Multi-layer thin film spectrum inversion model training method, inversion method, device and medium

Abstract

The application provides a multilayer thin film spectrum inversion model training method, an inversion method, an electronic device and a medium. The inversion model training method comprises: obtaining spectrum data for training, the spectrum data comprising ellipsometric parameters of a thin film structure. The ellipsometric parameters are encoded, and the encoded information is input into a spectrum inversion model to obtain a prediction result of the spectrum inversion model. A loss function is determined based on the prediction result of the spectrum inversion model; and the spectrum inversion model is trained based on the loss function, the prediction result comprising an ellipsometric parameter prediction result of the thin film structure and a structure parameter prediction result of each layer of the thin film structure, thereby effectively ensuring the physical interpretability of the prediction result of the spectrum inversion model while reducing the model parameter dimension, realizing end-to-end training of the spectrum inversion model based on the spectrum data for training, and ensuring that the spectrum inversion model obtained by training can effectively invert the multilayer thin film structure.

Description

Technical Field

[0001] This application relates to the field of semiconductor technology, specifically to a training method, inversion method, electronic device, and medium for a multilayer thin film spectral inversion model. Background Technology

[0002] An ellipsometer is an optical instrument that characterizes the properties of thin films by measuring the amplitude ratio and phase difference (parameters Ψ and Δ) of polarized light reflected from a sample surface. However, it does not directly provide optical constants; these must be solved through inverse problems. Inverse problems refer to inferring system parameters from observational data. In thin film spectral inversion, it is necessary to infer the material and structural properties of the thin film, such as refractive index n, extinction coefficient k, and film thickness d, from (Ψ, Δ).

[0003] Traditional methods require manual intervention and selection of dispersion models (such as Cauchy and Sellmeier), and rely on iterative fitting, which is time-consuming and prone to getting trapped in local optima. Existing data-driven spectral inversion methods, such as parametric models (such as Tauc-Lorentz), require small sample sizes, but the model form is fixed and difficult to adapt to unknown materials. While the more flexible high-dimensional KKB-spline (Kramers–Kronig) model can universally characterize various dielectric functions, in the case of multilayer thin films, the parameter correlations are complex, the inversion is unstable, and it is not conducive to deployment at the equipment end or production line environment.

[0004] Accordingly, there is a need in the field for a new multilayer thin film spectral inversion scheme to solve the above problems. Summary of the Invention

[0005] In order to overcome the above-mentioned deficiencies, this application is made to solve, or at least partially solve, the technical problem of how to achieve effective inversion of multilayer thin film structures.

[0006] In a first aspect, a method for training a multilayer thin film spectral inversion model is provided, the method comprising: Acquire spectral data for model training; wherein the spectral data includes ellipticity parameters of the thin film structure; and encode the ellipticity parameters to obtain encoded information; The encoded information is input into a spectral inversion model to obtain the prediction results of the spectral inversion model. The spectral inversion model includes a backbone network, a structural parameter prediction branch, and an elliptic parameter prediction branch. The prediction results include the elliptic parameter prediction results of the thin film structure and the structural parameter prediction results of each thin film layer. The backbone network is used to extract spectral features from the encoded information and obtain a global feature vector. The elliptic parameter prediction branch and the structural parameter prediction branch predict the elliptic parameter prediction results and the structural parameter prediction results, respectively, based on the global feature vector. The loss function is determined based on the prediction results; and the spectral inversion model is trained based on the loss function.

[0007] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion model, the elliptic parameter prediction branch includes a physical parameter prediction sub-branch, a physical mapping layer sub-branch, and an optical forward calculation layer; The prediction result of the elliptic deviation parameter based on the global feature vector includes: The physical parameter prediction sub-branch predicts the physical parameters of the thin film structure based on the global feature vector; The physical mapping layer sub-branch maps the physical parameters to the dielectric function of each thin film structure; wherein the dielectric function satisfies the Kramer-Kronig constraint; The optical forward computing layer calculates the elliptic parameter prediction result based on the dielectric function of each thin film structure and the predicted structural parameters.

[0008] In one technical solution of the above-mentioned training method for the spectral inversion model of multilayer thin films, the structural parameter prediction branch is used to predict the structural and interface features of each thin film structure based on the global feature vector, as the structural parameter prediction result; wherein, the structural parameter prediction result includes the thickness of each thin film structure.

[0009] In one technical solution of the above-mentioned multilayer thin film spectral inversion model training method, the structural parameter prediction results also include the roughness and dielectric function gradient of each thin film structure.

[0010] In one technical solution of the above-mentioned multilayer thin film spectral inversion model training method, the physical parameters include the oscillator parameters of each thin film structure; The physical mapping layer sub-branch maps the physical parameters to the dielectric function of each thin film structure, including: The physical mapping layer sub-branch determines the absorption spectrum of each thin film structure based on the oscillator strength, resonance energy, damping coefficient, and band gap energy in the oscillator parameters; The absorption spectrum was subjected to Hilbert transform to obtain the dispersive spectrum of each thin film structure; Based on the absorption spectrum and the dispersion spectrum, the dielectric function of each thin film structure is obtained using the oscillator model.

[0011] In one technical solution of the above-mentioned training method for the spectral inversion model of multilayer thin films, the physical parameters include latent variables of each thin film structure; The physical mapping layer sub-branch maps the physical parameters to the dielectric function of each thin film structure, including: The physical mapping layer sub-branch decodes the latent variables to obtain the dielectric function of each thin film structure.

[0012] In one technical solution of the above-mentioned training method for the spectral inversion model of multilayer thin films, the physical parameters include the oscillator parameters and latent variables of each layer of thin film structure; The physical mapping layer sub-branch maps the physical parameters to the dielectric function of each thin film structure, including: The physical mapping layer sub-branch obtains the dispersion spectrum and absorption spectrum corresponding to the oscillator parameters and the dispersion spectrum and absorption spectrum corresponding to the latent variables, respectively, based on the oscillator parameters and latent variables. Based on the dispersion spectrum corresponding to the oscillator parameters and the dispersion spectrum corresponding to the latent variable, a mixed dispersion spectrum is obtained; Based on the absorption spectrum corresponding to the oscillator parameters and the absorption spectrum corresponding to the latent variable, a mixed absorption spectrum is obtained; The dielectric function of each thin film structure is obtained based on the mixed dispersion spectrum and the mixed absorption spectrum.

[0013] In one technical solution of the above-mentioned training method for the spectral inversion model of multilayer thin films, the loss function includes fitting error loss; the fitting error loss is determined based on the difference between the predicted elliptic parameter and the corresponding elliptic parameter in the spectral data.

[0014] In one technical solution of the above-mentioned training method for the spectral inversion model of multilayer thin films, the loss function further includes a physical constraint loss, which is determined based on the dielectric function of each thin film structure obtained from the prediction result; the physical constraint loss is used to constrain the dielectric function to satisfy the Kramer-Kronig constraint.

[0015] In one technical solution of the above-mentioned training method for the multilayer thin film spectral inversion model, the loss function further includes a smoothness loss, which is used to constrain the continuity and smoothness of the dielectric function corresponding to the structural parameter prediction results of each thin film structure predicted by the structural parameter prediction branch in the wavelength dimension.

[0016] In one technical solution of the above-mentioned multilayer thin film spectral inversion model training method, the step of encoding the ellipticity parameters to obtain encoded information includes: The amplitude ratio angle and phase difference angle in the elliptic parameters are normalized, and the cosine code of the amplitude ratio angle cos2Ψ, the cosine component of the phase difference angle sin2Ψ×cosΔ, and the sine component of the phase difference angle sin2Ψ×sinΔ are calculated; where Ψ is the amplitude ratio angle and Δ is the phase difference angle. The cosine encoding, the cosine component, the sine component, and the transmittance in the elliptic parameter are used as the encoding information.

