidden defect. For the actual acquired discrete-time resolved spectral data X (which is also a vector of length L), the model calculates the amplitude coefficients s1, s2, and s3 of the three components, such that X ≈ s1 * B_noise + s2 * B_abs + s3 * B_defect. The amplitude coefficients are solved using least-squares projection: s1 = X * B_noise, s2 = X * B_abs, s3 = X * B_defect, where the dot product represents the vector inner product. After obtaining the three coefficients, the first component signal is reconstructed as s1 * B_noise, the second component as s2 * B_abs, and the third component as s3 * B_defect. For example, at a certain position X = [15, 42, 87, 105, 72, 28, 10], B_noise=[0.05,0.08,0.10,0.09,0.07,0.05,0.03]; B_abs=[0.02,0.06,0.15,0.20,0.18,0.10,0.05]; B_defect=[0.10,0.30,0.70,0.85,0.55,0.20,0.07]; After calculating the inner product, we may get s1=200, s2=180, s3=120, then the third component signal is 120*B_defect=[12,36,84,102,66,24,8.4].

[0041] The amplitude of the third component in each time window is calculated, and this amplitude sequence is associated with the simultaneously acquired spatial coordinates to obtain the temporal and spatial distribution of the hidden defect feature signal: the third component s3*B_defect itself is a sequence, and its i-th element represents the amplitude of the i-th time window. This amplitude sequence is directly used as the temporal distribution of the hidden defect feature signal. At the same time, the X and Y coordinates of the current scanning position are bound to this sequence to form four-dimensional data (X, Y, time window number, amplitude). For example, at the wafer coordinates (12.5mm, 8.3mm), the amplitude sequence of the third component is [12, 36, 84, 102, 66, 24, 8.4], which means that the amplitude of the defect signal at this position is 12 in the first window, 36 in the second window, and so on, until the seventh window is 8.4.

[0042] Subtracting the first and second components from the discrete-time resolved spectral data yields the decoupled hidden defect characteristic signal: Subtraction is performed at each time window: Decoupled signal = X - s1*B_noise - s2*B_abs. Since X = s1*B_noise + s2*B_abs + s3B_defect plus the residual, the result after subtraction is s3B_defect plus a small residual, which is the decoupled hidden defect characteristic signal. Continuing with the example above, X=[15,42,87,105,72,28,10], subtracting the first component [10,16,20,18,14,10,6] and the second component [3.6,10.8,27,36,32.4,18,9], we get [1.4,15.2,40,51,25.6,0,-5]. After removing the negative values, it is approximately [0,15.2,40,51,25.6,0,0], which differs from the third component [12,36,84,102,66,24,8.4]. This is because there are residuals in the actual decomposition, but the main peak features have been preserved. This decoupled signal is output as a pure hidden defect feature signal for use in the dynamic threshold tracking step.

[0043] Real-time monitoring of signal distribution and noise floor during the scanning process includes the following steps: Within the current scan line or frame, the original photoluminescence signal intensity values ​​corresponding to each spatial location are continuously acquired, forming a one-dimensional or two-dimensional signal distribution map. The silicon carbide wafer is scanned row by row, and an original photoluminescence signal intensity value is recorded at each spatial location. This intensity value is the undecoupled raw value obtained directly after adaptive excitation and initial acquisition. The intensity values ​​at all locations on the same scan line are arranged in spatial order to obtain a one-dimensional distribution map, and the data from all rows within the entire scan frame are arranged by rows and columns to obtain a two-dimensional distribution map. For example, when scanning the central region of a six-inch wafer, the current scan line starts at coordinates (10mm, 20mm) and ends at (10mm, 30mm), acquiring one point every one micrometer, for a total of 10,000 points. The intensity values ​​of these points are connected to form the one-dimensional signal distribution map for that row. If the entire rectangular area is scanned, a two-dimensional matrix is ​​obtained, where each element represents the original photoluminescence signal intensity at the corresponding wafer coordinates.

