Fiber optic spectrometer for film thickness measurement and measurement method

Abstract

The invention discloses a fiber optic spectrometer for film thickness measurement. The fiber optic spectrometer comprises a control and data processing module, a detection module, a light source module and a light splitting module, the optical fiber coupling module couples wide-spectrum light of the light source module and then guides the light to the light splitting module; the light splitting module adopts a cross-type Czerny-Turner light path structure, and performs collimation, dispersion and focusing processing on light beams in sequence; the detection module is used for receiving the optical signal processed by the light splitting module and forming an interference pattern; and the control and data processing module is used for driving acquisition, completing pixel-wavelength calibration and calculating the film thickness based on an acquired interference pattern. According to the optical fiber spectrometer measurement method for film thickness measurement, the problem of low accuracy of real-time film thickness detection in the prior art is solved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of optical measurement, and relates to a fiber spectrometer for film thickness measurement, and to a fiber spectrometer measurement method for film thickness measurement. BACKGROUND

[0002] Spectroscopic film thickness measurement systems are widely used for the thickness detection of transparent or semi-transparent films, and the typical measurement range is 40-120 microns. The measurement principle is based on optical transmission and interference effect: when broadband continuous light passes through the film layer to be measured, part of the light wave is reflected and interfered between the multiple interfaces of the film layer, and the transmission light intensity changes periodically with the wavelength, forming an interference fringe spectrum. By analyzing the interference period in the interference spectrum, the optical path difference can be calculated and the film layer thickness can be further derived.

[0003] In the prior art, film thickness measurement relies on high-precision instruments such as ellipsometers and white light interferometers, which have high measurement accuracy but are expensive, complex in structure, bulky, and require high operating environment and personnel, making it difficult to promote to small and medium-sized enterprises and production lines. Although some portable spectrometers have been compressed in cost and size, they still have the following problems: first, to ensure measurement accuracy, the system needs to integrate high-precision gratings, integrating spheres, special optical windows, etc., making it difficult to reduce costs; second, the mainstream CCD array detector is expensive, limiting its integration capability on low-cost platforms; third, there is a lack of control mechanism for light source stability and acquisition synchronization, making the system lack adaptability in industrial interference environments (such as vibration, strong light, temperature drift), and having poor measurement stability.

[0004] The current commonly used spectral structures include direct incidence, Rowland circle and Czerny-Turner configuration, etc. Among them, the Czerny-Turner optical path is the preferred solution for small spectrometers due to its high resolution, foldable optical path and compact structure. The present application adopts a cross-type Czerny-Turner optical path design, which significantly compresses the volume while maintaining spectral performance, improving structural integration and portability.

[0005] Although some documents attempt to use miniature spectral detection modules to build a simplified system, the measurement results are limited due to the lack of systematic optimization of the optical path, signal acquisition accuracy and film thickness calculation model, making it difficult to meet the three goals of "low cost, high accuracy, strong adaptability". Especially in online real-time detection scenarios, if there are problems such as low spectral signal-to-noise ratio or nonlinear detection response, the film thickness inversion accuracy will be significantly reduced, or even deviated.

[0006] Therefore, it is urgent to propose a film thickness spectral measurement system that is simple in structure, easy to integrate, low in hardware cost, and has good anti-interference ability, which can work stably in actual industrial environments and also consider real-time and measurement accuracy. SUMMARY

[0007] The application aims to provide a fiber-optic spectrometer for film thickness measurement, which solves the problem of low accuracy in real-time detection of film thickness in the prior art.

[0008] Another object of the application is to provide a fiber-optic spectrometer measurement method for film thickness measurement.

[0009] The technical scheme adopted by the application is a fiber-optic spectrometer for film thickness measurement, comprising a control and data processing module, a detection module, and a light source module and a light splitting module connected through a fiber coupling module; the light splitting module and the detection module are arranged in a housing; The fiber coupling module couples the wide-spectrum light of the light source module and guides it to the light splitting module; The light splitting module adopts a cross-type Czerny-Turner optical path structure to sequentially perform collimation, dispersion, and focusing on the light beam; The detection module is used to receive the light signal processed by the light splitting module and form an interference spectrum; The control and data processing module drives the collection, completes the pixel-wavelength calibration, and calculates the film thickness based on the convolutional neural network wide-spectrum thickness prediction model.

[0010] The application also has the characteristics that The light splitting module comprises a slit, a collimating mirror, a blazed grating, and a focusing mirror arranged in the housing in sequence according to the cross-type optical path, the collimating mirror and the focusing mirror are arranged adjacent to or connected to each other, and the collimating mirror and the focusing mirror are located on the two sides of the blazed grating to form a V-shaped optical path, the slit is provided on one side of the housing as a light inlet, a slit piece is arranged in the slit, and the detection module is located at the end of the light path of the focusing mirror on the housing.

[0011] The slit width is 25pm±5pm, the collimating mirror and the focusing mirror are both spherical mirrors, and the curvature radius of the focusing mirror is greater than or equal to the curvature radius of the collimating mirror.

[0012] The housing adopts a 3D printing integrated structure, comprising a top cover and a box body, the top cover and the box body are detachably connected, and the inner wall of the housing is provided with a black matte light-absorbing coating.

[0013] The light source module comprises a calibration light source and a measurement light source, the calibration light source adopts a fluorescent lamp, a mercury-argon lamp, and a monochromatic laser pen for calibration, the measurement light source adopts an LED wide-spectrum light source to provide wide-spectrum continuous light covering the 380-780nm waveband.

