Small sample migration-based multilayer film spectrum inversion model training method, multilayer film spectrum inversion method, device and medium

Abstract

The application provides a multilayer film spectrum inversion model training method based on small sample migration, a multilayer film spectrum inversion method, equipment and a medium, aiming to solve the technical problem that the spectrum inversion method has a large demand for sample data. For this purpose, the application pre-trains a spectrum inversion model based on a spectrum synthesis data set, and obtains a sample data set containing multiple spectrum sample data in a preset migration scene, wherein the spectrum sample data includes ellipsometric parameters of a multilayer film structure measured in the migration scene. According to the sample data set, the pre-trained spectrum inversion model is adjusted to obtain a spectrum inversion model adapted to the migration scene. When the spectrum inversion model is adjusted, the model parameters of the backbone network are frozen, which can improve the speed of model adjustment, realize efficient and high-precision model adjustment of the spectrum inversion model adapted to the migration scene, and improve the adaptability and inference efficiency of the spectrum inversion model for different migration scenes.

Description

Technical Field

[0001] This application relates to the field of semiconductor technology, specifically to a training method for a multilayer thin film spectral inversion model based on small sample migration, a multilayer thin film spectral inversion method, equipment, and medium. Background Technology

[0002] Ellipsometry: An optical instrument that characterizes thin film properties 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 an inverse problem. The inverse problem is the inference of system parameters from observational data. In ellipsometer analysis, it involves inferring (n, k, d) from (Ψ, Δ). Traditional methods rely on iterative fitting, require manual intervention and selection of dispersion models (such as Cauchy and Sellmeier), are time-consuming, and are prone to getting trapped in local optima.

[0003] Existing spectral inversion methods typically rely on a large number of training samples to ensure generalization performance and inversion accuracy. For scenarios involving new material systems, multilayer structures, and new measurement environments (AOI / band / substrate), the cost of re-acquiring large-scale training data is prohibitive. This problem is particularly pronounced when employing complex models such as high-dimensional KK B-splines—because these models introduce high-dimensional parameter spaces to pursue extreme flexibility, their training requires a large amount of data to constrain them, which further amplifies the demand for a large number of samples.

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

[0005] In order to overcome the above-mentioned shortcomings, this application is made to solve or at least partially solve the technical problem that the existing spectral inversion methods require a large amount of sample data.

[0006] In a first aspect, a training method for a multilayer thin film spectral inversion model based on small-sample transfer is provided, the method comprising:

[0007] The spectral inversion model is pre-trained based on a spectral synthesis dataset; wherein, the spectral synthesis dataset includes multiple spectral synthesis data; the spectral synthesis data includes the ellipticity parameters of a multilayer thin film structure; the spectral inversion model includes a backbone network, an ellipticity parameter prediction branch, and a structural parameter prediction branch; the backbone network is used to extract the spectral features of the multilayer thin film structure and obtain a global feature vector;

[0008] Obtain a sample dataset containing multiple spectral sample data in a preset migration scenario; wherein, the spectral sample data includes ellipticity parameters of a multilayer thin film structure measured in the migration scenario;

[0009] Based on the sample dataset, the pre-trained spectral inversion model is adjusted to obtain a spectral inversion model adapted to the migration scenario; wherein, when adjusting the spectral inversion model, the model parameters of the backbone network are frozen.

[0010] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion model based on small sample transfer, the elliptic parameter prediction branch includes a physical parameter prediction sub-branch and a physical mapping layer sub-branch.

[0011] The physical parameter prediction sub-branch is used to predict the physical parameters of each thin film structure based on the global feature vector, and output the physical parameter prediction results;

[0012] The physical mapping sub-branch is used to map the physical parameter prediction results to the dielectric function of each thin film structure; wherein the dielectric function is constrained by the Cramer-Kronig relation.

[0013] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion model based on small sample transfer, the structural parameter prediction branch is used to predict the structural and interface features of each thin film structure according to the global feature vector, and output the structural parameter prediction results; wherein, the structural parameter prediction results include: the thickness of each thin film structure.

[0014] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion model based on small sample migration, the predicted structural parameters also include the gradient of the roughness and / or dielectric function of each thin film structure.

[0015] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion models based on small sample transfer, the adjustment of the pre-trained spectral inversion model includes:

[0016] Adjust the model parameters of the predicted sub-branch based on the physical parameters, and / or,

[0017] Adjust the model parameters of the physical mapping layer sub-branch, and / or,

[0018] Adjust the model parameters of the predicted branch of the structural parameters.

[0019] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion model based on small sample transfer, the total loss of the pre-training and model adjustment process of the spectral inversion model includes the fitting error loss; the fitting error loss is determined based on the ellipticity parameter prediction results output by the spectral inversion model and the ellipticity parameters in the corresponding spectral synthesis data.

[0020] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion model based on small sample transfer, the total loss of the pre-training and model adjustment process of the spectral inversion model also includes Cramer-Kronig consistency constraint loss; the Cramer-Kronig consistency constraint loss is used to represent the Cramer-Kronig relation constraint of the dielectric function of each thin film structure predicted by the spectral inversion model.

[0021] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion model based on small sample transfer, the total loss of the pre-training and model adjustment process of the spectral inversion model also includes a search range constraint loss; the search range constraint loss is used to constrain the search range of the physical parameter prediction results and / or the structural parameter prediction results.

[0022] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion model based on small sample migration, the weight of the fitting error loss in the total loss is the first weight;

[0023] The weight of the Kramer-Kronig consistency constraint loss in the total loss is the second weight.

