Photovoltaic cell film thickness real-time monitoring method and system

Abstract

The invention provides a photovoltaic cell film thickness real-time monitoring method and system, and relates to the technical field of photovoltaic cells, and the method comprises the steps: setting a multi-channel optical fiber sensor array, enabling a receiving end optical fiber to collect a reflected light signal of a film, and enabling the reflected light signal to be connected into a photoelectric detector, and converting the reflected light signal into a reflected light intensity signal; preprocessing the obtained reflected light intensity signal, estimating the phase difference between adjacent cycles by using a phase unwrapping algorithm to obtain continuous thin film thickness change information, introducing a refractive index online estimation mechanism to compensate the refractive index change of the thin film material, and counting the oscillation amplitude in each cycle to obtain the refractive index of the thin film material. The real-time thickness value of the thin film is obtained in combination with compensation on the change of the refractive index; the thin film thickness distribution of the surface of the whole substrate is obtained through an intelligent interpolation algorithm, a thin film three-dimensional thickness distribution diagram is generated based on the obtained thin film thickness distribution through the graph drawing and rendering technology, and the dynamic evolution of the thin film growth process is presented in real time.

Description

Technical Field

[0001] The invention relates to photovoltaic cell technology, and in particular to a method and system for real-time monitoring of the thickness of a photovoltaic cell film. Background Art

[0002] Photovoltaic cells are key devices that convert solar energy into electrical energy. Their performance and conversion efficiency depend to a large extent on the thickness and uniformity of photovoltaic material films. In the process of preparing photovoltaic cell films, real-time and accurate monitoring of film thickness is of great significance for controlling the deposition process and optimizing process parameters. With the rapid development of the photovoltaic industry, large-scale production of thin-film solar cells has put forward higher requirements for the quality control of thin-film deposition processes. As one of the key parameters affecting cell performance, the uniformity and consistency of film thickness directly determine the photoelectric conversion efficiency of the cell. Traditional film thickness measurement methods, such as ellipsometry and step measurement, mostly adopt offline or intermittent measurement modes, which makes it difficult to achieve real-time monitoring and dynamic feedback of the film growth process. In addition, these methods can usually only obtain thickness information of local areas and cannot fully reflect the thickness distribution of the entire substrate surface. Therefore, there is an urgent need for a monitoring method and system that can measure film thickness in real time, in situ, and at multiple points to meet the dynamic optimization control needs of the photovoltaic cell film preparation process. Summary of the invention

[0003] The embodiments of the present invention provide a method and system for real-time monitoring of the thickness of a photovoltaic cell film, which can solve the problems in the prior art.

[0004] According to a first aspect of the embodiments of the present invention,

[0005] A method for real-time monitoring of the thickness of a photovoltaic cell film is provided, comprising:

[0006] A multi-channel optical fiber sensor array is provided, wherein each optical fiber sensing channel includes an emitting optical fiber and a receiving optical fiber, the emitting optical fiber is coupled to a monochromatic laser to generate a collimated light beam, the receiving optical fiber collects a reflected light signal and connects it to a photodetector, a collimated light beam is generated by a monochromatic laser, and is irradiated to the surface of a thin film grown on a substrate through the emitting optical fiber, while the receiving optical fiber collects a reflected light signal of the thin film, and connects the reflected light signal to a photodetector to convert it into a reflected light intensity signal;

[0007] The obtained reflected light intensity signal is preprocessed, and the preprocessed reflected light intensity signal is subjected to time-frequency analysis to extract characteristic parameters of the reflected light intensity signal, including oscillation frequency and oscillation amplitude, and a mapping relationship between the oscillation frequency and oscillation amplitude and the film thickness growth is established. The phase unwrapping algorithm is used to estimate the phase difference between adjacent periods to obtain continuous film thickness change information, and a refractive index online estimation mechanism is introduced to compensate for the refractive index change of the film material. By counting the oscillation amplitude in each period and combining the compensation for the refractive index change, the real-time thickness value of the film is obtained;

[0008] Using the real-time thickness value of the film and the pre-established numerical model of film growth, the film thickness distribution on the entire substrate surface is obtained through an intelligent interpolation algorithm. Based on the obtained film thickness distribution, a three-dimensional film thickness distribution map is generated using graphics drawing and rendering technology to present the dynamic evolution of the film growth process in real time.

[0009] In an optional embodiment,

[0010] The phase unwrapping algorithm is used to estimate the phase difference between adjacent cycles, including:

[0011] Performing statistical analysis on the phase differences of adjacent cycles, calculating the mean and variance of the phase differences, estimating the distribution characteristics of the phase differences according to the mean and variance, and obtaining an adaptive threshold using an adaptive threshold selection strategy based on the distribution characteristics of the phase differences, wherein the adaptive threshold selection strategy is one or more combinations of a threshold based on a mean and a standard deviation, a threshold based on a quantile, and a threshold based on entropy;

[0012] Constructing an undirected weighted phase difference network graph, wherein vertices represent cycles, edges represent phase differences between adjacent cycles, and edge weights are reliability metrics of the phase differences, wherein the reliability metrics comprehensively consider the amplitude, time correlation, and frequency domain characteristics of the phase differences, and performing edge detection on the undirected weighted phase difference network graph according to the obtained adaptive threshold to obtain a binary network graph;

[0013] The maximum flow minimum cut is searched in the binary network graph, and the network graph is divided into two parts: the source point set and the sink point set. The cut edge connects the two sets. The sum of the capacity of the cut edge is the capacity of the minimum cut. The cut edge corresponds to the position of the phase jump. An adaptive phase compensation mechanism is introduced for the vertices at both ends of the cut edge corresponding to the maximum flow minimum cut. The phase compensation value is adaptively determined to achieve continuous phase reconstruction and obtain the final phase difference between adjacent cycles.

[0014] In an optional embodiment,

[0015] An adaptive phase compensation mechanism is introduced for the vertices at both ends of the cut edge corresponding to the maximum flow minimum cut, and the phase compensation value is adaptively determined to achieve continuous phase reconstruction. The final phase difference between adjacent cycles includes:

[0016] Extract the phase difference of the period where the vertices at both ends of the cut edge are located and the phase difference of the adjacent period, calculate the weighted average of the phase difference of the adjacent period, and at the same time, measure the reliability of the phase difference of the adjacent period according to the amplitude, time correlation and frequency domain characteristics of the phase difference of the adjacent period, use the reliability as the weight of the phase difference of the adjacent period, dynamically adjust the weight through the adaptive weight strategy, obtain the phase compensation value, and compensate the phase of the vertices at both ends of the cut edge to correct the phase jump;

[0017] Repeat the phase compensation for all the cut edges in the binary network graph to obtain a continuous unwrapped phase, realize continuous phase reconstruction, and obtain the final phase difference between adjacent cycles;

[0018] The calculation formula of the phase compensation value is as follows:

[0019]

[0020] Among them, △η c represents the phase compensation value, i represents the edge cutting start index, j represents the edge cutting end index, M represents the number of adjacent cycles, and w k represents the weight of the phase difference of the kth cycle, △η k represents the phase difference of the kth cycle, △η i Indicates the phase difference of the period where the cutting edge starting point is located, △η j Indicates the phase difference of the cycle where the end point of the cutting edge is located.

[0021] In an optional embodiment,

[0022] The refractive index online estimation mechanism is introduced to compensate for the refractive index change of the thin film material. By counting the oscillation amplitude in each cycle and combining the compensation for the refractive index change, the real-time thickness value of the film is obtained, including:

[0023] Using the characteristic parameters of the reflected light intensity signal extracted in advance as input and the refractive index of the thin film material to be measured as output, a refractive index estimation model is established and trained;

[0024] During the thin film deposition process, the reflected light intensity signal is collected in real time, and the collected reflected light intensity signal is decomposed into multiple scales using wavelet transform to obtain sub-band signals of different frequency bands. The characteristic parameters of each sub-band signal are extracted and input into the trained refractive index estimation model to obtain the refractive index estimation value of each scale, and the refractive index estimation value of each scale is combined to obtain the final refractive index estimation value;

[0025] The final estimated refractive index value is combined with the phase difference of adjacent periods obtained by the phase unwrapping algorithm to calculate the film thickness increment in each period, and the initial film thickness estimation value is obtained by accumulating the film thickness increment in each period;

[0026] The Kalman filter algorithm is used to filter and optimize the initial film thickness estimate, and a state space model of the film thickness is established. The filter parameters are adaptively adjusted according to the change of the refractive index estimate. The iteration is repeated until the preset maximum number of iterations is reached to obtain the final real-time thickness value of the film.