[0017] In one technical solution of the above-mentioned multilayer thin film spectral inversion model training method, the step of acquiring spectral data for training includes: Based on a preset sampling range, the material and structural parameters of the thin film structure are randomly sampled; Based on the physical model, the ellipticity parameters of the thin film structure are obtained from the material parameters and structural parameters obtained by random sampling, and used as the spectral data for training.

[0018] In one technical solution of the above-mentioned multilayer thin film spectral inversion model training method, the thin film structure includes at least one optical thin film.

[0019] In a second aspect, a method for spectral inversion of multilayer thin films is provided, the method comprising: The ellipticity parameters of the thin film structure to be tested are encoded to obtain the coded information to be tested; The coded information to be detected is input into the trained spectral inversion model, and inversion reasoning is performed based on the preset structural parameter range to obtain the material and structural properties of each thin film structure in the thin film structure to be detected. The spectral inversion model is obtained by training the multilayer thin film spectral inversion model based on any one of the above-mentioned multilayer thin film spectral inversion model training methods.

[0020] In one technical solution of the above-mentioned multilayer thin film spectral inversion method, the step of performing inversion inference based on a preset structural parameter range to obtain the material and structural properties of each layer of the thin film structure to be detected includes: A preset number of structural parameter sample values ​​are uniformly sampled within the range of the structural parameters; Based on the loss function value corresponding to each structural parameter sample value, determine the target structural parameter sample value with the minimum loss function value; Inversion reasoning is performed within a preset range of the target structural parameter sampling values ​​to obtain the material and structural properties of each thin film structure.

[0021] In one technical solution of the above-mentioned multilayer thin film spectral inversion method, the method further includes: Based on the range of structural parameters and the coding information to be detected, the inversion reasoning result is obtained; Calculate the loss function value corresponding to the inversion inference result; and update the model parameters of the structure parameter prediction branch and / or ellipticity parameter prediction branch of the spectral inversion model in reverse according to the loss function value.

[0022] In a third aspect, an electronic device is provided, comprising at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program, which, when executed by the at least one processor, implements the method described in any one of the above-described technical solutions for training a multilayer thin film spectral inversion model and / or the method described in any one of the above-described technical solutions for multilayer thin film spectral inversion methods.

[0023] In a fourth aspect, a computer-readable storage medium is provided, wherein a plurality of program codes are stored therein, the program codes being adapted to be loaded and run by a processor to perform the method described in any one of the above-described technical solutions of the multilayer thin film spectral inversion model training method and / or the method described in any one of the above-described technical solutions of the multilayer thin film spectral inversion method.

[0024] The above-described technical solutions of this application have at least one or more of the following beneficial effects: In implementing the multilayer thin film spectral inversion model training method provided in this application, this application acquires spectral data for training, including the ellipticity parameters of the thin film structure. The ellipticity parameters are encoded, and the encoded information is input into the spectral inversion model to obtain the prediction results. The spectral inversion model includes a backbone network, a structural parameter prediction branch, and an ellipticity parameter prediction branch. The backbone network is used to extract spectral features from the encoded information and obtain a global feature vector; the ellipticity parameter prediction branch and the structural parameter prediction branch predict the ellipticity parameter prediction results and the structural parameter prediction results, respectively, based on the global feature vector; a loss function is determined based on the prediction results of the spectral inversion model; and the spectral inversion model is trained based on the loss function. Through the above configuration, the encoding information used for model training in this application is obtained by encoding the ellipticity parameters of the thin film structure. Furthermore, since the spectral inversion model consists of a backbone network, an ellipticity parameter prediction branch, and a structural parameter prediction branch, it enables end-to-end training of the spectral inversion model based on the spectral data used for training. The prediction results include the ellipticity parameter prediction results of the thin film structure and the structural parameter prediction results of each layer of the thin film structure. This effectively ensures the physical interpretability of the prediction results of the spectral inversion model while reducing the dimensionality of the model parameters. Consequently, it ensures that the trained spectral inversion model can effectively invert multilayer thin film structures. Attached Figure Description

[0025] The disclosure of this application will become more readily understood with reference to the accompanying drawings. It will be readily understood by those skilled in the art that these drawings are for illustrative purposes only and are not intended to limit the scope of protection of this application. Wherein: Figure 1 This is a schematic flowchart of the main steps of a multilayer thin film spectral inversion model training method according to an embodiment of this application; Figure 2 This is a schematic diagram of the main components of a spectral inversion model according to one embodiment of the present application; Figure 3 This is a schematic flowchart of the main steps of a multilayer thin film spectral inversion method according to an embodiment of this application. Detailed Implementation

[0026] Some embodiments of this application are described below with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of this application and are not intended to limit the scope of protection of this application.

[0027] In the description of this application, "module" and "processor" can include hardware, software, or a combination of both. A module can include hardware circuitry, various suitable sensors, communication ports, and memory, and may also include software components, such as program code, or a combination of software and hardware. The term "A and / or B" means all possible combinations of A and B, such as only A, only B, or A and B. The terms "at least one A or B" or "at least one of A and B" have a similar meaning to "A and / or B" and can include only A, only B, or A and B. The singular forms of the terms "a" and "this" can also include plural forms.

[0028] Here we will first explain some of the terms used in this application.

[0029] Ellipticity parameters are key physical quantities used in elliptic polarization spectroscopy to describe the change in polarization state of polarized light after reflection (or transmission) on a sample surface. They are typically denoted as Ψ (Psi, amplitude ratio angle) and Δ (Delta, phase difference angle).

[0030] Ψ(Psi): Defined as the arctangent of the ratio of the amplitude of the p component to the s component in the reflected light. The change of Ψ mainly reflects the change of the intensity ratio of the reflected light, and its range is [0°, 90°].

[0031] Δ(Delta): Represents the phase difference between the p component and the s component in the reflected light. The change of Δ is directly related to the phase delay experienced by the light wave when it propagates at the interface, and its range is [0°, 360°].

[0032] By accurately measuring Ψ and Δ, key optical constants such as the thickness, refractive index, and dielectric constant of optical thin films can be determined.

[0033] Kramer–Kronig relations (KK relations for short) are a mathematical theory describing the dispersion relationship between the real and imaginary parts of a linear response function.

[0034] The Hilbert transform is a linear operator widely used in mathematics and signal processing that transforms one real-valued function into another function with the same domain. See appendix Figure 1 , Figure 1 This is a schematic flowchart illustrating the main steps of a multilayer thin film spectral inversion model training method according to an embodiment of this application. Figure 1 As shown, the multilayer thin film spectral inversion model training method in this application embodiment mainly includes the following steps S101 to S103.

[0035] Step S101: Obtain spectral data for model training; wherein the spectral data includes the ellipticity parameters of the thin film structure; and encode the ellipticity parameters to obtain encoded information.

[0036] In this embodiment, spectral data for model training can be obtained, and the spectral data may include the ellipticity parameters of the thin film structure.

[0037] In one embodiment, the thin film structure may include at least one optical thin film. Optical thin films, as materials with diverse optical properties, are indispensable components in modern integrated circuit manufacturing processes. Optical measurement is a crucial means of monitoring the formation of optical thin film layers. Thin film structures typically include one or more optical thin films stacked on a substrate. That is, the thin film structure used for model training may include one optical thin film or multiple optical thin films. Spectral data may include multiple sample data points. For a single optical thin film, these multiple sample data points may include multiple ellipsometric parameters of that optical thin film under different measurement conditions. In the case where the thin film structure includes multiple optical thin films, based on parameters such as the stacking method of the multiple optical thin films and the thickness of each optical thin film, the multiple sample data points may include ellipsometric parameters obtained by performing multiple measurements on each semiconductor structure composed of multiple optical thin films under different measurement conditions.