[0014] The fiber coupling module comprises a fiber and a standard SMA905 connector.

[0015] The detection module comprises a TCD1304 linear array CCD detector.

[0016] Pixel-wavelength calibration is specifically, wavelength calibration is completed by fluorescent lamp standard spectrum line, and the measured pixel-wavelength fitting relationship is as follows: λ = 0.2408·P + 1076.5328, wherein λ is wavelength, nm; P is pixel number.

[0017] Another technical solution adopted by the application is a measurement method of a fiber spectrometer for film thickness measurement, and the steps are as follows: The wide-spectrum continuous light of the light source module is coupled into the light splitting module after coupling through the fiber coupling module, collimated, dispersed and focused, and finally imaged on the detection module to form an interference spectrum, the control and data processing module converts the collected pixel-light intensity into wavelength-light intensity through a pixel-wavelength mapping function, and then a wide-spectrum thickness prediction model based on a convolutional neural network is used for thin film thickness prediction.

[0018] The feature of another technical solution of the application is also characterized in that: The construction and training process of the wide-spectrum thickness prediction model based on the convolutional neural network are specifically as follows: Step 1, data set construction, including measurement data and simulation data, content is "transmittance spectrum-thickness", transmittance T(λ) of different wavelengths under thickness d, all data are subjected to mean normalization pretreatment to obtain the processed data set , and divided into a training set and a validation set; Step 2, film thickness initial value calculation based on the physical calculation model of interference fringes FWHM:

[0019] , wherein, is the thickness initial value, n is the refractive index of the thin film, FSR is the wavelength difference of adjacent two same-order interference peaks , unit: nm, is the central wavelength of adjacent interference peaks, and are the wavelengths of adjacent two same-order interference peaks, m is the interference order, ; Step 3, wide-spectrum thickness prediction model based on convolutional neural network is built and trained: The convolutional neural network structure includes an input layer, two convolutional layers, two pooling layers, a reshaping layer and a fully connected layer, and the specific configuration is as follows: Input layer: receiving one-dimensional spectrum data with a dimension of 1*300; Convolutional layer: containing two layers of convolution, both using a 3*3 convolution kernel, a step of 1, and a ReLU activation function, for extracting local spectral features, i.e., interference peaks and absorption valleys; Pooling layers: Alternate with convolutional layers, two layers in total, both using 2×2 max pooling with a stride of 2, downsampling the convolutional features to retain core features while reducing data dimensionality and computational cost; Fully connected layer: Expands the pooled high-dimensional features into a one-dimensional vector and outputs the prediction result, i.e., the predicted value d of the thin film optical thickness. z ; Network training specifically involves: Data and initial thickness values Input to the neural network, using a three-level hybrid loss function Convergence, of which, For data fitting loss, For physical constraint loss, The initial value guides the loss. The original mean square error term is used to measure the predicted thickness. Compared with actual thickness The direct deviation is used to ensure the accuracy of data fitting, where N is the total number of samples in the dataset. These are the weighting coefficients. These are the initial weights;

[0020] This formula is the physical loss function in the field of thin-film interferometry, used to measure the deviation between model predictions and theoretical laws. Among them, It is the model's predicted value for the film thickness of the i-th sample; These are theoretically calculated values ​​for thin-film interference. The actual wavelength corresponding to the i-th interference peak Where n is the refractive index of thin film sample i, is the constrained predicted value of the incident angle; N is the total number of samples.

[0021] , It is a squared loss function that averages the deviations between the predicted values ​​and the initial baseline values ​​of N samples, where, The model's predicted value for the film thickness of the i-th sample. The initial thickness value is obtained through a physical calculation model; The Adam optimizer was selected, and the network parameters were iteratively adjusted using gradient descent and backpropagation algorithms to minimize the loss function. A staged training strategy was used, and an initial value correction sub-network was set. After training, a well-trained broadband thickness prediction model was obtained.

[0022] The beneficial effects of this invention are: The application discloses a fiber spectrometer for film thickness measurement, adopts a cross-type Czerny-Turner folded optical path structure, effectively shortens an optical path and compresses a system volume, is suitable for production line embedded or portable use scenarios, and improves integration and adaptability; precise components such as a grating, a mirror frame and a detector are integrated on a same reference base surface by using an integrally formed shell structure, component installation errors are eliminated by a clamping groove positioning mode, and system assembly precision and long-term stability are improved; a high light absorption rate blackening coating treatment is adopted inside, and stray light and internal cavity reflection interference can be significantly inhibited. The application discloses a measurement method of the fiber spectrometer for film thickness measurement, selects a physical calculation model based on interference fringe FWHM for priori, and then adds a wide spectrum thickness prediction model of a convolutional neural network to perform accurate thickness calculation, the latter learns a nonlinear relationship between spectral characteristics and film thickness through a training set to improve robustness, the whole method does not need complex physical parameter inversion, realizes thickness prediction through an end-to-end neural network mapping, is simple and convenient to operate, has strong anti-interference ability, and is suitable for online detection scenes in industrial production. BRIEF DESCRIPTION OF DRAWINGS

[0023] Figure 1 FIG. 1 is a structural schematic diagram of the fiber spectrometer for film thickness measurement of the application; Figure 2 FIG. 2 is a shell structure schematic diagram in the fiber spectrometer for film thickness measurement of the application; Figure 3 FIG. 3 is a structure schematic diagram of a light splitting module in the fiber spectrometer for film thickness measurement of the application; Figure 4 FIG. 4 is a Czerny-Turner optical path structure diagram adopted in the fiber spectrometer of the application; Figure 5 FIG. 5 is a convolutional network structure schematic diagram of a control and data processing module in the fiber spectrometer of the application; Figure 6 FIG. 6 is a data processing diagram of a neural network in the fiber spectrometer of the application; Figure 7 FIG. 7 is a film thickness calculation value fluctuation range diagram of the film thickness measurement method of the fiber spectrometer of the application; Figure 8 FIG. 8 is a wavelength calibration linear fitting diagram of the fiber spectrometer of the application; Figure 9 FIG. 9 is a spectrum diagram of a fluorescent lamp converted by a fitting function in the application; Figure 10 FIG. 10 is a real spectrum diagram of a fluorescent lamp used in a verification test of the application.