[0027] In an optional embodiment,

[0028] The final estimated refractive index value is combined with the phase difference of adjacent periods obtained by the phase unwrapping algorithm to calculate the film thickness increment in each period. By accumulating the film thickness increment in each period, the calculation formula for the initial film thickness estimate is obtained as follows:

[0029]

[0030] in, represents the initial film thickness estimate, K represents the number of cycles, λ represents the wavelength of the incident light, m ​​represents the number of scales, and σ a represents the attention weight of the a-th scale, represents the refractive index estimate of the ath scale, β represents the nonlinear fusion coefficient, △η k represents the phase difference of the kth cycle, γ represents the nonlinear compensation coefficient, σ b represents the attention weight of the b-th scale, represents the refractive index estimate at the b-th scale.

[0031] In an optional embodiment,

[0032] Using the obtained real-time thickness value of the film, combined with the pre-established numerical model of film growth, the film thickness distribution on the entire substrate surface is obtained through an intelligent interpolation algorithm, including:

[0033] A thin film growth numerical model including deposition rate, substrate temperature and chamber pressure is established to describe the temporal and spatial evolution of film thickness. The thin film growth numerical model is discretized using the finite element method to divide the thin film growth region into a plurality of discrete point grids at equal intervals, and the film thickness at each discrete point is interpolated and approximated using shape functions. The thin film growth numerical model is converted into a set of algebraic equations by a weighted residual method, and the algebraic equations are solved to obtain the numerical solution of the film thickness at each discrete point.

[0034] Using the obtained real-time thickness value of the film, parameter identification and dynamic correction are performed on the numerical solution of the film thickness at each discrete point to obtain a corrected numerical solution of the film thickness;

[0035] Based on the corrected numerical solution of film thickness, a deep learning network model is constructed with the spatial coordinates of discrete points in the film growth area as input and the numerical solution of film thickness at the discrete points as output, and a training data set including the spatial coordinates of the discrete points and the corresponding numerical solution of film thickness is constructed, the deep learning network model is trained using the training data set, and the network structure and training hyperparameters are optimized to obtain a trained film thickness prediction model;

[0036] The spatial coordinates of any position in the film growth area are input into the trained film thickness prediction model to obtain the predicted value of the film thickness at the current position. The predicted values ​​of the film thickness at all positions in the film growth area and the corresponding numerical solutions of the film thickness are weighted averaged to obtain the film thickness distribution on the entire substrate surface.

[0037] In an optional embodiment,

[0038] Using the obtained real-time thickness value of the film, the numerical solution of the film thickness at each discrete point is parameter identified and dynamically corrected, and the corrected numerical solution of the film thickness is obtained, including:

[0039] Preprocessing the obtained real-time thickness value of the film to obtain a preprocessed real-time thickness value, wherein the preprocessing includes outlier removal, data smoothing, and interpolation point filling;

[0040] The preprocessed real-time thickness values ​​are fused using a preset spatial correlation model to obtain fused real-time thickness values, and an error objective function is constructed based on the fused real-time thickness values, wherein the error objective function includes a square sum term and a regularization term between the fused real-time thickness value and the numerical solution of the film thickness at each discrete point;

[0041] A nonlinear least squares optimization algorithm is used to solve the error objective function to obtain the optimal solution of the parameters to be identified in the thin film growth numerical model, and the thin film growth numerical model is corrected using the optimal solution, and the finite element solution is re-performed to obtain a corrected numerical solution for the film thickness.

[0042] According to a second aspect of the embodiments of the present invention,

[0043] A photovoltaic cell film thickness real-time monitoring system is provided, comprising:

[0044] The first unit is used to set up a multi-channel optical fiber sensor array, wherein each optical fiber sensing channel includes a transmitting end optical fiber and a receiving end optical fiber, the transmitting end optical fiber is coupled to a monochromatic laser to generate a collimated light beam, the receiving end optical fiber collects a reflected light signal and connects it to a photodetector, a collimated light beam is generated by a monochromatic laser, and is irradiated to the surface of a thin film grown on a substrate through the transmitting end optical fiber, and at the same time, the receiving end optical fiber collects a reflected light signal of the thin film, and connects the reflected light signal to the photodetector to convert it into a reflected light intensity signal;

[0045] The second unit is used to pre-process the obtained reflected light intensity signal, perform time-frequency analysis on the pre-processed reflected light intensity signal, extract characteristic parameters of the reflected light intensity signal, including oscillation frequency and oscillation amplitude, establish a mapping relationship between the oscillation frequency and oscillation amplitude and the film thickness growth, use the phase unwrapping algorithm to estimate the phase difference between adjacent periods, obtain continuous film thickness change information, and introduce a refractive index online estimation mechanism to compensate for the refractive index change of the film material. By counting the oscillation amplitude in each period and combining the compensation for the refractive index change, the real-time thickness value of the film is obtained;

[0046] The third unit is used to utilize the real-time thickness value of the film, combined with the pre-established numerical model of film growth, to obtain the film thickness distribution on the entire substrate surface through an intelligent interpolation algorithm. Based on the obtained film thickness distribution, a three-dimensional film thickness distribution map is generated using graphics drawing and rendering technology to present the dynamic evolution of the film growth process in real time.

[0047] According to a third aspect of the embodiments of the present invention,

[0048] An electronic device is provided, comprising:

[0049] processor;

[0050] a memory for storing processor-executable instructions;

[0051] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0052] A fourth aspect of the embodiments of the present invention is:

[0053] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0054] In this embodiment, by setting a multi-channel optical fiber sensor array, the film thickness information of multiple positions can be obtained simultaneously, and the thickness distribution of the entire substrate surface can be fully reflected, realizing the real-time, in-situ, and multi-point monitoring of the photovoltaic cell film thickness. Through the time-frequency analysis and feature extraction of the reflected light intensity signal, the mapping relationship between the oscillation frequency, oscillation amplitude and film thickness growth is established. Combined with the phase unwrapping algorithm and the refractive index online estimation mechanism, the continuous and high-precision monitoring of the film thickness change is realized. Using the intelligent interpolation algorithm, the real-time thickness value of the discrete measuring point is extended to the entire substrate surface, and the complete film thickness distribution information is obtained. The three-dimensional thickness distribution map of the film is generated by graphics drawing and rendering technology, and the dynamic evolution of the film growth process is intuitively presented. Real-time and dynamic thickness monitoring information can be used to optimize the film growth process parameters, and the deposition process is controlled by real-time feedback to ensure that the film thickness and uniformity meet the design requirements, improve the performance and conversion efficiency of photovoltaic cells, and realize the intelligence and automation of film thickness monitoring, reduce the influence of manual operation and subjective judgment, and improve measurement efficiency and reliability. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 A schematic diagram of the process of a method for real-time monitoring of the thickness of a photovoltaic cell thin film according to an embodiment of the present invention;

[0056] Figure 2 It is a schematic structural diagram of a photovoltaic cell film thickness real-time monitoring system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0059] Figure 1 FIG. 1 is a flow chart of a method for real-time monitoring of the thickness of a photovoltaic cell film according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0060] S101. Set up a multi-channel fiber optic sensor array, wherein each fiber optic sensing channel includes a transmitting end fiber and a receiving end fiber, the transmitting end fiber is coupled to a monochromatic laser to generate a collimated light beam, the receiving end fiber collects a reflected light signal and connects it to a photodetector, a monochromatic laser is used to generate a collimated light beam, and the collimated light beam is irradiated to the surface of a thin film grown on a substrate through the transmitting end fiber, and at the same time, the receiving end fiber collects the reflected light signal of the thin film, and connects the reflected light signal to a photodetector to convert it into a reflected light intensity signal.

