Full-automatic wafer film thickness detection method and system based on spectral reflection

Abstract

The invention relates to the related technical field of semiconductor wafers, in particular to a full-automatic wafer film thickness detection method and system based on spectral reflection, and the method comprises the steps: obtaining a wafer surface reflection spectral signal, and extracting spectral feature information and wavelength parameters; identifying a polarization characteristic parameter and an interface interference characteristic parameter, introducing the wavelength parameter as an adjusting variable, determining a wavelength dependence term, and generating a global film thickness distribution map, thereby solving the problems that in a detection scene with a multi-layer film or a rough interface, the influence of an incident light polarization state on a reflection signal is not fully considered, the reflection signal characteristic extraction is distorted, and the detection accuracy is poor. According to the method, the technical problem that the film thickness obtained through reflection characteristic backstepping deviates from the actual film thickness is solved, polarization characteristic parameters and interface interference characteristic parameters are recognized, the thickness of each detection point is determined after the film refractive index is corrected, the thickness detection precision is effectively improved, and the influence of spectral signal noise and interface interference is reduced; the technical effect of accurately presenting the global film thickness distribution of the wafer film is achieved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor wafer technology, specifically to a fully automated wafer film thickness detection method and system based on spectral reflectance. Background Technology

[0002] In the semiconductor manufacturing field, the uniformity of wafer thin film thickness directly affects the electrical performance and yield of chips. As chip manufacturing processes break through to the micro-nano level, the requirements for detection accuracy, efficiency and automation continue to increase. Spectral reflectance refers to inferring the thickness by analyzing the reflection characteristics of thin films to different wavelengths of light.

[0003] Current wafer film thickness detection based on spectral reflection does not fully consider the influence of the polarization state of the incident light on the reflected signal and ignores the correlation between the interlayer interference characteristics and wavelength. As a result, the detection results are easily interfered with in multilayer thin films or high-roughness interface scenarios. In addition, factors such as spectral signal noise and equipment power supply ripple can introduce systematic errors during the detection process, making it difficult to achieve accurate characterization of the film thickness across the entire domain through a single spectral feature. The generated film thickness distribution map often has local deviations, which increases the risk of yield loss caused by inaccurate film thickness detection in chip production.

[0004] In summary, existing technologies suffer from several technical problems in detection scenarios involving multilayer thin films or rough interfaces. These problems include insufficient consideration of the influence of incident light polarization on the reflected signal, resulting in distorted feature extraction of the reflected signal and discrepancies between the film thickness derived from the reflection characteristics and the actual film thickness. Summary of the Invention

[0005] This application provides a fully automated wafer film thickness detection method and system based on spectral reflectance, aiming to solve the technical problems in the prior art where, in detection scenarios with multilayer thin films or rough interfaces, the influence of the polarization state of the incident light on the reflected signal is not fully considered, resulting in distortion of the reflected signal feature extraction and deviation between the film thickness obtained by back-calculation of the reflection characteristics and the actual film thickness.

[0006] In view of the above problems, the technical solution to achieve the present application is as follows: In a first aspect, this application provides a fully automated wafer film thickness detection method based on spectral reflectance, wherein the method includes: emitting incident light to a wafer thin film using a laser source, simultaneously acquiring the wafer surface reflectance spectral signal, and extracting spectral feature information and wavelength parameters; identifying polarization feature parameters in the wafer surface reflectance spectral signal based on the polarization state of the incident light and the wavelength parameters; identifying interface interference feature parameters in the wafer surface reflectance spectral signal based on the interlayer interface of the wafer thin film and the spectral feature information; introducing the wavelength parameters as an adjustment variable based on the polarization feature parameters and the interface interference feature parameters in the wafer surface reflectance spectral signal to determine a wavelength dependence term; and determining the thickness value of each detection point of the wafer thin film according to the thin film refractive index corrected by the wavelength dependence term, thereby generating a global film thickness distribution map.

[0007] Preferably, the spectral feature information includes reflection peak intensity, reflection valley depth, characteristic peak position, and spectral half-width, and the wavelength parameter covers the tunable wavelength range of the laser source.

[0008] Preferably, the polarization characteristic parameters include s-polarized light reflectivity, p-polarized light reflectivity, and polarization state extinction ratio. The s-polarized light reflectivity and p-polarized light reflectivity are determined by acquiring the reflected light intensity in the corresponding polarization direction using a polarization beam splitter, and the polarization state extinction ratio is the ratio of s-polarized light reflectivity to p-polarized light reflectivity.

[0009] Preferably, the interface interference characteristic parameters include interlayer reflection peak position shift, interface scattering intensity factor, and interference fringe contrast. The interlayer reflection peak position shift is determined by the difference in the position of the spectral peaks formed by the reflected light from the interfaces of different thin film layers. The interface scattering intensity factor is positively correlated with the interface roughness and is determined by conversion through the baseline drift of the reflection spectral signal. The interference fringe contrast is the ratio of the difference between the reflection peak intensity and the reflection valley depth.

[0010] Preferably, based on the wavelength-dependent corrected thin film refractive index, a joint control analysis is performed with the filter array corresponding to the beam splitter under the grating + prism combination to set a full-band spectral feature matrix; the matching degree of the full-band spectral feature matrix is ​​introduced, and there is only one and only one condition that the multi-mode denoising network converges to the global optimal solution when the matching degree meets the standard and the convergence termination condition of the mean square error loss function is satisfied.

[0011] Preferably, the noise suppression of the wafer surface reflectance spectral signal adopts a Transformer-based multimodal denoising network: the original spectral signal and the wafer surface topography image are used as dual-modal inputs, and time-series modeling is performed by a Transformer encoder. At the same time, the edge features and texture features of the wafer surface topography image are extracted by a convolutional neural network, and the dual-modal features are concatenated and input into a fully connected layer. The voltage ripple coefficient is determined using the voltage closed-loop feedback data of the Cuk transformer. The correlation between the voltage ripple coefficient and the mean square error loss function is analyzed, and the convergence termination condition of the mean square error loss function is set with the spectral peak position shift error and reflectance measurement error.

[0012] Preferably, the voltage ripple coefficient is determined using the voltage closed-loop feedback data of the Cuk converter, and the dynamic adjustment threshold of the Cuk converter is set: when the spectral peak position shift error is detected to exceed the preset threshold, the duty cycle adjustment of the Cuk converter is automatically triggered, and the suppression of high-frequency ripple is enhanced by reducing the cutoff frequency of the LC filter network, so that the voltage ripple coefficient is controlled within the ripple tolerance threshold, and the corresponding reflection peak position measurement error is corrected to the peak position accuracy reference.

[0013] Preferably, for systematic errors in multi-batch wafer inspection, the long-term voltage stability parameter of the Cuk converter is incorporated into the error tracing system. When the inspection error of any batch exhibits a systematic shift, the influence weight of power supply stability on the thin film refractive index correction term is determined by comparing the inspection data in different voltage drift intervals, and the calibration coefficient of the wavelength-dependent term is optimized. A dual error compensation mechanism is set up through the power supply stability control of the Cuk converter and the noise suppression of the multi-mode denoising network.

