1. A method for determining an optical constant of a material, comprising:

acquiring ellipsometric parameters of the material; and

inputting the ellipsometric parameters into a machine learning model to obtain the optical constant of the material corresponding to the ellipsometric parameters; the machine learning model comprising a mapping relationship between the ellipsometric parameters and the optical constant of the material;

wherein the machine learning model is established by:

constructing each neural network layer of an initial model based on the mapping relationship between ellipsometric parameters and the optical constant of the material corresponding to the ellipsometric parameters;

constructing sample data which comprises the ellipsometric parameters and the optical constant of the material; and

training the each neural network layer of the initial model with the sample data to obtain the machine learning model;

wherein the training the each neural network layer of the initial model with the sample data to obtain the machine learning model comprises:

dividing the sample data into a training set, a validation set, and a test set;

inputting the training set and the validation set into the initial model respectively, optimizing the initial model to obtain optimization models and recording the optimization models constantly, meanwhile generating an error curve of the training set and an error curve of the validation set;

selecting M optimization models which perform best on the validation set from all optimization models, and averaging parameters of the M optimization models to obtain a second optimization model; and

inputting input parameters in the test set into the second optimization model to obtain a second output parameter which is obtained by an operation via the second optimization model, determining an error between the second output parameter and a reference output parameter in the test set, and taking the second optimization model as the machine learning model when the error is within a preset range;

or

wherein the training the each neural network layer of the initial model with the sample data to obtain the machine learning model comprises:

dividing the sample data into a training set, a validation set, and a test set;

inputting the training set and the validation set into the initial model respectively, optimizing the initial model to obtain optimization models and recording the optimization models constantly, meanwhile generating an error curve of the training set and an error curve of the validation set;

randomly selecting N optimization models from all optimization models, and averaging parameters of the N optimization models to obtain a third optimization model; and

inputting input parameters in the test set into the third optimization model to obtain a third output parameter which is obtained by an operation via the third optimization model, determining an error between the third output parameter and a reference output parameter in the test set, and taking the third optimization model as the machine learning model when the error is within a preset range.

2. The method according to claim 1, wherein the training the each neural network layer of the initial model with the sample data to obtain the machine learning model comprises:

inputting the ellipsometric parameters to the initial model to obtain current output data;

calculating a difference value between the current output data and a corresponding optical constant of the material, and when the difference value is greater than or equal to a preset value, adjusting network parameters of the each neural network layer of the initial model according to the difference value; and

calculating the difference value by an iterative operation until the difference value is less than the preset value, determining that each neural network layer of the initial machine learning model reaches convergence and obtaining the machine learning model.

3. The method according to claim 1, wherein the training the each neural network layer of the initial model with the sample data to obtain the machine learning model comprises:

dividing the sample data into a training set, a validation set and a test set;

inputting the training set and the validation set into the initial model respectively, optimizing the initial model to obtain optimization models and recording the optimization models constantly, while generating an error curve of the training set and an error curve of the validation set;

taking an optimization model as a first optimization model at a moment when an error value of the error curve of the verification data no longer decreases; and

inputting input parameters in the test set into the first optimization model to obtain a first output parameter obtained by an operation via the first optimization model, determining an error between the first output parameter and a reference output parameter in the test set, and taking the first optimization model as the machine learning model when the error is within a preset range.

4. The method according to claim 3, wherein the sample data of the training set, the validation set and the test set is homogeneous data.

5. The method according to claim 1, further comprising:

adjusting the sample data such that input parameters corresponding to different wavelength ranges correspond to a same wavelength range; and

normalizing the input parameters in the sample data such that values of the input parameters corresponding to the different wavelength ranges are in a same order of magnitude.

6. The method according to claim 1, further comprising:

taking the ellipsometric parameters as input parameters of updated data and the optical constant of the material obtained by operation via the machine learning model as an output parameter of the updated data, respectively;

inputting the input parameters of the updated data into the machine learning model to obtain current update output data;

calculating a difference value between the current updated output data and the output parameter of the updated data, and when the difference value is greater than or equal to the preset value, adjusting network parameters of the each nerve network layer of the machine learning model according to the difference value; and

calculating the difference value with an iterative operation until the difference value is less than a preset value, determining that each neural network layer of the machine learning model reaches convergence thereby completing update of the machine learning model.