[0061] The design and integration of the fiber optic sensor array includes the use of finite element analysis to simulate the illumination uniformity and sampling blind area under different sensor layout schemes based on the geometric structure of the sputtering deposition cavity and the size of the substrate. Through iterative optimization, the optimal fiber optic sensor array arrangement is determined to achieve full coverage scanning of the substrate surface.

[0062] Select appropriate fiber models and optical parameters, such as numerical aperture, fiber diameter, and cutoff wavelength, to meet the spectral range and light intensity requirements for film thickness measurement. Design fiber collimators and focusing lenses to control the spot size to match the characteristic scale of the film, and improve fiber coupling efficiency and measurement sensitivity.

[0063] According to the reflective spectral characteristics of the thin film material, select a suitable photodetector, such as a silicon photodiode, avalanche photodiode, or photomultiplier tube, etc. Evaluate the detector's response speed, gain linearity, and dynamic range to meet the speed and accuracy requirements of real-time measurement.

[0064] Design an integrated packaging structure of the fiber array and substrate support frame to ensure the relative position of the sensor and the substrate surface is stable. Develop an automatic alignment and calibration mechanism to achieve precise positioning of the fiber probe and sputtering source, reducing installation errors and measurement drift.

[0065] An array of multi-channel fiber optic sensors is used as the measurement probe. Each fiber optic sensing channel includes a transmitting fiber and a receiving fiber. The transmitting fiber is coupled with a monochromatic laser to generate a collimated light beam, and the receiving fiber collects the reflected light signal and connects it to the photodetector. By using a monochromatic laser to generate a collimated light beam, and irradiating the surface of the film grown on the substrate through the transmitting fiber, while the receiving fiber collects the reflected light signal of the film, and connects the reflected light signal to the photodetector to convert it into a reflected light intensity signal, real-time monitoring of the film growth process is achieved.

[0066] In this embodiment, by optimizing the arrangement of the optical fiber sensor array, full coverage scanning of the substrate surface is achieved to ensure monitoring and measurement of the entire film surface. Finite element analysis is used to simulate the uniformity of illumination under different sensor layout schemes to ensure uniform distribution of the light spot on the substrate surface and improve measurement accuracy. By designing the optimal sensor layout, the existence of sampling blind areas is reduced to ensure data collection on the entire substrate surface and avoid information loss. Fiber collimators and focusing lenses are designed to control the spot size to match the characteristic scale of the film, improve the fiber coupling efficiency, and enhance the measurement sensitivity. Appropriate fiber models and optical parameters, such as numerical aperture, fiber diameter, and cutoff wavelength are selected to meet the spectral range and light intensity requirements of film thickness measurement to ensure measurement accuracy and reliability. Multi-channel fiber optic sensors are used to realize real-time monitoring of the film growth process, and real-time monitoring and analysis of film properties are realized by collecting and processing reflected light signals.

[0067] S102. Preprocess the obtained reflected light intensity signal, perform time-frequency analysis on the preprocessed reflected light intensity signal, extract characteristic parameters of the reflected light intensity signal, including oscillation frequency and oscillation amplitude, establish a mapping relationship between the oscillation frequency and oscillation amplitude and the film thickness growth, use the phase unwrapping algorithm to estimate the phase difference between adjacent periods, obtain continuous film thickness change information, and introduce a refractive index online estimation mechanism to compensate for the refractive index change of the film material. By counting the oscillation amplitude in each period and combining the compensation for the refractive index change, the real-time thickness value of the film is obtained.

[0068] Among them, the collected raw signal is preprocessed, including background noise removal, baseline drift correction, signal smoothing, etc. Commonly used preprocessing methods include moving average filtering, wavelet denoising, polynomial fitting, etc. The preprocessed reflected light intensity signal has a higher signal-to-noise ratio and stability, laying the foundation for subsequent time-frequency analysis.

[0069] Perform time-frequency analysis on the preprocessed reflected light intensity signal to extract the characteristic parameters of periodic oscillation in the signal. Use time-frequency analysis methods such as short-time Fourier transform (STFT) or wavelet transform to convert the time domain signal into time-frequency domain representation. By analyzing the energy distribution and frequency changes in the time-frequency domain, extract the oscillation frequency and oscillation amplitude of the reflected light intensity signal.

[0070] Through theoretical analysis and experimental calibration, the mapping relationship between oscillation frequency and oscillation amplitude and film thickness is established. According to the physical mechanism of the film growth process, such as molecular beam epitaxy and magnetron sputtering, the theoretical model of the reflected light intensity signal is derived. By fitting the experimental data, the quantitative relationship between the characteristic parameters and the thickness is obtained. The accuracy of the mapping relationship directly affects the accuracy of thickness monitoring, and it is necessary to comprehensively consider factors such as the film material properties and growth conditions.

[0071] In the process of feature parameter extraction, due to the periodicity of the reflected light intensity signal, there is a phase difference between adjacent periods. In order to accurately estimate the phase difference, the phase unwrapping algorithm is introduced. Commonly used phase unwrapping algorithms include least squares method, iterative algorithm, branch cutting method, etc. By analyzing the phase changes of adjacent periods, the starting phase of each period is estimated to achieve continuous phase unwrapping. The phase unwrapping algorithm can eliminate the influence of phase jump on thickness estimation and improve the continuity and stability of measurement.

[0072] The refractive index of the thin film material is an important parameter that affects the reflected light intensity signal and may change during the growth process. In order to compensate for the impact of refractive index changes on thickness monitoring, an online refractive index estimation mechanism is introduced. By analyzing the amplitude and phase information of the reflected light intensity signal and combining the dispersion relation of the thin film material, the refractive index of the film is estimated in real time. Algorithms such as recursive least squares method and Kalman filtering can be used to dynamically update the refractive index estimate.

[0073] According to the established mapping relationship between characteristic parameters and thickness, the oscillation amplitude in each cycle is counted, and combined with the refractive index estimation results, real-time film thickness monitoring data is obtained. In order to further improve the estimation accuracy, multi-sensor data fusion technology can be introduced. By installing multiple fiber optic sensors at different positions or angles, multiple reflected light intensity signals can be obtained. Using data fusion algorithms such as weighted average and Kalman filtering, the estimation results of multiple sensors are optimized and combined to reduce the measurement error of a single sensor and improve the reliability of thickness monitoring.

[0074] The estimated real-time thickness monitoring data is fed back to the film growth control system to achieve closed-loop feedback control. By comparing the deviation between the real-time thickness and the target thickness, the growth process parameters such as substrate temperature, sputtering power, gas flow rate, etc. are dynamically adjusted to achieve precise control of the film thickness.

[0075] In an optional embodiment,

[0076] The phase unwrapping algorithm is used to estimate the phase difference between adjacent cycles, including:

[0077] Performing statistical analysis on the phase differences of adjacent cycles, calculating the mean and variance of the phase differences, estimating the distribution characteristics of the phase differences according to the mean and variance, and obtaining an adaptive threshold using an adaptive threshold selection strategy based on the distribution characteristics of the phase differences, wherein the adaptive threshold selection strategy is one or more combinations of a threshold based on a mean and a standard deviation, a threshold based on a quantile, and a threshold based on entropy;

[0078] Constructing an undirected weighted phase difference network graph, wherein vertices represent cycles, edges represent phase differences between adjacent cycles, and edge weights are reliability metrics of the phase differences, wherein the reliability metrics comprehensively consider the amplitude, time correlation, and frequency domain characteristics of the phase differences, and performing edge detection on the undirected weighted phase difference network graph according to the obtained adaptive threshold to obtain a binary network graph;

[0079] The maximum flow minimum cut is searched in the binary network graph, and the network graph is divided into two parts: the source point set and the sink point set. The cut edge connects the two sets. The sum of the capacity of the cut edge is the capacity of the minimum cut. The cut edge corresponds to the position of the phase jump. An adaptive phase compensation mechanism is introduced for the vertices at both ends of the cut edge corresponding to the maximum flow minimum cut. The phase compensation value is adaptively determined to achieve continuous phase reconstruction and obtain the final phase difference between adjacent cycles.

[0080] For example, first, the phase difference between adjacent cycles is statistically analyzed. By calculating the mean and variance of the phase difference, the distribution characteristics of the phase difference can be estimated. The mean reflects the central trend of the phase difference, and the variance indicates the degree of discreteness of the phase difference. Based on the mean and variance, a probability density function of the phase difference can be constructed. Common probability density functions include Gaussian distribution, Laplace distribution, etc. Through the probability density function, the distribution law of the phase difference can be characterized.