[0024] The search range constraint loss has the third weight in the total loss;

[0025] The first weight is greater than the second weight, and the second weight is greater than the third weight.

[0026] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion models based on small sample transfer,

[0027] The physical parameters include the oscillator parameters of each thin film structure and / or the latent variables of each thin film structure.

[0028] In one technical solution of the above-mentioned training method for multilayer thin film spectral inversion models based on small-sample transfer, after obtaining the spectral inversion model adapted to the transfer scenario, the method further includes:

[0029] Perform spectral inversion inference on unlabeled spectral sample data to obtain the inference results;

[0030] If the prediction error of the inference result is greater than the preset error threshold, the corresponding unlabeled spectral sample data is added to the sample dataset; and the updated sample dataset is used to continue training the spectral inversion model.

[0031] In a second aspect, a method for spectral inversion of multilayer thin films is provided, the method comprising:

[0032] The ellipticity parameters obtained by ellipticity measurement of the multilayer thin film structure under the preset migration scenario are input into the spectral inversion model adapted to the migration scenario to perform spectral inversion inference and obtain the material and structural properties of each layer of the multilayer thin film structure.

[0033] The spectral inversion model adapted to the migration scenario is obtained by training the multilayer thin film spectral inversion model based on small sample migration, as described in any of the above-mentioned technical solutions for training multilayer thin film spectral inversion models based on small sample migration.

[0034] In one technical solution of the above-mentioned multilayer thin film spectral inversion method, the method further includes:

[0035] In response to the detection of a performance degradation of the spectral inversion model and / or a change in measurement environment parameters, spectral measurement data of the multilayer thin film structure under the small sample migration scenario are collected, and the spectral measurement data includes the ellipticity parameters of the multilayer thin film structure;

[0036] Based on the spectral measurement data, the spectral inversion model is adjusted; wherein, when adjusting the spectral inversion model, the model parameters of the backbone network of the spectral inversion model are frozen, and only the model parameters of the structural parameter prediction branch and / or elliptic parameter prediction branch of the spectral inversion model are adjusted.

[0037] 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 based on small sample transfer and / or the method described in any one of the above-described technical solutions for multilayer thin film spectral inversion methods.

[0038] 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 based on small sample transfer and / or the method described in any one of the above-described technical solutions of the multilayer thin film spectral inversion method.

[0039] The above-described technical solutions of this application have at least one or more of the following beneficial effects:

[0040] In implementing the training method for a multilayer thin film spectral inversion model based on small-sample transfer provided in this application, the spectral inversion model is pre-trained based on a spectral synthesis dataset. This dataset includes multiple spectral synthesis data points, each containing ellipticity parameters of a multilayer thin film structure. The spectral inversion model comprises a backbone network, an ellipticity parameter prediction branch, and a structural parameter prediction branch. The backbone network is used to extract the spectral features of the multilayer thin film structure and obtain a global feature vector. A sample dataset containing multiple spectral sample data points within a pre-defined transfer scenario is obtained. These sample data points include the ellipticity parameters of the multilayer thin film structure measured under the transfer scenario. Based on the sample dataset, the pre-trained spectral inversion model is adjusted to obtain a spectral inversion model adapted to the transfer scenario. During the adjustment, the model parameters of the backbone network are frozen. Through this configuration, this application enables the pre-training of the spectral inversion model based on a spectral synthesis dataset. Furthermore, for transfer scenarios, the pre-trained spectral inversion model is adjusted based on the sample dataset within the transfer scenario, allowing the spectral inversion model to adapt to the transfer scenario based solely on the sample dataset. Meanwhile, because the model parameters of the backbone network of the spectral inversion model are frozen during the model adjustment process to adapt to the migration scenario, only other model parameters besides the backbone network need to be adjusted during the model adjustment process. This can effectively improve the speed of model adjustment, thereby achieving efficient and high-precision model adjustment of the spectral inversion model to adapt to the migration scenario. It can also effectively reduce the data dependence of the spectral inversion model when adapting to the migration scenario and improve the adaptability and inference efficiency of the spectral inversion model for different migration scenarios. Attached Figure Description

[0041] 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:

[0042] Figure 1 This is a schematic flowchart of the main steps of a training method for a multilayer thin film spectral inversion model based on small sample transfer, according to an embodiment of this application.

[0043] Figure 2 This is a schematic flowchart of the main steps of a multilayer thin film spectral inversion model training method based on small sample transfer, according to one embodiment of the present application.

[0044] 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

[0045] 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.

[0046] 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.

[0047] Here we will first explain some of the terms used in this application.

[0048] 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).

[0049] Ψ(Psi): Defined as the arctangent of the ratio of the amplitudes of the p and s components in the reflected light. The change of Ψ mainly reflects the change in the intensity ratio of the reflected light, and its range is [0°, 90°].

[0050] 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°].

[0051] 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.

[0052] 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.

[0053] The Hilbert transform is a linear operator widely used in mathematics and signal processing. It transforms a real-valued function or signal into another function with the same domain, its core function being to generate an analytical representation of the signal.

[0054] See appendix Figure 1, Figure 1 This is a schematic flowchart illustrating the main steps of a training method for a multilayer thin film spectral inversion model based on small-sample transfer learning, according to an embodiment of this application. Figure 1 As shown, the training method for the multilayer thin film spectral inversion model based on small sample migration in this application embodiment mainly includes the following steps S101 to S103.