[0024] In the figure: 1. slit plate; 2. collimating mirror; 3. grating; 4. focusing mirror; 5. detection module; 6. housing, 7 light source module, 8. fiber coupling module, 9. light splitting module, 10. control and data processing module. DETAILED DESCRIPTION

[0025] The application will be described in greater detail with reference to the drawings and specific embodiments, which are described by way of example only and therefore are not exhaustive of all embodiments according to the application. Any and all embodiments not explicitly described are intended to be included within the scope of the application, as understood by those of ordinary skill in the art.

[0026] In the present application, unless otherwise explicitly specified and limited, the terms "mounting", "connection", "connecting", "fixing" and the like should be understood in a broad sense, for example, can be fixed connection, or detachable connection, or integral; can be mechanical connection, or electrical connection; can be directly connected, or indirectly connected through an intermediate medium, can be the internal communication of two elements or the interaction relationship between two elements. For those of ordinary skill in the art, the specific meaning of the above terms in the present application can be understood according to the specific circumstances.

[0027] Unless otherwise defined, all technical and scientific terms used in the present application have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs. The terms used in the present application are only for the purpose of describing the embodiments of the present application and are not intended to limit the present application.

[0028] Example 1 The fiber spectrometer for film thickness measurement of the present embodiment has a structure as shown in Figure 1 which comprises a housing 6, a light source module 7, a fiber coupling module 8, a light splitting module 9, a detection module 5, and a control and data processing module 10. The light splitting module 9 and the detection module 5 are arranged in the housing 6.

[0029] The light source module 7 provides a wide-spectrum continuous light; The fiber coupling module 8 couples the wide-spectrum light from the light source module 7 and guides it to the light splitting module 9; The light splitting module 9 adopts a cross-type Czerny-Turner optical path structure as shown in Figure 3 and Figure 4 which sequentially collimates, disperses and focuses the light beam; The cross-type Czerny-Turner folded optical path structure effectively shortens the optical path and compresses the system volume, which is suitable for production line embedded or portable use scenarios, and improves the integration and adaptability; The detection module 5 is used to receive the light signal processed by the light splitting module 9 and form an interference spectrum; The control and data processing module 10 drives the collection, completes the pixel-wavelength calibration, and calculates the film thickness based on the convolutional neural network-based wide-spectrum thickness prediction model.

[0030] Embodiment 2 On the basis of embodiment 1, as shown in Figure 2 、 Figure 3 and Figure 4 The light splitting module 9 includes the slit sheet 1, the collimating mirror 2, the blaze grating 3, and the focusing mirror 4 arranged in sequence in the cross-type light path in the shell 6; the collimating mirror 2 and the focusing mirror 4 are arranged adjacently or in connection, and the collimating mirror 2 and the focusing mirror 4 are respectively located on the two sides of the blaze grating 3 to form a V-shaped light path.

[0031] The slit is opened on one side of the shell 6 as an entrance for light, and an external light source is connected, and there is a 1-micron slit sheet 1 inside.

[0032] The detection module 5 is located at the end of the light path of the focusing mirror 4 on the shell 6.

[0033] Embodiment 3 On the basis of embodiment 2, the slit width range is preferably 25 μm ± 5 μm. It is used to limit the incident light flux and ensure that the appropriate light intensity passes through the system; when the width is 25 μm, it meets the design of the light path structure, so that the imaging is clear.

[0034] The collimating mirror 2 is a spherical mirror with a radius of curvature of 150 mm. The collimating mirror 2 is arranged opposite to the slit to ensure that the light beam is collimated after passing through the slit. The collimating mirror 2 converts the straight collimation of the divergent light emitted from the slit into parallel light and projects it onto the blaze grating 3.

[0035] The blaze grating 3 has a blaze wavelength of 500 nm and a grating density of 300 lines / mm and works in a first-order diffraction mode.

[0036] The grating 3 is arranged between the collimating mirror 2 and the focusing mirror 4 and forms an acute angle relationship with the collimating mirror 2 and the focusing mirror 4, and the light path forms a V-shaped light path.

[0037] The focusing mirror 4 is a spherical mirror with a radius of curvature of 200 mm.

[0038] The focusing mirror 4 is arranged after the blaze grating 3 to focus the light beam diffracted by the grating onto the detection module 5. The radius of curvature of the focusing mirror 4 is greater than or equal to the radius of curvature of the collimating mirror 2 to ensure good light focusing effect.

[0039] Embodiment 4 On the basis of embodiment 3, the light source module 7 is divided into a calibration light source and a measurement light source, and a fluorescent lamp, a mercury argon lamp, a green laser pen (wavelength 515 nm) and a red laser pen (wavelength 650 nm) are used for calibration. An LED wide-spectrum light source is used as an actual measurement light source to provide wide-spectrum continuous light covering the 380-780 nm band.