[0038] As previously described, in one embodiment, the spectral data can be ellipsometric parameters obtained from actual ellipsometric measurements for each thin film structure. Correspondingly, to increase the number of training samples and improve the richness of the spectral data used for training, the spectral data can also be synthetic spectral data.

[0039] Specifically, material and structural parameters of the thin film structure can be randomly sampled based on a preset sampling range. Based on a physical model, the ellipticity parameters of the thin film structure are obtained from the randomly sampled material and structural parameters, thereby generating synthetic spectral data, which can then be used as spectral data for training.

[0040] That is, by randomly sampling material and structural parameters and combining them with a physical model of the thin film structure, synthetic spectral data is obtained through forward optical calculations. To obtain more reasonable synthetic spectral data, the principles of random sampling can include: The structural parameters of each thin film structure conform to the principle of physical rationality: Thickness range [5,200] nm: This allows the thickness range to cover typical optical thin film thicknesses; Surface roughness [0,5] nm: This ensures that the surface roughness is within the typical range for optical thin film surfaces; Gradient factor [0, 0.4]: ensures that the number of sub-layers after discretization is between 20 and 40.

[0041] In this scheme, the surface roughness of each thin film structure affects interface reflection, and the gradient factor is used to describe the continuous change of dielectric function within each thin film structure layer, and is used to model gradient layers (such as gradual changes in alloy composition, changes in doping concentration, etc.). In this scheme, the continuous change is approximated by discretizing the structure into 20-40 sub-layers.

[0042] Uniform sampling principle: The sampling process uses uniform distribution sampling to ensure full coverage of the parameter space and avoid bias towards certain specific values.

[0043] Principle of randomness: The number of layers in the thin film structure is random: for example, it can be set to 2-5 layers; When using an oscillator model to represent the physical model of each thin film structure: The oscillator type of the thin film structure is random: for example, it can be a Lorentz oscillator model / a Tauc-Lorentz oscillator model / a Gaussian oscillator model; The number of oscillators in a thin-film structure is random: each layer can have 1-3 oscillators.

[0044] The synthesized spectral data in this embodiment of the invention can be generated based on an oscillator model, possessing high physical plausibility and enabling more realistic simulation of multi-material, multi-spectral environments under measurement conditions. The spectral data obtained through the above method for model training may include the thin film structure's Ψ, Δ, T (transmittance), the angle of incidence (AOI) of the input spectrum, wavelength (λ), substrate / polarization information, etc.

[0045] In one implementation, the ellipticity parameters in the spectral data can be encoded to obtain encoded information, which can then be used as input data for the spectral inversion model.

[0046] In this embodiment, the original angles of the amplitude ratio angle and phase difference angle are represented as Ψ∈[0°,90°] and Δ∈[0°,360°]. Since these have periodic boundaries, they need to be mapped to a continuous space through encoding to avoid boundary discontinuities. Fixing the range of the encoded ellipticity parameters to around [-1,1] also facilitates gradient propagation.

[0047] Specifically, parameters such as amplitude ratio and phase difference angle in the elliptic parameters can be normalized to ensure the numerical stability of the loss function during subsequent model training, thereby accelerating the convergence of the loss function and achieving high-precision model fitting.

[0048] For example, the cosine code of the amplitude ratio angle cos(2Ψ), the cosine component of the phase difference angle sin(2Ψ)×cos(Δ), and the sine component of the phase difference angle sin(2Ψ)×sin(Δ) can be calculated; the cosine code, cosine component, sine component, and transmittance in the elliptic parameter are used as coding information and input into the spectral inversion model.

[0049] Among them, the cosine encoding of the amplitude ratio Ψ, cos(2Ψ), can map the amplitude ratio Ψ to the interval [-1,1], avoiding the singularity problem caused by the periodicity of the angle; The cosine component of the phase difference angle, sin(2Ψ)·cos(Δ), can couple the cosine component of the phase difference angle Δ with the amplitude ratio, linearize the phase information, avoid gradient discontinuity and training imbalance, and improve network stability. The sinusoidal component of the phase difference angle, sin(2Ψ)·sin(Δ), can couple the sinusoidal component of the phase difference angle Δ with the amplitude ratio, thus preserving the phase information completely, avoiding gradient discontinuity and training imbalance, and improving network stability.

[0050] Step S102: Input the encoded information into the spectral inversion model to obtain the prediction results of the spectral inversion model; the spectral inversion model includes a backbone network, a structural parameter prediction branch, and an elliptic parameter prediction branch; the prediction results include the elliptic parameter prediction results of the thin film structure and the structural parameter prediction results of each thin film layer; the backbone network is used to extract spectral features from the encoded information and obtain a global feature vector; the elliptic parameter prediction branch and the structural parameter prediction branch predict the elliptic parameter prediction results and the structural parameter prediction results respectively based on the global feature vector.

[0051] In this embodiment, the encoded information of the thin film structure can be input into the spectral inversion model, and the elliptic parameter and structural parameter can be predicted by the structural parameter prediction branch and the elliptic parameter prediction branch, respectively, thereby obtaining the elliptic parameter prediction result and the structural parameter prediction result.

[0052] In one implementation, the ellipticity parameter prediction results may include Ψ, Δ, and T.

[0053] In one implementation, the structural parameter prediction branch can be used to predict the structural and interface features of each thin film layer based on the global feature vector, as the structural parameter prediction result. The structural parameter prediction result may include the thickness t_i of each thin film layer; in another implementation, the structural parameter prediction result may further include the roughness r_i and / or the gradient h_i of the dielectric function of each thin film layer. Here, i represents the i-th thin film layer.

[0054] In one implementation, the backbone network is used to extract spectral features from the encoded information to obtain a global feature vector.

[0055] In one implementation, the input data of the spectral inversion model may also include an experimental condition vector, which is used to characterize wavelength, incident angle, substrate / polarization information, etc., during the spectral data measurement process.

[0056] In one implementation, the backbone network can be a one-dimensional residual convolutional backbone network (ResNet-1D).

[0057] In this embodiment, the input of the encoded information of the thin film structure is a high-dimensional input. This is because: In the art, spectral data is typically acquired through random sampling, with uniform sampling along the wavelength dimension; for example, there could be 512 sampling points along the wavelength dimension. Simultaneously, as mentioned above, the encoded information can include four channel dimensions (cos(2Ψ), sin(2Ψ)·cos(Δ), sin(2Ψ)·sin(Δ), T). Therefore, the total input dimension of the spectral inversion model is 512 × 4 = 2048 dimensions. Using a fully connected layer at this point would lead to an explosion of parameters; furthermore, fully connected layers cannot effectively utilize the local correlations between spectral data.

[0058] Meanwhile, high-dimensional data can easily lead to overfitting, and the training process requires a large amount of training data. Therefore, the backbone network can use ResNet-1D to reduce the number of parameters through local receptive fields and parameter sharing. At the same time, it can also utilize the translation invariance of convolution to improve the generalization ability of the spectral inversion model. Applying ResNet-1D as the backbone network can extract robust spectral features from the encoded information of multi-channel one-dimensional spectral data and output global feature vectors to the downstream branches of the spectral inversion model.

