A full-range film thickness spectral inversion method and system of adaptive hierarchical processing

Abstract

The application provides a full-range film thickness spectral inversion method and system of adaptive hierarchical processing, applied to the technical field of film thickness measurement, by obtaining and preprocessing the reflection spectrum data of the measured film layer, extracting the frequency domain characteristic parameters and spectral oscillation characteristic parameters; based on these characteristic parameters, the film layer thickness magnitude interval is pre-classified; multiple parallel thickness inversion channels are constructed to calculate multiple thickness candidate values; the confidence score is calculated for each candidate value; finally, the decision fusion logic is executed combined with the pre-classification result and the confidence score, to determine and output the final film thickness measurement result. This method can effectively solve the problems of periodic ambiguity and insufficient accuracy of traditional methods in thick film measurement, and has the beneficial effects of realizing accurate, fast and non-destructive measurement of full-range film thickness.

Description

Technical Field

[0001] This application relates to the field of film thickness measurement technology, and in particular to an adaptive layered processing full-range film thickness spectral inversion method and system. Background Technology

[0002] In the field of flexible electronics and semiconductor packaging, polyimide (PI) film is a key insulating and supporting material, and its thickness uniformity directly affects product performance. Currently, PI film thickness measurement faces two main challenges: first, the material itself has high transmittance, weak absorption, and anisotropy, making its optical constant model complex and subject to fluctuations with process technology, rendering traditional spectral fitting algorithms inaccurate; second, its application thickness range is wide (from nanometers to micrometers), making it difficult for a single model to balance accuracy and reliability across the entire measurement range. Existing technologies largely rely on contact profilometers or single-point ellipsometrists. The former, being contact-based, easily damages the film, while the latter has a small detection range, low efficiency for large-size films, and exhibits significant periodic ambiguity in thick film regions, resulting in large measurement errors. In summary, existing detection technologies cannot simultaneously meet the demands for online high precision, non-destructive testing, a large detection range, and rapid detection.

[0003] Therefore, existing technologies urgently need to be improved to address the aforementioned problems. Summary of the Invention

[0004] In view of the shortcomings of the prior art, this application provides an adaptive layered processing full-range film thickness spectral inversion method and system, which is applied to the field of film thickness measurement technology. It can realize accurate, fast and non-destructive measurement of film thickness across the entire range, and effectively solve the problems of periodic ambiguity and insufficient accuracy of traditional methods in thick film measurement.

[0005] A first aspect is a method for adaptively layered full-range film thickness spectral inversion, the method comprising the following steps: S1: Acquire the reflectance spectrum data of the film to be tested, and preprocess the reflectance spectrum data to extract frequency domain feature parameters and spectral oscillation feature parameters; S2: Based on the frequency domain feature parameters and the spectral oscillation feature parameters, the thickness range of the film to be tested is pre-classified to obtain the pre-classification result; S3: Construct multiple parallel thickness inversion paths to calculate the thickness of the film layer to be measured, and obtain multiple candidate thickness values; S4: Calculate a confidence score for each of the thickness candidate values; S5: Combining the pre-classification results and the confidence scores of each candidate thickness value, execute the decision fusion logic to determine and output the final film thickness measurement result.

[0006] Furthermore, step S1 includes: S11: Obtain the reflectance spectrum data of the film to be tested; S12: Perform empirical mode decomposition on the reflectance spectral data, extract the interference signal, and remove noise and baseline drift to obtain preprocessed spectral data, wherein the preprocessed spectral data is a spectral curve dominated by the interference signal; S13: Convert the preprocessed spectral data to the wavenumber domain, calculate the normalized power spectral density of the preprocessed spectral data in the wavenumber domain, and use it as the frequency domain characteristic parameter; S14: Calculate the first or second derivative of the preprocessed spectral curve, count the number of data points or local extrema of the first or second derivative that cross zero, and define them as the spectral oscillation characteristic parameters.

[0007] Furthermore, step S2 includes: S21: Based on the frequency domain feature parameters and the spectral oscillation feature parameters, a lightweight decision tree or logistic regression model is used to preclassify the thickness range of the film to be tested into a thin film region, a transition region, or a thick film region, which is taken as the preclassification result.

[0008] Furthermore, in step S3, there are three thickness inversion pathways: a first thickness inversion pathway based on a data-driven regression model, a second thickness inversion pathway based on a physical optics fitting model, and a third thickness inversion pathway based on a spectral periodicity analysis model.

[0009] Furthermore, step S3 includes: S31: Use a data-driven regression model to perform regression calculations on the preprocessed spectral data to obtain an initial estimate, which serves as the first thickness candidate value. The data-driven regression model is a pre-trained neural network model or a gradient boosting tree model. S32: Using the initial estimated value as the initial value of the physical optics fitting model, for the thin film region in the pre-classification result, perform global optimization in the physical optics fitting model to determine the second thickness candidate value corresponding to the physical optics fitting model; S33: For the thick film region in the pre-classification results, the periodogram analysis path of the spectral period analysis model is used to perform fast Fourier transform or Lomb-Scargle periodogram analysis on the preprocessed spectral data, extract the main frequency component of the interference signal to calculate the thickness estimate, and use it as the third thickness candidate value. The thickness candidate value includes the first thickness candidate value, the second thickness candidate value, and the third thickness candidate value.

[0010] Furthermore, step S4 includes: S41: Obtain the historical accuracy of the data-driven regression model on the validation set, and calculate the first confidence score corresponding to the data-driven regression model by combining it with the current output probability distribution; S42: Calculate the fitting residuals of the physical optics fitting model and the covariance matrix of the parameter estimates, and calculate the second confidence score corresponding to the physical optics fitting model based on the fitting residuals and the covariance matrix. S43: Extract the signal-to-noise ratio and full width at half maximum (FWHM) of the main peak of the power spectral density in the spectrum periodic analysis model, and calculate the third confidence score corresponding to the spectrum periodic analysis model based on the signal-to-noise ratio and the FWHM.

[0011] Furthermore, step S5 includes: S51: When the confidence score of any of the inversion pathways is higher than a preset threshold, and the corresponding candidate thickness value conforms to the pre-classification result, the candidate thickness value is determined as the final film thickness measurement result. S52: When the difference between the candidate thickness values ​​of the multiple inversion paths is within a preset accuracy range, the corresponding candidate thickness values ​​are weighted and averaged according to each confidence score to obtain the final film thickness measurement result.