[0040] The fiber coupling module includes a fiber and a standard SMA905 connector, and the SMA905 fiber connects the light source module 7 and the incident slit interface on the outer wall of the shell 6. The light signal is transmitted through a multi-mode optical fiber, making the system structure compact and the connection flexible, and facilitating integration.

[0041] The detection module 5 is a TCD1304 linear array CCD detector, which has not less than 3600 pixels and an effective length of about 29.2 mm, and can collect a spectroscopic image.

[0042] The wide-spectrum continuous light of the light source module 7 is coupled by the fiber coupling module 8 and then input into the slit. Then the light beam is collimated, dispersed and focused by the collimating mirror 2, the blazed grating 3 and the focusing mirror 4 in turn, and finally forms an interference spectrum on the detection module 5.

[0043] Embodiment 5 On the basis of embodiment 4, as shown in Figure 2 and Figure 3 The shell 6 adopts an SLA photocuring 3D printing integrated structure, and the overall size is 130 mm x 98 mm x 40 mm. A black matte high-absorption coating is arranged on the inner wall of the shell to absorb stray light, and the surface is blackened as a whole to improve the signal-to-noise ratio, which can significantly suppress the interference of stray light and internal cavity reflection. The shell 6 is externally provided with a USB transmission interface and a ventilation hole. The USB transmission interface is used to connect the optical fiber to the control and data processing module 10 at the computer end for light signal processing and heat release. The shell 6 includes a top cover and a box body, and the top cover and the box body are detachably connected, which is convenient for later debugging and maintenance.

[0044] The shell 6 is provided with a semi-enclosed collimating mirror card slot, a focusing mirror card slot and a detector card slot. The collimating mirror 2, the focusing mirror 4 and the detection module 5 are respectively located in the collimating mirror card slot, the focusing mirror card slot and the detector card slot. The collimating mirror 2 and the focusing mirror 4 are limited to slide radially by the semi-enclosed card slot. Positioning ribs are arranged on both sides of the detector card slot to guide the angle self-calibration during installation, so as to ensure that the photosensitive surface coincides with the optical focusing surface, improve the image acquisition quality, and the size of the card slot is consistent with the optical simulation model, and the structure tolerance is not less than 1 mm in the vertical direction.

[0045] The blazed grating 3 is fixed to the middle of the housing 6 via a grating holder. A limiting post is located behind the grating holder to secure the blazed grating 3 and prevent displacement during transportation or vibration, ensuring the grating's stability. A preset tilt angle of 20° is maintained between the grating's mounting surface and the base of the housing 6 to ensure optimal diffraction efficiency. A one-piece molded housing structure integrates precision components such as the grating, lens mount, and detector onto the same reference surface. A slot-based positioning method eliminates component installation errors, improving system assembly accuracy and long-term stability.

[0046] The spectrometer module 9 is installed in the housing 6, and the control and data processing module 10 is located on the computer outside the housing 6. The control and data processing module 10 drives the CCD detector to sample through the control circuit and transmits the spectral data to the computer. After the spectral acquisition is completed, pixel-wavelength calibration is performed based on the known standard spectral lines. A pixel-wavelength mapping function is established through linear fitting to minimize the difference between the known spectrum and the measured spectrum, thus completing the calibration of the spectrometer.

[0047] Pixel-wavelength calibration specifically involves wavelength calibration using standard fluorescent lamp spectral lines. The resulting pixel-wavelength fitting relationship is: λ = 0.2408·P + 1076.5328, where λ is the wavelength in nm and P is the pixel number.

[0048] The wavelength calibration and resolution verification process is as follows: Based on the pixel number and corresponding wavelength data of the feature peak, the pixel number and wavelength at the feature peak are extracted as shown in Table 1. The data in Table 1 are then subjected to preliminary linear fitting to obtain the following results: Figure 8 The linear fit plot shown has a coefficient of determination (R²) exceeding 0.998. The R² is a core statistical indicator for measuring the goodness of fit of a wavelength-pixel linear fit model, used to quantify the degree of fit between pixel numbers and actual wavelengths. The calibration function for the prototype is λ = 0.2408·P+1076.5328. This function can be used to directly convert the collected fluorescent light intensity-pixel data into light intensity-wavelength data. The resulting fluorescent light intensity-wavelength spectrum is shown below. Figure 9 As shown, the actual image resolution is calculated to be 2 nm based on the half-peak height at the center wavelength; and... Figure 10 The comparison of the actual fluorescent lamp spectra shown demonstrates that the spectral variation trends of the two are completely consistent, verifying the accuracy of the conversion function and the reliability of the data collected by the spectrometer.

[0049] Table 1. Characteristic Wavelength and Pixel Correspondence

[0050] The fiber spectrometer for film thickness measurement of the application has a spectral resolution of 4-7 nm, a central wavelength of about 6.3 nm, and a wavelength-pixel fitting R² of more than 0.998.

[0051] In the thickness calculation process of the control and data processing module 10, the wide-spectrum thickness prediction model based on the convolutional neural network is used to calculate the film thickness, the initial value of the thickness is calculated based on the physical calculation model of the interference fringe FWHM, and then the initial value of the thickness is input into the neural network model for accurate thickness calculation. The latter learns the nonlinear relationship between the spectral features and the film thickness through the training set to improve the robustness. In this way, the prediction accuracy can be increased and the speed can be improved on the basis of the physical model.