7. The method according to claim 6, wherein the ellipsometric parameters comprise Δ and Ψ, and the ellipsometric parameters (Δ, Ψ) satisfy the following equation:

rprs=tan⁢Ψ⁢ej⁢Δ.

wherein rP and rs respectively represent reflectance of the film for p-polarized light and s-polarized light.

8. The method according to claim 1, wherein the machine learning model comprises one or more of a convolutional neural network, a fully-connected neural network, or a recurrent neural network.

CROSS REFERENCE TO RELATED APPLICATIONS

[0001]This application claims priority to Chinese Patent Application No. 201910500995.0, entitled “Method and Apparatus for Determining Optical Constant of Material, and Method and Apparatus for Extending Material Database”, filed on Jun. 11, 2019, the content of which is expressly incorporated by reference in its entirety.

TECHNICAL FIELD

[0002]The present disclosure relates to the field of material science and technology, and particularly to a method and an apparatus for determining an optical constant of a material, and a method and an apparatus for extending a material database.

BACKGROUND

[0003]An ellipsometer is an instrument that utilizes a polarization dependence of reflection characteristics of light waves on a surface of a material to measure an optical constant (such as a refractive index, a dielectric constant and so on) and a thickness of a thin film. The light with different polarizations reflected by the surface of the material may have different reflectances, which are related to the optical constants and the thickness of the thin film. By measuring the reflectance difference characteristic curve of incident lights with different polarization states and different wavelengths, and performing a regression fitting to the curves, the optical constant (such as the refractive index, the dielectric constant, etc.) and the thickness of the thin film can be obtained. Fast and accurate regression fitting of the curve is an important part of the ellipsometer, and is also a crucial step for obtaining optical constant information of the thin film.

[0004]In the conventional solution, the experimenters rely heavily on their own experiences and need to constantly try to adjust various fitting parameters to implement the fitting to obtain the optical constants or thin-film thickness of the material. This process requires the experimenters to have extremely rich experiences, and the regression fitting of the curve is slow and inefficient.

SUMMARY

[0005]A method for determining an optical constant of a material, includes:

[0006]acquiring ellipsometric parameters of the material; and

[0007]inputting the ellipsometric parameters into a machine learning model to obtain the optical constant of the material corresponding to the ellipsometric parameters; the machine learning model including a mapping relationship between the ellipsometric parameters and the optical constant of the material.

[0008]A method for extending a material database, includes:

[0009]acquiring ellipsometric parameters of a material;

[0010]inputting the ellipsometric parameters into a machine learning model to obtain an optical constant of the material corresponding to the ellipsometric parameters;

[0011]looking up the optical constants of the material in the material database; and

[0012]if the optical constant of the material is not included in the material database, adding the ellipsometric parameters of the material and the optical constant of the material corresponding to the ellipsometric parameters into the material database.

[0013]In an embodiment, the method further includes:

[0014]determining whether the optical constant of the material corresponding to the ellipsometric parameters is included in the material database;

[0015]if the optical constant of the material corresponding to the ellipsometric parameters is not included in the material database, adding a name of the material into the known material database.

[0016]In an embodiment, the method further includes:

[0017]further updating the machine learning model if the optical constant of the material corresponding to the ellipsometric parameters is not in a preset optical constant range of the material.

[0018]An apparatus for determining an optical constant of a material, includes:

[0019]an ellipsometric parameter acquisition module, configured to acquire ellipsometric parameters of the material; and

[0020]a material optical constant acquisition module, configured to input the ellipsometric parameters into a machine learning model to obtain the optical constant of the material corresponding to the ellipsometric parameters; the machine learning model including a mapping relationship between the ellipsometric parameter and the optical constant of the material.

[0021]An apparatus for extending a material database, includes:

[0022]an ellipsometric parameter acquisition module, configured to acquire ellipsometric parameters of the material;

[0023]a material optical constant acquisition module, configured to input the ellipsometric parameters into a machine learning model to obtain an optical constant of the material corresponding to the ellipsometric parameters;

[0024]a lookup module, configured to look up the optical constant of the material in a material database; and

[0025]an extension module, configured to add the ellipsometric parameters of the material and the optical constants of the material corresponding to the ellipsometric parameters into the material database, if the optical constant of the material is not included in the material database.