[0014] Preferably, based on the local features of the global film thickness distribution map and combined with the anomalies of the wafer surface reflection spectrum signal, defect candidate regions are identified, and a defect classifier is trained. The inputs are the film thickness gradient and spectral absorption peak shift of the defect candidate region, and the output is the defect type. The average thickness, thickness standard deviation, and maximum thickness deviation of the wafer thin film are determined through the global film thickness distribution map. Defect classification and location are marked in combination with the defect candidate regions and the corresponding defect types, and visual reminders are provided by the defect classification and location marking.

[0015] In a second aspect, this application provides a fully automated wafer film thickness detection system based on spectral reflectance, wherein the system comprises: a feature extraction module: emitting incident light to a wafer thin film using a laser source, simultaneously acquiring the wafer surface reflectance spectral signal, and extracting spectral feature information and wavelength parameters; a feature parameter identification module: identifying polarization feature parameters in the wafer surface reflectance spectral signal based on the polarization state of the incident light and the wavelength parameters; identifying interface interference feature parameters in the wafer surface reflectance spectral signal based on the interlayer interface of the wafer thin film and the spectral feature information; a wavelength dependence term determination module: determining a wavelength dependence term by introducing the wavelength parameters as an adjustment variable based on the polarization feature parameters and the interface interference feature parameters in the wafer surface reflectance spectral signal; and a thickness value determination module: determining the thickness value of each detection point of the wafer thin film according to the thin film refractive index corrected by the wavelength dependence term, and generating a global film thickness distribution map.

[0016] In summary, one or more technical solutions provided in this application extract spectral features and wavelength parameters, combine the polarization state of the incident light and the interlayer characteristics of the thin film, identify polarization feature parameters and interfacial interference feature parameters, introduce wavelength parameters as adjustment variables to determine wavelength dependence, and determine the thickness of each detection point after correcting the refractive index of the thin film. This effectively improves the accuracy of thickness detection, reduces the influence of spectral signal noise and interfacial interference, and achieves the technical effect of accurately presenting the global thickness distribution of wafer thin films. Attached Figure Description

[0017] Figure 1 This application provides a flowchart of a fully automated wafer film thickness detection method based on spectral reflectance.

[0018] Figure 2 This application provides a schematic diagram of the structure of a fully automated wafer film thickness detection system based on spectral reflectance.

[0019] Explanation of reference numerals in the attached figures: Feature extraction module M100, Feature parameter recognition module M200, Wavelength dependence determination module M300, Thickness value determination module M400. Detailed Implementation

[0020] Example 1: The present application will be described in detail below with reference to the accompanying drawings, as follows... Figure 1 As shown, this application provides a fully automated wafer film thickness detection method based on spectral reflectance, wherein the method includes: S1: Use a laser light source to emit incident light onto the wafer thin film, simultaneously acquire the reflection spectrum signal of the wafer surface, and extract spectral feature information and wavelength parameters.

[0021] Specifically, laser light sources have the characteristics of good monochromaticity and strong directionality, making them suitable as incident light sources for spectral reflectance detection. They can ensure the stability of the incident light wavelength and the concentration of energy, reducing interference to the reflected signal caused by the instability of the light source characteristics. Simultaneous acquisition of the wafer surface reflectance spectrum signal refers to the real-time acquisition of the light signal reflected back from the wafer thin film surface through devices such as spectral detectors at the same time as the laser is incident. This includes spectral information related to the film thickness. Spectral feature information is the key indicators that reflect the characteristics of the thin film extracted from the reflectance spectrum signal, such as the intensity of the reflection peak and the depth of the reflection valley. Wavelength parameters refer to the wavelength range and specific wavelength values ​​covered by the incident and reflected light.

[0022] Execution steps: An incident light is emitted onto the wafer thin film using a laser source, for example, a tunable laser source with a wavelength range of 300nm-1000nm, ensuring coverage of the key wavelength range required for wafer thin film detection. When the incident light illuminates the wafer thin film surface, a spectrometer and other detection equipment are simultaneously activated to acquire the reflection spectrum signal. Spectral characteristic information such as reflection peak intensity, reflection valley depth, characteristic peak position, and full width at half maximum (FWHM) is extracted from the reflection spectrum signal. Simultaneously, the tunable wavelength range of the laser source is recorded as a wavelength parameter; commonly, the tunable wavelength range of the laser source is 300nm-1000nm. By acquiring the original reflection spectrum signal and key parameters, data support is provided for subsequent characteristic parameter identification and thickness calculation.

[0023] S2: Based on the polarization state of the incident light and in conjunction with the wavelength parameter, identify the polarization characteristic parameters in the reflection spectrum signal of the wafer surface; based on the interlayer interface of the wafer thin film and in conjunction with the spectral characteristic information, identify the interface interference characteristic parameters in the reflection spectrum signal of the wafer surface.

[0024] Specifically, the polarization state of incident light refers to the vibration direction characteristics of the incident light, such as linear polarization and circular polarization; polarization characteristic parameters are parameters that reflect the polarization characteristics of the reflected spectral signal, including s-polarized light reflectivity, p-polarized light reflectivity, and polarization state extinction ratio; thin film interlayer interface refers to the interface between different layers of a wafer thin film; interface interference characteristic parameters are parameters that reflect the interference characteristics of reflected light at the thin film interlayer interface, such as interlayer reflection peak position shift, interface scattering intensity factor, and interference fringe contrast.

[0025] Execution steps: When identifying polarization characteristic parameters, based on the known polarization state of the incident light, such as linearly polarized light, its vibration direction is 90° to the incident plane (s-polarization) and parallel to the incident plane (p-polarization). Combined with the extracted wavelength parameters, the intensity of reflected light in the s-polarization direction and the p-polarization direction are collected by a polarization beam splitter. For example, at a wavelength of 500nm, the reflection intensity of s-polarized light is 0.6 and the reflection intensity of p-polarized light is 0.3. Then, the reflectivity of s-polarized light is calculated to be 0.6 and the reflectivity of p-polarized light is 0.3, and the extinction ratio of polarization state is 0.6 / 0.3=2.

[0026] When identifying interface interference characteristic parameters, based on the interlayer interface of the wafer thin film, such as the interface between two thin films, and combined with spectral characteristic information, the interlayer reflection peak position shift is determined. For example, if the reflection peak position of the first layer interface is at 520nm and the reflection peak position of the second layer interface is at 530nm, the shift is 10nm. The interface scattering intensity factor is obtained by converting the baseline drift of the reflection spectrum signal, and the interference fringe contrast is determined. For example, if the reflection peak intensity is 0.7 and the reflection valley depth is 0.3, the contrast is (0.7-0.3) / 0.7. By deeply exploring the polarization characteristics and interface interference characteristics in the reflection spectrum signal, the limitations of conventional methods that rely only on a single spectral feature are overcome. This provides a more comprehensive and accurate characteristic basis for subsequent refractive index correction and thickness calculation, effectively reducing the detection deviation caused by not considering polarization and interface interference factors, and improving the adaptability to complex thin film structures.

[0027] S3: Based on the polarization characteristic parameters and interface interference characteristic parameters in the wafer surface reflection spectrum signal, the wavelength parameter is introduced as an adjustment variable to determine the wavelength dependence term; S4: Based on the thin film refractive index corrected by the wavelength dependence term, the thickness value of each detection point of the wafer thin film is determined to generate a global film thickness distribution map.