[0012] Furthermore, step S5 also includes: S53: When there is a conflict among the candidate thickness values ​​and the confidence scores of each value are lower than the preset threshold, the iterative fine process is initiated.

[0013] Furthermore, in step S53, initiating the iterative refinement process includes the following steps: S531: Use the first thickness candidate value output by the data-driven regression model or the third thickness candidate value output by the spectral periodic analysis model as the new initial value for the physical optics fitting model; S532: Based on the new initial values, re-execute the optimization calculation of the physical optics fitting model until the final film thickness measurement result is output.

[0014] Secondly, an adaptive hierarchical processing full-range film thickness spectrum inversion system is provided, the system being used to implement the steps of any of the methods described above, the system comprising: Acquisition module: Acquires the reflectance spectrum data of the film to be tested, and preprocesses the reflectance spectrum data to extract frequency domain feature parameters and spectral oscillation feature parameters; Classification module: Based on the frequency domain feature parameters and the spectral oscillation feature parameters, pre-classify the thickness range of the film to be tested to obtain the pre-classification result; First calculation module: Constructs multiple parallel thickness inversion paths to calculate the thickness of the film layer to be measured, and obtains multiple candidate thickness values; The second calculation module calculates a confidence score for each of the candidate thickness values. Result determination module: Combining the pre-classification results and the confidence scores of each candidate thickness value, the module executes decision fusion logic to determine and output the final film thickness measurement result.

[0015] Beneficial Effects: This application proposes an adaptive hierarchical processing method and system for full-range film thickness spectral inversion. It acquires and preprocesses the reflectance spectrum data of the film to be measured, extracting frequency domain feature parameters and spectral oscillation feature parameters. Based on these feature parameters, it pre-classifies the film thickness range. Multiple parallel thickness inversion pathways are constructed to calculate multiple candidate thickness values. A confidence score is calculated for each candidate value. Finally, combining the pre-classification results and confidence scores, a decision fusion logic is executed to determine and output the final film thickness measurement result. This method effectively solves the problems of periodic ambiguity and insufficient accuracy in traditional methods for thick film measurement, and has the beneficial effect of achieving accurate, rapid, and non-destructive measurement of film thickness across the entire range. Attached Figure Description

[0016] Figure 1 This is a flowchart of an adaptive layered processing full-range film thickness spectrum inversion method proposed in this application.

[0017] Figure 2 This is a structural diagram of the full-range film thickness spectrum inversion method for adaptive layering processing proposed in this application.

[0018] Labeling Explanation: 201. Acquisition Module; 202. Classification Module; 203. First Calculation Module; 204. Second Calculation Module; 205. Result Determination Module. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and marked in the accompanying drawings can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0020] In the manufacturing processes of modern flexible electronic devices and advanced semiconductor packaging, polyimide films, as a core insulating and supporting material, are crucial for determining the final product's performance and yield through precise and uniform thickness control. However, rapid, non-destructive, and accurate online measurement of this material's thickness has always been a technical challenge. On the one hand, polyimide materials themselves possess high light transmittance, weak absorption, and optical anisotropy, leading to complex optical constant models that vary with minute fluctuations in the manufacturing process. This makes it difficult for traditional spectral fitting-based algorithms to obtain stable and accurate results. On the other hand, polyimide films have extremely wide applications, with thicknesses ranging from ultrathin layers of tens of nanometers to thick layers of tens or even hundreds of micrometers. This vast measurement range makes it difficult for any single measurement model to simultaneously guarantee high accuracy and high reliability across the entire range. For example, physical optics fitting methods are highly accurate in thin film regions, but in thick film regions, they easily fall into the trap of periodic ambiguity, meaning that multiple thickness values ​​may produce similar spectra, leading to misjudgments. Conversely, spectral analysis-based methods perform excellently in thick film regions, but in thin film regions, the lack of interference fringes makes it difficult to effectively extract periodic information.

[0021] Existing detection methods, such as contact profilometers, while direct, are destructive and unsuitable for large-scale online production because their probes may scratch or contaminate the delicate film surface. Optical methods, such as single-point ellipsometry, are non-destructive, but their detection range is limited, resulting in low efficiency for measuring the uniformity of large-area films. Furthermore, they face significant periodic ambiguity issues when processing thick film samples, leading to substantial measurement errors. Therefore, the industry urgently needs a film thickness measurement solution that simultaneously meets the requirements of high-precision, non-destructive, full-range coverage, and rapid online detection.

[0022] To address the aforementioned challenges, this application proposes an adaptive, layered, full-range film thickness spectral inversion method, which includes the following steps: S1: Acquire the reflectance spectrum data of the film to be tested, and preprocess the reflectance spectrum data to extract frequency domain feature parameters and spectral oscillation feature parameters; S2: Based on frequency domain characteristic parameters and spectral oscillation characteristic parameters, the thickness range of the film to be tested is pre-classified to obtain the pre-classification results; S3: Construct multiple parallel thickness inversion pathways to calculate the thickness of the film to be tested and obtain multiple candidate thickness values; S4: Calculate the confidence score for each thickness candidate value; S5: Combining the pre-classification results and the confidence scores of each thickness candidate value, execute the decision fusion logic to determine and output the final film thickness measurement result.

[0023] The core idea of ​​this method is to no longer rely on a single fixed algorithm, but to make a rapid prediction of the thickness order based on the characteristics of the reflectance spectrum data of the film to be measured. Then, multiple dedicated inversion models for different thickness ranges are launched in parallel for calculation, and each calculation result is assigned a quantitative confidence score. Finally, through a set of dynamic decision fusion logic, all information is integrated to intelligently select, correct or fuse the results of each model, thereby outputting the optimal and most reliable film thickness measurement value in the entire range.

[0024] Specifically, in the steps of acquiring and preprocessing reflectance spectral data, the raw reflectance spectral data of the film under test is first obtained through a spectral measurement device. This device typically includes a broadband light source, such as a halogen lamp or a xenon lamp, which guides the light to a probe via an optical fiber. The probe illuminates the surface of the film under test perpendicularly or at a specific angle, then collects the light reflected from the film surface and the substrate interface, and sends it through another optical fiber to a spectrometer for spectral dispersion and detection, ultimately obtaining a raw spectral image with wavelength as the abscissa and reflectance as the ordinate.