[0026]A computer device including a processor and a memory storing a computer program executable on the processor, the steps in the method of any one of the above embodiments are implemented when the processor executes the computer program.

[0027]A computer-readable storage medium on which a computer program is stored, the steps in the method of any one of the above embodiments are implemented when the computer program is executed by a processor.

[0028]In an embodiment, an ellipsometer is provided, which includes:

[0029]a light source configured to emit light;

[0030]a light selection device configured to adjust wavelength and polarization of the light;

[0031]a light polarization compensation device configured to modify polarization state of the light to form probe light, the probe light irradiating a surface of a material sample, the material including a substrate and a thin film on a surface of the substrate;

[0032]a light polarization analysis device configured to acquire the probe light reflected or refracted by the material and obtain a measurement result; the measurement result comprising ellipsometric parameters Δ and Ψ of the material; and

[0033]a material optical constant determination device, configured to determine an optical constant of the material by the method for determining an optical constant of a material according to any one of the above embodiments based on the measurement result, a wavelength of the probe light, an incident angle of the probe light, a real part of the refractive index of the thin film, and a imaginary part of the refractive index of the substrate; the optical constant including one or both of the real part and the imaginary part of the refractive index of the thin film.

[0034]The present disclosure provides a method and an apparatus for determining an optical constant of a material, and a method and an apparatus for extending a material database. The method for determining an optical constant of a material includes: acquiring ellipsometric parameters of the material; inputting the ellipsometric parameters into a machine learning model to obtain the optical constant of the material corresponding to the ellipsometric parameters; the machine learning model includes a mapping relationship between the ellipsometric parameters and the optical constant of the material. The method for determining an optical constant of a material uses the machine learning model to implement an automatic fitting of ellipsometric parameters. Since in the method for determining the optical constant of the material, the optical constant of the material is calculated by a machine learning model, which no longer depends on the experiences of the experimenters, thereby reducing the requirements for the operators, accelerating the fitting of the data curve when calculating the optical constant of the material and improving the operation efficiency.

BRIEF DESCRIPTION OF DRAWINGS

[0035]FIG. 1 is a schematic view showing a working principle of an ellipsometer according to an embodiment of the present disclosure.

[0036]FIG. 2 shows a theoretical model for determining ellipsometric parameters Δ and Ψ according to an embodiment of the present disclosure.

[0037]FIG. 3 is a flow chart of a method for determining an optical constant of a material according to an embodiment of the present disclosure.

[0038]FIG. 4 is a flow chart of acquiring a machine learning model according to an embodiment of the present disclosure.

[0039]FIG. 5 is a flow chart of acquiring a machine learning model according to an embodiment of the present disclosure.

[0040]FIG. 6 is a flow chart of acquiring a machine learning model according to an embodiment of the present disclosure.

[0041]FIG. 7 is a flow chart of acquiring a machine learning model according to an embodiment of the present disclosure.

[0042]FIG. 8 is a flow chart of acquiring a machine learning model according to an embodiment of the present disclosure.

[0043]FIG. 9 is a flow chart of a method for determining an optical constant of a material and updating a machine learning model according to an embodiment of the present disclosure.

[0044]FIG. 10 is a flow chart of a method for extending a material database according to an embodiment of the present disclosure.

[0045]FIG. 11 is a schematic view of an apparatus for determining an optical constant of a material according to an embodiment of the present disclosure.

[0046]FIG. 12 is a schematic view of an apparatus for extending a material database according to an embodiment of the present disclosure.

[0047]FIG. 13 is a data diagram of an error test in a process of acquiring a machine learning model according to an embodiment of the present disclosure.

[0048]FIG. 14 is a comparison diagram between a conventional method and a method for determining an optical constant of a material according to an embodiment of the present disclosure, in which silicon is taken as a material to be tested.

[0049]FIG. 15 is a comparison diagram between a conventional method and a method for determining an optical constant of a material according to an embodiment of the present disclosure, in which only titanium dioxide is taken as a material to be tested.

[0050]FIG. 16 is a comparison diagram between a conventional method and a method for determining an optical constant of a material according to an embodiment of the present disclosure, in which gold is taken as a material to be tested.

[0051]FIG. 17 is a schematic structure diagram of an ellipsometer according to an embodiment of the present disclosure.