[0028] Specifically, wavelength dependence refers to the regular relationship between the physical properties of the thin film (such as refractive index) and the wavelength. By integrating polarization characteristic parameters (such as s / p polarized light reflectivity and polarization state extinction ratio) and interface interference characteristic parameters (such as interlayer reflection peak position shift and interference fringe contrast), and using wavelength parameters as dynamic adjustment variables, the correlation between these characteristic parameters and wavelength is established, thereby determining the wavelength dependence. The thin film refractive index is a key parameter for calculating the thin film thickness. Since it changes with wavelength, correcting it through wavelength dependence can improve accuracy. The global film thickness distribution map presents the thickness values ​​of all detection points on the wafer in a visual way, intuitively reflecting the overall distribution of the film thickness.

[0029] Execution steps: When determining wavelength-dependent terms, the identified polarization characteristic parameters and interface interference characteristic parameters are integrated as adjustment variables. For polarization characteristic parameters, such as the linear increase in reflectivity of s-polarized light from 0.5 to 0.7 and the increase in reflectivity of p-polarized light from 0.3 to 0.5 within the wavelength range of 400nm-800nm, the polarization extinction ratio remains stable at 1.67. For interface interference characteristic parameters, such as the increase in interlayer reflection peak position shift from 5nm to 15nm with increasing wavelength, and the increase in interference fringe contrast from 0.4 to 0.6. By using wavelength parameters (400nm-800nm) as adjustment variables and establishing mathematical models through algorithms such as multiple regression analysis, parameter fitting can be completed in a short time, ensuring efficiency.

[0030] When determining the thickness value and generating the spectrum, the refractive index of the thin film is corrected using a wavelength-dependent term. For example, at a wavelength of 500 nm, the uncorrected refractive index is 1.45, and the corrected refractive index is 1.47. Combining the thin film thickness calculation formula in the principle of spectral reflection, and further based on the interference equation: thickness = (λ × phase difference) / (4π × corrected refractive index), the thickness value of each 10 μm × 10 μm detection point on the wafer is obtained. The thickness data of several detection points are plotted into a global film thickness distribution spectrum using image processing technology. By establishing a wavelength-dependent term to correct the refractive index, the calculation deviation caused by the fixed refractive index in conventional methods is solved. The global film thickness distribution spectrum effectively improves the detection accuracy and global characterization capability of complex thin film structures.

[0031] Furthermore, the method for extracting spectral feature information and wavelength parameters in this application includes: The spectral feature information includes reflection peak intensity, reflection valley depth, characteristic peak position, and spectral half-width at half maximum (FWHM). The wavelength parameters cover the tunable wavelength range of the laser source.

[0032] Specifically, spectral feature information is a key feature index extracted from the reflection spectrum signal of the wafer surface. Among them, the reflection peak intensity refers to the maximum value of the reflected light intensity in the spectrum, reflecting the reflectivity of light at a specific wavelength; the reflection valley depth is the minimum value of the reflected light intensity in the spectrum, corresponding to the reflection peak intensity, and together they reflect the fluctuation characteristics of the spectrum; the characteristic peak position is the wavelength position corresponding to the reflection peak, which is an important reference point for analyzing the optical properties of thin films; the full width at half maximum (FWHM) of the spectrum is the wavelength width of the reflection peak at half the peak intensity, used to describe the width of the peak, and is related to the uniformity and other properties of the thin film; the wavelength parameter refers to the adjustable wavelength range of the laser source and the specific wavelength values ​​within that range. Covering the tunable wavelength range of the laser source means that the extracted wavelength parameters completely include all wavelength ranges that the source can output, ensuring that subsequent analysis can be carried out based on all available wavelength information of the source.

[0033] Execution steps: When the incident light emitted by the laser source illuminates the wafer thin film, the spectral detection equipment simultaneously acquires the reflection spectrum signal. The signal is analyzed using a signal processing algorithm to extract the aforementioned spectral feature information. Simultaneously, the tunable wavelength range of the laser source is recorded, for example, using a tunable wavelength range of 300nm-1000nm as the wavelength parameter, providing basic data support for subsequent feature parameter identification. Spectral feature information such as reflection peak intensity and reflection valley depth are directly related to the optical properties of the thin film. The position of the feature peak is related to the film thickness, while the full width at half maximum (FWHM) of the spectrum reflects the quality state of the thin film. The wavelength parameter covers the tunable range of the light source, ensuring that subsequent analysis based on complete wavelength information can be performed when combining polarization state and thin film interlayer interface identification of feature parameters. This avoids incomplete feature extraction due to missing wavelength information, which would affect detection accuracy. This is the initial data acquisition and extraction step for realizing a multi-feature fusion detection system.

[0034] Furthermore, the method of this application also includes: The polarization characteristic parameters include s-polarized light reflectivity, p-polarized light reflectivity, and polarization state extinction ratio. The s-polarized light reflectivity and p-polarized light reflectivity are determined by acquiring the reflected light intensity in the corresponding polarization direction using a polarization beam splitter. The polarization state extinction ratio is the ratio of s-polarized light reflectivity to p-polarized light reflectivity.

[0035] Specifically, polarization characteristic parameters are key indicators used to describe the polarization characteristics of the reflected spectral signals on a wafer surface. Among them, s-polarized light reflectivity refers to the ratio of the intensity of reflected light from s-polarized light with its vibration direction perpendicular to the incident plane to the intensity of incident light on the wafer surface, while p-polarized light reflectivity is the ratio of the intensity of reflected light from p-polarized light with its vibration direction parallel to the incident plane to the intensity of incident light. The polarization state extinction ratio is the ratio of s-polarized light reflectivity to p-polarized light reflectivity, used to quantify the difference in reflection characteristics between the two polarization states. A polarization beam splitter is an optical element that can separate light with different polarization directions. It is used to collect the reflection intensities of s-polarized light and p-polarized light separately, thereby obtaining the corresponding reflectivity.

[0036] Execution steps: When incident light (containing both s and p polarization states) irradiates the surface of the wafer thin film, the reflected light carrying polarization information propagates to a polarization beamsplitter. The beamsplitter separates the s-polarized and p-polarized light in the reflected light, allowing each to enter its corresponding detector. For example, if the incident light intensity is I0, after passing through the polarization beamsplitter, the detected s-polarized light reflection intensity is Is = 0.6I0, so the s-polarized light reflectivity is 0.6I0 / I0 = 0.6; the detected p-polarized light reflection intensity is Ip = 0.3I0, so the p-polarized light reflectivity is 0.3I0 / I0 = 0.3. The extinction ratio of the polarization state is 0.6 / 0.3=2. In the above steps, the polarization characteristics in the reflection spectrum are accurately extracted, and the reflection light of different polarization states is separated and collected by a polarization beam splitter. This provides data for subsequent analysis of the correlation between polarization characteristics and film thickness by combining wavelength parameters. Furthermore, in multilayer thin films or high-roughness interface scenarios, the reflection characteristics of s-polarized light and p-polarized light are significantly different. These parameters can effectively reflect the optical anisotropy and interface characteristics of the thin film, making up for the feature extraction distortion caused by polarization influence, and laying a key polarization characteristic foundation for improving the accuracy of film thickness detection.

