CN118602958A - A method for measuring film thickness online - Google Patents

Abstract

The invention discloses an online film thickness measurement method, comprising the following steps: calibrating a film thickness measurement device; selecting the substrate of a sample to be measured and the material type of the sample; placing the substrate of the sample to be measured on a sample placement table, detecting the reflected bright light field spectrum and the reflected dark light field spectrum of the substrate; placing the substrate with the film to be measured attached on the sample placement table, subtracting the difference between the reflected bright light field spectrum of the substrate and the reflected dark light field spectrum of the substrate from the detected reflected spectrum, and obtaining the reflection spectrum of the film; after completing the acquisition of the reflection spectrum of the film, a microcomputer automatically analyzes the reflection spectrum of the film and outputs the measurement result of the film thickness; changing the position of the sample placement table to realize multi-point measurement of the film thickness. The method of the invention uses a spectrometer to collect the film reflection spectrum, and can realize the online measurement of the film thickness through a film thickness measurement network model, with fast measurement speed, accurate results and low cost.

Description

A method for measuring film thickness online

Technical Field

The invention relates to the technical field of film thickness measurement, and in particular to an online film thickness measurement method.

Background Art

Optoelectronic devices are generally composed of thin films. During their manufacturing process, the thickness of the film is a key factor affecting the optoelectronic devices. Therefore, it is necessary to accurately detect the thickness of the film when preparing the film.

At present, the thin film thickness measurement technology is mainly divided into mechanical scanning method and optical measurement method. The mechanical scanning method, such as the step meter, has high accuracy and measurement range, but it is subject to many limitations in actual production due to factors such as low measurement speed, only single-point film thickness measurement and the need to destroy the film. At present, the optical measurement method has become the mainstream method for thin film thickness measurement, and can be divided into ellipsometer, structured illumination microscope, etc. according to the experimental principle.

The ellipsometer measures film thickness based on the fact that when incident light of a certain polarization state is reflected on the sample surface, the amplitude attenuation ratio and phase difference of the s-wave and p-wave of the reflected light change, which eventually causes the polarization state of the reflected light to change. Then, the incident light vector is multiplied point by point through the Jones matrix to obtain the electromagnetic field vectors of the s-wave and p-wave and calculate the corresponding reflection coefficient. According to the Fresnel formula, the ratio of the two reflection coefficients can be used to inversely solve the theoretical expression of the phase difference of the s-wave and p-wave amplitude attenuation coefficients, where the variables are the film thickness and refractive index. Therefore, the thickness can be calculated when the refractive index is known, or the film refractive index can be calculated when the thickness is known.

Using structured light illumination microscopy (MSIM) to measure film thickness, the thickness profile of a certain range of the film can be obtained. This method projects the sinusoidal fringe pattern generated by the digital micromirror device onto the film, uses CCD to obtain the film substrate and surface reflection image, and then moves the sample to the Z direction position, determines the focus position according to the change in the contrast ratio of the interference fringes, and compares the theoretical light intensity distribution with the film optical thickness as a variable at this position, so as to achieve film thickness measurement.

However, the above optical measurement method and mechanical scanning method have the disadvantages of slow measurement speed, few measurement points, high cost, etc., which are difficult to meet the actual production needs.

Summary of the invention

In order to overcome the above-mentioned shortcomings and deficiencies of the prior art, the purpose of the present invention is to provide a method for online measurement of thin film thickness, which adopts a spectrometer to collect the thin film reflection spectrum, and can realize online measurement of thin film thickness through a thin film thickness measurement network model, with fast measurement speed, accurate results and low cost.

Another object of the present invention is to provide a method for generating a film thickness measurement network model, which has fast iterative convergence speed and high measurement accuracy during training.

Another object of the present invention is to provide a device for generating the above-mentioned film thickness measurement network model.