DETAILED DESCRIPTION

[0052]In order to make the objectives, technical solution, and advantages of the present disclosure clearer, a method and an apparatus for determining an optical constant of a material, and a method and an apparatus for extending a material database detailed with reference to the accompanying drawings and embodiments. It should be noted that the specific embodiments described herein are merely used for explaining the present disclosure, and are not intended to limit the present disclosure.

[0053]An ellipsometer is an optical instrument for measuring an optical constant of a material meanwhile a thickness of a film. Due to the advantages such as high accuracy, usefulness for an ultra-thin film, no contact with the sample and no damage to the sample, the ellipsometer becomes an attractive instrument. Ellipsometry is high accurate, contactless, non-destructive. As shown in FIG. 1, an ellipsometer generally includes a light source, a polarizer P, a compensator C, an analyzer A, and a detector D (e.g., single-point or area array detector).

[0054]In use of the ellipsometer, light output from a continuous-spectrum white light source passes through a grating for a frequency selection to form monochromatic light. After the monochromatic light passes through the polarizer P and the compensator C, it is formed into elliptically polarized light and irradiates the sample. The differences in the reflectance rs of the s-polarization and the reflectance rP of the p-polarization of the sample change the polarization state of the reflected light. The intensity of the reflected light passed through the analyzer A is measured by the detector D.

[0055]FIG. 2 schematically shows directions of lights propagating in the sample of the ellisometry test. As exemplified in FIG. 2, the material sample includes a substrate and a thin film on the substrate. The light emitted from the light source passes through the polarizer P and the compensator C and then irridates the thin film from an interface between air and the thin film. Amplitude of the incident light is represented by E0, an angle between the incident light and the normal of the interface is represented by ϕ1, and the thickness of the thin film layer is represented by d.

[0056]A complex refractive index is defined as Ñi=ni+iki, where i=1, 2, 3 respectively represent air, the thin film, and the substrate. Ñ1 is the complex refractive index of air, Ñ2 is the complex refractive index of the thin film, Ñ3 is the complex refractive index of the substrate, and ni and κi are the real and imaginary parts of the complex refractive index respectively. The reflectance of the film for p-polarized light and s-polarized light are rp and rs respectively. ϕ1 is the incident angle of light, ϕ2 is a refraction angle of light at the interface between the air and the thin film, and ϕ3 is a refraction angle of light at the interface between the thin film and the substrate.

[0057]r12p and r12s are reflection coefficients of the p-polarized light and the s-polarized light respectively at the interface between air and the thin film, and satisfy the following equations:

[0058]r1⁢2p=N~2⁢cos⁢⁢ϕ1-N~1⁢cos⁢⁢ϕ2N~2⁢cos⁢⁢ϕ1+N~1⁢cos⁢⁢ϕ2,r1⁢2s=N~1⁢cos⁢⁢ϕ1-N~2⁢cos⁢⁢ϕ2N~1⁢cos⁢⁢ϕ1+N~2⁢cos⁢⁢ϕ2.

[0059]r23p and r23s are reflection coefficients of the p-polarized light and the s-polarized light respectively at the interface between the thin film and the substrate, and satisfy the following equations:

[0060]r2⁢3p=N~3⁢cos⁢⁢ϕ2-N~2⁢cos⁢⁢ϕ3N~3⁢cos⁢⁢ϕ2+N~⁢cos⁢⁢ϕ3,r2⁢3s=N~2⁢cos⁢⁢ϕ2-N~3⁢cos⁢⁢ϕ3N~1⁢cos⁢⁢ϕ1+N~3⁢cos⁢⁢ϕ3.

[0061]The total reflection is

[0062]r=r1⁢2+r2⁢3⁢e2⁢i⁢β1+r1⁢2⁢r2⁢3⁢e2⁢i⁢β

[0063]in which β=(n2+iκ2)kd cos ϕ2. Ellipsometric parameters (Δ, Ψ) satisfy the following equation:

[0064]rprs=tan⁢Ψ⁢ej⁢Δ.

[0065]Theoretical operations of relationships between the optical constants of a material and the ellipsometric parameters are shown above.

[0066]A measurement method by the ellipsometer and a conventional fitting method are described below.