[0037] Furthermore, the method of this application also includes: The interface interference characteristic parameters include interlayer reflection peak position shift, interface scattering intensity factor, and interference fringe contrast. The interlayer reflection peak position shift is determined by the difference in the position of the spectral peaks formed by the reflected light from the interfaces of different thin film layers. The interface scattering intensity factor is positively correlated with the interface roughness and is determined by conversion through the baseline drift of the reflection spectral signal. The interference fringe contrast is the percentage difference between the reflection peak intensity and the reflection valley depth.

[0038] Specifically, interface interference characteristic parameters are key parameters reflecting the interference characteristics of reflected light at different interlayer interfaces of wafer thin films. Among them, the interlayer reflection peak position shift refers to the wavelength position difference between the characteristic peaks formed in the spectrum by the reflected light from different thin film layers, which directly reflects the interference difference of the reflected light from each layer interface; the interface scattering intensity factor is an indicator used to quantify the degree of scattering at the interlayer interface of the thin film, and its value is positively correlated with the interface roughness, that is, the rougher the interface, the larger the value of the factor, which can be obtained by converting the baseline drift in the reflection spectrum signal; the interference fringe contrast is the ratio of the difference between the reflection peak intensity and the reflection valley depth to the reflection peak intensity (or to the reflection valley depth), which is used to describe the clarity of the interference fringes. The higher the contrast, the more obvious the interference characteristics.

[0039] Execution steps: When a laser is incident on a multilayer wafer thin film, the interfaces between different layers reflect the light. The reflected light from each interface forms corresponding reflection peaks in the spectrum. By analyzing the spectral signal, the wavelength positions of these peaks can be determined. For example, if the reflection peak of the first layer interface is located at 450nm and the reflection peak of the second layer interface is located at 460nm, then the interlayer reflection peak position offset is 10nm. This parameter can reflect the difference in optical path between different layer interfaces and provide a basis for analyzing the interlayer thickness relationship.

[0040] For the interface scattering intensity factor, if the baseline of the reflected spectral signal drifts by 0.1 due to interface scattering, the factor can be calculated as 0.05 using a preset conversion relationship, such as scattering intensity factor = 0.5 × baseline drift. As interface roughness increases, the baseline drift increases, and the factor value also rises, thus indirectly characterizing the interface roughness. The calculation of interference fringe contrast is based on the extracted reflection peak intensity and reflection valley depth. Assuming the reflection peak intensity is 0.8 and the reflection valley depth is 0.2, the difference is 0.6. Using the reflection peak intensity as a benchmark, its proportion is 0.6 / 0.8 = 0.75. The higher this value, the more significant the interference phenomenon, and the more beneficial it is to infer the film thickness through interference features. Preferably, in scenarios with multilayer films or rough interfaces, fully exploring the optical properties of the interlayer interfaces to determine interface interference details provides crucial interface feature basis for subsequent refractive index correction by combining spectral feature information and wavelength parameters. This reduces detection deviations caused by interface interference and significantly improves the detection adaptability of complex film structures.

[0041] Furthermore, based on the thin film refractive index corrected for the wavelength dependence term, the thickness value of each detection point of the wafer thin film is determined. The method of this application also includes: Based on the wavelength-dependent corrected thin film refractive index, a joint control analysis is performed with the filter array corresponding to the beam splitter under the grating + prism combination to set a full-band spectral feature matrix; the matching degree of the full-band spectral feature matrix is ​​introduced, and there is only one and only one. When the matching degree meets the standard and the convergence termination condition of the mean square error loss function is satisfied, the multi-mode denoising network converges to the global optimal solution.

[0042] Specifically, the wavelength-dependent corrected thin film refractive index refers to the refractive index value obtained after correcting the original thin film refractive index by using the wavelength parameter as an adjustment variable. This value is dynamically adjusted with the wavelength, which is more in line with the actual optical characteristics. The beam splitter in the combination of grating and prism is an optical component that decomposes composite light into monochromatic light of different wavelengths. The grating splits light through diffraction, and the prism splits light through refraction. The combination of the two can improve the beam splitting accuracy and wavelength coverage. The filter array consists of multiple filters of different wavelengths, which are used to selectively transmit light of specific wavelengths. It works with the beam splitter to achieve fine screening of the spectrum.

[0043] The full-band spectral feature matrix is ​​a data structure that integrates spectral feature information (such as reflection peak intensity and reflection valley depth at each wavelength) across the entire spectral range in matrix form, used to comprehensively characterize spectral properties. The matching degree of the full-band spectral feature matrix refers to the similarity between the actual acquired spectral feature matrix and the preset standard spectral feature matrix; a higher matching degree indicates more reliable spectral data. The convergence termination condition of the mean square error loss function refers to the condition under which the network stops training when the mean square error loss value of the multimodal denoising network decreases below a preset threshold and no longer changes significantly, used to ensure that the network's denoising effect reaches the expected level. The multimodal denoising network is a neural network that can simultaneously process multiple types of input data (such as spectral signals and image signals) to suppress noise; convergence to the global optimum indicates that the network model has achieved the best denoising effect.

[0044] Execution steps: The wavelength-dependent corrected thin film refractive index is linked with the beam-splitting component of the grating + prism combination, for example, the corrected refractive index value at 10nm intervals within the wavelength range of 400nm-1000nm; the beam-splitting component performs beam splitting on the reflection spectrum, and in conjunction with the filter array, filters are used to filter light of different wavelengths, such as filters containing wavelengths of 400nm, 450nm, ..., 1000nm; through joint control and analysis, the spectral feature information corresponding to each wavelength in the entire band is integrated into a full-band spectral feature matrix, for example, a 100×5 matrix, where 100 represents 100 wavelength points and 5 represents the corresponding 5 spectral features; the matching degree between the full-band spectral feature matrix and the standard matrix is ​​determined. If the matching degree reaches a preset threshold (e.g., 90%), and the mean square error loss function value of the multimodal denoising network drops below 0.001 (satisfying the convergence termination condition), then the network converges to the global optimal solution.

[0045] By combining the analysis of the beam splitter and the filter array, the accuracy and comprehensiveness of the spectral feature matrix are improved. The full-band spectral feature matrix provides richer spectral basis for subsequent thickness calculation. The introduction of convergence conditions for matching degree and mean square error loss function ensures that the multi-modal denoising network achieves the best denoising effect, reduces the interference of spectral signal noise on the detection results, improves the signal-to-noise ratio of the spectral signal, and thus improves the matching degree compliance rate of the full-band spectral feature matrix, laying a high-quality data foundation for the subsequent accurate determination of wafer thin film thickness.

[0046] Furthermore, the method of this application also includes: The noise suppression of the wafer surface reflectance spectral signal adopts a Transformer-based multimodal denoising network: the original spectral signal and the wafer surface topography image are used as dual-modal inputs, and time-series modeling is performed by a Transformer encoder. At the same time, edge features and texture features of the wafer surface topography image are extracted by a convolutional neural network, and the dual-modal features are concatenated and input into a fully connected layer. The voltage ripple coefficient is determined using the voltage closed-loop feedback data of the Cuk transformer. The correlation between the voltage ripple coefficient and the mean square error loss function is analyzed. The convergence termination condition of the mean square error loss function is set based on the spectral peak position shift error and reflectance measurement error.