[0055] Step S101: Pre-train the spectral inversion model based on the spectral synthesis dataset; wherein, the spectral synthesis dataset includes multiple spectral synthesis data; the spectral synthesis data includes the ellipticity parameters of the multilayer thin film structure; the spectral inversion model includes a backbone network, an ellipticity parameter prediction branch, and a structural parameter prediction branch; the backbone network is used to extract the spectral features of the multilayer thin film structure and obtain the global feature vector.

[0056] In this embodiment, a spectral synthesis dataset containing multiple spectral synthesis data can be constructed. The spectral synthesis dataset is then used to pre-train the spectral inversion model, enabling the backbone network to effectively learn the general mapping relationship between "material-structure-spectrum" of multilayer thin film structures. It can also learn the spectral statistical characteristics and interlayer correlations of multilayer thin film structures to obtain global feature vectors, providing strong feature extraction capabilities for subsequent transfer processing.

[0057] In one embodiment, the ellipsoidal parameters may include ellipticity angle Ψ, phase difference Δ, and transmittance T. Ellipticity angle Ψ describes the ratio of the major and minor axes of elliptically polarized light; its unit is degrees (°); its value ranges from 0° to Ψ to 45°; and it is measured directly using an ellipsometer. Phase difference Δ represents the phase difference between p-polarized and s-polarized light; its unit is degrees (°); its value ranges from -180° to Δ to 180°; and it is measured directly using an ellipsometer. Transmittance T represents the ratio of transmitted light intensity to incident light intensity; its value ranges from 0 to T to 1; and it is measured using a spectrophotometer or an ellipsometer.

[0058] In one implementation, the spectral inversion model may include a backbone network, an elliptic parameter prediction branch, and a structural parameter prediction branch. During pre-training, the parameters in the backbone network, elliptic parameter prediction branch, and structural parameter prediction branch of the spectral inversion model can all be trained. The backbone network is used to extract the spectral features of the multilayer thin film structure and obtain a global feature vector. The elliptic parameter prediction branch and the structural parameter prediction branch can be used to predict the elliptic parameter and structural parameters respectively based on the global feature vector, to obtain the elliptic parameter prediction results and structural parameter prediction results respectively.

[0059] In one implementation, the backbone network can be ResNet-1D (a one-dimensional residual network). In a specific example, the backbone network may include five residual blocks and a global pooling layer. The global feature vector output by the backbone network can be a 128-dimensional feature vector, which is simultaneously input to both the ellipticity parameter prediction branch and the structure parameter prediction branch to achieve parameter sharing and improve parameter processing efficiency. In another specific implementation, a self-attention layer can also be added to the backbone network, allowing it to focus on important spectral features.

[0060] In one implementation, the ellipticity parameter prediction branch may include a physical parameter prediction sub-branch and a physical mapping layer sub-branch. The physical parameter prediction sub-branch can be used to predict the physical parameters of each thin film structure based on the global eigenvector and output the predicted physical parameters. The physical mapping layer sub-branch can be used to map the predicted physical parameters to the dielectric function of each thin film structure; wherein the dielectric function is constrained by the Cramer-Kronig relation, which is a causal mathematical expression. For example, in the calculation of the dielectric function, the dispersive spectrum ε1(λ) can be uniquely determined from the absorption spectrum ε2(λ) through the Hilbert transform, meaning the response cannot precede the excitation, ensuring that the predicted dielectric function of each thin film structure satisfies physical laws.

[0061] In one implementation, the physical parameters can be the oscillator parameters of the thin film structure. An oscillator model is a physical model used to fit the complex dielectric function of a thin film material; it uses a set of constrained peaks to fit the material's response at different wavelengths. It is also fundamental to understanding the interaction between light and matter; the superposition of multiple oscillator models can describe the entire optical response spectrum of the material. The physical parameter prediction sub-branch can predict the oscillator parameters of each thin film layer based on the global eigenvector. The physical mapping layer sub-branch can map the dielectric function of each thin film layer based on the oscillator parameters of the thin film structure.

[0062] In a specific example, oscillator parameters may include oscillator strength, resonant energy, damping coefficient, and bandgap energy. For instance, when using the Lorentz oscillator model to describe the dielectric constant, the dielectric function of each thin film structure can be obtained based on the following formulas (1)-(3), according to the oscillator parameters of each thin film structure:

[0063] ε2(λ) = Σ A·E0·Γ / [(E²-E0²)² + (Γ·E)²] (1)

[0064] ε1(λ)=Hilbert[ε2] (2)

[0065] ε(λ) = ε1(λ) + i×ε2(λ) (3)

[0066] Where ε(λ) is the dielectric function, ε1(λ) is the dispersive spectrum, ε2(λ) is the absorption spectrum, A is the oscillator strength, E0 is the resonance energy, Γ is the damping coefficient, E is the photon energy, and the material can only produce interband coefficients when E is greater than Eg, where Eg is the band gap energy. Hilbert[] is the Hilbert transform. Through the Hilbert transform, the dielectric function of the obtained thin film structure is constrained by the Cramer-Kronig relation.

[0067] The advantage of obtaining the dielectric function of thin film structures by using oscillator parameter mapping is that:

[0068] - It has strong physical interpretability, and the oscillator parameters A, E0, Γ, and Eg all have clear physical meanings;

[0069] - Good stability: Satisfies the Kramer-Kronig relation;

[0070] - Few parameters: Only 4 parameters per layer.