[0044] In the signal distribution map, a preset sliding window is used to traverse the entire map. Within each window, the minimum signal strength is calculated and used as the local noise floor estimate for that window. Background noise levels vary at different locations in the signal distribution map; for example, noise is higher in areas with greater warping at the wafer edge and lower in the flat central region. Therefore, a single global noise value cannot be used. The sliding window method is employed: a preset window size is used, for example, 50 pixels wide and 50 pixels high (corresponding to actual dimensions of 50 micrometers by 50 micrometers). Starting from the top left corner of the distribution map, the window moves one step to the right or down each time (the step size is usually set to half the window width, i.e., 25 pixels). Within the area covered by each window, the minimum signal strength value among all signal strength values ​​is found. The minimum value is chosen instead of the average or median because it is closest to the background signal level when the local area is truly defect-free—defect signals typically exhibit increased intensity, and choosing the minimum value eliminates interference from defects and abnormal spikes. For example, if the intensity values ​​collected in a window are [52,48,55,47,53,49,51,46,54,50], and the minimum value is 46, then the local noise basis estimate for that window is 46. Repeat this operation for all windows in the entire image.

[0045] The local noise floor estimates of all windows are arranged spatially to form a noise floor distribution map. Each sliding window corresponds to a spatial region (e.g., the coordinates of the window center point). Placing the local noise floor estimate calculated for each window onto the center coordinates of that window results in a sparse grid. For coordinates not covered by the window center, bilinear interpolation or nearest neighbor interpolation is used to fill the gaps, ultimately obtaining a noise floor distribution map of the same size as the original signal distribution map. The value of each pixel in the map represents the estimated background noise level at that location. For example, if the original signal distribution map is 1000×1000 pixels and the sliding window step size is 25 pixels, then the window center points form a 40×40 grid, with a noise floor value at each grid point. After interpolation, a complete 1000×1000 noise floor distribution map is obtained. The noise floor value at coordinates (10.5mm, 20.3mm) in the map might be 47.2, while at (10.5mm, 20.5mm) it might be 47.8.

[0046] The noise floor distribution map is updated in real time. Each time a new raw signal strength value is acquired, the noise floor of its corresponding window is recalculated, and old data exceeding the preset historical duration is removed. During online inspection, the wafer moves continuously, generating new signal strength values, and the noise floor needs to be updated accordingly. A sliding time window or sliding spatial window mechanism is used: a preset historical duration is set, such as ten seconds or one hundred scan lines. Each time a new raw signal strength value is acquired, the sliding window to which that value belongs is identified (its spatial coordinates determine the window position). The portion of the old data stored within that window that exceeds the preset historical duration is removed (e.g., only points acquired within the last ten seconds are retained). Then, the minimum remaining signal strength value within that window is recalculated to obtain the updated local noise floor estimate, and the corresponding value in the noise floor distribution map is updated. For example: assuming the preset historical duration is five seconds, and the current acquisition value is 58 at coordinates (12.0mm, 15.0mm), this location belongs to window W (covering the area from 11.8mm to 12.2mm and from 14.8mm to 15.2mm). The window W already stores the intensity values ​​of 200 points collected in the past five seconds. Adding the new value, there are a total of 201 points. After removing the 30 old points collected more than five seconds ago, 171 points remain. The minimum value of these 171 points is calculated to be 45. Therefore, the value at the center of the window in the noise floor distribution map is updated to 45. This allows the noise floor distribution map to reflect the slow changes in wafer surface characteristics in real time. For example, the background noise gradually increases after the wafer is heated by a laser from room temperature, or the noise level changes when scanning different crystal orientation regions; the system can adjust accordingly.