[0025] The raw spectrum often contains multiple components. Besides the periodic oscillation signal generated by thin-film interference, it also contains random noise introduced by light source instability and detector dark current, as well as slowly changing baseline drift caused by factors such as sample tilt and stray light. To accurately extract the features required for subsequent analysis, the raw spectral data must undergo meticulous preprocessing. To this end, this application employs Empirical Mode Decomposition (EMD) to extract the interference signal and remove noise and baseline drift. EMD is an adaptive signal processing method that can decompose a complex nonlinear, non-stationary signal into a finite number of eigenmode functions. Each eigenmode function represents an inherent vibrational mode of the signal at different time scales.

[0026] Furthermore, step S1 includes: S11: Obtain the reflectance spectrum data of the film to be tested; S12: Perform empirical mode decomposition on the reflectance spectral data, extract the interference signal, and remove noise and baseline drift to obtain preprocessed spectral data. The preprocessed spectral data is a spectral curve dominated by the interference signal. S13: Convert the preprocessed spectral data to the wavenumber domain, calculate the normalized power spectral density of the preprocessed spectral data in the wavenumber domain, and use it as a frequency domain characteristic parameter. S14: Calculate the first or second derivative of the preprocessed spectral curve, count the number of data points or local extrema where the first or second derivative crosses zero, and define them as spectral oscillation characteristic parameters.

[0027] In a specific implementation scenario, the acquired raw reflectance spectrum data is a sequence containing 1024 data points. Performing empirical mode decomposition on this sequence may yield 5 to 8 intrinsic mode function (EMF) components. Typically, the first or first two EMF components exhibit high-frequency, irregular oscillations, corresponding to high-frequency noise in the system. The last EMF component, on the other hand, appears as a very smooth, monotonic curve or a curve with only one or two extrema, representing the overall baseline shift of the spectrum.

[0028] The middle few eigenmode function components clearly exhibit periodic oscillation characteristics, which is exactly the core signal generated by thin-film interference that we need. By discarding the eigenmode function components that represent noise and baseline drift, and then reconstructing and superimposing the remaining eigenmode function components that represent the interference signal, we can obtain a clean, baseline-corrected preprocessed spectral data. This spectral curve is dominated by the pure interference signal.

[0029] Empirical mode decomposition (EMD) is performed on the reflectance spectral data to extract the interference signal and remove noise and baseline drift. The original reflectance spectral data sequence is used as input signal, and adaptive EMD is performed to decompose the original data into a finite number of intrinsic mode function components and a residual component.

[0030] The average instantaneous frequency and energy percentage of each intrinsic mode function (IMF) component are calculated. Frequency and energy thresholds are set, and the first few components with an average instantaneous frequency higher than the high-frequency threshold and an energy percentage lower than the preset ratio are identified as high-frequency noise components and removed. The remaining components and the last IMF component with an extremely low and monotonically changing average instantaneous frequency are identified as baseline drift and removed. The remaining IMF components with obvious periodic oscillations and intermediate frequencies are linearly superimposed and reconstructed to obtain the clean interference signal curve after removing noise and baseline drift.

[0031] After obtaining the preprocessed spectral data, it is necessary to extract quantized feature parameters for subsequent thickness-level pre-classification. This application extracts two key feature parameters. The first is the frequency domain feature parameter. To calculate this parameter, the preprocessed spectral data needs to be converted from the wavelength domain to the wavenumber domain. The wavenumber is the reciprocal of the wavelength. The advantage of this is that in the wavenumber domain, the oscillation frequency of the thin-film interference signal is proportional to the optical thickness of the film. This linear relationship makes frequency domain analysis more intuitive and effective. After the conversion, the normalized power spectral density of the spectral data in the wavenumber domain is calculated. The power spectral density describes the distribution of signal power across frequencies. For a spectrum with obvious interference fringes, one or more significant peaks will appear on its power spectral density plot. The position of the main peak corresponds to the dominant frequency of the interference signal, which is directly related to the thickness of the film. Therefore, the normalized power spectral density curve itself, or the position, height, width, etc., of its main peak, can be used as a frequency domain feature parameter.

[0032] The second type is the spectral oscillation characteristic parameter. This parameter describes the density of oscillations in the spectral curve from another perspective. The calculation method involves taking the first or second derivative of the preprocessed spectral curve. Each peak and trough on the spectral curve corresponds to a zero-crossing point of the first derivative. Therefore, by statistically analyzing the number of times the first derivative curve crosses zero across the entire spectral range, the number of spectral oscillations can be directly quantified. The thicker the film layer, the denser the interference fringes, the more oscillations per unit wavenumber, and the greater the number of zero-crossing points.

[0033] In some embodiments, the second derivative can also be used for calculation. Inflection points on the spectral curve correspond to zero-crossing points of the second derivative, while local extrema—peaks and valleys—correspond to local extrema on the second derivative curve. By statistically analyzing the number of local extrema of the second derivative, the oscillation frequency of the spectrum can also be effectively measured. In some cases, the second derivative is more sensitive to small curvature changes in the spectral curve and may provide richer oscillation information. Ultimately, the statistically obtained zero-crossing data or the number of local extrema are defined as spectral oscillation characteristic parameters.

[0034] After obtaining the two key indicators, frequency domain characteristic parameters and spectral oscillation characteristic parameters, the next step is to pre-classify the thickness range of the film under test. The purpose of this step is to have a rough judgment on the film thickness before performing complex and time-consuming precise calculations, thereby providing guidance for the subsequent adaptive selection of the most suitable inversion model.

[0035] Furthermore, step S2 includes: S21: Based on frequency domain characteristic parameters and spectral oscillation characteristic parameters, a lightweight decision tree or logistic regression model is used to preclassify the thickness range of the film to be measured into a thin film region, a transition region, or a thick film region, which is used as the preclassification result.

[0036] The reason for choosing a lightweight model is to ensure extremely high efficiency in the pre-classification process, making its time cost negligible in the entire measurement process. Commonly used lightweight models include decision trees or logistic regression models.

[0037] In one specific embodiment, a pre-trained lightweight decision tree model is used for classification. The input to this decision tree consists of two features: frequency domain feature parameters, specifically the wave value of the main peak of the power spectral density; and spectral oscillation feature parameters, specifically the total number of zero-crossing points of the first derivative. The internal structure of the decision tree comprises a series of simple decision nodes. For example, the first node might determine whether the total number of zero-crossing points of the first derivative is less than 5. If the answer is yes, it is directly classified as a thin film region. If no, it proceeds to the next node, which might determine whether the wave value of the main peak of the power spectral density is greater than a certain preset threshold.