[0071] Obtaining the dielectric function of thin film structures using oscillator parameter mapping is more suitable for material scenarios where the optical characteristics of the material can be described by a small number of oscillators. Examples include standard semiconductor materials (Si, Ge, GaAs, etc.); oxide materials (SiO2, TiO2, Al2); and metal thin films (Au, Ag, Cu, etc.).

[0072] In one implementation, the physical parameters can be latent variables of the thin film structure. The physical parameter prediction sub-branch can predict the latent variables of each thin film layer based on the global feature vector. The physical mapping layer sub-branch can map the dielectric function of each thin film layer based on the latent variables. The dielectric function can be constrained to satisfy the Kramer-Kronig constraint based on the Lemer-Kronig consistency constraint loss.

[0073] In a specific example, the physical parameter prediction sub-branch can be based on an MLP (Multi-layer Perceptron) decoder to decode the global feature vector and obtain the latent variables of the thin film structure.

[0074] The advantage of using latent variable mapping to obtain the dielectric function of thin film structures is that:

[0075] - Strong expressive power: It can describe the optical properties of any material;

[0076] - High flexibility: adaptable to special materials and non-standard structures;

[0077] - No prior knowledge required: It does not depend on material-specific parameter ranges.

[0078] Using latent variable mapping to obtain the dielectric function of thin film structures is more suitable for material scenarios with complex optical properties that are difficult to describe with a simple oscillator model. Examples include novel two-dimensional materials (MoS2, WS2, hBN, etc.); complex multilayer structures; and non-standard material combinations.

[0079] In one implementation, the physical parameters can be the oscillator parameters and latent variables of the thin-film structure. The physical parameter prediction sub-branch can predict the oscillator parameters and latent variables of each thin-film layer based on the global feature vector. The physical mapping layer sub-branch can combine the oscillator parameters and latent variables of the thin-film structure to map and obtain the dielectric function of each thin-film layer. The dielectric function can be constrained to meet the Kramer-Kronig constraint based on the Lemer-Kronig consistency constraint loss.

[0080] In a specific example, the dielectric function of each thin film structure can be obtained by mapping the oscillator parameters and latent variables of the thin film structure according to the following formula (4):

[0081] ε = α×ε 振子 + (1-α)×ε 潜变量 (4)

[0082] Where ε is the dielectric function of the thin film structure, ε 振子 Let ε be the dielectric function corresponding to the oscillator parameters. 潜变量 Let be the dielectric function corresponding to the latent variable, and α be the weighting coefficient of the dielectric function corresponding to the oscillator parameters. α is learnable, and its initial value can be set to 0.5. Those skilled in the art can also set the initial value of α according to actual needs.

[0083] The advantage of obtaining the dielectric function of a thin film structure by combining oscillator parameters and latent variable mapping is that:

[0084] - Balancing interpretability and flexibility;

[0085] - Suitable for complex materials, the oscillator section provides the physical basis, and the latent variable section supplements the details.

[0086] The dielectric function of a thin film structure is obtained by combining oscillator parameters and latent variable mapping. It is suitable for scenarios requiring high-precision inversion; material scenarios with properties between standard and non-standard ones; and material scenarios that need to balance physical interpretability and inversion accuracy. Its advantages are high versatility, but the computational cost is slightly higher.

[0087] It should be noted that, regardless of whether oscillator parameters, latent variables, or a combination of the two are used, the core purpose is to construct a dielectric function characterization that satisfies the Cramer-Kronig consistency constraint, thereby supporting the spectral inversion of multilayer thin film structures; the above methods are different implementations or equivalent schemes under the same inventive concept.

[0088] In one embodiment, 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, and output the structural parameter prediction results; wherein, the structural parameter prediction results include the thickness of each thin film layer. In another embodiment, the structural parameter prediction results also include the gradient of the roughness and / or dielectric function of each thin film layer.

[0089] In one embodiment, the ellipticity prediction branch may further include an optical forward calculation layer. The optical forward calculation layer calculates the ellipticity prediction results based on the dielectric function and structural parameter prediction results of each thin film structure.

[0090] In this embodiment, the optical forward calculation layer can calculate the ellipticity parameter prediction result based on the dielectric function and structural parameter prediction results of each thin film structure.

[0091] In some specific examples, the optical forward computation layer can obtain ellipsoid 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. The dielectric function and structural parameter prediction results of the thin film structure can be input into the optical forward computation layer to obtain the ellipsoid parameter prediction results based on the parallel transfer matrix method. The structural parameter prediction results include the thickness, roughness, and gradient of the dielectric function of each thin film layer. The input data for the optical forward computation layer consists of the dielectric function, thickness, and incident angle AOI of the thin film structure; the output data is the ellipsoid parameter prediction results (including Ψ, Δ, and T).

[0092] In one implementation, the ellipticity parameter prediction branch can be implemented based on a neural network, which can directly predict the ellipticity parameters based on global feature vectors. Applying this ellipticity parameter prediction branch, which directly predicts ellipticity parameters based on global feature vectors, bypasses 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. Furthermore, it can achieve high-precision prediction of ellipticity parameters through a large amount of training data and an end-to-end optimization process.

[0093] In one embodiment, the generation of spectral synthesis data may include: randomly sampling material and structural parameters of the thin film structure based on a preset sampling range; obtaining ellipticity parameters of the thin film structure based on a physical model and the randomly sampled material and structural parameters, thereby obtaining spectral synthesis data, and then generating a spectral synthesis dataset for pre-training the spectral inversion model.