[0047] The defect discrimination threshold is dynamically adjusted based on the characteristic signals of hidden defects to obtain dynamically discriminated defect candidate regions and their signal characteristics. This process includes the following steps: Multiplying the value at the corresponding position in the noise floor distribution map by a preset multiplier coefficient, and then by a preset mapping coefficient, yields a baseline value for the dynamic defect discrimination threshold applicable to the hidden defect feature signal. The mapping coefficient is pre-calibrated based on the statistical relationship between signal strength before and after decoupling. The noise floor distribution map is calculated based on the original acquired photoluminescence signal intensity value, but the actual comparison involves the intensity of the hidden defect feature signal after decoupling. The signal amplitude range differs before and after decoupling—the original signal contains substrate background and self-absorption components, resulting in larger values; after decoupling, these interferences are removed, leading to smaller values. Therefore, the noise floor value cannot be used directly; it needs to be multiplied by a mapping coefficient to transform the noise floor in the original domain to the decoupling domain. The mapping coefficient calibration method involves taking a silicon carbide epitaxial wafer containing various defects, recording the original acquired photoluminescence signal intensity value and the decoupling hidden defect feature signal intensity value at the same location, collecting at least one thousand points covering an intensity range from low to high. Linear regression is performed on this data set, with the decoupled intensity as the dependent variable y and the original intensity as the independent variable x, fitting the equation y = k*x + b. Since the intensity in the background region approaches zero after decoupling, the intercept b can be ignored; therefore, the mapping coefficient k is the slope of the fitted line. For example, during calibration, k = 0.35, indicating that the decoupled intensity is approximately 35% of the original intensity. A preset multiplier coefficient is used to control detection sensitivity, typically set between 3 and 5. A smaller coefficient results in a lower threshold, making it easier to detect weak defects but also increasing false positives; a larger coefficient results in a higher threshold, increasing the risk of missed detections. In engineering, 3 times is commonly used as the initial value because for Gaussian noise, 3 times the standard deviation corresponds to a false alarm rate of approximately one in a thousand. The final dynamic defect discrimination threshold baseline value = value at the corresponding position in the noise baseline distribution map × multiplier coefficient × mapping coefficient. For example: if the noise baseline estimate at a certain position is 50 (original domain), the multiplier coefficient is set to 3, and the mapping coefficient is 0.35, then the baseline value = 50 × 3 × 0.35 = 52.5. This means that only when the intensity of the hidden defect feature signal after decoupling reaches 52.5 or higher will it be considered a candidate defect point.

[0048] The intensity of the decoupled hidden defect feature signal, obtained through the preceding nonlinear separation model, is a pure numerical value. For each spatial location, the decoupled intensity at that location is taken and compared with the dynamic benchmark value calculated above. If the intensity is greater than or equal to the benchmark value, it indicates that the optical signal at that location is significantly higher than the local background noise, suggesting a high probability of a defect, and it is marked as a defect candidate point; if it is less than the benchmark value, the signal is considered insufficient, and it is marked as a non-defect point. Continuing with the example above, with a benchmark value of 52.5, if the decoupled feature signal intensity at that location is 68, which is greater than 52.5, it is marked as a defect candidate point; if the intensity is only 45, which is less than 52.5, it is marked as a non-defect point. This threshold changes with the spatial location—the benchmark value automatically increases when the noise floor is high at the wafer edge and automatically decreases when the noise floor is low in the flat central region, achieving dynamic adaptive discrimination.

[0049] Connectivity analysis is performed on adjacent defect candidate points to merge spatially connected candidate points into a single defect candidate region, and the boundary coordinates of this region are recorded. A single candidate point may be just a noise spike or part of a larger defect. Spatially adjacent (up, down, left, right, or diagonally adjacent) candidate points need to be merged into a continuous region. Eight-connectivity criterion is commonly used: if other candidate points exist in the eight directions surrounding a candidate point, they are considered to belong to the same defect. The entire wafer is scanned, and all connected candidate point clusters are extracted; each cluster is a defect candidate region. The bounding rectangle boundary coordinates (minimum row number, maximum row number, minimum column number, maximum column number) or polygon contour coordinates of this region are recorded. For example, on a two-dimensional signal distribution map, the four points with coordinates (100, 200), (100, 201), (101, 200), and (101, 201) are all candidate points and are mutually adjacent. They are merged into a single candidate region, with the boundary rows from 100 to 101 and the columns from 200 to 201. If there is another isolated candidate point at (200, 500) and there are no other candidate points around it, then it becomes a separate candidate region.