[0038] Through these simple, data-feature-based judgments, the model can quickly classify the current spectral data into one of the following categories: thin film region, transition region, or thick film region. The thresholds for these judgments are obtained through learning and optimization on a large number of standard samples of known thickness.

[0039] In another embodiment, a logistic regression model can be used. The logistic regression model calculates the probability that the current sample belongs to the thin film region, transition region, or thick film region based on the input frequency domain feature parameters and spectral oscillation feature parameters. For example, for an input, the model might output: a probability of 0.85 for the thin film region, a probability of 0.10 for the transition region, and a probability of 0.05 for the thick film region. In this case, the model selects the category with the highest probability, i.e., the thin film region, as the final pre-classification result. Based on the frequency domain feature parameters and spectral oscillation feature parameters, a lightweight decision tree or logistic regression model is used to pre-classify the thickness range of the film to be measured into a thin film region, transition region, or thick film region as the pre-classification result.

[0040] To further illustrate the flexible application of the pre-classification process in different industrial scenarios, we will take the detection of polyimide films at different stages of semiconductor wafer manufacturing as an example. In the front-end coating process of advanced packaging, it is necessary to monitor the extremely thin polyimide insulating layer online in real time, which places extremely high demands on computational efficiency. In this online monitoring scenario, a large amount of known sample spectral feature data with thicknesses in the tens of nanometers range is extracted in advance. Using this data, a shallow, lightweight decision tree model is trained on an offline central computing server. The decision tree model's decision nodes are set to be extremely concise, typically containing no more than three levels of conditional branches. For example, simply judging whether the number of zero-crossing points of the first derivative is less than a specific minimum value can immediately block subsequent complex judgments. The trained model parameters are directly burned into the microcontroller chip attached to the coating equipment. When the coating equipment is running and acquiring spectra in real time, the microcontroller only needs to consume minimal computing resources to output the determination of the thin film area within milliseconds through this concise decision tree, ensuring that the high-speed operation of the production line is not hindered by the detection process.

[0041] In the flexible display panel substrate manufacturing scenario, the thickness of polyimide films often spans tens or even hundreds of micrometers, and due to different doping components, their spectral characteristics exhibit complex nonlinear variations. Relying on a single decision tree with rigid thresholds can easily lead to misclassification. In this offline sampling laboratory environment, a logistic regression model is used for pre-classification. The logistic regression model runs on a powerful industrial control computer, receiving frequency domain characteristic parameters and spectral oscillation characteristic parameters. Instead of providing an absolute category, it calculates three independent probability values ​​for the current sample falling into the thin-film region, transition region, and thick-film region. When the probability values ​​of a sample in the thick-film region and transition region are found to be very close—for example, 48% in the thick-film region and 45% in the transition region—the logistic regression model accurately reflects this ambiguity to subsequent processes. This allows subsequent inversion calculations to selectively use weighted comparisons of the dual-channel results corresponding to the thick-film and transition regions, rather than crudely drawing boundaries. This probability-based logistic regression method demonstrates strong robustness and adaptive adjustment capabilities when dealing with complex and variable materials with large thickness ranges.

[0042] Furthermore, in step S3, there are three thickness inversion pathways: a first thickness inversion pathway based on a data-driven regression model, a second thickness inversion pathway based on a physical optics fitting model, and a third thickness inversion pathway based on a spectral periodicity analysis model.

[0043] These three pathways are based on completely different principles and each has its own strengths and applications. By using parallel computing, their respective advantages can be maximized, ensuring reliable computational results in any thickness range.

[0044] Furthermore, step S3 includes: S31: Use a data-driven regression model to perform regression calculations on the preprocessed spectral data to obtain an initial estimate, which serves as the first thickness candidate value. The data-driven regression model is a pre-trained neural network model or a gradient boosting tree model. S32: Use the initial estimated value as the initial value of the physical optics fitting model. For the thin film region in the pre-classification results, perform global optimization in the physical optics fitting model to determine the second thickness candidate value corresponding to the physical optics fitting model. S33: For the thick film region in the pre-classification results, the periodogram analysis path of the spectral period analysis model is used to perform fast Fourier transform or Lomb-Scargle periodogram analysis on the preprocessed spectral data, extract the main frequency component of the interference signal to calculate the thickness estimate, and use it as the third thickness candidate value. The thickness candidate values ​​include a first thickness candidate value, a second thickness candidate value, and a third thickness candidate value.

[0045] Specifically, the first thickness inversion pathway utilizes a pre-trained data-driven regression model to directly perform regression calculations on the preprocessed spectral data, quickly obtaining an initial thickness estimate as a candidate value for the first thickness. This data-driven regression model can be a deep neural network model or a gradient boosting tree model.

[0046] In one embodiment employing a neural network model, specifically a multilayer perceptron, the input layer has 1024 neurons, corresponding precisely to the 1024 sampling points of the preprocessed spectral data. The network contains three hidden layers with 512, 256, and 128 neurons respectively, using rectified linear units as the activation function. The output layer is a single neuron that directly outputs the predicted membrane thickness value. This network, trained on hundreds of thousands of simulated spectral data points and thousands of real-world measured standard sample spectral data points, learns a complex nonlinear mapping from spectral morphology to membrane thickness. Its advantage lies in its extremely fast computation speed, providing a fairly accurate initial estimate within milliseconds.

[0047] In another embodiment employing a gradient boosting tree model, such as using the XGBoost or LightGBM algorithm library, the model consists of hundreds or thousands of tiny decision trees forming a powerful ensemble model. Each new decision tree learns how to correct the prediction residuals of all the previous trees. This type of model has a very strong fitting capability for tabular data, i.e., spectral data points, and its training process is generally faster and less sensitive to hyperparameters compared to neural networks.

[0048] The second thickness inversion pathway is based on a physical optics fitting model. The core of this model is to establish an accurate physical model to simulate the propagation, reflection, and interference processes of light in a multilayer film structure. This model is typically based on the Fresnel equations and the transfer matrix method, requiring inputs of the optical constants of the film material, namely the relationship between refractive index and extinction coefficient as a function of wavelength, and the film thickness. By changing the thickness parameters in the model, the theoretical reflection spectrum can be calculated.