The purpose of the present invention is achieved through the following technical solutions:

One embodiment of the present invention provides a method for online measurement of film thickness, wherein the method is based on a film thickness measurement device; the film thickness measurement device includes a sample unit, an optical path unit and a data processing unit;

The sample unit comprises a sample placement platform and a platform moving motor; the platform moving motor is used to drive the sample placement platform to move on a two-dimensional plane;

The optical path unit comprises a tungsten halogen lamp light source, an optical fiber probe, a spectrometer, a first optical fiber and a second optical fiber; the tungsten halogen lamp light source is connected to the optical fiber probe via the first optical fiber, and the optical fiber probe is connected to the spectrometer via the second optical fiber; the optical fiber probe is located above the sample placement table;

When measuring a thin film, the emitted light of the halogen tungsten lamp light source is output to the optical fiber probe through the first optical fiber, incident into the thin film and causing multi-beam interference at the upper and lower interfaces of the thin film. The reflected light carrying the film thickness information is received by the optical fiber probe and transmitted to the spectrometer through the second optical fiber. The spectrometer collects the reflected spectrum data, and after being analyzed and processed by the data processing unit, the measurement result of the film thickness is output;

The data processing unit is equipped with a film thickness measurement network model;

The film thickness measurement network model is obtained by the following method:

S11 dataset generation:

S111 obtains the reflectivity fitting function of the refractive index of the film and the substrate as it changes with the wavelength through polynomial nonlinear fitting of mathematical analysis software according to the refractive index of the film and the substrate;

S112 selects a wavelength band with a light coefficient close to zero, and generates a theoretical reflectivity spectrum of a thin film with a thickness between 500 nm and 10 μm according to a reflectivity fitting function of the refractive index of the thin film and substrate in the wavelength band and a set film thickness training step;

S113 uses Python built-in functions to introduce normal distribution noise into the theoretical reflectance spectrum of thin films, generate multiple samples, and divide the multiple samples into training sets and validation sets;

S114 changes the materials of the film and the substrate, repeats steps S111 to S114, and obtains data sets of films and substrates of different materials;

S12 convolutional neural network model construction:

The convolutional neural network model includes a first convolutional activation layer, a first pooling layer, a second convolutional activation layer, a second pooling layer, an expansion layer and a fully connected layer connected in sequence; the first convolutional activation layer includes a one-dimensional convolution operation module and a nonlinear activation function module; the first convolutional activation layer includes a first one-dimensional convolution operation module and a first nonlinear activation function module; the second convolutional activation layer includes a second one-dimensional convolution operation module and a second nonlinear activation function module; the convolutional neural network model takes the reflectance spectrum of the film as input and the thickness of the film as output;

S13 convolutional neural network model training;

S14 Convolutional neural network model verification;

The film thickness online measurement method comprises the following steps:

S1 calibrates the film thickness measuring device using a standard sample;

S2 selects the substrate of the sample to be tested and the material type of the sample;

S3 places the substrate of the sample to be tested on the sample placement table, and controls the platform to move the motor so that it is located directly below the optical fiber probe, and detects the reflected bright light field spectrum and reflected dark light field spectrum of the substrate;

S4 places the substrate with the film to be tested on the sample placement table, and controls the platform moving motor to make it directly below the optical fiber probe, and subtracts the difference between the reflected bright light field spectrum of the substrate and the reflected dark light field spectrum of the substrate from the detected reflection spectrum to obtain the reflection spectrum of the film;

After S5 completes the acquisition of the reflection spectrum of the film, the data processing unit automatically analyzes the reflection spectrum of the film and outputs the measurement result of the film thickness;

S6 controls the platform movement motor to change the position of the sample placement table to achieve multi-point measurement of film thickness.

In one embodiment of the present invention, the convolutional neural network model training in step S13 is specifically carried out in the following manner:

The Adam optimization method is used, with a learning rate of 0.005 and cross entropy as the loss function. If the loss value decreases within multiple iterations, the learning rate decreases by 10%.

In one embodiment of the present invention, in the convolutional neural network model building step described in step S12, the input is set to 50 200×1 matrices, the first one-dimensional convolution operation module is set to 64 convolution kernels, and the second one-dimensional convolution operation module is set to 128 convolution kernels. After the convolution operation is completed, it is flattened to obtain a 6400×1 matrix, and finally all connected to the film thickness label.

In one of the embodiments, the convolution kernel is set to 3×1, the stride is set to 1, and the edge zero padding is set to 1; the pooling kernel in the first pooling layer and the second pooling layer is set to 3×1, and the stride is set to 2.