[0050] Calculate the arithmetic mean and maximum value of the characteristic signal intensities of all candidate points within each defect candidate region, and use these as the signal feature output for that defect candidate region. A defect candidate region may contain multiple candidate points, each with its own decoupled characteristic signal intensity. Statistically calculate the intensity values ​​of all candidate points within the region, calculate the arithmetic mean to obtain the average intensity, reflecting the overall signal level of the defect; and take the maximum value to obtain the peak intensity, reflecting the most severe local area of ​​the defect. These two values ​​are used as the signal feature output for the region in subsequent multimodal verification steps. For example, if a candidate region contains five candidate points with characteristic signal intensities of 68, 72, 65, 70, and 75 respectively, then the arithmetic mean = (68 + 72 + 65 + 70 + 75) / 5 = 70, and the maximum value = 75. The output result is the region boundary coordinates plus the average intensity of 70 and the maximum intensity of 75.

[0051] For the defect candidate region, multiple detection modes are sequentially activated from photoluminescence detection mode, Raman spectroscopy detection mode, and nonlinear optical detection mode, specifically including the following steps: The candidate defect regions are sorted according to their spatial coordinates to form a queue to be detected; After dynamic threshold discrimination and connected component analysis, several candidate defect regions have been marked on the wafer. Each region has its spatial coordinate boundaries (e.g., minimum row number, maximum row number, minimum column number, maximum column number). To improve detection efficiency and avoid repeated back-and-forth movement of the scanning head, these regions need to be arranged in order according to their spatial location. A raster scanning order is typically used: arranged by row number from smallest to largest, and within the same row by column number from smallest to largest, forming a queue to be detected. For example: Suppose three candidate defect regions are detected on the wafer. Region 1's center coordinates are 10.2 mm in the row direction and 20.5 mm in the column direction; Region 2's are 10.5 mm in the row direction and 10.1 mm in the column direction; and Region 3's are 10.3 mm in the row direction and 30.0 mm in the column direction. After sorting by row number, Region 2, with the smallest row number, is placed first, followed by Region 1 and Region 3 with the same row number (then arranged by column number from smallest to largest; Region 1's column number of 20.5 is smaller than Region 3's column number of 30.0, so the order is Region 2, Region 1, Region 3). In this way, during subsequent sequential detections, the scanning head can move from region two to region one and then to region three, resulting in the shortest path.

[0052] For each candidate defect region in the queue, the system sequentially employs three different detection methods for precise measurement to ensure sufficient information is obtained for final defect type confirmation. The photoluminescence detection mode is the most basic, illuminating the region with adaptively adjusted laser parameters and acquiring photoluminescence spectra in the wavelength range of 350 nm to 800 nm using a spectrometer, focusing on characteristic peaks in the 500-600 nm band. The Raman spectroscopy detection mode uses the same laser (typically switching the wavelength to 266 nm or 355 nm) to measure the Raman shift spectrum of the region. The characteristic Raman peaks of silicon carbide are located near 776 nm (transverse optical mode) and 796 nm (longitudinal optical mode), and defects alter the intensity or position of these peaks. The nonlinear optical detection mode uses a higher-intensity ultrashort pulse laser (e.g., pulse width less than 100 femtoseconds, wavelength 700 nm) to excite the region to generate a second harmonic signal (wavelength changed to 350 nm) or a two-photon excited fluorescence signal. The three modes are executed sequentially, with three sets of data acquired for each region before moving to the next.

[0053] For example, consider the first region in the current inspection queue (region two, previously sorted, with coordinates of 10.5 mm in the row direction and 10.1 mm in the column direction). The system first switches the laser parameters to photoluminescence detection mode, acquires the spectrum of this region, and records a distinct emission peak at 520 nm. Then, it switches to Raman detection mode, acquires the Raman shift spectrum, and finds that the peak intensity at 776 cm / s is 30% lower than that of the defect-free region. Finally, it switches to nonlinear optics mode, irradiates with a femtosecond laser, and detects that the second harmonic signal intensity is five times that of the background region. After completing these three acquisitions, the data is stored and bound to the coordinates of the region, and then the process is repeated for the next region (region one). Data from all three modes are indispensable for each defect candidate region.