[0081] After estimating the distribution characteristics of the phase difference, an adaptive threshold selection strategy is used to determine the threshold. The adaptive threshold selection strategy dynamically adjusts the threshold according to the distribution characteristics of the phase difference to adapt to different noise levels and phase change characteristics. Commonly used adaptive threshold selection strategies include thresholds based on mean and standard deviation, thresholds based on quantiles, and thresholds based on entropy. The threshold based on mean and standard deviation takes into account the central tendency and dispersion of the phase difference, and determines the threshold by adding the mean to the multiple of the standard deviation. The quantile-based threshold selects a specific quantile of the phase difference distribution as the threshold, which can adapt to asymmetric distribution situations. The entropy-based threshold uses the entropy of the phase difference distribution to measure uncertainty and selects the optimal threshold by maximizing the entropy.

[0082] After determining the adaptive threshold, an undirected weighted phase difference network graph is constructed. The vertices in the network graph represent cycles, and the edges represent the phase differences between adjacent cycles. In order to better measure the reliability of the phase difference, a weight is assigned to each edge. The calculation of the weight comprehensively considers the amplitude, time correlation, and frequency domain characteristics of the phase difference. The amplitude reflects the size of the phase difference. The larger the amplitude, the higher the reliability of the phase difference. The time correlation measures the continuity of the phase difference in the time domain. The stronger the correlation, the higher the reliability of the phase difference. The frequency domain characteristics characterize the distribution of the phase difference in the frequency domain. The more concentrated the frequency domain distribution, the higher the reliability of the phase difference. By combining these three factors, the reliability measure of the phase difference is obtained as the weight of the edge.

[0083] The undirected weighted phase difference network graph is edge detected using an adaptive threshold to obtain a binary network graph. The purpose of edge detection is to find the location of the phase difference jump, that is, the edge where the phase difference exceeds the threshold. By comparing the edge weight with the threshold, the edge that exceeds the threshold is marked as 1, indicating that there is a phase jump; the edge that does not exceed the threshold is marked as 0, indicating that the phase is continuous. After edge detection, a binary network graph is obtained, in which the edge marked as 1 indicates the potential phase jump location.

[0084] After obtaining the binary network graph, the next step is to search for the maximum flow minimum cut in the graph. The goal of the maximum flow minimum cut algorithm is to divide the binary network graph into two subgraphs, the source and sink, so that the capacity of the cut edge between the two subgraphs is minimized. The cut edge represents the phase jump between the subgraphs, and the minimum capacity means the total number of phase jumps is the least.

[0085] In order to achieve maximum flow minimum cut, we first need to define the source and sink in the binary network graph. The source represents the starting point of phase continuity, and the sink represents the ending point of phase continuity. Usually, the first cycle is set as the source, and the last cycle is set as the sink. Then, add a directed edge from the source to the sink in the network graph, and set the capacity to infinity. The role of this edge is to ensure that the network graph meets the conditions of the maximum flow minimum cut theorem.

[0086] Next, the greedy algorithm is used to solve the maximum flow problem. The greedy algorithm gradually increases the network flow by continuously searching for augmenting paths until no augmenting paths can be found. An augmenting path is a path from a source to a sink where the residual capacity of the edges along the path is greater than zero. By increasing the flow on the augmenting path, the total flow of the network can be increased. The process of searching for augmenting paths and increasing the flow is repeated until no augmenting paths can be found, at which point the maximum flow is obtained.

[0087] According to the maximum flow minimum cut theorem, the maximum flow corresponds to the minimum cut. The minimum cut divides the network graph into two parts: the source point set and the sink point set. The cut edge connects the two sets. The sum of the capacities of the cut edges is the capacity of the minimum cut. In the binary network graph, the cut edge corresponds to the position of the phase jump.

[0088] After obtaining the maximum flow minimum cut, an adaptive phase compensation mechanism is introduced for the vertices at both ends of the cut edge. The purpose of phase compensation is to correct phase jumps and achieve continuous phase reconstruction. The phase compensation value is adaptively determined for the vertices at both ends of the cut edge. The calculation of the phase compensation value takes into account the phase change trend of adjacent cycles, and estimates the phase compensation value of the current cycle by fitting the phase difference of adjacent cycles.

[0089] In an optional embodiment,

[0090] An adaptive phase compensation mechanism is introduced for the vertices at both ends of the cut edge corresponding to the maximum flow minimum cut, and the phase compensation value is adaptively determined to achieve continuous phase reconstruction. The final phase difference between adjacent cycles includes:

[0091] Extract the phase difference of the period where the vertices at both ends of the cut edge are located and the phase difference of the adjacent period, calculate the weighted average of the phase difference of the adjacent period, and at the same time, measure the reliability of the phase difference of the adjacent period according to the amplitude, time correlation and frequency domain characteristics of the phase difference of the adjacent period, use the reliability as the weight of the phase difference of the adjacent period, dynamically adjust the weight through the adaptive weight strategy, obtain the phase compensation value, and compensate the phase of the vertices at both ends of the cut edge to correct the phase jump;

[0092] Repeat the phase compensation for all the cut edges in the binary network graph to obtain a continuous unwrapped phase, realize continuous phase reconstruction, and obtain the final phase difference between adjacent cycles;

[0093] The calculation formula of the phase compensation value is as follows:

[0094]

[0095] Among them, △η c represents the phase compensation value, i represents the edge cutting start index, j represents the edge cutting end index, M represents the number of adjacent cycles, and w k represents the weight of the phase difference of the kth cycle, △η k represents the phase difference of the kth cycle, △η i Indicates the phase difference of the period where the cutting edge starting point is located, △η j Indicates the phase difference of the cycle where the end point of the cutting edge is located.

[0096] Optionally, in practical applications, in order to improve the efficiency of the algorithm, an approximate algorithm can be used to solve the maximum flow minimum cut problem, such as the push-relabel algorithm. This algorithm updates the network flow through local operations, avoiding the process of repeatedly searching for augmented paths, and has lower time complexity. In addition, in the process of phase compensation, an adaptive weight strategy can be introduced to dynamically adjust the weight according to the reliability of the phase difference between adjacent cycles. For example, the reliability of the phase difference can be measured using indicators such as the amplitude, time correlation, and frequency domain characteristics of the phase difference, and used as the basis for the weight. In this way, the phase compensation value can be estimated more accurately and the accuracy of phase reconstruction can be improved.

[0097] In this embodiment, by statistically analyzing the mean, variance, and distribution characteristics of the phase difference of adjacent cycles, an adaptive threshold selection strategy such as a mean standard deviation threshold, a quantile threshold, and an entropy threshold is adopted, and a reasonable threshold can be adaptively determined according to the specific situation of the data, thereby improving the applicability and robustness of the threshold selection. Taking into account a variety of characteristics such as the amplitude, time correlation, and frequency domain characteristics of the phase difference, a reliability metric for the phase difference is constructed, which can more comprehensively evaluate the quality and credibility of the phase difference and lay the foundation for subsequent phase unpacking. The phase difference problem is modeled as an undirected weighted network graph, where the vertex represents the cycle, the edge represents the phase difference, and the edge weight is a reliability metric. This network topology model can well characterize the phase relationship and reliability difference between cycles. Using the maximum flow minimum cut theory in graph theory, the position of the phase jump can be efficiently searched in the network graph, and the network can be divided into two sets to create conditions for phase reconstruction. For the detected phase jump position, an adaptive phase compensation mechanism is introduced to adaptively determine a reasonable compensation value according to the specific situation, thereby achieving accurate continuous phase reconstruction.