[0047] Specifically, the Transformer-based multimodal denoising network is a noise suppression model that integrates the Transformer architecture and multimodal data processing capabilities. Multimodal processing refers to simultaneously handling two different types of input: the original spectral signal and the wafer surface topography image. The Transformer encoder models the temporal data (such as the sequence of spectral signals varying with wavelength) using a self-attention mechanism, capturing long-range dependencies in the data. The convolutional neural network excels at extracting local image features, used to extract edge and texture features, including the contours of thin film layers, from the wafer surface topography image. Dual-modal feature concatenation merges the spectral features processed by the Transformer encoder with the image features extracted by the convolutional neural network into a unified feature vector, achieving the fusion of multi-source information. Fully connected layers are layers in the neural network that connect each neuron to all neurons in the previous layer, used to perform nonlinear transformations on the concatenated features and the final denoising output.

[0048] The Cuk converter is a DC-DC power converter with buck-boost functionality. Its voltage closed-loop feedback data refers to the process data of real-time monitoring and adjustment of the output voltage through a feedback mechanism. The voltage ripple coefficient is an indicator that measures the magnitude of the AC component in the output voltage; the smaller the ripple coefficient, the better the voltage stability. Correlation analysis is used to determine the degree of correlation between the voltage ripple coefficient and the mean square error loss function to assess the impact of power supply ripple on network errors. Spectral peak position shift error refers to the deviation between the detected wavelength position of the reflection peak and the true position, and reflectivity measurement error refers to the deviation between the detected reflectivity value and the true value. Both serve as the basis for setting the convergence termination condition of the mean square error loss function; that is, when these two errors in the network output fall within a preset range, the loss function stops converging.

[0049] Execution steps: Acquire the original spectral signal and wafer surface morphology image, and feed them as dual-modal inputs into a multimodal denoising network. The original spectral signal is input into a Transformer encoder, which learns the temporal correlation of reflection intensity at different wavelengths through a self-attention mechanism, such as the dependence of the reflection peak intensity change of any band on adjacent bands. Simultaneously, the wafer surface morphology image is input into a convolutional neural network, which extracts edge features and texture features through convolutional layers and pooling layers in sequence. For example, the edge feature is the clear outline of the interface between thin film layers, and the texture feature is the distribution of light and dark textures caused by surface roughness. The spectral feature vector output by the Transformer and the image feature vector output by the convolutional neural network are concatenated to form a fused feature vector, which is then input into a fully connected layer for processing, and the denoised spectral signal is output.

[0050] It is important to know that the voltage ripple of a laser source affects the light source intensity and spectral characteristics, which in turn leads to spectral peak shift and affects the accuracy of film thickness measurement. During the emission of incident light from the laser source, a Cuk converter is used to collect voltage closed-loop feedback data in real time to obtain the voltage ripple coefficient. For example, the fluctuation data of the output voltage within the range of 12V±0.1V. Specifically, for sinusoidal ripple, its effective value is approximately 1 / 3 of the peak value. Therefore, the effective value of the AC ripple is (0.1V ÷ 0.1V). Therefore, the voltage ripple factor is determined as: AC ripple RMS value ÷ DC component = (0.1V ÷ V). ) ÷ 12V × 100%, therefore, correspondingly, through correlation analysis, it was found that when the ripple coefficient increases, the mean square error loss function value increases, indicating that the power supply ripple has a significant impact on the detection error. Based on this, the preferred convergence termination condition of the mean square error loss function is set as follows: spectral peak position shift error ≤ 0.5nm and reflectivity measurement error ≤ 1%; when the network training reaches the condition, the iteration stops. At this time, in the denoised spectral signal output, noise interference such as reflection peak jitter caused by power supply ripple is effectively suppressed.

[0051] Furthermore, the voltage ripple coefficient is determined using the voltage closed-loop feedback data of the Cuk converter. The method of this application includes: The voltage ripple coefficient is determined using the voltage closed-loop feedback data of the Cuk converter, and the dynamic adjustment threshold of the Cuk converter is set: when the spectral peak position shift error is detected to exceed the preset threshold, the duty cycle adjustment of the Cuk converter is automatically triggered. By reducing the cutoff frequency of the LC filter network, the suppression of high-frequency ripple is enhanced, so that the voltage ripple coefficient is controlled within the ripple tolerance threshold, and the corresponding reflection peak position measurement error is corrected to the peak position accuracy reference.

[0052] Specifically, the voltage closed-loop feedback data of the Cuk converter refers to the data that the Cuk converter monitors the output voltage in real time through the feedback loop during operation, and sends the deviation information between the actual voltage and the target voltage back to the control unit. This data is used to dynamically adjust the output voltage to maintain stability. The voltage ripple coefficient is the ratio of AC ripple component to DC component in the output voltage, and is a key indicator for measuring voltage stability. The dynamic adjustment threshold is a pre-set critical value for the spectral peak position shift error used to trigger the adjustment mechanism of the Cuk converter. When the actual error exceeds this value, the adjustment process is initiated. Duty cycle adjustment refers to adjusting the magnitude and stability of the output voltage by changing the ratio of the conduction time to the period of the switching transistor in the Cuk converter. The LC filter network is a filter circuit composed of inductors and capacitors, used to filter out AC ripple in the output voltage. Lowering the cutoff frequency of the LC filter network means improving the low-pass performance of the filter, thereby suppressing high-frequency ripple. The ripple tolerance threshold is the maximum allowable voltage ripple coefficient. Within this range, the detection system can be guaranteed to work normally. The peak position accuracy reference is the target control value for the reflection peak position measurement error, that is, the accuracy standard that needs to be achieved after correction.

[0053] Execution steps: The Cuk converter continuously acquires voltage closed-loop feedback data to obtain the current voltage ripple coefficient. Simultaneously, if the preset dynamic adjustment threshold for spectral peak position shift error is 1nm, the ripple tolerance threshold is 0.005, and the peak position accuracy reference is ±0.3nm; when the spectral peak position shift error at any wavelength reaches 1.2nm (exceeding 1nm), the duty cycle adjustment of the Cuk converter is triggered, adjusting the duty cycle of the switching transistor, reducing the cutoff frequency of the LC filter network, and enhancing the filtering effect on high-frequency ripple. Specifically, the lower the cutoff frequency, the stronger the attenuation capability for high-frequency ripple above the cutoff frequency. Although the duty cycle of the Cuk converter does not directly determine the inductor and capacitor parameters, it can indirectly affect the frequency distribution of ripple by changing the switching frequency or energy transfer efficiency. After adjustment, the voltage ripple coefficient is reduced until it is below the ripple tolerance threshold of 0.005, and the corresponding reflection peak position measurement error is corrected to within the ±0.3nm peak position accuracy reference limit.

[0054] Preferably, in environments with severe high-frequency ripple interference, the ripple coefficient alone cannot fully reflect the impact of high-frequency ripple. A spectrum analyzer is used to detect the frequency distribution of the ripple, and combined with the spectral analysis of the ripple, it is ensured that the high-frequency components are effectively suppressed, thus effectively guaranteeing the consistency of detection accuracy. By linking the spectral peak position shift error with the voltage adjustment of the Cuk converter, a real-time voltage ripple suppression closed loop is formed, which solves the problem of reflection peak position measurement deviation caused by equipment power supply ripple, significantly improves the stability of spectral feature extraction, and provides a more reliable basis for film thickness conversion based on peak position parameters.