[0054] Selecting the most sensitive excitation method for different types of hidden defects, and confirming the type, density, and location of the hidden defects, specifically includes the following steps: The data collected from the three detection modes are compared one by one with a pre-built standard library to calculate the matching degree of each mode for each defect type. The preset standard feature database is obtained through measurements of a large number of known defect samples. For example, photoluminescence spectra are collected at the location of a basal plane dislocation semi-ring confirmed by transmission electron microscopy, and the average curve of multiple measurements is used as the standard feature curve of the defect. Standard curves for shallow stacking faults, microtubules, triangular defects, etc., are established in the same way. For the photoluminescence detection mode, the signal feature curve is the relationship between wavelength and intensity; for the Raman shift spectrum, the signal feature curve is the relationship between Raman shift and intensity; for nonlinear optical signals, the signal feature curve can be the relationship between time delay and signal intensity or the relationship between wavelength and intensity. The matching degree calculation takes the photoluminescence mode as an example: assuming that the collected signal feature curve has N wavelength points (for example, one point is collected every nanometer from 500 nanometers to 600 nanometers, for a total of 101 points), the standard feature curve of the basal plane dislocation semi-ring in the database has a corresponding standard intensity value at the same 101 wavelength point. Subtract the standard intensity from the intensity collected at each wavelength point, take the absolute value, add up these 101 absolute values, and divide by 101 to obtain the mean absolute deviation. The preset maximum permissible deviation is a pre-defined constant, such as 50 intensity counts. Subtract (mean absolute deviation divided by the preset maximum permissible deviation) from one. If the mean absolute deviation is 10, the matching degree is 1 - 10 divided by 50 = 0.8; if the mean absolute deviation is 40, the matching degree is 1 - 40 divided by 50 = 0.2; if the mean absolute deviation exceeds 50, the matching degree becomes negative and is taken as zero. The closer the matching degree is to 1, the better the acquired signal matches the standard defect characteristics. This calculation is performed for each detection mode and each defect type.

[0055] For each candidate defect region, the defect type corresponding to the detection mode with the highest matching degree is selected as the final defect type for that region, and the sensitive excitation mode corresponding to this type is recorded as a recommended parameter for subsequent review. If the final defect type is inconsistent with the previous judgment result, the current final confirmation result shall prevail. Each candidate defect region has data from three detection modes, and each mode will calculate the matching degree of multiple defect types. All matching degrees are compared together, and the one with the highest value is identified as the final confirmed defect type. For example, in a certain region, the matching degree of the plane dislocation half-ring in photoluminescence mode is 0.85, and that of the shallow stacking fault is 0.3; in Raman mode, the matching degree of the plane dislocation half-ring is 0.9, and that of the shallow stacking fault is 0.2; in nonlinear optics mode, the matching degree of the plane dislocation half-ring is 0.75, and that of the shallow stacking fault is 0.4. The highest matching degree is 0.9 (plane dislocation half-ring in Raman mode), so the final defect type is the plane dislocation half-ring. Simultaneously record the excitation mode most sensitive to this defect type—for example, a basal plane dislocation half-ring responds best to excitation with a wavelength of 266 nm, high power, and a short pulse. Record this parameter for use in subsequent reviews. "Previous judgment result" refers to the preliminary judgment result obtained by comparing the average decay time with the reference decay time. Since the preliminary judgment only used photoluminescence decay time information, while the final confirmation here uses complete data from three detection modes, it is more accurate. Therefore, if the two are inconsistent, the current final confirmation result shall prevail. For example: the preliminary judgment is that the region is a shallow stacking fault, but multimodal matching degree calculation shows that the matching degree of the basal plane dislocation half-ring is significantly higher, then it is ultimately judged to be a basal plane dislocation half-ring.

[0056] The density of a defect type is obtained by counting the number of the same defect type within all candidate defect regions and dividing the result by the total area of ​​the batch of wafers. After confirming the types of all candidate defect regions, the number of regions with the same defect type is summed up. For example, in a batch of wafers, there are 30 plane dislocation half-rings and 15 shallow stacking faults. If the total area of ​​this batch of wafers is the effective area corresponding to a 6-inch diameter (approximately 176.7 square centimeters), then the density of plane dislocation half-rings is approximately 0.17 per square centimeter (30 divided by 176.7), and the density of shallow stacking faults is approximately 0.085 per square centimeter (15 divided by 176.7). The density values ​​are used to evaluate the quality level of the entire batch of wafers and as a basis for subsequent process adjustments.