[0098] In an optional embodiment,

[0099] The refractive index online estimation mechanism is introduced to compensate for the refractive index change of the thin film material. By counting the oscillation amplitude in each cycle and combining the compensation for the refractive index change, the real-time thickness value of the film is obtained, including:

[0100] Using the characteristic parameters of the reflected light intensity signal extracted in advance as input and the refractive index of the thin film material to be measured as output, a refractive index estimation model is established and trained;

[0101] During the thin film deposition process, the reflected light intensity signal is collected in real time, and the collected reflected light intensity signal is decomposed into multiple scales using wavelet transform to obtain sub-band signals of different frequency bands. The characteristic parameters of each sub-band signal are extracted and input into the trained refractive index estimation model to obtain the refractive index estimation value of each scale, and the refractive index estimation value of each scale is combined to obtain the final refractive index estimation value;

[0102] The final estimated refractive index value is combined with the phase difference of adjacent periods obtained by the phase unwrapping algorithm to calculate the film thickness increment in each period, and the initial film thickness estimation value is obtained by accumulating the film thickness increment in each period;

[0103] The Kalman filter algorithm is used to filter and optimize the initial film thickness estimate, and a state space model of the film thickness is established. The filter parameters are adaptively adjusted according to the change of the refractive index estimate. The iteration is repeated until the preset maximum number of iterations is reached to obtain the final real-time thickness value of the film.

[0104] Exemplarily, first, during the thin film deposition process, the reflected light intensity signal is collected in real time, and the collected reflected light intensity signal is decomposed into multi-scales using wavelet transform. Wavelet transform is a time-frequency analysis tool. By selecting a suitable mother wavelet function, the reflected light intensity signal can be decomposed into sub-band signals of different frequency bands. Feature parameters are extracted for each sub-band signal, including statistical features such as the mean, variance, skewness, kurtosis, and frequency domain features of the sub-band signal, such as frequency, amplitude, and phase. The extracted feature parameters are input into the trained refractive index estimation model to obtain the refractive index estimation values ​​of each scale. In order to comprehensively utilize the refractive index information of different scales, an adaptive weighted fusion strategy is introduced. By learning the association weights between the refractive index estimation values ​​of each scale and the final refractive index, the importance of the estimated values ​​of each scale is adaptively adjusted to obtain the final refractive index estimation value.

[0105] The film thickness increment in each period is calculated by combining the phase difference between adjacent periods obtained by the phase unwrapping algorithm and the estimated refractive index obtained by the refractive index estimation model. The initial film thickness estimate is obtained by accumulating the film thickness increment in each period.

[0106] In order to further improve the accuracy and robustness of film thickness estimation, the Kalman filter algorithm is introduced to filter and optimize the initial film thickness estimation value. Kalman filtering is a recursive state estimation method. It dynamically estimates the film thickness by establishing a state space model of film thickness and combining the measured value and state prediction value. In the Kalman filtering process, the measurement noise and process noise covariance matrix of the filter are adaptively adjusted according to the change of the refractive index estimation value to cope with the dynamic change of the refractive index of the film material. Through the prediction and update steps of Kalman filtering, the optimal estimate of the film thickness is obtained as the final real-time film thickness monitoring result.

[0107] In an optional embodiment,

[0108] The final estimated refractive index value is combined with the phase difference of adjacent periods obtained by the phase unwrapping algorithm to calculate the film thickness increment in each period. By accumulating the film thickness increment in each period, the calculation formula for the initial film thickness estimate is obtained as follows:

[0109]

[0110] in, represents the initial film thickness estimate, K represents the number of cycles, λ represents the wavelength of the incident light, m ​​represents the number of scales, and σ a represents the attention weight of the a-th scale, represents the refractive index estimate of the ath scale, β represents the nonlinear fusion coefficient, △η krepresents the phase difference of the kth cycle, γ represents the nonlinear compensation coefficient, σ b represents the attention weight of the b-th scale, represents the refractive index estimate at the b-th scale.

[0111] In this embodiment, the refractive index estimation model trained by the characteristic parameters of the reflected light intensity signal can estimate the refractive index change of the thin film material at different deposition stages in real time, providing key parameter input for subsequent film thickness measurement. The reflected light intensity signal is subjected to wavelet multi-scale decomposition, and the refractive index of each scale is estimated respectively, and then fused, which can effectively extract the multi-scale characteristic information of the signal and improve the robustness and accuracy of the refractive index estimation. The phase difference obtained by phase unpacking is combined with the estimated refractive index to calculate the film thickness increment in each cycle, laying the foundation for further obtaining the absolute thickness of the film. Using the Kalman filter algorithm, a film thickness state space model is established, and the filter parameters are adaptively adjusted according to the refractive index change, which can effectively eliminate measurement noise, optimize the initial thickness estimate, and obtain a smooth and accurate real-time thickness value. The data-driven model based on the reflected light intensity signal and the analytical model based on the phase difference are integrated to achieve the complementarity of the two approaches and improve the reliability and stability of thickness measurement. The obtained real-time thickness value of the film can be fed back to the control system and compared with the set thickness target value to achieve real-time closed-loop monitoring and precise control of the film thickness.

[0112] S103. Using the obtained real-time thickness value of the thin film in combination with the pre-established numerical model of thin film growth, the thin film thickness distribution on the entire substrate surface is obtained through an intelligent interpolation algorithm. Based on the obtained thin film thickness distribution, a three-dimensional thickness distribution map of the thin film is generated using graphics drawing and rendering technology to present the dynamic evolution of the thin film growth process in real time.

[0113] After obtaining the real-time thickness value of the film, in order to comprehensively analyze and grasp the film growth situation on the entire substrate surface, it is necessary to further obtain the film thickness distribution on the entire substrate surface.

[0114] First, a numerical model of film growth is established through numerical simulation and experimental data fitting. This model takes into account various physical and chemical factors in the film growth process, such as deposition rate, substrate temperature, chamber pressure, etc., and describes the temporal and spatial evolution of film thickness through partial differential equations. The partial differential equations are discretized using the finite element method to obtain the numerical solution of the film thickness distribution. In practical applications, the real-time thickness monitoring values ​​obtained are used to perform real-time corrections on the numerical model to improve the prediction accuracy of the model.

[0115] Then, based on the pre-established numerical model of thin film growth, the film thickness distribution on the entire substrate surface is obtained through an intelligent interpolation algorithm. Specifically,

[0116] Based on the position and thickness value of the real-time thickness monitoring point, the initial thickness distribution surface is constructed using radial basis function interpolation. Radial basis function interpolation establishes the mapping relationship between spatial position and thickness value by selecting appropriate basis functions, such as Gaussian function, multi-quadratic function, etc.

[0117] The initial thickness distribution surface is constrained and optimized using the numerical model of thin film growth. The initial thickness distribution obtained by radial basis function interpolation is fitted with the predicted results of the numerical model by the least squares method to obtain the optimized thickness distribution surface.

[0118] Machine learning technology is introduced to further refine the optimized thickness distribution surface. The support vector machine regression algorithm is used to train the prediction model of film thickness distribution with spatial position as input and optimized thickness value as output. Through model training, the spatial variation law of film thickness distribution is learned.

[0119] Using the trained film thickness distribution prediction model, the entire substrate surface is densely sampled to obtain a high-resolution film thickness distribution map. Through the combination of interpolation and machine learning, a high-precision estimation of film thickness distribution is achieved.

[0120] Finally, based on the obtained film thickness distribution, the three-dimensional film thickness distribution map is generated using graphics drawing and rendering technology to present the dynamic evolution of the film growth process in real time.

[0121] The substrate surface is divided into regular grids, each grid point corresponds to a film thickness value. Using the triangulation algorithm, a triangular mesh model of the three-dimensional film surface is generated based on the grid points.

[0122] The illumination model and texture mapping technology are used to render the three-dimensional film surface realistically. By setting the appropriate light source position and material properties, the optical properties of the film surface, such as reflection, refraction and scattering, are simulated. The color and texture information of the film surface is mapped to the triangular mesh model using texture mapping technology to enhance the realism of the film surface.

[0123] As the film growth process progresses, the film thickness distribution data is updated in real time, and the 3D film surface model is adjusted dynamically. Through real-time rendering and animation technology, a dynamic evolution video of the film growth process is generated to intuitively present the changes in the film surface morphology. An interactive 3D visualization interface is designed to allow users to rotate, scale, and translate the 3D film surface model through interactive devices such as mouse and keyboard, and observe the detailed features of the film surface in all directions.