[0049] The fitting process involves using an optimization algorithm to continuously adjust the thickness value until the difference between the calculated theoretical spectrum and the actual measured spectrum, typically measured as root mean square error, is minimized. The thickness value that minimizes this error is then considered the true film thickness. This method offers extremely high accuracy in thin-film regions, but its drawbacks include high computational cost and extreme sensitivity to the choice of initial values. If the initial values ​​deviate too far from the true values, the optimization algorithm can easily get trapped in local optima, leading to periodic ambiguity.

[0050] To address this issue, this application uses the first candidate thickness value obtained from the first path, i.e., the fast initial estimate, as the initial value for optimizing the physical optics fitting model. The advantage of this approach is that using an initial value close to the true value can significantly shorten the convergence time of the optimization algorithm and effectively help the algorithm skip erroneous local optima and directly approach the global optimum.

[0051] Specifically, this pathway focuses on calculating the thickness of samples pre-classified as thin film regions. Through global optimization, a high-precision thickness value is ultimately determined as the second thickness candidate value.

[0052] Depending on the application environment of different physical model fitting and optimization strategies, the optimization calculation method for thin film regions can exhibit various forms. In an independent testing station of a high-precision optical coating workshop, the thin film layer under test has extremely stable optical constants, and the environmental noise is extremely low. In this scenario, the physical optics fitting model is configured to perform a global search using a very rigorous simulated annealing algorithm. Since the goal is to pursue absolute accuracy, the model generates a large number of random probe steps within the upper and lower limits set throughout the entire thin film region, using initial estimates as seed points. These step steps are relatively large in the early stages of the algorithm, creating a trap. As the number of iterations increases, the step steps decrease exponentially, gradually allowing the search range to escape any possible local optimum and approach the theoretically smallest root mean square error position. This process may take several seconds of computation time, but it can provide nanometer-level limit accuracy, fully meeting laboratory-level quality inspection requirements.

[0053] Conversely, on high-speed roll-to-roll thin film production lines, the films are constantly in a state of high-speed movement, and external mechanical vibrations inevitably cause slight distortions in the spectral signals. Continuing to use time-consuming simulated annealing algorithms would result in delayed detection results, failing to provide timely guidance for production adjustments. In this high-speed scenario, the global optimization within the physical optics fitting model is replaced by a particle swarm optimization algorithm. Furthermore, the search scope is no longer the broad entire thin film region, but rather a very narrow dynamic search window is constructed, tightly focused on the initial estimates provided by the data-driven regression model. Within this narrow window, dozens of virtual particles simultaneously begin searching for the minimum residual solution. This optimization method sacrifices extremely small absolute precision for a response speed on the order of hundreds of milliseconds. The collective cooperative characteristics of the particle swarm algorithm also effectively smooth out single-point spectral distortions caused by mechanical vibrations, ensuring stable and reliable second thickness candidate values ​​are output even in harsh industrial environments.

[0054] The third thickness inversion pathway is based on a spectral periodicity analysis model. This pathway is specifically designed for measurements in thick film regions. When the film is thick, its reflection spectrum exhibits very dense and highly periodic interference fringes in the wavenumber domain. The spectral periodicity analysis model utilizes this characteristic. This pathway employs periodogram analysis to perform Fast Fourier Transform or Lomb-Scargle periodogram analysis on the preprocessed spectral data.

[0055] In one specific embodiment, if the spectral data is uniformly sampled in the wavenumber domain, a Fast Fourier Transform (FFT) is directly applied. The result of the transform is a spectrum, with the horizontal axis representing frequency. The periodicity of the interference signal will form a very sharp peak on this graph, and the location of this peak corresponds to the dominant frequency of the interference signal. According to thin-film interference theory, this dominant frequency is proportional to the optical thickness of the film. A thickness estimate can be calculated from this frequency using a simple formula, serving as a third candidate thickness value. This method can easily achieve a computational range down to the millimeter level, effectively solving the problem of thick-film measurement.

[0056] In another embodiment, if the spectrometer sampling is not strictly uniform in the wavenumber domain, or if there are some bad or missing points in the data, Lomb-Scargle periodogram analysis is more robust. The Lomb-Scargle periodogram is a generalization of the Fourier transform, specifically designed for processing non-uniformly sampled data, and can also accurately extract the dominant frequency components of the signal.

[0057] After computation through three parallel pathways, we obtained three potentially different but well-founded candidate thickness values. To make an informed choice among them, a confidence score needs to be calculated for each candidate value.

[0058] Furthermore, step S4 includes: S41: Obtain the historical accuracy of the data-driven regression model on the validation set, and calculate the first confidence score corresponding to the data-driven regression model by combining it with the probability distribution of the current output. S42: Calculate the fitting residuals and covariance matrix of the parameter estimates of the physical optics fitting model, and calculate the second confidence score corresponding to the physical optics fitting model based on the fitting residuals and covariance matrix. S43: Extract the signal-to-noise ratio and full width at half maximum (FWHM) of the main peak of the power spectral density in the spectral periodic analysis model, and calculate the third confidence score corresponding to the spectral periodic analysis model based on the signal-to-noise ratio and FWHM.

[0059] For the first thickness inversion pathway, the calculation of its first confidence score considers two aspects: First, the historical accuracy of the data-driven regression model on the offline validation set. After model training, an independent validation set is used to evaluate the model's performance across different thickness ranges. This historical accuracy reflects the model's generalization ability and inherent reliability. Second, the probability distribution of the model's output for the current specific spectrum. Some regression models, especially neural network-based models, can output a probability distribution rather than a single value; the sharpness of this distribution reflects the model's confidence in the current prediction. Combining these two factors yields a comprehensive first confidence score.

[0060] In one specific implementation, the accuracy can be evaluated using the Softmax probability distribution or the historical accuracy Acc(d1) of the model on the validation set for the corresponding thickness interval. In this formula, the first part This reflects the model's confidence in the current prediction. Here, P is the probability distribution of the model output, Δ is the allowable error window, d1 is the first candidate thickness value, and d represents the possible true thickness value, a variable that takes values ​​on the x-axis of the probability distribution P. Part Two It represents the historical accuracy of the data-driven regression model on the offline validation set, where Acc represents the historical accuracy function. This indicates that the historical accuracy depends on the current first thickness candidate value. The first confidence level score is given.