In one embodiment, the film thickness training step size is 50 nm.

In one embodiment, the film thickness measuring device also includes a display and control unit; the display and control unit includes a display screen and control buttons; the display screen is used to display the reflection spectrum of the sample and the measurement results of the film thickness; the control unit is used to set sample parameters and measurement parameters.

In one of the embodiments, the sample placement table is also provided with an adjustment knob for adjusting the angle between the sample placement table and the horizontal plane so that the sample placement table can be placed horizontally; the halogen tungsten lamp light source is also connected to the light source adjustment knob; the light source adjustment knob is used to adjust the light intensity of the halogen tungsten lamp light source.

In one embodiment, the data processing unit is a microcomputer; the film thickness measurement network model is written in Python and is imported into the microcomputer after training and verification.

One of the embodiments of the present invention further provides a method for generating a film thickness measurement network model, comprising the following steps: S11: generating a data set:

S111 obtains the reflectivity fitting function of the refractive index of the film and the substrate as it changes with the wavelength through polynomial nonlinear fitting of mathematical analysis software according to the refractive index of the film and the substrate;

S112 selects a wavelength band with a light coefficient close to zero, and generates a theoretical reflectivity spectrum of a thin film with a thickness between 500 nm and 10 μm according to a reflectivity fitting function of the refractive index of the thin film and substrate in the wavelength band and a set film thickness training step;

S113 uses Python built-in functions to introduce normally distributed noise into the theoretical reflectance spectrum of thin films, generate multiple samples, and divide the samples into training sets and validation sets;

S114 changes the materials of the film and the substrate, repeats steps S111 to S113, and obtains data sets of films and substrates of different materials;

S12 convolutional neural network model construction:

The convolutional neural network model includes a first convolutional activation layer, a first pooling layer, a second convolutional activation layer, a second pooling layer, a fully connected layer and an unfolding layer connected in sequence; the first convolutional activation layer includes a first one-dimensional convolution operation module and a first nonlinear activation function module; the second convolutional activation layer includes a second one-dimensional convolution operation module and a second nonlinear activation function module; the convolutional neural network model takes the reflectance spectrum of the film as input and the thickness of the film as output;

S13 Convolutional Neural Network Model Training: Adopt Adam optimization method, set learning rate as 0.005, and use cross entropy as loss function. If the loss value decreases within multiple iterations, the learning rate decreases by 10%.

S14 Convolutional neural network model verification.

One of the embodiments of the present invention further provides a device for generating a film thickness measurement network model, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to perform the steps of the method.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

(1) The thin film thickness online measurement method of the present invention uses a spectrometer to collect the thin film reflection spectrum and generates a thin film thickness measurement network model with the thin film reflection spectrum as input, which can realize the online measurement of the thin film thickness with fast measurement speed, accurate results and low cost.

(2) In the method for generating a network model for measuring film thickness of the present invention, in the step of generating a data set, a theoretical reflectance spectrum of a film with a thickness between 500nm and 10μm is first generated, and normally distributed noise is introduced into the theoretical reflectance spectrum of the film to generate multiple samples, thereby improving the diversity of the data and preventing overfitting. At the same time, it ensures that a large number of training samples can be generated under the same film thickness, thereby ensuring the accuracy of the measurement network model.

(3) The thin film thickness measurement network model generation method of the present invention builds a neural network consisting of two convolution activation layers, two pooling layers, one flattening layer and one fully connected layer. By optimizing the parameters and training of the neural network, after 20 iterations, the classification accuracy of the neural network can reach 99%.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a schematic diagram showing the composition of a thin film thickness measuring device according to an embodiment of the present invention.

FIG. 2 is a schematic cross-sectional view of a control cabinet of a film thickness measuring device according to an embodiment of the present invention.

FIG. 3 is a flow chart of a network model for measuring film thickness according to an embodiment of the present invention.

FIG. 4 is a schematic diagram of a spectrum generation process according to an embodiment of the present invention.

FIG. 5 is a reflectivity fitting function of the refractive index of the film and substrate obtained according to an embodiment of the present invention as a function of wavelength.

FIG. 6 is a theoretical reflectivity spectrum before and after the introduction of noise in an embodiment of the present invention.