[0124] The film thickness distribution of the entire substrate surface is obtained through an intelligent interpolation algorithm, and a three-dimensional film thickness distribution map is generated by combining graphics drawing and rendering technology, realizing real-time monitoring and visual analysis of the film growth process. This method can provide intuitive decision support for the optimization control of the film growth process, which helps to improve the quality and efficiency of film deposition. At the same time, by dynamically presenting the evolution of the film surface morphology, the understanding of the film growth mechanism is deepened, providing a new perspective and means for the research and application of thin film materials.

[0125] In an optional embodiment,

[0126] Using the obtained real-time thickness value of the film, combined with the pre-established numerical model of film growth, the film thickness distribution on the entire substrate surface is obtained through an intelligent interpolation algorithm, including:

[0127] A thin film growth numerical model including deposition rate, substrate temperature and chamber pressure is established to describe the temporal and spatial evolution of film thickness. The thin film growth numerical model is discretized using the finite element method to divide the thin film growth region into a plurality of discrete point grids at equal intervals, and the film thickness at each discrete point is interpolated and approximated using shape functions. The thin film growth numerical model is converted into a set of algebraic equations by a weighted residual method, and the algebraic equations are solved to obtain the numerical solution of the film thickness at each discrete point.

[0128] Using the obtained real-time thickness value of the film, parameter identification and dynamic correction are performed on the numerical solution of the film thickness at each discrete point to obtain a corrected numerical solution of the film thickness;

[0129] Based on the corrected numerical solution of film thickness, a deep learning network model is constructed with the spatial coordinates of discrete points in the film growth area as input and the numerical solution of film thickness at the discrete points as output, and a training data set including the spatial coordinates of the discrete points and the corresponding numerical solution of film thickness is constructed, the deep learning network model is trained using the training data set, and the network structure and training hyperparameters are optimized to obtain a trained film thickness prediction model;

[0130] The spatial coordinates of any position in the thin film growth area are input into the trained thin film thickness prediction model to obtain the predicted value of the thin film thickness at the current position. The predicted values ​​of the thin film thickness at all positions in the thin film growth area and the corresponding numerical solutions of the thin film thickness are weighted averaged to obtain the thin film thickness distribution on the entire substrate surface.

[0131] For example, thin film growth is a complex multi-physics coupling process involving multiple physical phenomena such as heat conduction, fluid flow, and chemical reactions. In order to accurately describe the spatiotemporal evolution of film thickness, a suitable numerical model needs to be established. First, a numerical model describing the film growth process is established. The model takes film thickness as the dependent variable, time, deposition rate, substrate temperature, and chamber pressure as independent variables, and describes the spatiotemporal evolution of film thickness through partial differential equations. Among them, the time derivative of film thickness characterizes the film growth rate, the spatial derivative characterizes the curvature change of the film surface, and the deposition rate, substrate temperature, and chamber pressure are associated with the derivative of film thickness through function terms that reflect the film growth mechanism. This partial differential equation numerical model can comprehensively characterize the multi-physics coupling effect in the film growth process, providing a solid theoretical basis for subsequent algorithm design.

[0132] Then, the finite element method is used to solve the established numerical model of partial differential equations for film growth. First, the film growth area is divided into regular rectangular grids, and each grid point corresponds to a discrete spatial position. In each grid unit, the bilinear shape function is used to interpolate and approximate the film thickness, and the continuous film thickness distribution is converted into discrete node thickness values. The partial differential equation numerical model is converted into a set of algebraic equations by the weighted residual method, and the unknown quantities of the equations are the numerical solutions of the film thickness at each grid node. By using a stable and efficient linear solution algorithm, such as the conjugate gradient method or the multigrid method, the numerical solution of the film thickness of the discrete nodes can be quickly obtained, thereby obtaining the film thickness distribution in the entire film growth area.

[0133] In order to further improve the accuracy of film thickness prediction, real-time thickness monitoring data is introduced to dynamically correct the finite element calculation results. Using the real-time thickness measurement values ​​obtained during the film growth process, the parameter identification method is used to dynamically adjust the model parameters in the finite element equation. Specifically, the sum of the squares of the difference between the measured thickness and the finite element calculated thickness is constructed as the objective function, and the least squares method is used to solve the optimal model parameter correction. Substituting the identified parameter correction into the original finite element equation, the corrected film thickness numerical solution can be obtained, realizing the real-time synchronization of the numerical model and the actual growth process.

[0134] On the basis of obtaining the corrected numerical solution of film thickness, the deep learning technology is further used to construct a film thickness prediction model. The spatial coordinates of discrete grid nodes are used as input, and the numerical solution of film thickness at the nodes is used as output to train the multilayer perceptron neural network. By designing a reasonable network structure and selecting appropriate activation functions, such as the ReLU function, and using regularization techniques such as Dropout, a deep learning film thickness prediction model with strong nonlinear fitting ability and good generalization performance can be established. During the training process, the corrected finite element film thickness solution is used as the label and the discrete grid node coordinates are used as the feature to construct a training data set. At the same time, data enhancement techniques such as adding Gaussian noise and rotation transformation are used to expand the training sample space. The network weights are continuously updated using the stochastic gradient descent method and the back propagation algorithm to minimize the error between the predicted thickness and the finite element thickness solution until the optimal model parameter configuration is reached.

[0135] Finally, the trained deep learning film thickness prediction model is applied to the prediction of film thickness at any position. For any spatial position within the film growth area, its coordinates are input into the trained deep learning model to quickly obtain the predicted film thickness value at that position. In order to further improve the prediction accuracy, the deep learning predicted value and the finite element calculated value can also be weighted fused. Using a certain weight coefficient, the prediction results of the two methods are linearly combined to form the final predicted film thickness distribution. The predicted film thickness values ​​at all positions in the film growth area and the corresponding numerical solutions of the film thickness are weighted averaged to obtain the film thickness distribution of the entire substrate surface. By flexibly adjusting the weight coefficient, a balance can be achieved between the generalization ability of the deep learning model and the physical interpretability of the finite element method, achieving high-precision and high-reliability prediction of film thickness distribution.

[0136] In this embodiment, by establishing a numerical model of film growth including key parameters such as deposition rate, substrate temperature and chamber pressure, the evolution law of film thickness in time and space can be described more accurately, laying a theoretical foundation for subsequent model solving and prediction. The finite element method is used to discretize the numerical model of film growth, and the weighted residual method is used to convert the model into a set of algebraic equations, so that the accurate thickness numerical solution at each discrete point in the film growth area can be obtained. Using the film thickness data measured in real time, the numerical solution of the model is parameter identified and dynamically corrected, which effectively corrects the possible deviations in the model solving process and improves the accuracy of model prediction. Based on the corrected thickness numerical solution, a deep learning network model is constructed and effectively trained, which can fully explore the intrinsic mapping relationship between the analytical model and the data, and improve the nonlinear fitting and generalization ability of the model. By performing weighted average interpolation calculation on the predicted value of discrete point thickness, it can be extended from discrete points to the entire substrate surface to obtain a continuous film thickness distribution prediction result, which significantly expands the coverage of the prediction. The organic combination of analytical model solution, measured data fusion and deep learning model prediction not only utilizes the guiding role of the theoretical model, but also incorporates data-driven optimization correction, which can efficiently and accurately predict and map the film thickness distribution on the entire substrate surface.

[0137] In an optional embodiment,

[0138] Using the obtained real-time thickness value of the film, the numerical solution of the film thickness at each discrete point is parameter identified and dynamically corrected, and the corrected numerical solution of the film thickness is obtained, including:

[0139] Preprocessing the obtained real-time thickness value of the film to obtain a preprocessed real-time thickness value, wherein the preprocessing includes outlier removal, data smoothing, and interpolation point filling;

[0140] The preprocessed real-time thickness values ​​are fused using a preset spatial correlation model to obtain fused real-time thickness values, and an error objective function is constructed based on the fused real-time thickness values, wherein the error objective function includes a square sum term and a regularization term between the fused real-time thickness value and the numerical solution of the film thickness at each discrete point;

[0141] A nonlinear least squares optimization algorithm is used to solve the error objective function to obtain the optimal solution of the parameters to be identified in the thin film growth numerical model, and the thin film growth numerical model is corrected using the optimal solution, and the finite element solution is re-performed to obtain a corrected numerical solution for the film thickness.