[0052] The fiber spectrometer for film thickness measurement of the application has strong spectral line recognition and film thickness inversion capability, low hardware cost, strong universality of core components, and supports standard SMA fiber interface. The overall manufacturing cost is lower than that of the same resolution spectrometer on the market, and it is suitable for large-scale industrialization.

[0053] Example 6 The film thickness measurement method of the fiber spectrometer for film thickness measurement of the application uses the fiber spectrometer for film thickness measurement of example 5 and is implemented in the following steps: The wide-spectrum continuous light of the light source module 7 is coupled into the light splitting module 9 after being coupled by the fiber coupling module 8 for collimation, dispersion and focusing, and finally imaged on the detection module 5 to form an interference spectrum. The control and data processing module 10 converts the collected pixel-light intensity into wavelength-light intensity through a pixel-wavelength mapping function, and then uses a wide-spectrum thickness prediction model based on a convolutional neural network (CNN) to predict the film thickness: the initial value is provided by the physical model and the neural network is used for fine correction to realize the film thickness prediction: first, based on the physical calculation model of the interference fringe FWHM, the interference peak / valley features are extracted from the reflected spectrum to calculate the theoretical initial value of the film thickness; then, the initial value and the original spectral reflectance are used as the input of the neural network, and the prediction logic of "initial value as the benchmark and spectral feature fine-tuning" is realized through the "initial value correction sub-network"; during training, a hybrid loss function containing an initial value constraint term is used to strengthen the correlation between the initial value and the predicted value, and finally the thickness prediction result considering the physical rationality and spectral fitting accuracy is output.

[0054] The construction and training process of the wide-spectrum thickness prediction model based on the convolutional neural network is as follows: Step 1, data set construction, including measurement data and simulation data, in the format of csv, the file content is "transmittance spectrum - thickness", the transmittance T(λ) of different wavelengths under the thickness d. A total of 3000 groups.

[0055] The data set construction is as follows: ​1.1 Measurement Data Construction 1.1.1 Thin film preparation: Following the principle of controllable variables, by adjusting the coating speed, ink viscosity, or selecting different specifications of substrates, thin film samples with known thicknesses covering the target measurement range (40-120 micrometers) are prepared. For each thickness, more than 3 parallel samples are set for sampling to reduce random errors. 1.1.2 Measurement Experimental Environment Control: Spectral acquisition was conducted in a standard environment with a temperature of 23-25℃ and a humidity of 55%±5% to ensure measurement stability; 1.1.3 Multispectral Acquisition: Using the fiber optic spectrometer described in Example 5 (cross-type Czerny-Turner optical path structure, with a TCD1304 linear CCD detection module), three types of spectra were acquired sequentially: Dark spectrum, denoted as : Collect light intensity under conditions of no incident light to eliminate system background noise interference; Standard spectrum, denoted as Light intensity was collected using a blank substrate without a thin film as a standard sample to establish a benchmark for transmittance calculation; The spectrum after incident on the thin film is denoted as Fix the film to Figure 4 In the interlayer shown, light intensity information after transmission through the thin film is collected; The pixel-light intensity curve acquired at this time, where pixels are represented by p and light intensity values ​​are the acquired light intensity, in candela; 1.1.4. Complete the pixel-wavelength conversion. Based on the pixel-wavelength mapping function, convert the acquired pixel-light intensity into wavelength-light intensity. The measured pixel-wavelength fitting relationship is: λ = 0.2408·P + 1076.5328, where Where λ is the wavelength (nm) and P is the pixel number.

[0056] 1.1.5 Transmittance Calculation: The transmittance spectrum of the thin film is calculated using a standardized formula.

[0057] in, For thin film samples at wavelength The light intensity at that location, Dark spectral intensity The basic spectral intensity of the standard sample is used; after calculation, one-dimensional spectral transmittance data with a shape of 1×300 is obtained (corresponding to the transmittance values ​​of 300 wavelength points in the 380-780nm band).

[0058] 1.1.6, Thickness measurement: The actual thickness of each film sample is measured using a white light interferometer as label data. That is, the measured spectral transmittance corresponding to the actual film thickness, as d in the csv file. Form a "transmittance spectrum - thickness" csv file.

[0059] 1.2, Simulation data To make up for the shortcomings of not covering each gradient in the film preparation process, add simulation data to expand the film thickness coverage.

[0060] 1.2.1, Physical modeling: Use Fresnel formula and interference theory formula to derive the transmittance formula at each wavelength. Its transmittance formula is as follows: , where is the transmittance coefficient of the upper interface (air-film) and the lower interface (film-substrate) of the film is the reflection coefficient of the upper and lower interfaces of the film, is the total phase difference of the reflected light on the upper and lower surfaces of the film, , d is the thickness of the film, is the wavelength of the incident light, is the reflection phase change difference, n is the refractive index.

[0061] 1.2.2, Simulation sample generation: Based on the above model, the film parameters are brought in, including thickness d: 40-120μm, refractive index n: 1.3-1.8, range from 380 to 780, by , generate 300-dimensional transmittance spectrum, generate "transmittance spectrum - thickness" simulation data, stored as a csv file.

[0062] 1.3, Data storage: Store the transmittance data collected by the measurement experiment and the simulation experiment uniformly, in csv format, to form a complete data set.

[0063] Step 2, data set standardization and division The data set includes paired data collected by measuring transmittance and simulating spectral transmittance and corresponding film thickness.