[0057] The spatial coordinate boundaries, final defect type, and density data of each defect candidate region are summarized to output a confirmed defect information list. Finally, all detection results are compiled into a list, with each record containing the boundary coordinates of a defect candidate region (e.g., row direction from 10.2 mm to 10.5 mm, column direction from 20.1 mm to 20.4 mm), the final defect type of that region (e.g., plane dislocation semi-ring), and the statistical density of that type of defect across the entire wafer. This list can be directly output to the production management system for wafer grading, process feedback, or failure analysis. For example, the first entry in the output information list is "Region 1, coordinate boundaries (10.2—10.5 mm, 20.1—20.4 mm), type plane dislocation semi-ring, density 0.17 per square centimeter"; the second entry is "Region 2, coordinate boundaries (10.5—10.8 mm, 10.0—10.3 mm), type shallow stacking fault, density 0.085 per square centimeter."

[0058] The above algorithms or formulas are all dimensionless and numerical calculations, and the results are obtained by software simulation based on a large amount of collected data to obtain the most recent real-world results. The preset parameters are set by those skilled in the art according to the actual situation.

[0059] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0060] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0061] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0062] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for online defect detection during silicon carbide chip fabrication, characterized in that, Includes the following steps: Based on the quality of the real-time acquired photoluminescence signal, the wavelength, power, and pulse timing of the excitation laser are adaptively adjusted to excite the hidden defects in the silicon carbide chip to generate characteristic luminescence signals for resolution, thus obtaining the original photoluminescence signal. Time-resolved spectral acquisition of the original photoluminescence signal was performed, and the hidden defect signal was decoupled from the substrate background noise and self-absorption effect using a preset nonlinear separation model to extract the characteristic signal of the hidden defect. The signal distribution and noise floor are monitored in real time during the scanning process. The defect discrimination threshold is dynamically adjusted according to the characteristic signals of hidden defects to obtain the dynamically discriminated defect candidate region and the signal characteristics of the defect candidate region. For candidate defect regions, photoluminescence detection mode, Raman spectroscopy detection mode, and nonlinear optical detection mode are sequentially activated. The most sensitive excitation mode is selected for different types of hidden defects to confirm the type, density, and location of the hidden defects.

2. The online defect detection method in the silicon carbide chip fabrication process according to claim 1, characterized in that, Adaptive adjustment of the wavelength, power, and pulse timing of the excitation laser includes the following steps: The photoluminescence signal at the current scanning position is acquired in real time, and the signal-to-noise ratio of the signal is calculated by subtracting the common logarithm of the noise root mean square value from the common logarithm of the common logarithm of the signal peak intensity. At the same time, the defect response intensity is calculated by subtracting the background signal intensity of the adjacent defect-free area from the signal intensity at the current scanning position, and then dividing by the background signal intensity. The adjacent defect-free area refers to the area around the current position where the signal intensity is stable and there are no abnormal changes. When the signal-to-noise ratio (SNR) is lower than the first preset SNR threshold, or the defect response intensity is lower than the first preset response intensity threshold, the power of the excitation laser is increased step by step in a fixed increment of 5% of the current power value. After each increase, the SNR and defect response intensity are recalculated until the SNR reaches the second preset SNR threshold or the power reaches the preset maximum safe power. If the SNR still does not meet the standard after power adjustment, the wavelength of the excitation laser is switched to a preset backup wavelength of 266 nanometers. In pulse timing adjustment, the transient decay curve of the photoluminescence signal is acquired and fitted to obtain the carrier lifetime. The fitting method is as follows: with time as the independent variable and signal intensity as the dependent variable, the time value corresponding to the centroid of the area under the decay curve is calculated; then the laser pulse width is set to half of the carrier lifetime and the repetition frequency is set to one-third of the carrier lifetime, so that the excitation pulse matches the carrier lifetime of the hidden defect. The adjusted wavelength, power, and pulse timing parameters are used as the initial excitation conditions for the next scan position.