[0055] Furthermore, the method of this application includes: For systematic errors in multi-batch wafer inspection, the long-term voltage stability parameter of the Cuk converter is incorporated into the error tracing system. When the inspection error of any batch shows a systematic shift, the influence weight of power supply stability on the thin film refractive index correction term is located by comparing the inspection data of different voltage drift intervals, and the calibration coefficient of the wavelength-dependent term is optimized. A dual error compensation mechanism is set up through the power supply stability control of the Cuk converter and the noise suppression of the multi-mode denoising network.

[0056] Specifically, systematic errors in multi-batch wafer inspection refer to regular, repeatable errors caused by factors inherent in the inspection system itself when inspecting multiple batches of wafers, such as performance drift after long-term equipment operation, rather than random errors; the long-term voltage stability parameter of a Cuk converter is an indicator of the stability of the output voltage of the Cuk converter during long-term operation, usually expressed as the amount of voltage drift over a certain period of time; the error tracing system is a systematic framework for tracing the source of errors and analyzing the error propagation path. Incorporating the long-term voltage stability parameter into it can clarify the impact of power supply stability on inspection errors; systematic bias refers to the overall bias of the inspection error of a batch of wafers in a certain direction, such as being generally larger or smaller. The deviation is relatively small, rather than a random deviation at individual detection points; the voltage drift range divides the voltage drift over time into different ranges to analyze the differences in detection data under different drift degrees; the influence weight is a parameter that quantifies the degree of influence of power supply stability on the thin film refractive index correction term. The larger the weight, the more significant the influence of power supply stability on the correction term; the calibration coefficient of the wavelength dependence term is a parameter used to adjust the calculation accuracy of the wavelength dependence term. Optimizing this coefficient can improve the accuracy of the refractive index correction; the dual error compensation mechanism refers to the mechanism that simultaneously compensates for detection errors through power supply stability control of the Cuk converter (hardware level) and noise suppression of the multi-modal denoising network (algorithm level). The two work together to reduce the overall error.

[0057] Execution steps: Continuously record the long-term voltage stability parameters of the Cuk converter, such as the voltage drift of ±0.2V over 100 hours of continuous operation, and incorporate it into the error traceability system; when inspecting a batch of wafers, it is found that the film thickness values ​​at all inspection points of the batch are on average 2nm higher than the standard value, indicating a systematic shift. At this time, retrieve the voltage drift data during the inspection period of this batch and divide it into multiple intervals, such as 0-0.05V, 0.05-0.1V, and 0.1-0.2V, and compare the inspection data corresponding to different intervals: for example, when the voltage drift is 0V-0.05V, the average film thickness detection error is 0.5nm; when the drift is 0.1V-0.2V, the average error is 2.5nm. This determines the influence weight of power supply stability on the thin film refractive index correction term. Based on this, optimize the calibration coefficient of the wavelength dependence term to make the corrected refractive index more closely match the actual value.

[0058] Meanwhile, the dynamic voltage regulation of the Cuk converter reduces the impact of power supply ripple on detection; combined with the multi-modal denoising network to suppress noise in the spectral signal, a dual error compensation mechanism is formed. Preferably, the thickness detection system error between the same batch is consistent in terms of standard deviation. By setting a dual error compensation mechanism for different batches of wafer films, the system error of multi-batch wafer detection can be controlled, effectively solving the problem of error accumulation caused by equipment performance drift and environmental interference in long-term detection. This ensures the reliability and comparability of wafer film thickness detection results of different batches, meeting the high requirements for detection stability in large-scale mass production.

[0059] Furthermore, the method of this application also includes: Based on the local features of the global film thickness distribution map and the anomalies in the wafer surface reflectance spectrum signal, defect candidate regions are identified, and a defect classifier is trained. The inputs are the film thickness gradient and spectral absorption peak shift of the defect candidate region, and the output is the defect type. The average thickness, thickness standard deviation, and maximum thickness deviation of the wafer thin film are determined through the global film thickness distribution map. Defects are classified, located, and labeled in combination with the defect candidate regions and the corresponding defect types, and visual reminders are provided through the defect classification, location, and labeling.

[0060] Specifically, local features of the global film thickness distribution map refer to the film thickness distribution characteristics of local areas that differ significantly from the surrounding areas, such as sudden increases or decreases in local thickness or abnormal fluctuations; anomalies in the wafer surface reflection spectrum signal refer to points in the reflection spectrum signal that deviate from the normal pattern, such as a sudden abnormal increase or decrease in reflectivity at a certain wavelength, or an unexpected large shift in the reflection peak position; defect candidate regions refer to wafer thin film regions that may have defects based on the analysis of local features and anomalies; the defect classifier is a machine learning-based model used to identify and classify the types of defect candidate regions; film thickness gradient refers to the rate of change of film thickness within the defect candidate region, reflecting the speed of thickness change; spectral absorption peak shift refers to the wavelength position difference between the spectral absorption peak corresponding to the defect candidate region and the absorption peak in the normal region; defect types include uneven film thickness, local overthickness, local underthickness, and abnormal absorption caused by interface particle contamination.

[0061] Average thickness refers to the arithmetic mean of the thicknesses at all inspection points of the wafer thin film, reflecting the overall thickness level of the film; thickness standard deviation is used to measure the dispersion of the thickness at each inspection point from the average thickness, reflecting the uniformity of the thickness; maximum thickness deviation refers to the maximum difference between the thickness at all inspection points and the average thickness, reflecting extreme deviations in thickness; defect classification and location labeling refers to clearly marking the location, type, and related parameters of the defect candidate area; visual reminders present the labeled defect information on the display device in an intuitive way using graphics, colors, etc.

[0062] Execution steps: Based on the global film thickness distribution map, analyze local features. If a significant difference in film thickness is found between a certain region and the surrounding film thickness, such as a sudden drop in film thickness from 200nm to 150nm, forming a distinct local feature; simultaneously, combine the wafer surface reflection spectrum signal. If the reflection peak in the spectrum corresponding to this region suddenly shifts from 600nm to 620nm, it is identified as an anomaly. Combining both, this region is identified as a candidate defect region. Extract the film thickness gradient and spectral absorption peak shift of this candidate defect region. Use these data as input to train a defect classifier. The defect classifier learns from a large number of samples. When the above parameters are input, it can output the defect type as local abnormal thinning of film thickness.

[0063] The average thickness, standard deviation, and maximum thickness deviation of the wafer thin film are obtained by using a global film thickness distribution map. Combined with the identified defect candidate regions and their corresponding local thin film thickness aberration types, the regions are classified and located on the map. For example, the region location is marked with a yellow box, and local thinning is marked, such as a film thickness gradient of 10 nm / μm and an absorption peak shift of 20 nm. The marked map is then displayed on the screen in a visual manner, with reminders provided through color contrast and text descriptions.