[0067]I. Specific measurement method by the ellipsometer (Here, the use of the polarizer, analyzer, compensator, etc., is only an embodiment to measure the ellipsometric parameters (Δ, Ψ). There are many other embodiments; for example, the ellipsometric parameters (Δ, Ψ) can also be obtained by a photo-elastic modulator based configuration):

[0068](1) Acquiring an Angle θP of the Polarizer and an Angle θA of the Analyzer Respectively:

[0069]The compensator C and the analyzer A are fixed while the polarizer P is rotated, to find out the angle θP of the polarizer when a minimum light intensity is recorded by the detector; and the polarizer P and the compensator C are fixed while the analyzer A is rotated, to find out the angle θA of the analyzer when a minimum light intensity is recorded by the detector.

[0070](2) Parameter Transforming:

[0071]The ellipsometric parameters (Δ, Ψ), the angle θP of the polarizer and the angle θA of the analyzer, and a fixed angle θC of the compensator satisfy a function relationship: (Δ, Ψ)=f(θP, θA, θC). The values of the angle θP of the polarizer and the angle θA of the analyzer are directly read through the control program of the ellipsometer, and the value of the ellipsometer test parameter vector (Δ, Ψ) is obtained through the function relationship (Δ, Ψ)=f(θP, θA, θC).

[0072]II. Data fitting is performed on the ellipsometric parameters measured by the ellipsometer, to obtain the optical constants of the sample:

[0073](1) Modeling:

[0074]In the fitting procedure of the ellipsometry, a model of air-thin film-substrate is set up. The model of air-thin film-substrate is consistent with the configuration of the sample.

[0075](2) Determining Basic Parameters Before the Fitting:

[0076]The materials of the thin film and the substrate are specified respectively, and an initial value of the thickness and a range of the thickness of the thin film are set. The material of the substrate can be set as a known material (such as SiO2), and a refractive index file thereof can be imported directly from a material database. The refractive index of the material of the thin film is unknown, and the refractive index fitting is required, so the refractive index curve is set as a superposition of multiple optical dispersion models.

[0077](3) Manually Setting the Profile and the Relevant Parameters Based on the Model of Air-Thin Film-Substrate and the Basic Parameters Before the Fitting:

[0078]Appropriate optical dispersion models are adopted for the material of the thin film, and the relevant parameters of each optical dispersion model is manually set, such that the results of (Δ, Ψ) obtained by fitting under the model are as close as possible to the results of the measured ellipsometric parameters (Δ, Ψ).

[0079](4) Automatic Accurate Fitting:

[0080]The step (3) can be performed multiple times. In each time, after the step (3) is performed, a step of regression fitting can be performed. The step (4) is set because the manually set parameters in the step (3) cannot further make the results of (Δ, Ψ) obtained by the fitting approximate the measurement result. In the present step, the use of the automatic accurate fitting function of the program, can be implemented by the accurate fitting or called approximate fitting through an algorithm by the program.

[0081]The present disclosure provides a method and an apparatus for determining an optical constant of a material, and a method and an apparatus for extending a material database. The method for determining an optical constant of a material can be used in combination with a conventional ellipsometer. A novel ellipsometer may also be used to implement the method for determining an optical constant of a material in the present application.

[0082]Referring to FIG. 3, the present disclosure provides a method for determining an optical constant of a material, which includes the following steps.

[0083]S10: ellipsometric parameters is acquired. The ellipsometric parameters in this step can include ellipsometric parameters Δ and Ψ.

[0084]S20: the ellipsometric parameters are input to a machine learning model to obtain the optical constant of the material according to the ellipsometric parameters. The machine learning model includes a mapping relationship between the ellipsometric parameters and the optical constants of the material.

[0085]The optical constant of the material can include refractive index, an extinction coefficient, or a dielectric constant of the material. A specific mapping relationship between the ellipsometric parameters and the optical constant of the material can be provided based on corresponding theories of the basic physics. For example, the mapping relationship between the ellipsometric parameters and the optical constant corresponding to the ellipsometric parameters of the material can be determined according to the interference theory of single-layer film and the theories of optical reflection and transmission of the thin film.