[0064] In step 2, the data set data preprocessing is as follows: 2.1 Standardization: Normalize all spectral transmittance data by mean, formula is , where is the maximum value of the transmittance of this sample, as the preprocessed data set for storage; 2.2 Data set division: Divide the data according to the proportion of 80% training set and 20% validation set, avoid model overfitting through cross-validation.

[0065] Step 3: Based on the physical calculation model of the interference fringes FWHM, extract features such as interference peaks / valleys from the reflection spectrum and calculate the theoretical initial value of the film thickness; Specifically: 3.1. Transmittance spectral data of the preprocessed 1×300-dimensional thin film Interference peak identification is performed using the "local maxima method." A threshold is set, requiring transmittance to be greater than 5% of adjacent wavelengths. This method is used to traverse 300 wavelengths within the 380-780nm band, selecting local maxima that meet the criteria and recording them as the wavelengths of two adjacent interference peaks of the same order. and , m For the interference level, ; 3.2 Calculation of characteristic parameters center wavelength Take the average wavelength of two adjacent interference peaks.

[0066] Free Spectral Range (FSR): Calculates the wavelength difference between adjacent interference peaks, i.e. .

[0067] 3.3.2 Calculation of Initial Value Formula for Physical Thickness of Thin Film Calculate the initial thickness using the formula ,in , The center wavelength of adjacent interference peaks n Let be the refractive index of the thin film. FSR Wavelength difference between two adjacent interference peaks of the same order , Unit: nm.

[0068] Step 4: Building a broadband thickness prediction model based on convolutional neural networks; The process of building a convolutional neural network is as follows: 4.1 Neural network structure as follows Figure 5 As shown, it includes an input layer, two convolutional layers, two pooling layers, a reshaping layer, and a fully connected layer, with the following specific configuration: Input layer: Receives one-dimensional spectral data with a dimension of 1×300.

[0069] Convolutional layer: contains two layers of convolution, both using a 3x3 convolution kernel with a step size of 1. The first layer of convolution serves to perform convolution operations on the input features, outputting a 32@600x1 feature map (32 channels, size 600x1) for extracting local features. The second layer serves to extract more complex features by performing convolution again, outputting a 64@150x1 feature map. The activation function uses ReLU for extracting spectral local features, i.e., interference peaks and absorption valleys. Pooling layer: two layers are alternately arranged with the convolutional layer. Both use 2x2 max pooling with a step size of 2. The convolutional features are downsampled to reduce data dimension and computational load while preserving key features. Reshaping layer: flattens the feature map into a 4800x1 one-dimensional vector, preparing for the fully connected layer. The first pooling layer serves to downsample the convolutional output, outputting a 32@150x1 feature map, reducing computational load and preserving key features. The second pooling layer serves to downsample again, outputting a 64@150x1 feature map. Fully connected layer: expands the high-dimensional features after pooling into a one-dimensional vector, outputting the prediction result, i.e., the predicted film optical thickness d z ; Figure 6 Data processing diagram for neural network: one-dimensional feature map: input layer receives one-dimensional feature data; hidden layer includes layers between input layer and output layer, responsible for feature extraction and conversion; output layer outputs the final prediction result. Step 5, convolutional neural network model training 5.1 Input and output of convolutional neural network model Input is one-dimensional spectral transmittance data and initial thickness ; Output is the predicted film optical thickness d z ; 5.2 Loss function design A three-level hybrid loss function is used, i.e. , where is the data fitting loss, is the physical constraint loss, is the initial value guiding loss, the original mean square error term, used to measure the direct deviation between the predicted thickness and the actual thickness , to ensure data fitting accuracy, N is the total number of data set samples, is the weight coefficient (value 0.1-0.3, balancing data fitting and physical constraints), is the initial value weight (initial training = 0.2, the model is guided to the initial value; in the later = 0.05, the correction ability of the release spectrum feature);

[0070] N is the total number of samples, is the predicted value of the film thickness of the model for the ith sample; is the theoretical calculation value of the film interference, corresponding to the actual ith interference peak wavelength , n is the refractive index of the film sample i, is the constraint prediction value of the incident angle; By taking the average of the "absolute deviation of the predicted value from the theoretical value" of all samples, the The smaller, the higher the degree of fitting of the model to the physical law of film interference.

[0071] , is the square loss function of the "deviation of the predicted value from the initial reference value" of N samples, wherein, the predicted value of the film thickness of the model for the ith sample, is the initial value of the thickness obtained by the physical calculation model; 5.3 Optimizer and hyperparameter settings 5.3.1 Select Adam optimizer, adaptively adjust learning rate, consider training speed and stability, initial learning rate is set to

[0072] Through gradient descent method and back propagation algorithm to adjust network parameters (including convolution kernel weight, fully connected layer weight and bias term), minimize the loss function.

[0073] 5.3.2 Batch size is set to 32, balance training efficiency and gradient stability; 5.3.3 Set the maximum number of iterations: set to 1000 times, reserve sufficient training space, avoid underfitting caused by premature termination.

[0074] 5.4 Use the phased training strategy to set the initial value correction subnetwork:

[0075] In the formula, k , b is a learnable parameter: 5.4.1 The first stage, constrain the linear transformation parameters of the initial value correction subnetwork k = 1, b = 0, forced model output satisfies To prevent initial deviations from the physical reference by using small corrections based on spectral characteristics; set Strengthen the guiding role of physical constraints and initial value constraints, and prioritize ensuring the physical rationality of the prediction results.