[0061] For the second thickness inversion pathway, the calculation of its second confidence score is primarily based on the quality of the results from the physical optics fitting model. Evaluation metrics include the fitting residuals and the covariance matrix of the parameter estimates. The fitting residuals, such as the root mean square error, directly measure the degree of agreement between the theoretical and measured spectra; the smaller the residuals, the better the fit and the higher the confidence level. The covariance matrix of the parameter estimates reflects the uncertainty of the fitted thickness values. If the values ​​on the diagonal of the covariance matrix, i.e., the variance of the thickness parameters, are very small, it indicates that the fitting result is very stable; even with minor perturbations in the input spectrum, the result will not change drastically, resulting in high confidence. The second confidence score can be calculated using a formula that integrates the fitting residuals and parameter variances.

[0062] The fitting residuals and covariance matrices of the parameter estimates of the physical optics fitting model are calculated. Based on the fitting residuals and covariance matrices, the second confidence score corresponding to the physical optics fitting model is calculated. The root mean square error between the theoretically calculated spectrum and the actual measured spectrum output by the physical optics fitting model is extracted as the fitting residual. At the same time, the variance corresponding to the thickness parameter estimates during the model optimization process is extracted as the key element of the covariance matrix.

[0063] By introducing normalized weighting coefficients, the root mean square error and the estimated variance of the thickness parameter are multiplied by their respective weighting coefficients and then summed. One is added to the sum as the denominator, and one is used as the numerator for reciprocal calculation. In this way, a second confidence score ranging from 0 to 1 is obtained. The smaller the root mean square error and the estimated variance, the closer the second confidence score is to 1, which means that the physical optics fitting result is more reliable.

[0064] In one specific implementation, for example Where α and β are normalized weighting coefficients, To extract the estimated variance of the thickness parameter, RMSE is the root mean square error, where... ,in The intensity of the measured spectrum, The intensity of the theoretically calculated spectrum output by the physical optics fitting model, where λ is the wavelength. The parameter vector is the number of spectral data points. This is the second confidence level score.

[0065] For the third thickness inversion path, the calculation of its third confidence score focuses on the quality of the main peak of the power spectral density in the spectral periodicity analysis model. Two main metrics are evaluated: signal-to-noise ratio (SNR) and full width at half maximum (FWHM). SNR is the ratio of the height of the main peak to the average amplitude of the noise region far from the main peak. A higher SNR indicates a stronger interference signal, a more defined periodicity, and a more reliable result. FWHM is the width of the peak at half its height. A narrower FWHM indicates a sharper peak, more accurate frequency localization, and higher accuracy in thickness calculation. A peak with a high SNR and a narrow FWHM corresponds to a very high third confidence score. The SNR and FWHM of the main peak of the power spectral density in the spectral periodicity analysis model are extracted, and the third confidence score corresponding to the spectral periodicity analysis model is calculated based on these parameters. The main peak with the largest amplitude is identified in the power spectral density plot output by the spectral periodic analysis model. The ratio of the amplitude of the main peak to the average amplitude of the background noise region far from the main peak is calculated as the signal-to-noise ratio (SNR). Simultaneously, the frequency span at half the peak height of the main peak is measured as the full width at half maximum (FWHM). An adjustment coefficient is introduced, and the FWHM is multiplied by the adjustment coefficient, plus one as the denominator. The SNR is then divided by this denominator, and the resulting value is the third confidence score. A higher SNR and a narrower FWHM result in a larger third confidence score, indicating a more accurate determination of the interference frequency by the periodic analysis.

[0066] In one specific embodiment ,in For adjustment coefficients, FWHM is the main peak half-width at half-maximum, and SNR is the main peak signal-to-noise ratio. ,in The amplitude of the main peak The amplitude of the noise region far from the main peak; Yes The average value is taken as the average amplitude of the noise region far from the main peak.

[0067] In the steps described above, this application obtained a pre-classification result, three candidate thickness values, and three corresponding confidence scores. The final step is to execute the decision fusion logic to determine and output the final film thickness measurement result.

[0068] Furthermore, step S5 includes: S51: When the confidence score of any inversion path is higher than the preset threshold and the corresponding thickness candidate value meets the pre-classification result, the thickness candidate value is determined as the final film thickness measurement result. S52: When the difference between the thickness candidate values ​​of multiple inversion paths is within the preset accuracy range, the corresponding thickness candidate values ​​are weighted and averaged according to each confidence score to obtain the final film thickness measurement result.

[0069] The decision fusion logic includes several scenarios. The first scenario is when the confidence score of a certain inversion pathway is significantly higher than other pathways, for example, exceeding a preset high threshold, such as 0.9, and its calculated candidate thickness value matches the pre-classification result. In this case, the candidate thickness value is directly adopted as the final film thickness measurement result. For example, if the pre-classification result is a thick film region, and the thickness value calculated by the third pathway also falls within the thick film region, and its confidence score is as high as 0.95, far higher than the other two pathways, then the calculation result of the third pathway is directly output.

[0070] The second scenario occurs when the differences between candidate thickness values ​​from multiple inversion pathways are very small, within a preset accuracy range. This indicates that these models, based on different principles, all point to the same result. In this case, to obtain a more robust result, a weighted average of these candidate thickness values ​​is calculated based on their respective confidence scores. The weights are proportional to the confidence scores; higher confidence scores contribute more to the average. The weighted average obtained in this way is used as the final film thickness measurement result.

[0071] When the differences between candidate thickness values ​​from multiple inversion pathways are within a preset accuracy range, a weighted average is calculated based on each confidence score to obtain the final film thickness measurement result. Regarding the specific implementation of the weighted average in the decision fusion logic, differentiated fusion strategies can be implemented based on the measurement object's tolerance for consistency in results from different inversion pathways.

[0072] In batch measurements of typical consumer electronics protective films, due to the uniformity of the film structure, the candidate thickness values ​​given by various inversion pathways are usually very similar. In this application context, a linear weighted averaging method can be used. Specifically, the confidence scores output by each of the three inversion pathways are directly normalized so that the sum of the three scores equals one; the normalized value is then used as the weighting coefficient. Next, the candidate thickness value for each channel is multiplied by its corresponding weighting coefficient and summed. This conventional linear fusion method is simple and direct in computation, and when the channel results are relatively consistent, it can quickly smooth out minor computational fluctuations within each model, providing a stable and continuous film thickness measurement.