[0094] In a specific example, a spectral synthesis dataset can contain 50,000 to 100,000 spectral synthesis data samples. The sampling range of the spectral synthesis data can include:

[0095] - Oscillator parameters: A∈[1,6], E0∈[2,5]eV, Γ∈[0.1,0.6]eV, Eg∈[0,2]eV;

[0096] - Thickness: t_i∈[10, 200] nm;

[0097] - Angle of incidence: AOI∈[45°, 75°];

[0098] - Wavelength: λ = 300–1000 nm.

[0099] In one implementation, the ellipticity parameters can be encoded, and the encoded ellipticity parameters can be input into the spectral inversion model. Specifically, the input data of the spectral inversion model may include: cos2Ψ: [batch, Nλ] (representing the cosine of the ellipticity angle); sin2Ψ·cosΔ: [batch, Nλ] (representing the phase difference cosine term); sin2Ψ·sinΔ: [batch, Nλ] (representing the phase difference sine term); T: [batch, Nλ] (transmittance).

[0100] Step S102: Obtain a sample dataset containing multiple spectral sample data in a preset migration scenario; wherein, the spectral sample data includes the ellipticity parameters of the multilayer thin film structure measured in the migration scenario.

[0101] In this embodiment, a spectral sample dataset comprising multiple spectral sample data points within a preset migration scenario can be obtained. The spectral sample data in the dataset may include the ellipticity parameters of the multilayer thin film structure within the migration scenario. The ellipticity parameters of each multilayer thin film structure within this migration scenario can be measured using instruments such as an ellipsometry. That is, a small number (e.g., less than 200 samples) of actually measured spectral sample data from the migration scenario can be used to construct the sample dataset.

[0102] In one embodiment, the migration scenario can include migration of the thin film structure to a new material system (e.g., from SiO2 / TiO2 to MoS2 / ITO); migration to a new spectral band (e.g., from visible light to near-infrared light); migration of a new AOI or instrument (e.g., AOI=65°→AOI=70°); migration to a new substrate (e.g., glass→silicon-based); and changes in multilayer thin film structure (N=2→N=4, where N is the number of layers in the thin film structure). The migration scenario can also include the joint migration of multiple migration scenarios mentioned above.

[0103] Step S103: Adjust the pre-trained spectral inversion model according to the sample dataset to obtain a spectral inversion model that adapts to the transfer scenario; wherein, when adjusting the spectral inversion model, the model parameters of the backbone network are frozen.

[0104] In this embodiment, the pre-trained spectral inversion model can be adjusted based on a sample dataset to obtain an inversion model adapted to the transfer scenario. During the adjustment process, the model parameters of the backbone network of the spectral inversion model can be frozen, and only the model parameters of one or more of the physical parameter prediction sub-branch, physical mapping layer sub-branch, and structural parameter prediction branch can be adjusted. This allows the backbone network to transfer the general knowledge learned during pre-training, i.e., the spectral feature extraction strategy, to the new task in this transfer scenario, and to fine-tune the model parameters in the physical parameter prediction sub-branch, physical mapping layer sub-branch, and structural parameter prediction branch. This enables each branch network to adapt to new materials and environments during model adjustment, while reducing the risk of overfitting.

[0105] In other words, a small amount of real measurement data from the transfer scenario can be used to adjust the spectral inversion model, enabling it to quickly adapt to the transfer scenario. Since the backbone network has already learned general spectral representation capabilities through the pre-training phase, freezing the backbone network is to maintain these learned general spectral representation capabilities in the transfer scenario, which is a way to improve transfer stability and sample utilization efficiency.

[0106] In one implementation, the ellipticity parameters in the sample dataset can be encoded and then input into the spectral inversion model to adjust the model. The method for encoding the ellipticity parameters is similar to that in the model pre-training stage, and will not be elaborated further for simplicity.

[0107] In one implementation, the total loss during the pre-training and model tuning of the spectral inversion model may include fitting error loss, Cramer-Kronig consistency constraint loss, and search range constraint loss.

[0108] In this embodiment, the fitting error loss can be determined based on the ellipticity parameter prediction results output by the spectral inversion model and the corresponding ellipticity parameters in the synthesized spectral data. The Cramer-Kronig consistency constraint loss can be used to represent the Cramer-Kronig relation constraint on the dielectric function of each thin film structure predicted by the spectral inversion model. The search range constraint loss can be used to constrain the search range of the physical parameter prediction results and the structural parameter prediction results.

[0109] In one implementation, the fitting error loss can be obtained based on the RMSE (root mean square error) of Ψ, Δ, and T. For example, the fitting error loss can be obtained by weighting the RMSEs of Ψ, Δ, and T.

[0110] In a specific example, the weights of Ψ and Δ in the fitting error loss can be set to 1.0, and the weight of T can be set to 0.5. That is, Ψ and Δ are set as the primary inversion targets, and T is the secondary inversion target.

[0111] In one implementation, the dispersive spectrum ε1 of the thin film structure can be calculated first based on the Kramer-Kronig (KK) relation. Then, the dispersive spectrum predicted by the spectral inversion model is compared with the dispersive spectrum obtained based on the KK relation to obtain the Kramer-Kronig consistency constraint loss. Specifically, when the physical parameter prediction sub-branch predicts the physical parameters of the thin film structure based on the global eigenvector, including only the oscillator parameters of each thin film layer (i.e., using an analytical oscillator model to characterize the physical model of each thin film layer), the calculated dielectric function is already forced to satisfy the Kramer-Kronig constraint. In this case, there is no need to add a Kramer-Kronig consistency constraint loss term to the loss function.