FIG7 is a schematic diagram of a convolutional neural network model according to an embodiment of the present invention.

FIG8 shows the change in loss values of the training set and the test set in an embodiment of the present invention.

FIG. 9 is a flow chart of a method for measuring film thickness according to an embodiment of the present invention.

FIG. 10 is a comparison diagram of an actual spectrum and a predicted spectrum in an embodiment of the present invention.

DETAILED DESCRIPTION

The present invention will be further described in detail below in conjunction with the examples, but the embodiments of the present invention are not limited thereto.

Example

Please refer to FIG. 1 and FIG. 2 , one embodiment of the present invention provides a thin film thickness measuring device, including a data processing unit, an optical path unit and a sample unit.

The data processing unit includes a microcomputer 2; the sample unit includes a sample placement table 3, a platform moving motor 4 and an adjustment knob 5, wherein the platform moving motor 4 is used to drive the sample placement table 3 to move on a two-dimensional plane; the adjustment knob 5 is used to adjust the angle between the sample placement table 3 and the horizontal plane, so that the sample placement table 3 is adjusted to be placed horizontally. The optical path unit includes a halogen tungsten lamp light source 6, an optical fiber probe 7, a first optical fiber 8, a spectrometer 1 and a second optical fiber 9; the halogen tungsten lamp light source 6 is connected to the optical fiber probe 7 through the first optical fiber 8, and the optical fiber probe 7 is connected to the spectrometer 1 through the second optical fiber 9; the optical fiber probe 7 is located above the sample placement table 3. The halogen tungsten lamp light source 6, the spectrometer 1 and the microcomputer 2 are located in a control cabinet 10. The control cabinet 10 is provided with a display screen and control buttons connected to the microcomputer 2. The halogen tungsten lamp light source 6 is provided with a light source adjustment knob for adjusting the light intensity of the halogen tungsten lamp light source.

In the above embodiment, when the film thickness measuring device measures the film, the outgoing light of the halogen tungsten lamp light source is output to the fiber optic probe via the first optical fiber, incident on the film and causing multi-beam interference at the upper and lower interfaces of the film, and the reflected light carrying the film thickness information is received by the fiber optic probe and transmitted to the spectrometer via the second optical fiber. The spectrometer collects the reflected spectrum data, and after being analyzed and processed by the data processing unit, outputs the measurement result of the film thickness.

In one embodiment of the present invention, a film thickness measurement network model is built into a microcomputer. The film thickness measurement network model is written in Python and is imported into the microcomputer after training and verification.

Referring to FIG. 3 , in one embodiment of the present invention, a method for generating a film thickness measurement network model is as follows:

S11 dataset generation:

S111 obtains the reflectivity fitting function of the refractive index of the film and the substrate as it changes with the wavelength through polynomial nonlinear fitting of mathematical analysis software according to the refractive index of the film and the substrate;

S112 selects a wavelength band with a light coefficient close to zero, and uses a spectrum generation module to generate a theoretical reflectivity spectrum of a film with a thickness between 500 nm and 10 μm according to a reflectivity fitting function of the film and substrate refractive index in the wavelength band and a set film thickness training step;

S113 uses Python built-in functions to introduce normal distribution noise into the theoretical reflectance spectrum of thin films, generate multiple samples, and divide the multiple samples into training sets and validation sets;

S114 changes the materials of the film and the substrate, repeats steps S111 to S113, and obtains data sets of films and substrates of different materials;

S12 convolutional neural network model construction:

In one embodiment, the convolutional neural network model includes a first convolutional activation layer, a first pooling layer, a second convolutional activation layer, a second pooling layer, an expansion layer and a fully connected layer connected in sequence; the first convolutional activation layer includes a first one-dimensional convolution operation module and a first nonlinear activation function module; the second convolutional activation layer includes a second one-dimensional convolution operation module and a second nonlinear activation function module; the convolutional neural network model uses the reflectance spectrum of the film as input and the thickness of the film as output;

S13 convolutional neural network model training;

S14 Convolutional neural network model verification.