[0142] Exemplarily, first, the acquired real-time thickness monitoring value of the film is preprocessed, including removing outliers, smoothing data, and interpolating missing data. Outlier removal refers to the use of statistical methods to identify and remove thickness measurements that are significantly deviated from the normal range to eliminate the influence of factors such as sensor failure or environmental interference. Data smoothing uses techniques such as sliding average and local polynomial fitting to filter random fluctuations in thickness measurements and improve the signal-to-noise ratio of the data. Interpolation refers to the use of methods such as spline interpolation and radial basis function interpolation to reasonably estimate missing or sparse measurement points to ensure the integrity and continuity of the data.

[0143] The pre-processed real-time thickness values ​​are fused through the spatial correlation model to obtain a more accurate and robust thickness estimate. The spatial correlation model takes into account the spatial dependency between different measurement locations, and uses geo-statistical methods such as Kriging interpolation and Gaussian process regression to perform spatial interpolation and interpolation on scattered point measurements to generate a continuous thickness distribution map. The fusion process makes full use of the redundant information of multiple sensors, and through weighted averaging or Bayesian inference, the credibility of different measurement values ​​is integrated to obtain the optimal thickness estimate.

[0144] After obtaining the fused real-time thickness value, an error objective function is constructed to quantify the deviation between the fused estimate and the predicted value of the numerical model. The error objective function consists of two parts: the mean square error term and the regularization term. The mean square error term represents the mean square error between the fused real-time thickness value and the thickness numerical solution of the finite element model at each discrete point, reflecting the degree of fit between the model prediction and the actual observation. The regularization term introduces prior knowledge or constraints into the objective function, such as the smoothness and sparsity of the parameters, to avoid overfitting and improve the stability of the solution. Common regularization terms include L1 norm, L2 norm, total variation regularization, etc. The appropriate regularization method is selected according to the characteristics of the specific problem.

[0145] The error objective function is solved by nonlinear least squares optimization algorithms, such as the Levenberg-Marquardt algorithm and the trust region reflective algorithm. These algorithms continuously adjust the parameters to be identified in the thin film growth numerical model, such as growth rate and diffusion coefficient, through iterative search, so that the error objective function reaches the minimum value. During the optimization process, the key parameters are determined by model sensitivity analysis, and adaptive step size control, multi-scale optimization and other technologies are used to improve the convergence speed and robustness.

[0146] After obtaining the optimal solution of the parameters to be identified through nonlinear least squares optimization, it is substituted into the numerical model of thin film growth to correct the original finite element equation. The corrected model can more accurately describe the actual growth process, reflecting the feedback correction effect of real-time monitoring data on numerical simulation. Based on the corrected model, the finite element solution is re-performed to obtain an updated numerical solution for the film thickness. The new numerical solution not only takes into account the dominant role of the physical mechanism, but also integrates the dynamic information provided by the real-time monitoring data, which can better predict and control the film growth process.

[0147] In this embodiment, through a series of steps such as data preprocessing, spatial correlation model fusion, error objective function construction, and nonlinear least squares optimization, dynamic correction and feedback update of real-time measurement values ​​and numerical simulation results are achieved. The rich information provided by the real-time monitoring data is fully utilized, the limitations of traditional offline parameter identification are overcome, and it can adapt to the dynamic changes and uncertainties of the thin film growth process. At the same time, by introducing regularization terms and prior knowledge, the stability and interpretability of parameter identification are improved. It provides new ideas and means for real-time monitoring and optimization control of the thin film growth process. Through the deep integration of real-time monitoring data and numerical models, deviations in the growth process can be discovered and corrected in time, and closed-loop feedback control can be achieved. At the same time, the key parameters identified are helpful to optimize process conditions, improve growth strategies, and improve the quality and performance of thin film materials.

[0148] Figure 2 FIG. 1 is a schematic diagram of the structure of a real-time monitoring system for the thickness of a photovoltaic cell film according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0149] The first unit is used to set up a multi-channel optical fiber sensor array, wherein each optical fiber sensing channel includes a transmitting end optical fiber and a receiving end optical fiber, the transmitting end optical fiber is coupled to a monochromatic laser to generate a collimated light beam, the receiving end optical fiber collects a reflected light signal and connects it to a photodetector, a collimated light beam is generated by a monochromatic laser, and is irradiated to the surface of a thin film grown on a substrate through the transmitting end optical fiber, and at the same time, the receiving end optical fiber collects a reflected light signal of the thin film, and connects the reflected light signal to the photodetector to convert it into a reflected light intensity signal;

[0150] The second unit is used to pre-process the obtained reflected light intensity signal, perform time-frequency analysis on the pre-processed reflected light intensity signal, extract characteristic parameters of the reflected light intensity signal, including oscillation frequency and oscillation amplitude, establish a mapping relationship between the oscillation frequency and oscillation amplitude and the film thickness growth, use the phase unwrapping algorithm to estimate the phase difference between adjacent periods, obtain continuous film thickness change information, and introduce a refractive index online estimation mechanism to compensate for the refractive index change of the film material. By counting the oscillation amplitude in each period and combining the compensation for the refractive index change, the real-time thickness value of the film is obtained;

[0151] The third unit is used to utilize the real-time thickness value of the film, combined with the pre-established numerical model of film growth, to obtain the film thickness distribution on the entire substrate surface through an intelligent interpolation algorithm. Based on the obtained film thickness distribution, a three-dimensional film thickness distribution map is generated using graphics drawing and rendering technology to present the dynamic evolution of the film growth process in real time.

[0152] According to a third aspect of the embodiments of the present invention,

[0153] An electronic device is provided, comprising:

[0154] processor;

[0155] a memory for storing processor-executable instructions;

[0156] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0157] A fourth aspect of the embodiments of the present invention is:

[0158] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0159] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for real-time monitoring of the thickness of a photovoltaic cell film, characterized in that: include: A multi-channel optical fiber sensor array is provided, wherein each optical fiber sensing channel includes an emitting optical fiber and a receiving optical fiber, the emitting optical fiber is coupled to a monochromatic laser to generate a collimated light beam, the receiving optical fiber collects a reflected light signal and connects it to a photodetector, a collimated light beam is generated by a monochromatic laser, and is irradiated to the surface of a thin film grown on a substrate through the emitting optical fiber, while the receiving optical fiber collects a reflected light signal of the thin film, and connects the reflected light signal to a photodetector to convert it into a reflected light intensity signal; The obtained reflected light intensity signal is preprocessed, and the preprocessed reflected light intensity signal is subjected to time-frequency analysis to extract characteristic parameters of the reflected light intensity signal, including oscillation frequency and oscillation amplitude, and a mapping relationship between the oscillation frequency and oscillation amplitude and the film thickness growth is established. The phase unwrapping algorithm is used to estimate the phase difference between adjacent periods to obtain continuous film thickness change information, and a refractive index online estimation mechanism is introduced to compensate for the refractive index change of the film material. By counting the oscillation amplitude in each period and combining the compensation for the refractive index change, the real-time thickness value of the film is obtained; Using the real-time thickness value of the film and the pre-established numerical model of film growth, the film thickness distribution on the entire substrate surface is obtained through an intelligent interpolation algorithm. Based on the obtained film thickness distribution, a three-dimensional film thickness distribution map is generated using graphics drawing and rendering technology to present the dynamic evolution of the film growth process in real time.

2. The method according to claim 1, characterized in that The phase unwrapping algorithm is used to estimate the phase difference between adjacent cycles, including: Performing statistical analysis on the phase differences of adjacent cycles, calculating the mean and variance of the phase differences, estimating the distribution characteristics of the phase differences according to the mean and variance, and obtaining an adaptive threshold using an adaptive threshold selection strategy based on the distribution characteristics of the phase differences, wherein the adaptive threshold selection strategy is one or more combinations of a threshold based on a mean and a standard deviation, a threshold based on a quantile, and a threshold based on entropy; Constructing an undirected weighted phase difference network graph, wherein vertices represent cycles, edges represent phase differences between adjacent cycles, and edge weights are reliability metrics of the phase differences, wherein the reliability metrics comprehensively consider the amplitude, time correlation, and frequency domain characteristics of the phase differences, and performing edge detection on the undirected weighted phase difference network graph according to the obtained adaptive threshold to obtain a binary network graph; The maximum flow minimum cut is searched in the binary network graph, and the network graph is divided into two parts: the source point set and the sink point set. The cut edge connects the two sets. The sum of the capacity of the cut edge is the capacity of the minimum cut. The cut edge corresponds to the position of the phase jump. An adaptive phase compensation mechanism is introduced for the vertices at both ends of the cut edge corresponding to the maximum flow minimum cut. The phase compensation value is adaptively determined to achieve continuous phase reconstruction and obtain the final phase difference between adjacent cycles.