[0112] In a specific example, the Cramer-Kronig consistency constraint loss can be obtained based on the following formula (5):

[0113] L_kk = mean((ε1_pred -ε1_kk)²) (5)

[0114] Where L_kk is the Kramer-Kronig consistency constraint loss, ε1_pred is the predicted dispersion spectrum, ε1_kk is the dispersion spectrum calculated based on the KK relation, and mean() represents the averaging operation.

[0115] In one implementation, the search range constraint loss can constrain the search range of the physical parameter prediction results and the structural parameter prediction results. The constraint objects of the search range constraint loss can include the thickness of the thin film structure, oscillator parameters, and latent variables. This ensures the physical rationality of the physical parameter prediction results and the structural parameter prediction results.

[0116] In some specific examples, the search range for the thickness t_i of the thin film structure can be within ±20 nm of the initial thickness prediction result; the purpose of this setting is to avoid unreasonable thickness predictions caused by overfitting in migration scenarios; the implementation method is to penalize prediction values ​​that exceed the ±20 nm range of the initial thickness prediction result.

[0117] The search range for the oscillator parameters can be:

[0118] - A>0 (Oscillator strength must be positive)

[0119] - Γ>0 (damping coefficient must be positive)

[0120] - Eg ≥ 0 (band gap energy must be non-negative)

[0121] - The implementation method is to use the clamp function in the physical mapping layer to ensure non-negativity.

[0122] The search range of the latent variable z can be z ∈ [-3, 3]; the purpose of setting the search range of the latent variable is to limit the search space of the latent variable and improve stability; the implementation method is to penalize the latent variable that exceeds the range in the loss function.

[0123] In one implementation, the fitting error loss has the first weight in the total loss; the Cramer-Kronig consistency constraint loss has the second weight; and the search range constraint loss has the third weight. The first weight is greater than the second weight, and the second weight is greater than the third weight. That is, in the loss function, the fitting error loss is the dominant factor to ensure the accuracy of spectral fitting, while the Cramer-Kronig consistency constraint loss and the search range constraint loss are auxiliary factors to ensure the physical consistency of the prediction results and the rationality of the parameters.

[0124] In one implementation, the first weight can be 1, the second weight can be in the range of 0.1-0.3, and the third weight can be in the range of 0.01-0.1.

[0125] The principle for setting the second weight is to avoid it being too large or too small:

[0126] - Too small (<0.1): Cramer-Kronig consistency constraints are insufficient and may produce non-physical results;

[0127] - Too large (>0.3): Over-constraint, which may affect the fitting accuracy.

[0128] The principle for setting the third weight is to avoid it being too large or too small:

[0129] - Too small (<0.01): Insufficient range constraints, may lead to overfitting;

[0130] - Too large (>0.1): Over-constraint, limiting the model's expressive power.

[0131] Specifically, the loss function can be obtained according to the following formula (6):

[0132] L = RMSE(Ψ,Δ,T) + λ_KK×L_KK + λ_range×L_range (6)

[0133] Where L is the loss function, RMSE(Ψ,Δ,T) is the fitting error loss, λ_KK is the second weight, L_KK is the Kramer-Kronig consistency constraint loss, λ_rang is the third weight, and L_range is the search range constraint loss.

[0134] In a specific example, the second weight can be 0.2 and the third weight can be 0.05.

[0135] In one implementation, after obtaining a spectral inversion model adapted to the migration scenario, spectral inversion inference can be performed on unlabeled spectral sample data to obtain inference results; if the prediction error of the inference results is greater than a preset error threshold, the corresponding unlabeled spectral sample data is added to the sample dataset; and the updated sample dataset is used to continue training the spectral inversion model.

[0136] In this embodiment, after obtaining a spectral inversion model adapted to the migration scenario, predictions can be made on unlabeled spectral sample data based on the spectral inversion model to obtain inference results. Based on the inference results, the prediction error is determined. If the prediction error exceeds a preset error threshold, the unlabeled spectral sample data can be added to the sample dataset, and the updated sample dataset can be used to adjust the spectral inversion model, thereby achieving active learning of the spectral inversion model. During the active learning process of the spectral inversion model, the model parameters of the backbone network can be frozen, and only some or all of the model parameters of the physical parameter prediction sub-branch, the physical mapping layer sub-branch, and the structural parameter prediction branch can be adjusted. Those skilled in the art can set the error threshold value according to the needs of the actual application.

[0137] Specifically, when predicting unlabeled spectral sample data, uncertainty assessment can be performed on the prediction results. For example, prediction error can be calculated as an uncertainty index. Then, the most representative sample data, i.e., the sample data with the largest prediction error, can be selected and added to the sample dataset. Furthermore, after fine-tuning the model parameters of the physical parameter prediction sub-branch, the physical mapping layer sub-branch, and the structural parameter prediction branch based on the updated sample dataset, the uncertainty of the sample data with the largest prediction error can be reassessed.

[0138] It should be noted that the active learning process of the above-mentioned spectral inversion model is an adaptive optimization process in the application stage under the transfer scenario. It is mainly used to gradually improve the adaptability of the spectral inversion model to specific transfer scenarios, rather than a process of retraining the complete model.

[0139] In one implementation, see Appendix Figure 2 ,like Figure 2 As shown, this can be done based on the following steps S301 to S303:

[0140] Step S301: Pre-training of the spectral inversion model.

[0141] Step S302: Model adjustment of the spectral inversion model based on the migration scenario.