3. The online defect detection method in the silicon carbide chip fabrication process according to claim 2, characterized in that, Excite hidden defects in a silicon carbide chip to generate distinguishable characteristic light emission signals to obtain the original photoluminescence signal. The specific steps include: The surface of a silicon carbide chip is irradiated with adaptively adjusted laser parameters to excite electronic transitions in hidden defect regions. The photoluminescence signal intensity was recorded as a function of time after the laser pulse stopped. The luminescence intensity and attenuation characteristic parameters of the hidden defects were calculated. The luminescence intensity was calculated by taking the maximum intensity value on the curve. The attenuation characteristic parameter was the average attenuation time, which was calculated by integrating the product of time and signal intensity over a period of three times the time required for the signal to attenuate to one-eth of its peak value, starting from the moment the laser pulse stopped, and then dividing by the integral value of the signal intensity over the same time period. For a plane dislocation half-ring, the first average attenuation time was calculated; for a shallow stacking fault, the second average attenuation time was calculated. In the defect-free region of the silicon carbide chip, the reference decay time of the substrate band edge emission is obtained using the same laser parameters and calculation methods. The first average decay time is compared with the reference decay time: if the two are equal or the difference is less than the preset allowable deviation range, it is initially determined that the signal cannot be distinguished from the substrate background, the position is marked as the area to be re-examined and re-acquisition is triggered, and adaptive adjustment is re-executed after re-acquisition is triggered; if the two are not equal and the difference is greater than or equal to the preset allowable deviation range, it is initially confirmed that the hidden defect is a base plane dislocation half-ring. The second average decay time is compared with the reference decay time: if the two are equal or the difference is less than the preset allowable deviation range, it is initially determined that the signal cannot be distinguished from the substrate background, the position is marked as the area to be re-examined and re-acquisition is triggered, and adaptive adjustment is re-executed after re-acquisition is triggered; if the two are not equal and the difference is greater than or equal to the preset allowable deviation range, it is initially confirmed that the hidden defect is a shallow stacking fault. When the same location simultaneously meets the preliminary confirmation conditions for both a plane dislocation half-loop and a shallow stacking fault, the defect is initially identified as a plane dislocation half-loop according to the preset priority rules. Since a plane dislocation half-loop contains stacking faults in its structure and poses a greater threat to device reliability, it is not counted as a shallow stacking fault again. The photon count within the emission band of the hidden defect is extracted using a preset bandpass filter, and the emission intensity is obtained by dividing the photon count by the acquisition time. The calculated luminescence intensity and the first or second average decay time corresponding to the preliminary judgment result are correlated with the current spatial coordinates to generate the original photoluminescence signal containing the luminescence intensity and decay characteristics of the hidden defects.

4. The online defect detection method in the silicon carbide chip fabrication process according to claim 3, characterized in that, The time-resolved spectral acquisition of the original photoluminescence signal includes the following steps: At the same time the excitation laser pulse stops, the recorded continuous time change curve is sampled at a preset fixed time interval. The total sampling time is set to the time required for the signal intensity to decay to one percent of its peak value. The sampled discrete signal intensity values ​​are arranged in chronological order to form the original time series data. The original time series data is divided into a predetermined number of continuous and non-overlapping time windows, with each time window containing multiple sampling points; The arithmetic mean of the intensity values ​​of all sampling points within each time window is calculated to obtain the average signal intensity of that time window. The average signal intensities of all time windows are then arranged in chronological order to generate discrete time-resolved spectral data.

5. The online defect detection method in the silicon carbide chip fabrication process according to claim 4, characterized in that, The hidden defect signal is decoupled from the substrate background noise and self-absorption effect using a pre-defined nonlinear separation model, and the characteristic signal of the hidden defect is extracted. The specific steps include: Discrete-time resolved spectral data is used as input, and a preset nonlinear separation model is invoked. This model uses an independent component analysis algorithm and is trained in advance using mixed signal samples containing known hidden defect signals, substrate background noise and self-absorption effects to obtain three independent signal component basis functions. The input discrete time-resolved spectral data is decomposed into three signal components, where the first component corresponds to the substrate background noise, the second component corresponds to the signal distortion caused by the self-absorption effect, and the third component corresponds to the characteristic signal of the hidden defect. The amplitude of the third component in each time window is calculated, and the amplitude sequence is correlated with the spatial coordinates acquired simultaneously to obtain the temporal and spatial distribution of the hidden defect feature signal. Subtracting the first and second components from the discrete-time resolved spectral data yields the decoupled hidden defect feature signal.