3. The method according to claim 2, characterized in that An adaptive phase compensation mechanism is introduced for the vertices at both ends of the cut edge corresponding to the maximum flow minimum cut, and the phase compensation value is adaptively determined to achieve continuous phase reconstruction. The final phase difference between adjacent cycles includes: Extract the phase difference of the period where the vertices at both ends of the cut edge are located and the phase difference of the adjacent period, calculate the weighted average of the phase difference of the adjacent period, and at the same time, measure the reliability of the phase difference of the adjacent period according to the amplitude, time correlation and frequency domain characteristics of the phase difference of the adjacent period, use the reliability as the weight of the phase difference of the adjacent period, dynamically adjust the weight through the adaptive weight strategy, obtain the phase compensation value, and compensate the phase of the vertices at both ends of the cut edge to correct the phase jump; Repeat the phase compensation for all the cut edges in the binary network graph to obtain a continuous unwrapped phase, realize continuous phase reconstruction, and obtain the final phase difference between adjacent cycles; The calculation formula of the phase compensation value is as follows: Among them, △η c represents the phase compensation value, i represents the edge cutting start index, j represents the edge cutting end index, M represents the number of adjacent cycles, and w k represents the weight of the phase difference of the kth cycle, △η k represents the phase difference of the kth cycle, △η i Indicates the phase difference of the period where the cutting edge starting point is located, △η j Indicates the phase difference of the cycle where the end point of the cutting edge is located.

4. The method according to claim 1, characterized in that: The refractive index online estimation mechanism is introduced to compensate for the refractive index change of the thin film material. By counting the oscillation amplitude in each cycle and combining the compensation for the refractive index change, the real-time thickness value of the film is obtained, including: Using the characteristic parameters of the reflected light intensity signal extracted in advance as input and the refractive index of the thin film material to be measured as output, a refractive index estimation model is established and trained; During the thin film deposition process, the reflected light intensity signal is collected in real time, and the collected reflected light intensity signal is decomposed into multiple scales using wavelet transform to obtain sub-band signals of different frequency bands. The characteristic parameters of each sub-band signal are extracted and input into the trained refractive index estimation model to obtain the refractive index estimation value of each scale, and the refractive index estimation value of each scale is combined to obtain the final refractive index estimation value; The final estimated refractive index value is combined with the phase difference of adjacent periods obtained by the phase unwrapping algorithm to calculate the film thickness increment in each period, and the initial film thickness estimation value is obtained by accumulating the film thickness increment in each period; The Kalman filter algorithm is used to filter and optimize the initial film thickness estimate, and a state space model of the film thickness is established. The filter parameters are adaptively adjusted according to the change of the refractive index estimate. The iteration is repeated until the preset maximum number of iterations is reached to obtain the final real-time thickness value of the film.

5. The method according to claim 4, characterized in that The final estimated refractive index value is combined with the phase difference of adjacent periods obtained by the phase unwrapping algorithm to calculate the film thickness increment in each period. By accumulating the film thickness increment in each period, the calculation formula for the initial film thickness estimate is obtained as follows: in, represents the initial film thickness estimate, K represents the number of cycles, λ represents the wavelength of the incident light, m ​​represents the number of scales, and σ a represents the attention weight of the a-th scale, represents the refractive index estimate of the ath scale, β represents the nonlinear fusion coefficient, △η k represents the phase difference of the kth cycle, γ represents the nonlinear compensation coefficient, σ b represents the attention weight of the b-th scale, represents the refractive index estimate at the b-th scale.

6. The method according to claim 1, characterized in that Using the obtained real-time thickness value of the film, combined with the pre-established numerical model of film growth, the film thickness distribution on the entire substrate surface is obtained through an intelligent interpolation algorithm, including: A thin film growth numerical model including deposition rate, substrate temperature and chamber pressure is established to describe the temporal and spatial evolution of film thickness. The thin film growth numerical model is discretized using the finite element method to divide the thin film growth region into a plurality of discrete point grids at equal intervals, and the film thickness at each discrete point is interpolated and approximated using shape functions. The thin film growth numerical model is converted into a set of algebraic equations by a weighted residual method, and the algebraic equations are solved to obtain the numerical solution of the film thickness at each discrete point. Using the obtained real-time thickness value of the film, parameter identification and dynamic correction are performed on the numerical solution of the film thickness at each discrete point to obtain a corrected numerical solution of the film thickness; Based on the corrected numerical solution of film thickness, a deep learning network model is constructed with the spatial coordinates of discrete points in the film growth area as input and the numerical solution of film thickness at the discrete points as output, and a training data set including the spatial coordinates of the discrete points and the corresponding numerical solution of film thickness is constructed, the deep learning network model is trained using the training data set, and the network structure and training hyperparameters are optimized to obtain a trained film thickness prediction model; The spatial coordinates of any position in the film growth area are input into the trained film thickness prediction model to obtain the predicted value of the film thickness at the current position. The predicted values ​​of the film thickness at all positions in the film growth area and the corresponding numerical solutions of the film thickness are weighted averaged to obtain the film thickness distribution on the entire substrate surface.

7. The method according to claim 6, characterized in that Using the obtained real-time thickness value of the film, the numerical solution of the film thickness at each discrete point is parameter identified and dynamically corrected, and the corrected numerical solution of the film thickness is obtained, including: Preprocessing the obtained real-time thickness value of the film to obtain a preprocessed real-time thickness value, wherein the preprocessing includes outlier removal, data smoothing, and interpolation point filling; The preprocessed real-time thickness values ​​are fused using a preset spatial correlation model to obtain fused real-time thickness values, and an error objective function is constructed based on the fused real-time thickness values, wherein the error objective function includes a square sum term and a regularization term between the fused real-time thickness value and the numerical solution of the film thickness at each discrete point; A nonlinear least squares optimization algorithm is used to solve the error objective function to obtain the optimal solution of the parameters to be identified in the thin film growth numerical model, and the thin film growth numerical model is corrected using the optimal solution, and the finite element solution is re-performed to obtain a corrected numerical solution for the film thickness.

8. A photovoltaic cell film thickness real-time monitoring system, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to set up a multi-channel optical fiber sensor array, wherein each optical fiber sensing channel includes a transmitting end optical fiber and a receiving end optical fiber, the transmitting end optical fiber is coupled to a monochromatic laser to generate a collimated light beam, the receiving end optical fiber collects a reflected light signal and connects it to a photodetector, a collimated light beam is generated by a monochromatic laser, and is irradiated to the surface of a thin film grown on a substrate through the transmitting end optical fiber, and at the same time, the receiving end optical fiber collects a reflected light signal of the thin film, and connects the reflected light signal to the photodetector to convert it into a reflected light intensity signal; The second unit is used to pre-process the obtained reflected light intensity signal, perform time-frequency analysis on the pre-processed reflected light intensity signal, extract characteristic parameters of the reflected light intensity signal, including oscillation frequency and oscillation amplitude, establish a mapping relationship between the oscillation frequency and oscillation amplitude and the film thickness growth, use the phase unwrapping algorithm to estimate the phase difference between adjacent periods, obtain continuous film thickness change information, and introduce a refractive index online estimation mechanism to compensate for the refractive index change of the film material. By counting the oscillation amplitude in each period and combining the compensation for the refractive index change, the real-time thickness value of the film is obtained; The third unit is used to utilize the real-time thickness value of the film, combined with the pre-established numerical model of film growth, to obtain the film thickness distribution on the entire substrate surface through an intelligent interpolation algorithm. Based on the obtained film thickness distribution, a three-dimensional film thickness distribution map is generated using graphics drawing and rendering technology to present the dynamic evolution of the film growth process in real time.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images