[0073] Furthermore, step S5 also includes: S53: When there are conflicts among the candidate thickness values ​​and all confidence scores are lower than the preset threshold, start the iterative fine process.

[0074] In actual measurements, more complex situations may arise, such as significant conflicts between candidate thickness values ​​and low confidence scores for all pathways, falling below the preset threshold. This typically indicates poor spectral quality of the sample or that its thickness falls within a weak region of a particular model, preventing any model from providing a high-confidence result. In such challenging situations, instead of hastily outputting an unreliable result, an iterative and refined process is initiated.

[0075] Furthermore, in step S53, initiating the iterative refinement process includes the following steps: S531: Use the first thickness candidate value output by the data-driven regression model or the third thickness candidate value output by the spectral periodic analysis model as the new initial value for the physical optics fitting model; S532: Based on the new initial values, re-execute the optimization calculation of the physical optics fitting model until the final film thickness measurement result is output.

[0076] The specific approach of this iterative refinement process is to utilize the results from other pathways to help the most accurate physical optics fitting model undergo further optimization. Specifically, the first candidate thickness value output by the data-driven regression model, or the third candidate thickness value output by the spectral periodicity analysis model, is used as a new initial value for the physical optics fitting model. Then, based on this new initial value, the global optimization calculation of the physical optics fitting model is re-executed. This process essentially gives the physical model a second chance, allowing it to start the search anew from a completely new starting point that may be closer to the true solution. This iterative process can be repeated until the physical optics fitting model converges to a high-confidence result, or the results from different pathways become consistent, ultimately outputting a refined and validated film thickness measurement result.

[0077] Secondly, an adaptive hierarchical processing full-range film thickness spectral inversion system is provided, which implements the steps of any of the above methods, and the system includes: Acquisition module 201: Acquires the reflectance spectrum data of the film to be tested, and preprocesses the reflectance spectrum data to extract frequency domain feature parameters and spectral oscillation feature parameters; Classification module 202: Based on frequency domain feature parameters and spectral oscillation feature parameters, pre-classify the thickness range of the film to be tested and obtain the pre-classification results; First calculation module 203: Constructs multiple parallel thickness inversion paths to calculate the thickness of the film to be tested and obtain multiple candidate thickness values; Second calculation module 204: Calculates a confidence score for each thickness candidate value; Result determination module 205: Combines the pre-classification results and the confidence scores of each thickness candidate value, executes the decision fusion logic, determines and outputs the final film thickness measurement result.

[0078] This system employs a modular design, decomposing the film thickness spectral inversion process into several stages: data acquisition and preprocessing, thickness magnitude pre-classification, multi-path parallel inversion calculation, confidence assessment, and decision fusion. The acquisition module 201 collects raw spectral data and extracts key feature parameters, providing a foundation for subsequent processing. The classification module 202 uses these feature parameters to make a preliminary judgment on the film thickness, achieving adaptive layered processing. The first calculation module 203 runs multiple inversion models in parallel, generating multiple candidate thickness values ​​to improve the robustness of the inversion. The second calculation module 204 performs a reliability assessment on each candidate value, quantifying its confidence level. Finally, the result determination module 205 integrates the pre-classification results and confidence scores, intelligently selecting or fusing the optimal film thickness measurement result. This effectively solves the accuracy, reliability, and efficiency problems of traditional methods in full-range film thickness measurement, achieving high-precision, non-destructive, and large-range film thickness measurement.

[0079] In some embodiments of this application, the above-described adaptive layering full-range film thickness spectral inversion system achieves accurate measurement of film thickness through the coordinated operation of its various functional modules.

[0080] Specifically, the acquisition module 201 may consist of one or more spectral sensors, a data acquisition unit, and a signal processor. For example, the spectral sensor may be a broadband spectrometer, the data acquisition unit may be an analog-to-digital converter, and the signal processor may be a digital signal processor (DSP) or an embedded microcontroller.

[0081] The classification module 202 can execute a pre-trained lightweight decision tree model or logistic regression model by one or more processors. The processor can be a central processing unit (CPU), a graphics processing unit (GPU), or an application-specific integrated circuit (ASIC).

[0082] The first computing module 203 may consist of one or more high-performance computing units that are capable of executing inversion algorithms based on data-driven regression models, physical optics fitting models, and spectral periodic analysis models in parallel.

[0083] The second calculation module 204 can be integrated with the first calculation module 203 or exist as a separate evaluation unit. This module can utilize statistical analysis, machine learning models, or preset rules to evaluate the reliability of each thickness candidate value and output the corresponding confidence score. For example, a dedicated evaluation algorithm library can be configured to calculate the confidence score based on the characteristics of different inversion paths.

[0084] The result determination module 205 can be an intelligent decision-making unit that integrates multiple decision fusion strategies. This module can adaptively select the optimal fusion strategy based on real-time evaluation results and ultimately output the film thickness measurement results via a display, data interface, or control signals.

[0085] The adaptive layered processing full-range film thickness spectral inversion system proposed in this application aims to solve the problems of insufficient accuracy, poor reliability, and low efficiency of existing film thickness measurement systems across the entire range. Traditional measurement systems often rely on a single hardware configuration or algorithm model, making it difficult to meet the measurement needs of both nanoscale thin films and micrometer-scale thick films, and are easily affected by the complexity of material optical properties and process fluctuations.

[0086] The system in this application overcomes the aforementioned challenges through its modular design and intelligent processing flow. The acquisition module efficiently collects and preprocesses spectral data, providing high-quality input for subsequent analysis. The classification module achieves intelligent pre-classification across film thickness ranges, enabling the system to adaptively apply the most suitable inversion strategy for different thickness ranges, avoiding the limitations of a single model in full-range measurements. For example, based on the pre-classification results, the system can prioritize the use of a more accurate physical-optical fitting path for thin film regions or a more robust spectral periodicity analysis path for thick film regions.

[0087] Furthermore, this application constructs multiple parallel thickness inversion pathways and evaluates their reliability through an independent confidence scoring mechanism. Compared with single or cascaded inversion systems in the prior art, this system significantly improves the robustness and efficiency of measurement through multi-path parallel computation. The result determination module combines the pre-classification results and the confidence scores of each thickness candidate value, and executes intelligent decision fusion logic. It can fully utilize the advantages of different inversion pathways and perform weighted or selective calculations based on their reliability, thereby effectively solving the periodic ambiguity problem existing in the thick film region of traditional systems and significantly improving the accuracy and reliability of full-range measurement.