6. The online defect detection method in the silicon carbide chip fabrication process according to claim 5, characterized in that, Real-time monitoring of signal distribution and noise floor during the scanning process includes the following steps: Within the current scan line or scan frame, continuously acquire the original photoluminescence signal intensity values ​​corresponding to each spatial location to form a one-dimensional or two-dimensional signal distribution map; In the signal distribution map, the entire map is traversed with a preset sliding window size. The minimum signal strength is calculated in each window, and the minimum value is used as the local noise basis estimate for that window. Arrange the local noise base estimates of all windows according to their spatial location to form a noise base distribution map; The noise floor distribution map is updated in real time. For each new raw signal strength value collected, the noise floor of the window in which it is located is recalculated, and old data that has exceeded the preset historical time period is removed.

7. The online defect detection method in the silicon carbide chip fabrication process according to claim 6, characterized in that, The defect discrimination threshold is dynamically adjusted based on the characteristic signals of hidden defects to obtain dynamically discriminated defect candidate regions and their signal characteristics. This process includes the following steps: Multiply the value at the corresponding position in the noise floor distribution map by a preset multiplier coefficient, and then multiply by a preset mapping coefficient to obtain the benchmark value of the dynamic defect discrimination threshold applicable to the hidden defect feature signal. The intensity of the decoupled hidden defect feature signal is compared with the benchmark value of the dynamic defect discrimination threshold: if the feature signal intensity is greater than or equal to the benchmark value, the spatial location is marked as a defect candidate point; if it is less than the benchmark value, it is marked as a non-defect point. Connectivity analysis is performed on adjacent defect candidate points to merge spatially connected candidate points into a single defect candidate region, and the boundary coordinates of this region are recorded. Calculate the arithmetic mean and maximum value of the characteristic signal intensity of all candidate points within each defect candidate region, and use them as the signal feature output of that defect candidate region.

8. The online defect detection method in the silicon carbide chip fabrication process according to claim 7, characterized in that, For the defect candidate region, multiple detection modes are sequentially activated from photoluminescence detection mode, Raman spectroscopy detection mode, and nonlinear optical detection mode, specifically including the following steps: The candidate defect regions are sorted according to their spatial coordinates to form a queue to be detected; For each candidate defect region in the queue, perform the following operations in sequence: first, enable the photoluminescence detection mode and collect the photoluminescence spectrum of the region; then, enable the Raman spectroscopy detection mode and collect the Raman shift spectrum of the region; finally, enable the nonlinear optical detection mode and collect the second harmonic or two-photon excitation signal of the region. For each candidate defect region, data from all three detection modes must be collected.

9. The online defect detection method in the silicon carbide chip fabrication process according to claim 8, characterized in that, Selecting the most sensitive excitation method for different types of hidden defects, and confirming the type, density, and location of the hidden defects, specifically includes the following steps: Photoluminescence spectrum, Raman shift spectrum, and nonlinear optical signal are compared with a pre-defined standard feature database of various hidden defects. The matching degree is calculated as follows: for each detection mode, the acquired signal feature curve is subtracted point by point from the standard feature curve of the corresponding defect type in the database. The absolute values ​​of the differences at each point are summed and divided by the total number of points of the standard feature curve to obtain the average absolute deviation. Then, the ratio of the average absolute deviation to the pre-defined maximum allowable deviation is subtracted from one to obtain the matching degree between the signal and the defect type under the detection mode. The matching degree value is between zero and one. For each defect candidate region, the defect type corresponding to the detection mode with the highest matching degree is selected as the final defect type of the region, and the sensitive activation mode corresponding to the type is recorded as the recommended parameter for subsequent review; if the final defect type is inconsistent with the previous judgment result, the current final confirmation result shall prevail. The density of the same defect type is obtained by counting the number of the same defect type in all defect candidate regions and dividing it by the total area of ​​the batch of wafers. The spatial coordinate boundaries, final defect type, and density data of each defect candidate region are summarized, and a list of confirmed defect information is output.

Images