[0059] In one implementation, the one-dimensional residual convolutional backbone network (ResNet-1D) may include Conv1D (one-dimensional convolutional network), residual blocks, and global pooling layers. The residual blocks can be 4-6 blocks. For example, the structure of the one-dimensional residual convolutional backbone network (ResNet-1D) is: Conv1D (4D → 64D) → ResidualBlock (residual block) × 5 → GlobalAvgPool (global pooling layer) → FC (256D → 128D), thereby obtaining a 128-dimensional global feature vector, which is then input to the elliptic parameter prediction branch and the structural parameter prediction branch.

[0060] In one implementation, the elliptic parameter prediction branch can directly predict the elliptic parameters based on the global feature vector. That is, it bypasses the complex physical model calculation and uses a direct prediction method to predict the elliptic parameters based on the global feature vector.

[0061] In a specific example, the global feature vector is a 128-dimensional vector. The structure of the elliptic parameter prediction branch can be: FC (fully connected layer) (128-dimensional → 512-dimensional) → ReLU (activation function) → Dropout (dropout unit) → FC (fully connected layer) (512-dimensional → 256-dimensional) → ReLU (activation function) → FC (fully connected layer) (256-dimensional → n_wavelengths (number of wavelengths) × 3). The output mapping of the elliptic parameter prediction branch is: Psi ∈ [0, 90°], Delta ∈ [0, 360°], T ∈ [0, 1].

[0062] The elliptic parameter prediction branch, which directly predicts elliptic parameters based on global feature vectors, can bypass complex physical model calculations, simplifying the training process of the spectral inversion model and effectively improving the efficiency and convergence speed of the training process. At the same time, it can achieve high-precision prediction of elliptic parameters through a large amount of training data and end-to-end optimization process.

[0063] In one implementation, the ellipticity prediction branch may include a physical parameter prediction sub-branch, a physical mapping layer sub-branch, and an optical forward computation layer. The physical parameter prediction sub-branch predicts the physical parameters of each thin film structure based on the global feature vector. The physical mapping layer sub-branch maps the physical parameters to the dielectric function of each thin film structure; wherein the dielectric function satisfies the Cramer-Kronig constraint. The optical forward computation layer calculates the ellipticity prediction result based on the dielectric function of each thin film structure and the predicted structural parameters.

[0064] In this embodiment, the physical parameter prediction sub-branch can predict low-dimensional physical parameters of the thin film structure based on the global feature vector. For example, the physical model of each thin film structure can be represented by multiple oscillator models, and each oscillator model can include only 4-dimensional oscillator parameters. Similarly, low-dimensional latent variables can be used to characterize the physical parameters of the thin film structure, so the dimension of the physical parameters can be 8-16 dimensions. The physical mapping layer sub-branch can map the low-dimensional physical parameters to the dielectric function of each thin film structure.

[0065] The physical parameter prediction sub-branch enables the prediction of low-dimensional physical parameters for each thin film structure, significantly reducing the number of parameters. While ensuring the generalization ability of the spectral inversion model across materials, it also avoids the need for large amounts of training data required by inversion methods such as high-dimensional B-spline. The optical forward computation layer can combine the dielectric function and structural parameter prediction results of each thin film structure to obtain the ellipticity parameter prediction results.

[0066] In one embodiment, the physical parameters may include the oscillator parameters of each thin film structure. The physical mapping sub-branch can determine the absorption spectrum of each thin film structure based on the oscillator strength, resonant energy, damping coefficient, and bandgap energy in the oscillator parameters; perform a Hilbert transform on the absorption spectrum to obtain the dispersion spectrum of each thin film structure; and obtain the dielectric function of each thin film structure based on the absorption spectrum and dispersion spectrum, relying on the oscillator model.

[0067] In this embodiment, the physical parameter prediction sub-branch can predict the oscillator parameters of the thin film structure based on the global eigenvectors. The oscillator model is a physical model used to fit the complex permittivity of the thin film material. It uses a set of constrained peaks to fit the material's response in different wavelength bands. It is also the basis for understanding the interaction between light and matter. The superposition of multiple oscillator models can describe the entire optical response spectrum of the material. The oscillator parameters can include oscillator strength, resonance energy, damping coefficient, and bandgap energy.

[0068] The physical meaning and function of the oscillator parameters are explained below: 1. A_k (Oscillator Strength, Amplitude) Physical meaning: The strength of the k-th oscillator determines the peak value of the imaginary part ε2 of the dielectric function; Value range: [0.1, 5.0]; Function: To control the absorption intensity of a material for photons of a specific energy.

[0069] 2. E0_k (Resonance Energy) Physical meaning: The resonant energy of the k-th oscillator corresponds to the peak position of the imaginary part ε2 of the dielectric function; Unit: electron volt (eV); Value range: [1.0, 5.0] eV; Function: Determines at which wavelength (energy) a material has strong absorption.

[0070] 3. Γ_k (damping coefficient) Physical meaning: The damping (broadening) coefficient of the oscillator determines the width of the absorption peak; Unit: electron volt (eV); Value range: [0.05, 1.0] eV; Effect: The larger the Γ value, the wider the absorption peak, and the stronger the carrier scattering in the corresponding material.

[0071] 4. Eg_k (Band Gap Energy) Physical meaning: Band gap energy of the material (only for the Tauc-Lorentz model); Unit: electron volt (eV); Value range: [0.0, 3.0] eV; Function: Determines the absorption threshold of semiconductor materials.

[0072] Specifically, the physical mapping sub-branch can determine the absorption spectrum of each thin film structure based on the oscillator strength, resonant energy, damping coefficient, and bandgap energy. For example, the absorption spectrum of each thin film structure can be determined based on the following formula (1), i.e., the imaginary part ε2 of the dielectric function: ε2(λ)=Σ f(λ;A_k,E0_k,Γ_k,Eg_k) (1) Where ε2(λ) is the absorption spectrum of the thin film structure.

[0073] The dispersive spectrum of each thin film structure can be obtained based on the absorption spectrum according to the following formula (2): ε1(λ)=Hilbert[ε2] (2) Where ε1(λ) is the dispersive spectrum of the thin film structure, and Hilbert[] is the Hilbert transform.

[0074] The dielectric function of each thin film structure can be determined according to the following formula (3): ε(λ) = ε1(λ) + i×ε2(λ) (3) Where ε(λ) is the dielectric function of the thin film structure.

[0075] The role of the Hilbert transform is to calculate the dispersive spectrum ε1(λ) from the absorption spectrum ε2(λ), which ensures that the dielectric function satisfies the Kramer-Kronig (KK) constraint and that the absorption and dispersive spectra of the thin film structure are coupled.

[0076] Predicting oscillator parameters based on global eigenvectors and then predicting the dielectric function of thin film structures based on these oscillator parameters offers several advantages: Firstly, it enables the realization of an electronic transition or lattice vibration mode corresponding to each oscillator in the material, effectively describing the optical response of the thin film structure. Secondly, it allows for the description of the dielectric function across the entire spectral range using only a small number of parameters (4 parameters per oscillator). Thirdly, the parameters within each oscillator have clear physical meanings, facilitating understanding and analysis of the prediction process and improving its physical interpretability. Furthermore, since the physical model of each thin film structure can be composed of combinations of 1-3 oscillator models, the prediction dimension is reduced to 4-12 dimensions, significantly decreasing the number of parameters in the spectral inversion model. This ensures the model's generalization ability across materials, avoiding the need for large amounts of training data. Finally, because it conforms to physical laws, it effectively enhances the interpretability of the spectral inversion model.