[0142] Step S303: Output the spectral inversion model adapted to the migration scenario.

[0143] Based on the methods described in steps S101 to S103 above, this embodiment of the application pre-trains a spectral inversion model based on a spectral synthesis dataset. The spectral synthesis dataset includes multiple spectral synthesis data, which include ellipticity parameters of a multilayer thin film structure. The spectral inversion model includes a backbone network, an ellipticity parameter prediction branch, and a structural parameter prediction branch. The backbone network is used to extract the spectral features of the multilayer thin film structure and obtain a global feature vector. A sample dataset containing multiple spectral sample data is obtained within a preset migration scenario. The spectral sample data includes the ellipticity parameters of the multilayer thin film structure measured in the migration scenario. Based on the sample dataset, the pre-trained spectral inversion model is adjusted to obtain a spectral inversion model adapted to the migration scenario. When adjusting the spectral inversion model, the model parameters of the backbone network are frozen. Through the above configuration, this embodiment of the application can realize the pre-training process of the spectral inversion model based on a spectral synthesis dataset. Furthermore, for migration scenarios, the pre-trained spectral inversion model is adjusted based on the sample dataset within the migration scenario, enabling the spectral inversion model to adapt to the migration scenario based solely on the sample dataset. Meanwhile, because the model parameters of the backbone network of the spectral inversion model are frozen during the model adjustment process to adapt to the migration scenario, only other model parameters besides the backbone network need to be adjusted during the model adjustment process. This can effectively improve the speed of model adjustment, thereby achieving efficient and high-precision model adjustment of the spectral inversion model to adapt to the migration scenario. It can also effectively reduce the data dependence of the spectral inversion model when adapting to the migration scenario and improve the adaptability and inference efficiency of the spectral inversion model for different migration scenarios.

[0144] Furthermore, this application also provides a method for spectral inversion of multilayer thin films.

[0145] See appendix Figure 3 , 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:

[0146] Step S201: Input the ellipsoid parameters obtained by the ellipsometer measurement of the multilayer thin film structure under the preset migration scenario into the spectral inversion model adapted to the migration scenario, perform spectral inversion inference, and obtain the material and structural properties of each layer of the multilayer thin film structure; wherein, the spectral inversion model adapted to the migration scenario is trained based on the multilayer thin film spectral inversion model training method based on small sample migration described in the above embodiment of the multilayer thin film spectral inversion model training method based on small sample migration.

[0147] In this embodiment, the multilayer thin film spectral inversion model training method based on small sample migration described in the above embodiment can be used to train the multilayer thin film spectral inversion model. Based on the trained spectral inversion model, the ellipticity parameters obtained by the ellipticity meter measurement of the multilayer thin film structure in the migration scenario are inverted and inferred to obtain the material and structural properties of each layer of the multilayer thin film structure.

[0148] In one implementation, when performing inverse reasoning, the spectral inversion model freezes the backbone network and only fine-tunes some and all model parameters in the physical parameter prediction sub-branch, the physical mapping layer sub-branch, and the structural parameter prediction branch.

[0149] In one implementation, in response to the detection of performance degradation of the spectral inversion model, changes in measurement environment parameters, etc., spectral measurement data of multilayer thin film structures under sample migration scenarios can be collected. The spectral measurement data includes the ellipticity parameters of the multilayer thin film structures. Based on the spectral measurement data, the spectral inversion model is adjusted. When adjusting the spectral inversion model, the model parameters of the backbone network of the spectral inversion model are frozen, and only some or all of the model parameters of the structural parameter prediction branch and the ellipticity parameter prediction branch of the spectral inversion model are adjusted.

[0150] In this embodiment, if a performance degradation of the spectral inversion model or changes in measurement environment parameters are detected, spectral measurement data of the multilayer thin film structure under sample migration scenarios can be collected to incrementally train the spectral inversion model. For example, if instrument optical path drift is detected, requiring long-term online measurement correction, or if there are batch differences in the material manufacturing process, incremental training of the spectral inversion model can be performed. During the incremental training of the spectral inversion model, the model parameters of the backbone network can be frozen, and only some or all of the model parameters of the physical parameter prediction sub-branch, the physical mapping layer sub-branch, and the structural parameter prediction branch can be adjusted.

[0151] The multilayer thin film spectral inversion model training method and multilayer thin film spectral inversion method based on small-sample transfer provided in this invention can pre-train the spectral inversion model based on a spectral synthesis dataset. Furthermore, for transfer scenarios, the pre-trained spectral inversion model can be adjusted based on a sample dataset specific to the transfer scenario. This allows the spectral inversion model to adapt to transfer scenarios using only the sample dataset. This improves the adaptability and inference efficiency of the spectral inversion model for different transfer scenarios (such as different spectral bands, AOI, substrate, and number of layers), facilitating rapid deployment on production lines, instruments, or new projects.

[0152] 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.

[0153] 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.

[0154] Another aspect of this application provides a computer-readable storage medium.

[0155] 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 training method of the multilayer thin film spectral inversion model based on few-sample transfer based in the above-described method embodiments. This program can be loaded and run by a processor to implement the above-described training method of the multilayer thin film spectral inversion model based on few-sample transfer. 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.

[0156] 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.

[0157] Another aspect of this application provides an electronic device.

[0158] 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 of the embodiments of the above-described training method for multilayer thin film spectral inversion model based on small sample migration.

[0159] 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 of the embodiments of the above-described multilayer thin film spectral inversion method.

[0160] 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.

