Method for constructing dispersion models of non-standard complex materials using machine learning
Through machine learning methods, the principal component analysis and K-mean clustering algorithm are used to automatically analyze the primitives of complex materials, solving the problem of difficult simulation of dispersion performance parameters of complex materials in the existing technology, and achieving accurate detection of complex materials.
Patent Information
- Application Number
- CN202210252477.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-15
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-03-15
AI Technical Summary
In integrated circuit production, dispersion performance parameters of complex materials and doped mixed materials are difficult to effectively simulate through existing effective dielectric approximation models, especially when the material components are complex or unknown, resulting in difficulty in quality detection.
Using machine learning method, the primitives of complex materials are automatically analyzed through principal component analysis and K-mean clustering algorithm, and used as constituent elements of the effective medium approximation model to fit dispersion performance parameters.
Without artificially studying the components and optical properties of complex materials, it can automatically simulate the dispersion performance parameters of complex materials, improving the accuracy of quality detection of complex materials in integrated circuits.
Smart Images

Figure CN114898824B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a semiconductor production measurement technology, and in particular to a method for constructing a non-standard complex material dispersion model using machine learning. Background Art
[0002] In the field of integrated circuit production, ellipsometers are usually used to measure the thickness, key dimensions and other parameters of film materials for quality inspection in the production process of chip products. Ellipsometers are a non-destructive detection method based on optical principles, and their reliability has been verified by a lot of theory and practice. There are often some unconventional materials and mixed materials of multiple materials in integrated circuits, and their exact dispersion performance parameters cannot be obtained. For these materials, the effective medium approximation (EMA, Effective Medium Approximation) model is generally used to simulate the dispersion performance parameters of these new materials using the dispersion performance parameters of known materials. However, when the components of the material itself are relatively complex: such as organic materials, composite doped materials, etc.; or when it is impossible to write a clear material component, it is difficult to find a suitable physical model to fit the dispersion performance parameters of these new materials. Therefore, it is impossible to effectively use the EMA model for simulation analysis. Therefore, it has become a huge challenge facing the current integrated circuit production and testing links.
[0003] The effective medium approximation model is applied to mixed medium systems composed of two or more different components. When fitting, the effective medium model is usually used for rough surface and transition layer analysis of thin films. For composite materials composed of two materials, the unified formula for the dielectric constant is:
[0004]
[0005] Where ε is the effective dielectric constant; ε 1 and ε 2 are the dielectric constants of the two mixed media respectively; ε c is the dielectric constant of the medium containing the dopant; p is the volume percentage of material 2; L is the effective polarization coefficient of material 2, usually L = 1 / 3. The EMA model requires the dielectric constant ε of the mixed N materials. When the components of the material itself are complex or it is impossible to write out the components for analysis, the dielectric constants of these new materials are unknown, and it is difficult to find a suitable physical model to fit them. Summary of the invention
[0006] In order to solve the problem that the components of new materials are relatively complex and unclear and cannot be fitted by existing models, a method of using machine learning to construct a dispersion model of non-standard complex materials is proposed. The machine learning method is used to automatically analyze the "primitives" that make up the complex materials, and directly use them as the components of the EMA model to fit the dispersion performance parameters of these complex materials.
[0007] The technical solution of the present invention is: a method for constructing a dispersion model of non-standard complex materials using machine learning, which specifically includes the following steps:
[0008] 1) Data collection of unknown samples: According to the structure and unevenness of the unknown samples, the measurement range and the location and number of the sample points are limited, and all the sample points are measured to obtain the ellipsometric spectra or Fourier coefficients corresponding to all the sample points;
[0009] 2) Use the principal component analysis algorithm to compress the dimension of the data of each sampling point to obtain compressed dimensional data;
[0010] 3) Use K-means clustering algorithm to perform clustering calculation on the compressed dimension data of each sample point in step 2), and the k obtained by calculation is 1 The k cluster centers are used as the basic elements of the unknown samples to be tested. 1 The sample points corresponding to the cluster centers are used as the primitives of the EMA model. The EMA model is used to obtain the dispersion performance parameters of the unknown sample material to be tested, which are used to describe the basic elements of the unknown sample material to be tested.