[0088] In summary, the system of this application provides an advanced solution for high-precision, high-reliability, non-destructive, and efficient measurement of film thickness across the entire range through adaptive hierarchical processing, multi-path parallel inversion, and intelligent decision fusion, providing important technical support for fields such as flexible electronics and semiconductor packaging.

[0089] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for adaptively layered full-range film thickness spectral inversion, characterized in that, The method includes the following steps: S1: Acquire the reflectance spectrum data of the film to be tested, and preprocess the reflectance spectrum data to extract frequency domain feature parameters and spectral oscillation feature parameters; S2: Based on the frequency domain feature parameters and the spectral oscillation feature parameters, the thickness range of the film to be tested is pre-classified to obtain the pre-classification result; S3: Construct multiple parallel thickness inversion paths to calculate the thickness of the film layer to be measured, and obtain multiple candidate thickness values; S4: Calculate a confidence score for each of the thickness candidate values; S5: Combining the pre-classification results and the confidence scores of each candidate thickness value, execute the decision fusion logic to determine and output the final film thickness measurement result.

2. The adaptive layered processing full-range film thickness spectral inversion method according to claim 1, characterized in that, Step S1 includes: S11: Obtain the reflectance spectrum data of the film to be tested; S12: Perform empirical mode decomposition on the reflectance spectral data, extract the interference signal, and remove noise and baseline drift to obtain preprocessed spectral data, wherein the preprocessed spectral data is a spectral curve dominated by the interference signal; S13: Convert the preprocessed spectral data to the wavenumber domain, calculate the normalized power spectral density of the preprocessed spectral data in the wavenumber domain, and use it as the frequency domain characteristic parameter; S14: Calculate the first or second derivative of the preprocessed spectral curve, count the number of data points or local extrema of the first or second derivative that cross zero, and define them as the spectral oscillation characteristic parameters.

3. The adaptive layered processing full-range film thickness spectral inversion method according to claim 1, characterized in that, Step S2 includes: S21: Based on the frequency domain feature parameters and the spectral oscillation feature parameters, a lightweight decision tree or logistic regression model is used to preclassify the thickness range of the film to be tested into a thin film region, a transition region, or a thick film region, which is taken as the preclassification result.

4. The adaptive layered processing full-range film thickness spectral inversion method according to claim 1, characterized in that, In step S3, there are three thickness inversion pathways: a first thickness inversion pathway based on a data-driven regression model, a second thickness inversion pathway based on a physical optics fitting model, and a third thickness inversion pathway based on a spectral periodicity analysis model.

5. The adaptive layered processing full-range film thickness spectral inversion method according to claim 4, characterized in that, Step S3 includes: S31: Use a data-driven regression model to perform regression calculations on the preprocessed spectral data to obtain an initial estimate, which serves as the first thickness candidate value. The data-driven regression model is a pre-trained neural network model or a gradient boosting tree model. S32: Using the initial estimated value as the initial value of the physical optics fitting model, for the thin film region in the pre-classification result, perform global optimization in the physical optics fitting model to determine the second thickness candidate value corresponding to the physical optics fitting model; S33: For the thick film region in the pre-classification results, the periodogram analysis path of the spectral period analysis model is used to perform fast Fourier transform or Lomb-Scargle periodogram analysis on the preprocessed spectral data, extract the main frequency component of the interference signal to calculate the thickness estimate, and use it as the third thickness candidate value. The thickness candidate value includes the first thickness candidate value, the second thickness candidate value, and the third thickness candidate value.

6. The adaptive layered processing full-range film thickness spectral inversion method according to claim 5, characterized in that, Step S4 includes: S41: Obtain the historical accuracy of the data-driven regression model on the validation set, and calculate the first confidence score corresponding to the data-driven regression model by combining it with the current output probability distribution; S42: Calculate the fitting residuals of the physical optics fitting model and the covariance matrix of the parameter estimates, and calculate the second confidence score corresponding to the physical optics fitting model based on the fitting residuals and the covariance matrix. S43: Extract the signal-to-noise ratio and full width at half maximum (FWHM) of the main peak of the power spectral density in the spectrum periodic analysis model, and calculate the third confidence score corresponding to the spectrum periodic analysis model based on the signal-to-noise ratio and the FWHM.

7. The adaptive layered processing full-range film thickness spectral inversion method according to claim 1, characterized in that, Step S5 includes: S51: When the confidence score of any of the inversion pathways is higher than a preset threshold, and the corresponding candidate thickness value conforms to the pre-classification result, the candidate thickness value is determined as the final film thickness measurement result. S52: When the difference between the candidate thickness values ​​of the multiple inversion paths is within a preset accuracy range, the corresponding candidate thickness values ​​are weighted and averaged according to each confidence score to obtain the final film thickness measurement result.

8. The adaptive layered processing full-range film thickness spectral inversion method according to claim 7, characterized in that, Step S5 also includes: S53: When there is a conflict among the candidate thickness values ​​and the confidence scores of each value are lower than the preset threshold, the iterative fine process is initiated.

9. The adaptive layered processing full-range film thickness spectral inversion method according to claim 8, characterized in that, In step S53, initiating the iterative refinement process includes the following steps: S531: Use the first thickness candidate value output by the data-driven regression model or the third thickness candidate value output by the spectral periodic analysis model as the new initial value for the physical optics fitting model; S532: Based on the new initial values, re-execute the optimization calculation of the physical optics fitting model until the final film thickness measurement result is output.

10. A full-range film thickness spectral inversion system with adaptive layered processing, characterized in that, The system is used to implement the steps of the method according to any one of claims 1-9, the system comprising: Acquisition module: Acquires the reflectance spectrum data of the film to be tested, and preprocesses the reflectance spectrum data to extract frequency domain feature parameters and spectral oscillation feature parameters; Classification module: Based on the frequency domain feature parameters and the spectral oscillation feature parameters, pre-classify the thickness range of the film to be tested to obtain the pre-classification result; First calculation module: Constructs multiple parallel thickness inversion paths to calculate the thickness of the film layer to be measured, and obtains multiple candidate thickness values; The second calculation module calculates a confidence score for each of the candidate thickness values. Result determination module: Combining the pre-classification results and the confidence scores of each candidate thickness value, the module executes decision fusion logic to determine and output the final film thickness measurement result.

Images