[0161] 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 based on small-sample transfer, characterized in that, The method includes: The spectral inversion model is pre-trained based on a spectral synthesis dataset. The spectral synthesis dataset includes multiple spectral synthesis data points, each containing ellipticity parameters of a multilayer thin film structure. The spectral inversion model comprises a backbone network, an ellipticity parameter prediction branch, and a structural parameter prediction branch. The backbone network extracts the spectral features of the multilayer thin film structure and obtains a global feature vector. The structural parameter prediction branch outputs the structural parameter prediction results, which include at least the thickness of each thin film layer. Obtain a sample dataset containing multiple spectral sample data in a preset migration scenario; wherein, the spectral sample data includes ellipticity parameters of a multilayer thin film structure measured in the migration scenario; Based on the sample dataset, the pre-trained spectral inversion model is adjusted to obtain a spectral inversion model adapted to the migration scenario; wherein, when adjusting the spectral inversion model, the model parameters of the backbone network are frozen; The elliptic parameter prediction branch includes a physical parameter prediction sub-branch and a physical mapping layer sub-branch; The physical parameter prediction sub-branch is used to predict the physical parameters of each thin film structure based on the global feature vector, and output the physical parameter prediction results; The physical mapping sub-branch is used to map the physical parameter prediction results to the dielectric function of each thin film structure; wherein the dielectric function is constrained by the Cramer-Kronig relation.

2. The training method for multilayer thin film spectral inversion model based on small sample transfer as described in 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, and output the structural parameter prediction results.

3. The training method for multilayer thin film spectral inversion model based on small sample migration as described in 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 multilayer thin film spectral inversion model based on small sample transfer as described in claim 2, characterized in that, The adjustment of the pre-trained spectral inversion model includes: Adjust the model parameters of the predicted sub-branch based on the physical parameters, and / or, Adjust the model parameters of the physical mapping layer sub-branch, and / or, Adjust the model parameters of the predicted branch of the structural parameters.

5. The training method for multilayer thin film spectral inversion model based on small sample transfer as described in claim 2, characterized in that, The total loss of the pre-training and model adjustment process of the spectral inversion model includes the fitting error loss; the fitting error loss is determined based on the ellipticity parameter prediction results output by the spectral inversion model and the ellipticity parameters in the corresponding spectral composite data.

6. The training method for multilayer thin film spectral inversion model based on small sample transfer as described in claim 5, characterized in that, The total loss of the pre-training and model adjustment process of the spectral inversion model also includes the Cramer-Kronig consistency constraint loss; the Cramer-Kronig consistency constraint loss is used to represent the Cramer-Kronig relation constraint of the dielectric function of each thin film structure predicted by the spectral inversion model.

7. The training method for multilayer thin film spectral inversion model based on small sample transfer as described in claim 6, characterized in that, The total loss of the pre-training and model adjustment process of the spectral inversion model also includes the search range constraint loss; the search range constraint loss is used to constrain the search range of the physical parameter prediction results and / or the structural parameter prediction results.

8. The training method for multilayer thin film spectral inversion model based on small sample transfer as described in claim 7, characterized in that, The weight of the fitting error loss in the total loss is the first weight; The weight of the Kramer-Kronig consistency constraint loss in the total loss is the second weight. The search range constraint loss has the third weight in the total loss; The first weight is greater than the second weight, and the second weight is greater than the third weight.

9. The training method for a multilayer thin film spectral inversion model based on small sample transfer as described in claim 1, characterized in that, The physical parameters include the oscillator parameters of each thin film structure and / or the latent variables of each thin film structure.

10. The training method for a multilayer thin film spectral inversion model based on small sample transfer as described in claim 1, characterized in that, After obtaining the spectral inversion model adapted to the migration scenario, the method further includes: Perform spectral inversion inference on unlabeled spectral sample data to obtain the inference results; If the prediction error of the inference result is greater than the preset error threshold, the corresponding unlabeled spectral sample data is added to the sample dataset; and the updated sample dataset is used to continue training the spectral inversion model.

11. A method for spectral inversion of multilayer thin films, characterized in that, The method includes: The ellipticity parameters obtained by ellipticity measurement of the multilayer thin film structure under the preset migration scenario are input into the spectral inversion model adapted to the migration scenario to perform spectral inversion inference and obtain the material and structural properties of each layer of the multilayer thin film structure. The spectral inversion model adapted to the migration scenario is obtained by training the multilayer thin film spectral inversion model training method based on small sample migration as described in any one of claims 1 to 10.

12. The multilayer thin film spectral inversion method according to claim 11, characterized in that, The method further includes: In response to the detection of a performance degradation of the spectral inversion model and / or a change in measurement environment parameters, spectral measurement data of the multilayer thin film structure under the small sample migration scenario are collected, and the spectral measurement data includes the ellipticity parameters of the multilayer thin film structure; Based on the spectral measurement data, the spectral inversion model is adjusted; wherein, when adjusting the spectral inversion model, the model parameters of the backbone network of the spectral inversion model are frozen, and only the model parameters of the structural parameter prediction branch and / or elliptic parameter prediction branch of the spectral inversion model are adjusted.

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 that, when executed by the at least one processor, implements the multilayer thin film spectral inversion model training method based on small sample migration as described in any one of claims 1 to 10 and / or the multilayer thin film spectral inversion method as described in any one of claims 11 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 based on small sample migration as described in any one of claims 1 to 10 and / or the multilayer thin film spectral inversion method as described in any one of claims 11 to 12.

Images