In the above embodiment, the principle of the spectrum generation module is as follows:

As shown in FIG4 , the incident angle of the light in the plate is θ ₀ , and the optical path difference between the two consecutive light beams is D = 2n ₁ dcosθ ₀ . If θ ₀ = 0 in the case of normal incidence, the corresponding phase difference is Where n ₁ d is the optical thickness of the film, and λ is the wavelength of light in a vacuum. According to the Fresnel formula, the reflection coefficients on the upper and lower surfaces of the film are:

The first order reflected beam and other order reflected beams can be described as:

The total intensity of the reflected light can be obtained as follows:

Therefore, the film reflectivity can be expressed as:

It can be seen that there are two unknown quantities in this reflectivity generation model, namely the wavelength of light and the thickness of the film.

In one embodiment of the present invention, taking a silicon substrate and a Pi film as an example, the specific settings in the process of generating a film thickness measurement network model are as follows:

Dataset generation: After obtaining the reflectivity fitting function of the refractive index of the film and substrate with wavelength (as shown in Figure 5), select the part with extinction coefficient close to zero (such as less than 0.05), that is, the 600-800nm band, and generate the theoretical reflectivity spectrum of the film with thickness between 500nm-10μm and interval of 50nm. After introducing normal distribution noise according to Python built-in function, it is used as the model training set. The theoretical reflectivity spectra before and after the introduction of noise are shown in (a) and (b) in Figure 6 respectively. There are a total of 190 film thickness labels, each label has 30 training samples and 20 test samples. A total of 9500 samples in the training set and test set are imported into the model for training and verification.

Convolutional neural network model construction: Figure 7 shows the composition and settings of the convolutional neural network model. The convolutional neural network model consists of two layers of convolution activation layers (Conv and ReLU), two layers of pooling layers (Pooling), one flattening layer (Reshaping) and one fully connected layer (FC). The input is set to 50 200×1 matrices. The first convolution operation sets 64 convolution kernels, and the second convolution operation sets 128 convolution kernels. After the convolution operation is completed, it is flattened to obtain a 6400×1 matrix, and finally fully connected to 190 film thickness labels. According to many practices, the convolution kernel is set to 3×1, the stride is set to 1, the padding is set to 1, the pooling kernel is set to 3×1, and the stride is set to 2, which can obtain the best classification effect with the minimum number of iterations.

Convolutional neural network model training: Adopt Adam optimization method, with a learning rate of 0.005. If the loss value decreases within multiple iterations, the learning rate decreases by 10% to ensure that the model is efficient and close to the optimal solution. Using cross entropy as the loss function, after 20 iterations of the model, the loss values of both the training set and the test set have been significantly reduced (as shown in Figure 8). The accuracy of the classification of theoretical spectral data reached 99%, which fully demonstrated the model's ability to classify spectral data of different film thicknesses.

After training the film thickness measurement network model, the film thickness measurement is performed. The specific steps are as follows (as shown in Figure 9):

S1 calibrates the film thickness measuring device using a standard sample;

S2 selects the substrate of the sample to be tested and the material type of the sample;

S3 places the substrate of the sample to be tested on the sample placement table, and controls the platform to move the motor so that it is located directly below the optical fiber probe, and detects the reflected bright light field spectrum and reflected dark light field spectrum of the substrate;

S4 places the substrate with the film to be tested on the sample placement table, and controls the platform moving motor to make it directly below the optical fiber probe, and subtracts the difference between the reflected bright light field spectrum of the substrate and the reflected dark light field spectrum of the substrate from the detected reflection spectrum to obtain the reflection spectrum of the film;

After S5 completes the acquisition of the reflection spectrum of the film, the microcomputer automatically analyzes the reflection spectrum of the film and outputs the measurement result of the film thickness;

S6 controls the platform movement motor to change the position of the sample placement table to achieve multi-point measurement of film thickness.

The embodiments of the present invention also carry out the following verifications: first, the predicted theoretical spectrum data value is compared with the actual spectrum data; second, the actual film thickness is measured using a step profiler and compared with the predicted value of the present invention.

Verification 1:

The thin film sample (taking silicon substrate and Pi film as examples) is measured through the above steps S1 to S4, with four sampling points in the measurement process, and the actual spectrum data is obtained; the actual spectrum is compared with the theoretical spectrum (generated by steps S111 to S112), and the results are shown in (a) to (d) in Figure 10. As shown in Figure 10, in the range of 600-800nm, the actual spectrum is basically consistent with the predicted spectrum. When the refractive index remains unchanged, the thickness of the film depends on the size of the peak spacing of the reflectivity spectrum.