[0064] Through the above steps, a closed loop is achieved from film thickness detection to defect identification and presentation. By combining the global film thickness distribution map and spectral signal anomalies, defect candidate regions are accurately identified. The trained defect classifier can efficiently determine the defect type, while parameters such as average thickness and thickness standard deviation provide an overall reference for defect assessment. Defect classification, location labeling, and visualization alerts improve the efficiency of defect handling, effectively assisting in quality control during chip production and reducing yield losses caused by untimely defect handling.

[0065] In summary, the beneficial effects of the embodiments of this application are: This application utilizes a laser source to emit incident light onto a wafer thin film, simultaneously acquiring the wafer surface reflection spectrum signal and extracting spectral feature information and wavelength parameters. Based on the polarization state of the incident light and combined with the wavelength parameters, polarization feature parameters in the wafer surface reflection spectrum signal are identified. Based on the interlayer interfaces of the wafer thin film and combined with spectral feature information, interface interference feature parameters in the wafer surface reflection spectrum signal are identified. Based on the polarization feature parameters and interface interference feature parameters in the wafer surface reflection spectrum signal, wavelength parameters are introduced as adjustment variables to determine wavelength dependence terms. The thickness value of each detection point of the wafer thin film is determined using the thin film refractive index corrected for wavelength dependence terms, generating a global film thickness distribution map. This application provides a fully automated wafer film thickness detection method and system based on spectral reflection. By extracting spectral features and wavelength parameters, and combining the polarization state of the incident light and the interlayer characteristics of the thin film, polarization feature parameters and interfacial interference feature parameters are identified. Wavelength parameters are introduced as adjustment variables to determine wavelength dependence. After correcting the refractive index of the thin film, the thickness of each detection point is determined. This effectively improves the accuracy of thickness detection, reduces the influence of spectral signal noise and interfacial interference, and achieves the technical effect of accurately presenting the global thickness distribution of wafer thin films.

[0066] Example 2, based on the same inventive concept as the fully automated wafer film thickness detection method based on spectral reflectance in the foregoing examples, such as... Figure 2 As shown in the embodiment of this application, a fully automated wafer film thickness detection system based on spectral reflectance is provided, wherein the system includes: Feature extraction module M100: Uses a laser light source to emit incident light onto the wafer thin film, simultaneously acquires the reflection spectrum signal of the wafer surface, and extracts spectral feature information and wavelength parameters.

[0067] Feature parameter recognition module M200: Based on the polarization state of the incident light and in combination with the wavelength parameter, it identifies the polarization feature parameters in the reflection spectrum signal of the wafer surface; based on the interlayer interface of the wafer thin film and in combination with the spectral feature information, it identifies the interface interference feature parameters in the reflection spectrum signal of the wafer surface.

[0068] Wavelength Dependency Determination Module M300: Based on the polarization characteristic parameters in the wafer surface reflection spectrum signal and the interface interference characteristic parameters in the wafer surface reflection spectrum signal, the wavelength parameter is introduced as an adjustment variable to determine the wavelength dependency.

[0069] Thickness value determination module M400: Based on the thin film refractive index after wavelength dependence correction, determine the thickness value of each detection point of the wafer thin film and generate a global film thickness distribution map.

[0070] Furthermore, the feature extraction module M100 is used to perform the following method: The spectral feature information includes reflection peak intensity, reflection valley depth, characteristic peak position, and spectral half-width at half maximum (FWHM). The wavelength parameters cover the tunable wavelength range of the laser source.

[0071] Furthermore, the feature parameter recognition module M200 is used to perform the following method: The polarization characteristic parameters include s-polarized light reflectivity, p-polarized light reflectivity, and polarization state extinction ratio. The s-polarized light reflectivity and p-polarized light reflectivity are determined by acquiring the reflected light intensity in the corresponding polarization direction using a polarization beam splitter. The polarization state extinction ratio is the ratio of s-polarized light reflectivity to p-polarized light reflectivity.

[0072] Furthermore, the feature parameter recognition module M200 is used to perform the following method: The interface interference characteristic parameters include interlayer reflection peak position shift, interface scattering intensity factor, and interference fringe contrast. The interlayer reflection peak position shift is determined by the difference in the position of the spectral peaks formed by the reflected light from the interfaces of different thin film layers. The interface scattering intensity factor is positively correlated with the interface roughness and is determined by conversion through the baseline drift of the reflection spectral signal. The interference fringe contrast is the percentage difference between the reflection peak intensity and the reflection valley depth.

[0073] Furthermore, the thickness value determination module M400 is also used to perform the following method: Based on the wavelength-dependent corrected thin film refractive index, a joint control analysis is performed with the filter array corresponding to the beam splitter under the grating + prism combination to set a full-band spectral feature matrix; the matching degree of the full-band spectral feature matrix is ​​introduced, and there is only one and only one. When the matching degree meets the standard and the convergence termination condition of the mean square error loss function is satisfied, the multi-mode denoising network converges to the global optimal solution.

[0074] Furthermore, the thickness value determination module M400 is also used to perform the following method: The noise suppression of the wafer surface reflectance spectral signal adopts a Transformer-based multimodal denoising network: the original spectral signal and the wafer surface topography image are used as dual-modal inputs, and time-series modeling is performed by a Transformer encoder. At the same time, edge features and texture features of the wafer surface topography image are extracted by a convolutional neural network, and the dual-modal features are concatenated and input into a fully connected layer. The voltage ripple coefficient is determined using the voltage closed-loop feedback data of the Cuk transformer. The correlation between the voltage ripple coefficient and the mean square error loss function is analyzed. The convergence termination condition of the mean square error loss function is set based on the spectral peak position shift error and reflectance measurement error.

[0075] Furthermore, the thickness value determination module M400 is also used to perform the following method: The voltage ripple coefficient is determined using the voltage closed-loop feedback data of the Cuk converter, and the dynamic adjustment threshold of the Cuk converter is set: when the spectral peak position shift error is detected to exceed the preset threshold, the duty cycle adjustment of the Cuk converter is automatically triggered. By reducing the cutoff frequency of the LC filter network, the suppression of high-frequency ripple is enhanced, so that the voltage ripple coefficient is controlled within the ripple tolerance threshold, and the corresponding reflection peak position measurement error is corrected to the peak position accuracy reference.

[0076] Furthermore, the thickness value determination module M400 is also used to perform the following method: For systematic errors in multi-batch wafer inspection, the long-term voltage stability parameter of the Cuk converter is incorporated into the error tracing system. When the inspection error of any batch shows a systematic shift, the influence weight of power supply stability on the thin film refractive index correction term is located by comparing the inspection data of different voltage drift intervals, and the calibration coefficient of the wavelength-dependent term is optimized. A dual error compensation mechanism is set up through the power supply stability control of the Cuk converter and the noise suppression of the multi-mode denoising network.

[0077] Furthermore, the thickness value determination module M400 is also used to perform the following method: Based on the local features of the global film thickness distribution map and the anomalies in the wafer surface reflectance spectrum signal, defect candidate regions are identified, and a defect classifier is trained. The inputs are the film thickness gradient and spectral absorption peak shift of the defect candidate region, and the output is the defect type. The average thickness, thickness standard deviation, and maximum thickness deviation of the wafer thin film are determined through the global film thickness distribution map. Defects are classified, located, and labeled in combination with the defect candidate regions and the corresponding defect types, and visual reminders are provided through the defect classification, location, and labeling.