[0077] In one implementation, the physical parameters may include latent variables for each thin film structure. The physical parameter prediction sub-branch can predict the latent variables for each thin film structure based on the global feature vector; the physical mapping layer sub-branch can decode the latent variables to obtain the dielectric function of each thin film structure.

[0078] In this embodiment, the latent variables can be decoded using a Multi-layer Perceptron (MLP) to obtain the dielectric function of each thin film structure. Using latent variables as a low-dimensional representation of the physical model of each thin film structure and decoding them with an MLP to obtain the dielectric function offers several advantages: MLP can learn arbitrarily complex dielectric functions; it offers high flexibility, not limited to a specific physical model; and it is highly adaptable, capable of handling various complex materials. Furthermore, the dielectric function obtained from decoding the thin film structure using latent variables can be constrained to meet the Kramer-Kronig constraint by adding a corresponding physical constraint loss term to the loss function.

[0079] In one implementation, the physical parameters may include oscillator parameters and latent variables for each thin film structure. The physical mapping layer sub-branch can obtain the dispersion and absorption spectra corresponding to the oscillator parameters and the latent variables, respectively. A mixed dispersion spectrum is obtained based on the dispersion spectra corresponding to the oscillator parameters and the latent variables. A mixed absorption spectrum is obtained based on the absorption spectra corresponding to the oscillator parameters and the latent variables. The dielectric function of each thin film structure is obtained based on the mixed dispersion and mixed absorption spectra.

[0080] In this embodiment, the dielectric function of each thin film structure can be predicted by combining the oscillator parameters and the latent variables. Specifically, the dispersion and absorption spectra corresponding to the oscillator parameters and the latent variables can be obtained respectively based on the oscillator parameters and the latent variables. A mixed dispersion spectrum is obtained by combining the dispersion spectra corresponding to the oscillator parameters and the latent variables. A mixed absorption spectrum is obtained by combining the absorption spectra corresponding to the oscillator parameters and the latent variables. Then, the dielectric function of each thin film structure is obtained based on the mixed dispersion spectrum and the mixed absorption spectrum. Specifically, the dielectric function of each thin film structure can be obtained according to the following formulas (4) and (5): ε 1= ε 1_osc +ε 1_lat (4) ε 2= ε 2_osc +ε 2_lat (5) Where, ε 1_osc The dispersive spectrum corresponding to the oscillator parameters; ε 1_lat The dispersive spectrum corresponding to the latent variable; ε 2_osc The absorption spectrum corresponding to the oscillator parameters; ε 2_lat This represents the absorption spectrum corresponding to the latent variable.

[0081] Combining oscillator parameters and latent variables to predict the dielectric function of each thin film structure has the advantage of balancing the stability and expressive power of the spectral inversion model and supplementing the latent variables with complex features. The main features are described by the oscillator parameters, while the details are supplemented by the latent variables, which can achieve a balance between the interpretability and flexibility of the spectral inversion model. Even if some latent variables fail, the oscillator parameters can still provide reasonable prediction results, effectively improving the robustness of the spectral inversion model.

[0082] Similarly, the obtained dielectric function can be constrained to satisfy the Kramer-Kronig constraint based on the physical constraint loss term in the loss function.

[0083] It should be noted that the above-mentioned different parameterization methods are all used to construct dielectric functions that satisfy physical constraints (Kremer-Kronig constraints). Their technical objectives, implementation paths, and overall technical concepts are consistent, and they belong to different implementation methods or equivalent alternatives under the same inventive concept.

[0084] In one implementation, the optical forward computing layer can calculate the ellipticity parameter prediction results based on the dielectric function and structural parameter prediction results of each thin film structure. In some specific examples, the optical forward computing layer can obtain the ellipticity parameter prediction results based on the parallel transfer matrix method (TMM), combining the dielectric function and structural parameter prediction results of the thin film structure.

[0085] Specifically, the predicted dielectric function and structural parameters of the thin film structure can be input into the optical forward computation layer, and the predicted ellipticity parameters can be obtained based on the parallel transfer matrix method. The predicted structural parameters include the thickness, roughness, and gradient of the dielectric function of each thin film layer.

[0086] The following explains the role of each parameter in the structural parameter prediction results within the TMM: 1) Thickness t_i: - Used to calculate the phase factor for each thin film structure; - It affects the multilayer interference effect and determines the oscillation period of the reflection spectrum.

[0087] 2) Roughness r_i: - By using the EMA (Effective Medium Approximation) model, the rough surface is equivalent to a mixed layer of material and air, and its dielectric function can be expressed by the following formula (6): ε_eff = 0.5×ε_material + 0.5×ε_air (6) Where ε_eff is the dielectric function of the hybrid layer, ε_material is the dielectric function of the material, and ε_air is the dielectric function of air.

[0088] - Based on the above calculations, the rough layer is equivalent to a thin effective dielectric layer in the TMM calculation.

[0089] 3) Gradient factor h_i: - Discretize the gradient layer into multiple sub-layers, n_sub = 20 + h_i×20 sub-layers; - Dielectric function of each sublayer: ε(h) = ε0·(1 - h_i×h_ratio), where h_ratio represents the center position of the sublayer; - Calculate the transfer matrix layer by layer in TMM and multiply them together.

[0090] In the process of calculating the transfer matrix, if the gradient factor is greater than 0, discretization is performed. Discretizing the gradient layer effectively ensures the continuous change of the dielectric function between layers.

[0091] For the discretized gradient layers, the transmission matrix of each gradient layer can be calculated sequentially layer by layer; within each gradient layer, the transmission matrix is ​​calculated in parallel for all wavelength points. The multiplication of the calculated transmission matrices is used to calculate the reflectivity and transmission amplitude of the stacked thin film structure, obtaining the predicted results of reflectivity and transmission amplitude. Based on the predicted results of reflectivity and transmission amplitude, the predicted results of the ellipsometric parameters of the thin film structure are then calculated.

[0092] It should be noted that the elliptic parameter prediction method based on the transfer matrix is ​​merely exemplary. Those skilled in the art can also implement the function of the optical forward calculation layer according to other numerical implementation methods, and thus obtain the elliptic parameter prediction results. All of these are within the scope of protection of this application.

[0093] Step S103: Determine the loss function based on the prediction results; and train the spectral inversion model based on the loss function.

[0094] In this embodiment, the loss function can be determined based on the prediction results of the spectral inversion model, and the model parameters of the spectral inversion model can be updated based on backpropagation of the loss function, thereby realizing the training of the spectral inversion model.

[0095] In one implementation, the loss function of the spectral inversion model may include a fitting error loss. The fitting error loss can be determined based on the difference between the predicted ellipticity parameters and the corresponding ellipticity parameters in the spectral data.

[0096] In this embodiment, the fitting error loss L_RMSE can be determined according to the following formula (7): L_RMSE = RMSE(Ψ, Δ,T) (7) RMSE stands for root mean square error.

[0097] The fitting error loss ensures that the ellipticity parameter predictions obtained by the spectral inversion model are closer to the ellipticity parameters in the spectral data used for training. The fitting error loss is used to directly measure the accuracy of the spectral inversion model.

[0098] In one implementation, the loss function of the spectral inversion model may further include a physical constraint loss. The physical constraint loss can be determined based on the dielectric function of each thin film structure obtained from the prediction results; the physical constraint loss is used to constrain the dielectric function to satisfy the Kramer-Kronig constraint.