[0076] 5.4.2 Second Phase: Fine-tuning Period (80% of iterations remaining, 800 iterations) - Released k , b Fixed limitations; adjustment By reducing the constraint intensity, the model can fully learn the subtle features in the data while following the physical laws, thereby improving accuracy.

[0077] 5.5 Set termination conditions; any one of the following conditions must be met.

[0078] Convergence and stability conditions: Training on the training set, real-time validation on the validation set, and the change in validation set loss over 20 consecutive iterations ≤ a set threshold. In this embodiment, the threshold is set as follows: The model has been determined to have converged, and further training will not result in significant performance improvement. Maximum iteration condition: Training is terminated when the number of iterations reaches a preset 1000 to balance efficiency and performance; Overfitting protection condition: If the validation set loss increases for 15 consecutive iterations and the training set loss continues to decrease, it is determined to be overfitting. Training is terminated early and the optimal model parameters at this time (the parameters when the validation set loss is minimized) are saved.

[0079] 5.6 After training is terminated, save the complete parameters of the optimal model (including all weights and biases of the convolutional layer, fully connected layer, and initial value correction subnetwork) and generate a model file that can be directly used to predict the thickness of the sample to be tested, ensuring efficient use in subsequent prediction processes.

[0080] To verify the actual performance of the fiber optic spectrometer for film thickness measurement in Embodiment 5 of the present invention, a transparent PET film sample was selected for spectral acquisition and film thickness measurement experiments. The test results show that within the 400–780 nm wavelength range, the spectral signal-to-noise ratio (SNR) of this device is significantly improved under typical illumination conditions, effectively suppressing the influence of stray light and background noise. Through repeated measurements of the same sample, the calculated film thickness fluctuated within the following range: Figure 7 As shown, the fluctuation range of the calculated film thickness is controlled within ±2 nm, demonstrating good repeatability and stability.

[0081] The fiber optic spectrometer for film thickness measurement of the present invention can reduce the film thickness inversion error for typical thin films (50–120 μm range), meet the requirements of industrial testing for measurement accuracy and stability, and has good potential for widespread application.

[0082] The optical fiber spectrometer for film thickness measurement of the application supports an installation in a production pipeline pair of shot type configuration, is suitable for general film production manufacturing demand scenes, realizes overall optimization in structure layout, manufacturing process and system integration by introducing a cross type Czerny-Turner optical path structure and an integrated 3D printing box body design. Not only effectively compresses the spectral path, significantly reduces the whole machine volume and manufacturing cost, but also improves the assembly precision through the card slot positioning and optical simulation, ensures the optical axis precision and measurement stability. In addition, high light absorption coating and modular heat dissipation design are adopted inside, which effectively suppresses the influence of stray light interference and environmental temperature drift. Compared with existing high-end spectral instruments, the application has the comprehensive advantages of low cost, compact structure and strong anti-interference, especially suitable for transparent / semi-transparent film thickness non-contact online detection requirements in industrial field or resource limited environment.

[0083] The measurement method of the optical fiber spectrometer for film thickness measurement of the application has the advantages of: The physical calculation model based on interference fringe FWHM is selected for priori, and then the wide spectrum thickness prediction model of convolutional neural network is used for accurate thickness calculation. The latter learns the non-linear relationship between spectral features and film thickness through the training set to improve robustness. The whole method does not need complex physical parameter inversion, realizes thickness prediction through end-to-end neural network mapping, is simple to operate, has strong anti-interference ability, and is suitable for online detection scenes in industrial production. In addition, the method increases the initial value guidance. Firstly, it can reduce the difficulty of network learning: the physical initial value gives a rough value close to the true thickness, which reduces the search range of the model and significantly improves the convergence speed. Only the "fine correction amount" needs to be learned on this basis, instead of starting from the initial value, so as to greatly reduce the iteration times, reach the convergence state faster, reduce the search range of the model, improve the convergence speed, and also reduce the risk of falling into local optimum. Secondly, the robustness of complex scenes is improved: for samples with large spectral noise and fuzzy interference peaks (such as ultra-thin or highly scattering films), the physical constraint loss avoids outputting unreasonable results, and the prediction error is reduced; the physical interpretability is enhanced: the predicted thickness The physical initial value and the spectral correction amount are composed, and the size of the correction amount can reflect the degree of supplement of the spectral information to the initial value, solving the explanation difficulty of the "black box" output of the traditional neural network.

[0084] The embodiments of the application only represent the preferred embodiments of the optical fiber spectrometer for film thickness measurement of the application, and any optical fiber spectrometer designed by adopting similar technical features of the application will fall within the protection scope of the optical fiber spectrometer for film thickness measurement of the application.

Claims

1. Optical fiber spectrometer for film thickness measurement, characterized in that, It comprises a control and data processing module (10), a detection module (5), and a light source module (7) and a light splitting module (9) connected through a fiber coupling module (8), wherein the light splitting module (9) and the detection module (5) are arranged in a shell (6); The fiber coupling module (8) couples the wide-spectrum light of the light source module (7) and guides it to the light splitting module (9); The light splitting module (9) adopts a cross-type Czerny-Turner optical path structure to sequentially perform collimation, dispersion, and focusing on the light beam; The detection module (5) is used for receiving the light signal processed by the light splitting module (9) and forming an interference spectrum; The control and data processing module (10) drives the collection, completes the pixel-wavelength calibration, and calculates the film thickness based on the wide-spectrum thickness prediction model of the convolutional neural network.