[0078] In summary, any step can be stored as a computer instruction or program in an unrestricted computer memory and can be called and recognized by an unrestricted computer processor; no further restrictions are imposed here.

[0079] Furthermore, the above technical solutions only embody the preferred technical solutions of the embodiments of this application. Any changes that those skilled in the art may make to certain parts of these solutions embody the novel principles of the embodiments of this application. Obviously, those skilled in the art can make various modifications and variations to this application without departing from the scope of this application.

Claims

1. A fully automated wafer film thickness detection method based on spectral reflectance, characterized in that, The method includes: A laser light source is used to emit incident light onto a wafer thin film, and the reflection spectrum signal of the wafer surface is acquired simultaneously to extract spectral feature information and wavelength parameters. Based on the polarization state of the incident light and the wavelength parameter, polarization characteristic parameters in the wafer surface reflection spectrum signal are identified; based on the interlayer interface of the wafer thin film and the spectral characteristic information, interface interference characteristic parameters in the wafer surface reflection spectrum signal are identified. Based on the polarization characteristic parameters and interface interference characteristic parameters in the wafer surface reflection spectrum signal, the wavelength parameter is introduced as an adjustment variable to determine the wavelength dependence term; Based on the wavelength-dependent corrected thin film refractive index, the thickness value of each detection point of the wafer thin film is determined, and a global film thickness distribution map is generated.

2. The fully automated wafer film thickness detection method based on spectral reflectance as described in claim 1, characterized in that, The method for extracting spectral feature information and wavelength parameters includes: The spectral feature information includes reflection peak intensity, reflection valley depth, characteristic peak position, and spectral half-width at half maximum (FWHM). The wavelength parameters cover the tunable wavelength range of the laser source.

3. The fully automated wafer film thickness detection method based on spectral reflectance as described in claim 1, characterized in that, The polarization characteristic parameters include s-polarized light reflectivity, p-polarized light reflectivity, and polarization state extinction ratio. The s-polarized light reflectivity and p-polarized light reflectivity are determined by acquiring the reflected light intensity in the corresponding polarization direction using a polarization beam splitter. The polarization state extinction ratio is the ratio of s-polarized light reflectivity to p-polarized light reflectivity.

4. The fully automated wafer film thickness detection method based on spectral reflectance as described in claim 1, characterized in that, The interface interference characteristic parameters include interlayer reflection peak position shift, interface scattering intensity factor, and interference fringe contrast. The interlayer reflection peak position shift is determined by the difference in the position of the spectral peaks formed by the reflected light from the interfaces of different thin film layers. The interface scattering intensity factor is positively correlated with the interface roughness and is determined by conversion through the baseline drift of the reflection spectral signal. The interference fringe contrast is the percentage difference between the reflection peak intensity and the reflection valley depth.

5. The fully automated wafer film thickness detection method based on spectral reflectance as described in claim 1, characterized in that, The method further includes determining the thickness value of each detection point of the wafer thin film based on the thin film refractive index corrected for the wavelength dependence term. Based on the thin film refractive index corrected by the wavelength dependence term, a joint control analysis is performed with the filter array corresponding to the beam-splitting component under the grating + prism combination to set the full-band spectral feature matrix. By introducing the matching degree of the full-band spectral feature matrix, there exists one and only one condition that the multimodal denoising network converges to the global optimal solution when the matching degree meets the threshold and the mean square error loss function is satisfied.

6. The fully automated wafer film thickness detection method based on spectral reflectance as described in claim 5, characterized in that, The method further includes: The noise suppression of the wafer surface reflectance spectral signal adopts a Transformer-based multimodal denoising network: the original spectral signal and the wafer surface topography image are used as dual-modal inputs, and time-series modeling is performed by a Transformer encoder. At the same time, edge features and texture features of the wafer surface topography image are extracted by a convolutional neural network, and the dual-modal features are concatenated and input into a fully connected layer. The voltage ripple coefficient is determined using the voltage closed-loop feedback data of the Cuk transformer. The correlation between the voltage ripple coefficient and the mean square error loss function is analyzed. The convergence termination condition of the mean square error loss function is set based on the spectral peak position shift error and reflectance measurement error.

7. The fully automated wafer film thickness detection method based on spectral reflectance as described in claim 6, characterized in that, The method for determining the voltage ripple coefficient using voltage closed-loop feedback data from a Cuk converter includes: The voltage ripple coefficient is determined using the voltage closed-loop feedback data of the Cuk converter, and the dynamic adjustment threshold of the Cuk converter is set: when the spectral peak position shift error is detected to exceed the preset threshold, the duty cycle adjustment of the Cuk converter is automatically triggered. By reducing the cutoff frequency of the LC filter network, the suppression of high-frequency ripple is enhanced, so that the voltage ripple coefficient is controlled within the ripple tolerance threshold, and the corresponding reflection peak position measurement error is corrected to the peak position accuracy reference.

8. The fully automated wafer film thickness detection method based on spectral reflectance as described in claim 7, characterized in that, The method includes: For systematic errors in multi-batch wafer inspection, the long-term voltage stability parameter of the Cuk converter is incorporated into the error tracing system. When the inspection error of any batch shows a systematic shift, the influence weight of power supply stability on the thin film refractive index correction term is located by comparing the inspection data of different voltage drift intervals, and the calibration coefficient of the wavelength-dependent term is optimized. The power supply stability control of the Cuk converter and the noise suppression setting of the multimodal denoising network are used to establish a dual error compensation mechanism.

9. The fully automated wafer film thickness detection method based on spectral reflectance as described in claim 1, characterized in that, The method further includes: Based on the local features of the global film thickness distribution map, combined with the anomalies of the wafer surface reflection spectrum signal, defect candidate regions are identified, and a defect classifier is trained. The inputs are the film thickness gradient and spectral absorption peak shift of the defect candidate region, and the output is the defect type. The average thickness, standard deviation, and maximum thickness deviation of the wafer thin film are determined by the global film thickness distribution map. Defects are classified, located, and labeled by combining the defect candidate regions and corresponding defect types, and visual reminders are provided by the defect classification and location labels.

10. A fully automated wafer film thickness detection system based on spectral reflectance, characterized in that, The system is used to implement the fully automated wafer film thickness detection method based on spectral reflectance according to any one of claims 1-9, wherein the system comprises: Feature extraction module: Uses a laser light source to emit incident light onto the wafer thin film, simultaneously acquires the reflection spectrum signal of the wafer surface, and extracts spectral feature information and wavelength parameters; Feature parameter recognition module: Based on the polarization state of the incident light and in combination with the wavelength parameter, it identifies the polarization feature parameters in the reflection spectrum signal of the wafer surface; Based on the interlayer interface of the wafer thin film and in combination with the spectral feature information, it identifies the interface interference feature parameters in the reflection spectrum signal of the wafer surface. Wavelength dependence determination module: Based on the polarization characteristic parameters and interface interference characteristic parameters in the wafer surface reflection spectrum signal, the wavelength parameter is introduced as an adjustment variable to determine the wavelength dependence. Thickness value determination module: Based on the thin film refractive index after wavelength dependence correction, determine the thickness value of each detection point of the wafer thin film and generate a global film thickness distribution map.

Images