Verification 2:

The XP-200 step profiler from Ambios Technology Company of the United States was used as the film thickness verification instrument. The parameters of the step profiler are as follows: stage diameter: 6 inches; scanning length: 50μm to 55mm; vertical measurement range: 524μm; 3D scanning function module: including measurement software, color zoom camera, precision vibration isolation platform and micron standard calibration module.

First, the film thickness measurement network model of this embodiment is used to predict the thickness of four sampling points on the film sample (taking silicon substrate and Pi film as examples). Then, the vertical distance between the film surface and the substrate surface is measured three times at the sampling point using the step meter measurement method. The average is taken to calculate the film thickness at the four points to verify the accuracy of the measurement results of the present invention. The results are shown in Table 1.

The data in Table 1 show that the error is between 1.1% and 3.6%, which proves that the present invention has a good film thickness measurement level and has high reliability and repeatability.

Table 1 Sample data statistics

The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principles of the present invention shall be equivalent replacement modes and shall be included in the protection scope of the present invention.

Claims (10)

1. A thin film thickness online measurement method, characterized in that the thin film thickness online measurement method is based on a thin film thickness measurement device; the thin film thickness measurement device includes a sample unit, an optical path unit and a data processing unit; The sample unit comprises a sample placement platform and a platform moving motor; the platform moving motor is used to drive the sample placement platform to move on a two-dimensional plane; The optical path unit comprises a tungsten halogen lamp light source, an optical fiber probe, a spectrometer, a first optical fiber and a second optical fiber; the tungsten halogen lamp light source is connected to the optical fiber probe via the first optical fiber, and the optical fiber probe is connected to the spectrometer via the second optical fiber; the optical fiber probe is located above the sample placement table; When measuring a thin film, the emitted light of the halogen tungsten lamp light source is output to the optical fiber probe through the first optical fiber, incident into the thin film and causing multi-beam interference at the upper and lower interfaces of the thin film. The reflected light carrying the film thickness information is received by the optical fiber probe and transmitted to the spectrometer through the second optical fiber. The spectrometer collects the reflected spectrum data, and after being analyzed and processed by the data processing unit, the measurement result of the film thickness is output; The data processing unit is equipped with a film thickness measurement network model; The film thickness measurement network model is obtained by the following method: S11 dataset generation: S111 obtains the reflectivity fitting function of the refractive index of the film and the substrate as it changes with the wavelength through polynomial nonlinear fitting of mathematical analysis software according to the refractive index of the film and the substrate; S112 selects a wavelength band with a light coefficient close to zero, and generates a theoretical reflectivity spectrum of a thin film with a thickness between 500 nm and 10 μm according to a reflectivity fitting function of the refractive index of the thin film and substrate in the wavelength band and a set film thickness training step; S113 uses Python built-in functions to introduce normal distribution noise into the theoretical reflectance spectrum of thin films, generate multiple samples, and divide the multiple samples into training sets and validation sets; S114 changes the materials of the film and the substrate, repeats steps S111 to S114, and obtains data sets of films and substrates of different materials; S12 convolutional neural network model construction: The convolutional neural network model includes a first convolutional activation layer, a first pooling layer, a second convolutional activation layer, a second pooling layer, an expansion layer and a fully connected layer connected in sequence; the first convolutional activation layer includes a first one-dimensional convolution operation module and a first nonlinear activation function module; the second convolutional activation layer includes a second one-dimensional convolution operation module and a second nonlinear activation function module; the convolutional neural network model takes the reflectance spectrum of the film as input and the thickness of the film as output; S13 convolutional neural network model training; S14 Convolutional neural network model verification; The film thickness online measurement method comprises the following steps: S1 calibrates the film thickness measuring device using a standard sample; S2 selects the substrate of the sample to be tested and the material type of the sample; S3 places the substrate of the sample to be tested on the sample placement table, and controls the platform to move the motor so that it is located directly below the optical fiber probe, and detects the reflected bright light field spectrum and reflected dark light field spectrum of the substrate; S4 places the substrate with the film to be tested on the sample placement table, and controls the platform moving motor to make it directly below the optical fiber probe, and subtracts the difference between the reflected bright light field spectrum of the substrate and the reflected dark light field spectrum of the substrate from the detected reflection spectrum to obtain the reflection spectrum of the film; After S5 completes the acquisition of the reflection spectrum of the film, the data processing unit automatically analyzes the reflection spectrum of the film and outputs the measurement result of the film thickness; S6 controls the platform movement motor to change the position of the sample placement table to achieve multi-point measurement of film thickness. 