2. The fiber-optic spectrometer for film thickness measurements according to claim 1, characterized in that, The light splitting module (9) comprises a slit, a collimating mirror (2), a blazed grating (3), and a focusing mirror (4) arranged in the shell (6) in sequence according to the cross-type optical path, the collimating mirror (2) and the focusing mirror (4) are arranged adjacently or in connection, and the collimating mirror (2) and the focusing mirror (4) are respectively located on the two sides of the blazed grating (3) to form a V-shaped optical path, the slit is arranged on one side of the shell (6) as an entrance, and a slit piece (1) is arranged in the slit, and the detection module (5) is located at the end of the light path of the focusing mirror (4) on the shell (6).

3. The fiber optic spectrometer for film thickness measurements of claim 2, wherein, The slit width is 25μm±5μm, the collimating mirror (2) and the focusing mirror (4) are both spherical mirrors, and the curvature radius of the focusing mirror (4) is greater than or equal to that of the collimating mirror (2).

4. The fiber optic spectrometer for film thickness measurements of claim 1, wherein, The shell (6) adopts a 3D printing integrated structure, comprising a top cover and a box body, the top cover and the box body are detachably connected, and the inner wall of the shell (6) is provided with a black matte light-absorbing coating.

5. The fiber optic spectrometer for film thickness measurements of claim 1, wherein, The light source module (7) comprises a calibration light source and a measurement light source, the calibration light source adopts a fluorescent lamp, a mercury-argon lamp, and a monochromatic laser pen for calibration, and the measurement light source adopts an LED wide-spectrum light source to provide a wide-spectrum continuous light covering the 380-780nm waveband.

6. The fiber optic spectrometer for film thickness measurements of claim 1, wherein, The fiber coupling module comprises a fiber and a standard SMA905 connector.

7. The fiber optic spectrometer for film thickness measurements of claim 1, wherein, The detection module (5) comprises a TCD1304 linear array CCD detector.

8. The fiber optic spectrometer for film thickness measurements of claim 1, wherein, The pixel-wavelength calibration is specifically that the wavelength calibration is completed by a fluorescent lamp standard spectrum line, and a pixel-wavelength fitting relationship measured is: λ = 0.2408·P + 1076.5328, wherein λ is a wavelength, nm; and P is a pixel number. 0.2408·P + 1076.5328, wherein λ is a wavelength, nm; and P is a pixel number.

9. The measurement method of a fiber-optic spectrometer for film thickness measurement according to any one of claims 1 to 8, characterized in that, The steps are as follows: The wide-spectrum continuous light of the light source module (7) is coupled through the fiber coupling module (8) and then input into the light splitting module (9) for collimation, dispersion, and focusing, and finally imaged on the detection module (5) to form an interference spectrum, the control and data processing module (10) converts the collected pixel-light intensity into wavelength-light intensity through a pixel-wavelength mapping function, and then uses a wide-spectrum thickness prediction model based on a convolutional neural network to predict the film thickness.

10. The measurement method of a fiber-optic spectrometer for film thickness measurement according to claim 9, characterized in that, The wide-spectrum thickness prediction model based on the convolutional neural network is constructed and trained as follows: Step 1, dataset construction, including measurement data and simulation data, content is "transmittance spectrum-thickness", transmittance T(λ) of different wavelengths under thickness d, all data are mean normalized pretreatment, and the processed data set is obtained and divided into training set and validation set; Step 2, initial film thickness calculation based on the physical calculation model of the FWHM of interference fringes: wherein, is the initial thickness, n is the refractive index of the film, FSR is the wavelength difference between two adjacent interference peaks of the same order, in nm, is the center wavelength of the adjacent interference peaks, and is the wavelength of the adjacent two interference peaks of the same order, m is the interference order, ; Step 3, neural network model building and training: The convolutional neural network structure comprises an input layer, two convolutional layers, two pooling layers, a reshaping layer, and a fully connected layer, and the specific configuration is as follows: The input layer receives one-dimensional spectral data with a dimension of 1×300; Convolutional layer: contains two layers of convolution, both using 3x3 convolution kernel, step length is 1, and the activation function uses ReLU, which is used to extract local spectral features, i.e. interference peaks and absorption valleys; Pooling layer: two layers are alternately arranged with convolutional layer, both using 2x2 maximum pooling with step length of 2, which is used to down-sample the convolutional features, reduce data dimension and computational complexity while retaining core features; Fully connected layer: the high-dimensional features after pooling are unfolded into a one-dimensional vector, and the prediction result, i.e., the film optical thickness prediction value d, is output z ; The network training specifically comprises: data and thickness initial value inputting the neural network, and adopting a three-level hybrid loss function converging, wherein, is a data fitting loss, is a physical constraint loss, is an initial value guiding loss, is used for measuring a direct deviation of a predicted thickness from an actual thickness, N is a total number of data set samples, is a weight coefficient, is an initial value weight; wherein, is the predicted value of the film thickness of the model for the i-th sample; is the theoretical calculated value of the film interference, actually corresponds to the i-th interference peak wavelength , n is the refractive index of the film sample i, is the constraint prediction value of the incidence angle; ， is a squared loss function that averages the deviation of the "predicted value from the initial reference value" for N samples, where, the model's prediction of the film thickness for the i-th sample, is the initial value of the thickness derived from a physical calculation model; Adam optimizer is selected, and network parameters are iteratively adjusted by gradient descent method and back propagation algorithm to minimize loss function. A phased training strategy is used, and the initial value of the subnetwork is corrected. After training, the trained wide-spectrum thickness prediction model is obtained.

Images