[0086]In the present embodiment, a machine learning model is applied to a process of determining an optical constant of a material. Fast and accurate determination of optical constants of a single profile material, a multi-profile complex material, and an anisotropic material are implemented through the machine learning model. The optical constant of the material is determined through the machine learning model. The parameters of the regression fitting can be adjusted without relying too much on the experiences of the experimenters anymore, and a constant high-speed automatic correction is performed to implement the curve fitting of the measurement data, to finally obtain the optical constant of the material. For details, please refer to FIGS. 14-16, in which confirmatory tests are taken upon new materials.

[0087]FIG. 14 shows a comparison diagram between the conventional fitting method and the method for determining an optical constant of a material provided by the present disclosure (called hereinafter the present method), in which the material to be tested is silicon. It can be seen from FIG. 14 that the curves of n and κ and the curves of Δ and Ψ of the semiconductor silicon obtained by the present method substantially coincides with these obtained by the conventional fitting method. In FIG. 14, near the short-wave band from 500 nm to 600 nm, the results obtained by the present method are more consistent with the measured data than these obtained by the conventional fitting method. From FIG. 14, a significant difference can be seen between the curve of κ obtained by the present method and that by the conventional fitting method in the wavelength range of 500 nm to 600 nm. In the curves of Δ and Ψ, the data of Δ of the semiconductor silicon obtained by the present method is more consistent with the actual experimental measurement value (see the sub-figure in FIG. 14). Therefore, the data obtained by the present method is more accurate than that obtained by the conventional fitting method. FIG. 15 shows a comparison diagram between the conventional fitting method and the present method, in which the material to be tested is titanium dioxide (TiO2). FIG. 16 shows a comparison diagram between the conventional fitting method and the present method, in which the material to be tested is gold. The comparison results in FIGS. 14, 15 and 16 can prove that the present method can accurately calculate the refractive indices of the materials. In addition to the embodiments of silicon, titanium dioxide, and gold, the present method can also be applied to other materials to implement an automatic fitting of the ellipsometric measurement results by utilizing the machine learning model. Furthermore, the optical constants of the material obtained by the present method has the same or even higher accuracy than that obtained through the conventional fitting method.

[0088]Referring to FIG. 4, in an embodiment, a modeling method for the machine learning model specifically includes the following steps.

[0089]S210: each neural network layer of an initial model is constructed based on the mapping relationship between the ellipsometric parameters and the optical constant of the material corresponding to the ellipsometric parameters. The initial model can include an input layer, an intermediate layer, and an output layer. The intermediate layer can further include one or more of a convolutional layer, a pooling layer or a fully connected layer, or any can be used as intermediate layer.

[0090]S220: sample data is constructed; the sample data can include ellipsometric parameters and the optical constant of the material.

[0091]In the present step, the ellipsometric parameters included in the sample data is a parameter input into the machine learning model. In an embodiment, the input parameter can include ellipsometric parameters Δ and ψ, an incident angle of light Φ, a wavelength λ of the light incident on the thin film, a complex refractive index of the substrate Ñ3, a complex refractive index of the thin film Ñ2, and so on. In the present step, the optical constant of the material included in the sample data is an output of the machine learning model. In an embodiment, the output can include a refractive index (Ñ) or a dielectric constant (ε) of the thin film. For example, the output can be a vector consisting of (n, κ, d), where n and k are refractive index and extinction coefficient respectively, or a vector consisting of (εr, εi, d), where εr and εi are real and imaginary dielectric constants respectively. The process of constructing the sample data includes a process of data processing. For example, it is required to set the sample data corresponding to the same or similar wavelength range. The type of data included in the sample data should be comprehensive, and the quantity of data included in the sample data should be large enough. For example, the sample data includes ellipsometric parameters and optical constants of the material obtained from different optical dispersion models. The optical dispersion models includes (but not limited to) at least one of Flossie model, Gaussian model, Cauchy model, Lorentz model, Drude model, Sellmeier model, or Fano model.

[0092]S230: each neural network layer of the initial model is trained on the sample data to obtain the machine learning model. In this step, the process of training each neural network layer of the initial model can be implemented by executing computer programs.

[0093]In the present embodiment, a method for establishing a machine learning model is provided. The method specifically includes steps of constructing each neural network layer of an initial model, constructing sample data, and training each neural network layer of the initial model, to enable the initial model to carry basic functions that the machine learning model needs to implement.