[0099] In this embodiment, the physical constraint loss can be determined based on the dielectric function of each thin film structure predicted by the spectral inversion model. Specifically, the physical constraint loss L_KK can be determined according to the following formula (8): L_KK = ||ε1- Hilbert(ε2)|| 2 (8) Physical constraint loss is used to ensure that the dielectric function predicted by the spectral inversion model satisfies physical laws, i.e., the Cramer-Kronig constraint. Setting a physical constraint loss effectively ensures that the prediction results of the inversion model have physical meaning, avoiding non-physical interpretations. Specifically, when the physical parameter prediction sub-branch predicts the physical parameters of the thin film structure based on the global feature vector, including only the oscillator parameters of each thin film layer—that is, when an analytical oscillator model is used to characterize the physical model of each thin film layer—the calculated dielectric function is already forced to satisfy the Cramer-Kronig constraint. Therefore, no physical loss constraint term needs to be added to the loss function.

[0100] In one implementation, the loss function may further include a smoothness loss. The smoothness loss can be used to constrain the continuity and smoothness of the dielectric function corresponding to the predicted structural parameters of each thin film structure in the wavelength dimension.

[0101] In this embodiment, the smoothness loss can be used to constrain the continuity and smoothness of the dielectric function corresponding to the structural parameter prediction results of the spectral inversion model in the wavelength dimension. Specifically, the smoothness loss L_smooth can be determined according to the following formula (9): L_smooth = ||dε1 / dλ|| 2 + ||dε2 / dλ|| 2 (9) Where λ is the wavelength.

[0102] Setting a smoothness loss ensures that the dielectric function of the predicted results from the spectral inversion model is continuously smooth in the wavelength dimension, avoiding non-physical abrupt changes and improving the numerical stability of the predicted results from the spectral inversion model.

[0103] In one implementation, the loss function of the spectral inversion model includes a primary loss function for fitting error, used to optimize the accuracy of elliptic parameter prediction; a secondary loss function for physical constraint, used to balance physical constraints and fitting accuracy; and an auxiliary loss function for smoothness, used to improve the numerical stability of the prediction results. For example, the weight of the fitting error loss in the loss function can be 1; the weight of the physical constraint loss can be less than 1 and greater than the weight of the smoothness loss.

[0104] In one implementation, the loss function L of the spectral inversion model can be determined according to the following formula (10): L = RMSE(Ψ, Δ,T) + λ_KK×L_KK + λ_s×L_smooth (10) Where λ_KK is the weight of the physical constraint loss, and λ_s is the weight of the smoothness loss.

[0105] In one implementation, the fitting error loss can be normalized by dividing the mean square error (MSE) of the calculated elliptic parameter prediction results by the square mean of the target value as the fitting error loss. This ensures that the fitting error loss is within a reasonable range (<1.0), avoids numerical explosion, and improves the stability and convergence speed of the spectral inversion model.

[0106] In a specific example, when the elliptic parameter prediction branch includes a physical parameter prediction sub-branch, a physical mapping layer sub-branch, and an optical forward computation layer, the weight of the physical constraint loss can be 0.1, and the weight of the smoothness loss can be 0.01.

[0107] In another specific example, when the elliptic parameter prediction branch directly predicts the elliptic parameters based on the global feature vector, the weight of the physical constraint loss can be 0.001, and the weight of the smoothness loss can be 0.0001.

[0108] In one implementation, a training cutoff condition for the spectral inversion model can be set. When the training cutoff condition is met, the spectral inversion model can be considered to have converged.

[0109] In a specific example, the convergence metrics for the spectral inversion model can include a fitting error loss of <0.03 and a thickness error of <2nm. That is, when the fitting error loss is <0.03 and the thickness error is <2nm, the spectral inversion model can be considered to have met the training cutoff condition and has converged.

[0110] In one implementation, see Appendix Figure 2 ,like Figure 2 As shown, the main framework of the spectral inversion model in this embodiment may include: a spectral data input layer, ResNet-1D, a physical parameter prediction sub-branch, a physical mapping layer sub-branch, an optical forward computation layer, a structural parameter prediction branch, and a loss function optimization. The physical parameter prediction sub-branch, the physical mapping layer sub-branch, and the optical forward computation layer constitute the ellipticity parameter prediction branch.

[0111] Spectral data input layer: Its function is to collect and standardize the observation and prior data used for inversion; its input is: spectral data used for training, including: Ψ, Δ, T, angle of incidence (AOI), wavelength (λ), basis / polarization information, etc.; its output is: coded information of ellipsoid parameters and experimental condition vector.

[0112] ResNet-1D: Its function is to extract global feature vectors from the encoded information and experimental condition vectors corresponding to multi-channel one-dimensional spectral data. ResNet-1D can include Conv1D (one-dimensional convolutional network) + residual blocks + global pooling layers, which can achieve noise suppression and scale alignment. Its output data is a global feature vector.

[0113] The physical parameter prediction sub-branch is responsible for predicting the physical parameters of the thin film structure. These parameters can be predicted using latent variables z and / or oscillator parameters {A,E0,Γ,Eg}. Its output data consists of the physical parameters for each layer of the thin film structure.

[0114] The structural parameter prediction branch predicts the structural and interface characteristics of the thin film structure, including the thickness, surface roughness, and dielectric function gradient of each layer. Its output is the predicted structural parameters of the thin film structure.

[0115] The physical mapping sub-branch maps physical parameters to the dielectric function of each thin film layer. The prediction process of the physical mapping sub-branch satisfies the Cramer-Kronig constraint. Its output data are the dielectric functions of each thin film layer.

[0116] Optical forward calculation layer: Its function is to calculate the reflection / transmission amplitude of the multilayer stack based on the predicted dielectric function and structural parameters of each thin film structure. Its output data is the predicted ellipsometric parameters (including Ψ, Δ, and T).

[0117] Loss function optimization: The loss function is a combination of fitting error loss, physical constraint loss, and smoothness loss. Based on the loss function, end-to-end training of the spectral inversion model is achieved, jointly updating the model parameters of the backbone network, structural parameter prediction branch, and elliptic parameter prediction branch, thereby improving the physical interpretability and generalization ability of the spectral inversion model. The training strategy for the spectral inversion model is: early stopping on the validation set or optimal weight preservation method to prevent overfitting.

[0118] Based on the methods described in steps S101 to S103 above, this embodiment of the application obtains spectral data for training, including ellipticity parameters of the thin film structure. The ellipticity parameters are encoded to obtain encoding information, which is then input into the spectral inversion model to obtain the prediction results. The spectral inversion model includes a backbone network, a structural parameter prediction branch, and an ellipticity parameter prediction branch. The backbone network is used to extract spectral features from the encoded information and obtain a global feature vector; the ellipticity parameter prediction branch and the structural parameter prediction branch predict the ellipticity parameter prediction results and the structural parameter prediction results, respectively, based on the global feature vector; a loss function is determined based on the prediction results of the spectral inversion model; and the spectral inversion model is trained based on the loss function. With the above configuration, since the encoding information used for model training in this embodiment is obtained by encoding the ellipticity parameters of the thin film structure, and since the spectral inversion model consists of a backbone network, an ellipticity parameter prediction branch, and a structural parameter prediction branch, end-to-end training of the spectral inversion model based on the spectral data used for training can be achieved. The prediction results include the ellipticity parameter prediction results of the thin film structure and the structural parameter prediction results of each layer of the thin film structure. This effectively ensures the physical interpretability of the prediction results of the spectral inversion model while reducing the dimensionality of the model parameters. Furthermore, it ensures that the trained spectral inversion model can effectively invert multilayer thin film structures.

[0119] Furthermore, this application also provides a method for spectral inversion of multilayer thin films.