2. The online film thickness measurement method according to claim 1 is characterized in that the convolutional neural network model training in step S13 is specifically carried out in the following manner: The Adam optimization method is used, with a learning rate of 0.005 and cross entropy as the loss function. If the loss value decreases within multiple iterations, the learning rate decreases by 10%. 3. The online measurement method for thin film thickness according to claim 1 is characterized in that, in the convolutional neural network model building step described in step S12, the input is set to 50 200×1 matrices, the first one-dimensional convolution operation module is set to 64 convolution kernels, and the second one-dimensional convolution operation module is set to 128 convolution kernels. After the convolution operation is completed, it is flattened to obtain a 6400×1 matrix, and finally all connected to the film thickness label. 4. The online measurement method for thin film thickness according to claim 3 is characterized in that the convolution kernel is set to 3×1, the stride is set to 1, and the edge zero padding is set to 1; the pooling kernel in the first pooling layer and the second pooling layer is set to 3×1, and the stride is set to 2. 5 . The online film thickness measurement method according to claim 1 , wherein the film thickness training step length is 50 nm. 6. According to the online film thickness measurement method of claim 1, the film thickness measurement device also includes a display and control unit; the display and control unit includes a display screen and control buttons; the display screen is used to display the reflection spectrum of the sample and the measurement results of the film thickness; the control unit is used to set sample parameters and measurement parameters. 7. The online measurement method for thin film thickness according to claim 1 is characterized in that the sample placement table is also provided with an adjustment knob for adjusting the angle between the sample placement table and the horizontal plane so that the sample placement table can be placed horizontally; the halogen tungsten lamp light source is also connected to the light source adjustment knob; the light source adjustment knob is used to adjust the light intensity of the halogen tungsten lamp light source. 8. The online film thickness measurement method according to claim 1 is characterized in that the data processing unit is a microcomputer; the film thickness measurement network model is written in Python and is imported into the microcomputer after training and verification. 9. A method for generating a film thickness measurement network model, characterized in that it comprises the following steps: S11 dataset generation: S111 obtains the reflectivity fitting function of the refractive index of the film and the substrate as it changes with the wavelength through polynomial nonlinear fitting of mathematical analysis software according to the refractive index of the film and the substrate; S112 selects a wavelength band with a light coefficient close to zero, and generates a theoretical reflectivity spectrum of a thin film with a thickness between 500 nm and 10 μm according to a reflectivity fitting function of the refractive index of the thin film and substrate in the wavelength band and a set film thickness training step; S113 uses Python built-in functions to introduce normally distributed noise into the theoretical reflectance spectrum of thin films, generate multiple samples, and divide the samples into training sets and validation sets; S114 changes the materials of the film and the substrate, repeats steps S111 to S113, and obtains data sets of films and substrates of different materials; S12 convolutional neural network model construction: The convolutional neural network model includes a first convolution activation layer, a first pooling layer, a second convolution activation layer, a second pooling layer, a fully connected layer and an unfolding layer connected in sequence; the first convolution activation layer includes a first one-dimensional convolution operation module and a first nonlinear activation function module; the second convolution activation layer includes a second one-dimensional convolution operation module and a second nonlinear activation function module; the convolutional neural network model takes the reflectance spectrum of the film as input and the thickness of the film as output; S13 Convolutional Neural Network Model Training: Adopt Adam optimization method, set learning rate as 0.005, and use cross entropy as loss function. If the loss value decreases within multiple iterations, the learning rate decreases by 10%. S14 Convolutional neural network model verification. 10. A device for generating a network model for measuring film thickness, characterized in that it comprises a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to execute the steps of the method according to claim 9.