[0094]Referring to FIG. 5, in an embodiment, the step S230 of training each neural network layer of the initial model by the sample data to obtain the machine learning model may include the following steps.

[0095]S231: the ellipsometric parameters are input to the initial model to obtain a current output. In this step, the current output obtained is a data quantity in the process of training the initial model.

[0096]S232: a difference value between the current output and the corresponding optical constant of the material is calculated; if the difference value is greater than or equal to a preset value, network parameters of each neural network layer of the initial model are adjusted according to the difference value. In this step, the difference value between the current output and the optical constant of the material can reflect an extent to which the initial model is to be adjusted.

[0097]S233: the difference value is calculated by performing an iterative operation until the difference value is less than the preset value, then it is determined that each neural network layer of the initial machine learning model reaches convergence at which the machine learning model is obtained. In this step, the preset value can be a self-set value. For example, the preset value can be 10−6 or 10−8.

[0098]In the present embodiment, each neural network layer of the initial model is trained by the sample data to obtain the machine learning model. It should be appreciated that multiple different types of the machine learning models can be established according to the requirements of test accuracy. Multiple types of machine learning models can be trained on the same sample data, or each type of the machine learning model can be trained on different sample data. In this way, the trained machine learning model has a higher accuracy, and can satisfy the use of different types of data. In the present embodiment, the method provided can be applied to obtain one or more types of machine learning models according to the initial model, and the specific model can be selected according to the actual requirement.

[0099]Referring to FIG. 6, in an embodiment, the step S230 of training each neural network layer of the initial model on the sample data to obtain the machine learning model can include the following steps.

[0100]S234: the sample data is divided into a training set, a validation set and a test set. In this step, the training set, the validation set and the test set are three data sets that do not overlap each other. For example, 10000 entries are provided, a training set includes 6000 entries, a validation set includes 2000 entries, and a test set includes 2000 entries; each entry includes ellipsometric parameters Δ and Ψ, a thickness d of the thin film, a wavelength λ, an incident angle, the refractive index n and the extinction coefficient k. The data sets in the sample data described herein can also be allocated according to other proportions. The training set and the validation set are put into the initial model simultaneously, which can implement the effect of verification while training.

[0101]S235: the training set and the validation set are respectively input into the initial model to optimize the initial model, and optimization models are obtained and recorded constantly. Meanwhile an error curve of the training set and an error curve of the validation set are generated. In this step, the optimization models can be recorded constantly by a computer, and meanwhile an error curve can be generated by the computer.

[0102]Referring to FIG. 13, for a point on the error curve of the training set, the abscissa of the point represents the time or number of iterative operations for obtaining the point, and the ordinate of the point represents an error between an operation result obtained from the initial model under the corresponding abscissa and the actual optical constant of the material. For a point on the error curve of the validation set, the abscissa of the point represents the time or number of iterative operations for obtaining the point, and the ordinate of the point represents an error between an operation result obtained from the initial model under the corresponding abscissa and the actual optical constant of the material.

[0103]S236: an optimization model in S235 at a moment when an error value of the error curve of the validation set no longer decreases is taken as a first optimization model. In this step, there may be multiple specific methods for acquiring the first optimization model, such as a marker search method.

[0104]S237: input parameters in the test set are input into the first optimization model to obtain a first output parameter which is obtained by calculation via the first optimization model; an error between the first output parameter and a reference output parameter in the test set is determined; if the error is within a preset range, the first optimization model is determined as the machine learning model.

[0105]In the present embodiment, it should be noted that the machine learning system is trained by the training set and the performance of the machine learning system is monitored by the validation set. The training is stopped when the performance of the machine learning system on the validation set no longer improves. The machine learning model with the best performance on the validation set is selected as a training model. After repeating this process several times, a model with the best performance on the validation set is selected as a final machine learning system and the performance thereof is tested by the test set.

[0106]Referring to FIG. 7, in an embodiment, the step S230 of training each neural network layer of the initial model on the sample data to obtain the machine learning model may include the following steps.

[0107]S234: the sample data is divided into a training set, a validation set and a test set.

[0108]S235: the training set and the validation set are respectively input into the initial model, the initial model is optimized and optimization models are obtained and recorded constantly, meanwhile an error curve of the training set and an error curve of the validation set are generated.

[0109]S238: M optimization models performing best obta