[0120] See attached diagram. Figure 3 This is a schematic flowchart illustrating the main steps of a multilayer thin film spectral inversion method according to an embodiment of this application. Figure 3 As shown, the multilayer thin film spectral inversion method of this application mainly includes the following steps S201 and S202: Step S201: Encode the ellipticity parameters of the thin film structure to be detected to obtain the detection coding information.

[0121] In this embodiment, after training a spectral inversion model with satisfactory performance using the aforementioned multilayer thin film spectral inversion model training method, the spectral inversion model can be used to invert the spectra of multilayer thin films. Specifically, the ellipticity parameters of the thin film structure to be detected can first be encoded to obtain the detection encoding information. The ellipticity parameters of the thin film structure to be detected can be measured using instruments such as an ellipsometry. The process of encoding the ellipticity parameters of the thin film structure to be detected is similar to the aforementioned process of encoding the ellipticity parameters in the spectral data used for training; for simplicity, it will not be elaborated further here.

[0122] Step S202: Input the coded information to be detected into the trained spectral inversion model, perform inversion inference based on the preset structural parameter range, and obtain the material and structural properties of each layer of the thin film structure to be detected; wherein, the spectral inversion model is trained based on the multilayer thin film spectral inversion model training method described in the above embodiment of the multilayer thin film spectral inversion model training method.

[0123] In this embodiment, the encoded information to be detected can be input into the trained spectral inversion model, and inversion reasoning with physical constraints can be performed within the preset structural parameter range to obtain the material and structural properties of the thin film structure to be detected.

[0124] In one implementation, the range of structural parameters for inversion inference can be set based on the range of training data during the training phase of the spectral inversion model. For example, during the training phase, the thickness range of the thin film structure is t_i ∈ [5, 200] nm, and during the inversion inference process, the thickness range of the thin film structure can be set to t_i ± 20 nm. That is, within the thickness range of [5, 200] nm corresponding to the thin film structure in the training data, inversion inference is performed within ± 20 nm of the initial predicted thickness value of the spectral inversion model.

[0125] In one embodiment, a preset number of structural parameter sample values ​​can be uniformly sampled within the range of structural parameters; based on the loss function value corresponding to each structural parameter sample value, a target structural parameter sample value with the smallest loss function value is determined; and inversion reasoning is performed within a preset range of the target structural parameter sample value to obtain the material and structural properties of each thin film structure.

[0126] In this embodiment, a search strategy combining coarse and fine search can be employed within the structural parameter range. Specifically, uniform sampling is performed within the structural parameter range, with a preset number of structural parameter sample values, and the loss function value corresponding to each structural parameter sample value is calculated. Then, physical constraint inversion reasoning is performed within a preset range of the target structural parameter sample value with the latest loss function value to obtain the material and structural properties of each thin film structure. Since the search does not need to start from randomly initialized parameters but instead performs fine-tuning near local optima, the optimization objective is relatively simple, significantly improving the convergence speed.

[0127] In one example, when inverting the thickness of a certain thin film, the structural parameter range can be set to ±20 nm, with a preset number of 10. That is, in the coarse search process, 10 thickness samples are uniformly sampled within ±20 nm of the initially predicted thickness value from the thickness values ​​in the training data of the spectral inversion model, and the target thickness sample value is determined among these 10. The fine search process further refines the search around the target thickness sample value to obtain the accurate inversion result for the thickness of that thin film.

[0128] In one implementation, during the inversion inference process of the spectral inversion model, gradient descent or grid search methods can be used to obtain the material and structural properties of each thin film structure.

[0129] In one implementation, the inversion inference result is obtained based on the structural parameter range and the coding information to be detected; the loss function value corresponding to the inversion inference result is calculated; and the model parameters of the structural parameter prediction branch and the elliptic parameter prediction branch of the spectral inversion model are updated in reverse according to the loss function value.

[0130] In this embodiment, during the physical constraint inversion inference process based on the spectral inversion model, since the backbone network of the spectral inversion model has already learned the ability to extract stable features from spectral data during the training phase, and these features are universal for different material combinations, the model parameters of the backbone network can be frozen during the inversion inference process. Freezing the backbone network can significantly reduce the number of trainable parameters in the inversion inference stage (reducing the number of inversion inference parameters by at least 80% in most scenarios), effectively reducing the risk of overfitting during the inversion inference process and improving the generalization ability of the inversion inference model. By only fine-tuning the model parameters of the structural parameter prediction branch and the elliptic parameter prediction branch during the inversion inference process, the ability to quickly adapt to new samples can be achieved, enabling adaptation to new data to be detected within a second-level timeframe. Furthermore, freezing the model parameters of the backbone network can also avoid large changes in model parameters during the inversion inference process, effectively improving the stability of the process of obtaining the material and structural properties of thin film structures through inversion inference based on the spectral inversion model.

[0131] In one implementation, a large learning rate can be set for the spectral inversion model during the inversion inference process, which can ensure that the spectral inversion model converges after 5-10 iterations.

[0132] The multilayer thin film spectral inversion method provided in this invention offers an efficient deep learning solution. By optimizing the training strategy and simplifying the model architecture, it significantly improves computational efficiency and training stability while ensuring prediction accuracy. It also enhances the generalization ability to small samples or unknown materials, making it more suitable for deployment in equipment or production line environments.

[0133] It should be noted that although the steps in the above embodiments are described in a specific order, those skilled in the art will understand that in order to achieve the effect of this application, different steps do not necessarily have to be executed in such an order. They can be executed simultaneously (in parallel) or in other orders. These adjusted solutions are equivalent to the technical solutions described in this application and therefore will also fall within the protection scope of this application.

[0134] Those skilled in the art will understand that all or part of the processes in the method of the above-described embodiment can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0135] Another aspect of this application provides a computer-readable storage medium.

[0136] In one embodiment of a computer-readable storage medium according to this application, the computer-readable storage medium can be configured to store a program for performing the multilayer thin-film spectral inversion model training method of the above-described method embodiments. This program can be loaded and run by a processor to implement the above-described multilayer thin-film spectral inversion model training method. For ease of explanation, only the parts related to the embodiments of this application are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of this application. The computer-readable storage medium can be a storage device formed by various electronic devices, such as a magnetic disk, hard disk, optical disk, flash memory, read-only memory, random access memory, etc. Optionally, in the embodiments of this application, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0137] In another embodiment of the computer-readable storage medium according to this application, the computer-readable storage medium can be configured to store a program for performing the multilayer thin film spectral inversion method of the above-described method embodiments. This program can be loaded and run by a processor to implement the above-described multilayer thin film spectral inversion method. For ease of explanation, only the parts related to the embodiments of this application are shown; for specific technical details not disclosed, please refer to the method section of the embodiments of this application. The computer-readable storage medium can be a storage device formed by various electronic devices, such as a magnetic disk, hard disk, optical disk, flash memory, read-only memory, random access memory, etc. Optionally, in the embodiments of this application, the computer-readable storage medium is a non-transitory computer-readable storage medium.

[0138] Another aspect of this application provides an electronic device.

[0139] In an embodiment of an electronic device according to this application, the electronic device may include at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program that, when executed by the at least one processor, implements the method described in any embodiment of the multilayer thin film spectral inversion model training method described above.

[0140] In another embodiment of the electronic device according to this application, the electronic device may include at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program that, when executed by the at least one processor, implements the method described in any embodiment of the multilayer thin film spectral inversion model training method described above.

[0141] In some embodiments of this application, the processor may be a central processing unit, a microprocessor, a graphics processor, a digital signal processor, or any other suitable processor. The processor has data and / or signal processing functions. The processor may be implemented in software, in hardware, or a combination of both. The electronic device described in this application may be, but is not limited to, a mobile phone, tablet computer, desktop computer, laptop computer, handheld computer, notebook computer, in-vehicle device, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), augmented reality (AR) / virtual reality (VR) device, etc., and this application does not limit this.

[0142] The technical solution of this application has been described above with reference to one embodiment shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of this application is obviously not limited to these specific embodiments. Without departing from the principles of this application, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of this application.

Claims

1. A training method for a multilayer thin film spectral inversion model, characterized in that, The method includes: Acquire spectral data for model training; wherein the spectral data includes ellipticity parameters of the thin film structure; and encode the ellipticity parameters to obtain encoded information; the thin film structure includes one or more optical thin films stacked on a substrate; The encoded information is input into a spectral inversion model to obtain the prediction results of the spectral inversion model. The spectral inversion model includes a backbone network, a structural parameter prediction branch, and an elliptic parameter prediction branch. The prediction results include the elliptic parameter prediction results of the thin film structure and the structural parameter prediction results of each thin film layer. The backbone network is used to extract spectral features from the encoded information and obtain a global feature vector. The elliptic parameter prediction branch and the structural parameter prediction branch predict the elliptic parameter prediction results and the structural parameter prediction results, respectively, based on the global feature vector. The elliptic parameter prediction branch includes a physical parameter prediction sub-branch, a physical mapping layer sub-branch, and an optical forward calculation layer. The physical parameter prediction sub-branch predicts the physical parameters of the thin film structure based on the global feature vector. The physical parameters include the oscillator parameters and latent variables of each thin film layer. The physical mapping layer sub-branch maps the physical parameters to the dielectric function of each thin film layer. The dielectric function satisfies the Kramer-Kronig constraint. The optical forward calculation layer calculates the elliptic parameter prediction results based on the dielectric function of each thin film layer and the structural parameter prediction results. A loss function is determined based on the prediction results; and the spectral inversion model is trained based on the loss function; wherein the loss function includes fitting error loss, physical constraint loss, and smoothness loss; the fitting error loss is determined based on the difference between the predicted ellipticity parameter and the corresponding ellipticity parameter in the spectral data; the physical constraint loss is determined based on the dielectric function of each thin film structure obtained from the prediction results; the physical constraint loss is used to constrain the dielectric function to satisfy the Kramer-Kronig constraint; the smoothness loss is used to constrain the continuous smoothness of the dielectric function corresponding to the predicted structural parameters of each thin film structure in the wavelength dimension.

2. The training method for the multilayer thin film spectral inversion model according to claim 1, characterized in that, The structural parameter prediction branch is used to predict the structural and interface features of each thin film structure based on the global feature vector, as the structural parameter prediction result; wherein, the structural parameter prediction result includes the thickness of each thin film structure.

3. The training method for the multilayer thin film spectral inversion model according to claim 2, characterized in that, The structural parameter prediction results also include the gradient of the roughness and / or dielectric function of each thin film structure.

4. The training method for the multilayer thin film spectral inversion model according to claim 1, characterized in that, The physical mapping layer sub-branch maps the physical parameters to the dielectric function of each thin film structure, including: The physical mapping layer sub-branch determines the absorption spectrum of each thin film structure based on the oscillator strength, resonance energy, damping coefficient, and band gap energy in the oscillator parameters; The absorption spectrum was subjected to Hilbert transform to obtain the dispersive spectrum of each thin film structure; Based on the absorption spectrum and the dispersion spectrum, the dielectric function of each thin film structure is obtained using the oscillator model.

5. The training method for the multilayer thin film spectral inversion model according to claim 1, characterized in that, The physical mapping layer sub-branch maps the physical parameters to the dielectric function of each thin film structure, including: The physical mapping layer sub-branch decodes the latent variables to obtain the dielectric function of each thin film structure.

6. The training method for the multilayer thin film spectral inversion model according to claim 1, characterized in that, The physical mapping layer sub-branch maps the physical parameters to the dielectric function of each thin film structure, including: The physical mapping layer sub-branch obtains the dispersion spectrum and absorption spectrum corresponding to the oscillator parameters and the dispersion spectrum and absorption spectrum corresponding to the latent variables, respectively, based on the oscillator parameters and latent variables. Based on the dispersion spectrum corresponding to the oscillator parameters and the dispersion spectrum corresponding to the latent variable, a mixed dispersion spectrum is obtained; Based on the absorption spectrum corresponding to the oscillator parameters and the absorption spectrum corresponding to the latent variable, a mixed absorption spectrum is obtained; The dielectric function of each thin film structure is obtained based on the mixed dispersion spectrum and the mixed absorption spectrum.

7. The training method for the multilayer thin film spectral inversion model according to claim 1, characterized in that, The process of encoding the elliptic deviation parameters to obtain encoded information includes: The amplitude ratio angle and phase difference angle in the elliptic parameters are normalized, and the cosine code of the amplitude ratio angle cos2Ψ, the cosine component of the phase difference angle sin2Ψ×cosΔ, and the sine component of the phase difference angle sin2Ψ×sinΔ are calculated; where Ψ is the amplitude ratio angle and Δ is the phase difference angle. The cosine encoding, the cosine component, the sine component, and the transmittance in the elliptic parameter are used as the encoding information.

8. The training method for the multilayer thin film spectral inversion model according to claim 1, characterized in that, The acquisition of spectral data for training includes: Based on a preset sampling range, the material and structural parameters of the thin film structure are randomly sampled; Based on the physical model, the ellipticity parameters of the thin film structure are obtained from the material parameters and structural parameters obtained by random sampling, and used as the spectral data for training.

9. The training method for the multilayer thin film spectral inversion model according to claim 8, characterized in that, The thin film structure includes at least one optical thin film.

10. A method for spectral inversion of multilayer thin films, characterized in that, The method includes: The ellipticity parameters of the thin film structure to be tested are encoded to obtain the coded information to be tested; The coded information to be detected is input into the trained spectral inversion model, and inversion reasoning is performed based on the preset structural parameter range to obtain the material and structural properties of each thin film structure in the thin film structure to be detected. The spectral inversion model is obtained by training based on the multilayer thin film spectral inversion model training method according to any one of claims 1 to 9.

11. The multilayer thin film spectral inversion method according to claim 10, characterized in that, The inversion reasoning based on a preset structural parameter range to obtain the material and structural properties of each thin film layer in the thin film structure to be tested includes: A preset number of structural parameter sample values ​​are uniformly sampled within the range of the structural parameters; Based on the loss function value corresponding to each structural parameter sample value, determine the target structural parameter sample value with the minimum loss function value; Inversion reasoning is performed within a preset range of the target structural parameter sampling values ​​to obtain the material and structural properties of each thin film structure.

12. The multilayer thin film spectral inversion method according to claim 10, characterized in that, The method further includes: Based on the range of structural parameters and the coding information to be detected, the inversion reasoning result is obtained; Calculate the loss function value corresponding to the inversion inference result; and update the model parameters of the structure parameter prediction branch and / or ellipticity parameter prediction branch of the spectral inversion model in reverse according to the loss function value.

13. An electronic device, characterized in that, include: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores a computer program, which, when executed by the at least one processor, implements the multilayer thin film spectral inversion model training method according to any one of claims 1 to 9 and / or the multilayer thin film spectral inversion method according to any one of claims 10 to 12.

14. A computer-readable storage medium storing a plurality of program codes, characterized in that, The program code is adapted to be loaded and run by a processor to perform the multilayer thin film spectral inversion model training method as described in any one of claims 1 to 9 and / or the multilayer thin film spectral inversion method as described in any one of claims 10 to 12.

Images