[0011] Furthermore, all the sample measurement points in step 1) are sample measurement points at different positions on a single unknown sample to be measured, or sample measurement points at different positions on multiple samples.
[0012] Furthermore, the measurement in step 1) can be replaced by model parameters obtained by fitting each sample point with a certain material model.
[0013] Furthermore, the material model is an optical response model of the material, and the optical response model includes but is not limited to a Cauchy model, a Lorentz model, and a simple harmonic oscillator model.
[0014] Furthermore, the dimensional data after compression in step 2) is 2-dimensional or 3-dimensional data that can represent the original data.
[0015] Furthermore, the specific implementation steps of step 3) are as follows: first, randomly select k cluster centers; then, by calculating the Euclidean distance from each compressed sample point data in step 2) to each cluster center, the distance from each sample point data to each cluster center is compared in turn, and the sample point data is assigned to the cluster with the closest cluster center, and finally k cluster centers are obtained after clustering. 1 clusters; calculate k 1 The center of each cluster is called the new k 1 Cluster centers; repeat the above process until the final result converges; finally, the k 1 The k cluster centers are used as the basic elements of the unknown samples to be tested. 1The sample points corresponding to the cluster centers are used as primitives to form the EMA model. The EMA model is used to obtain the dispersion performance parameters of the unknown sample to be tested, which are used to describe the basic elements of the unknown sample to be tested.
[0016] The beneficial effect of the present invention is that the present invention uses machine learning to construct a method for a dispersion model of non-standard complex materials, and utilizes the principal component analysis (PCA) algorithm and the K-means clustering algorithm in the machine learning method, without the need to manually study the components and optical properties of the complex materials, to simulate the dispersion performance parameters of the complex materials in the integrated circuit. The method can automatically analyze the "element" that constitutes the complex material, and directly use it as the constituent element of the EMA model to fit the model. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 This is a schematic diagram of the distribution of each measurement point compressed into 2 dimensions according to the present invention;
[0018] Figure 2 It is a refractive index n-wavelength curve diagram of each sample point of the present invention;
[0019] Figure 3 It is the absorptivity k-wavelength curve diagram of each sample point of the present invention. DETAILED DESCRIPTION
[0020] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.
[0021] The method of constructing a non-standard complex material dispersion model using machine learning in the present invention specifically includes the following steps:
[0022] 1. Data preparation:
[0023] The traditional modeling method is to know the basic information of the sample to be tested, such as the chemical composition of the material and the applicable physical model, and then adjust the model parameters to fit the spectrum measured by the ellipsometer (a spectral measurement that uses the property of linear polarized light to convert into elliptically polarized light after reflection from the sample to obtain the optical constants of the sample), so as to obtain the parameters of the model through optimization methods. However, for non-standard complex materials, it is impossible to know their chemical composition and determine the physical model in advance. It is also impossible to find suitable primitives using the EMA model.
[0024] By measuring a series of sample points, the ellipsometric spectrum or Fourier coefficient corresponding to the sample point can be obtained. This series of measurement points can be sample points at different positions on a single sample or at different positions on multiple samples. The measurement range and quantity need to ensure that the sample unevenness is included.
[0025] The initial data used in the method of the present invention can be the original spectrum measured by the ellipsometer, or the Fourier coefficients obtained by spectrum calculation, or the model parameters obtained by fitting each sample point with a certain material model, for example: including but not limited to the optical response models of various materials such as the Cauchy model / Lorentz model / simple harmonic oscillator model.
[0026] 2. Data dimension compression:
[0027] The initial data corresponding to each sample point can range from a few to thousands. If a simple harmonic oscillator model is used, there will generally be at least six, and generally there will be more than a dozen or dozens of parameters. If the original spectrum or Fourier coefficients are used as the initial data, the data dimensions corresponding to each measurement point are several thousand.
[0028] The present invention uses a principal component analysis algorithm to perform dimensional compression on the data of each sample point, which can usually be compressed to 3 dimensions or 2 dimensions.
[0029] The principal component analysis method converts a set of variables that may be correlated into a set of linearly uncorrelated variables through orthogonal transformation. This set of variables after conversion is called principal components. In other words, some linearly correlated dimensions are reduced, and only the most linearly uncorrelated sets of dimensions are retained.
[0030] According to experience, if the data involved in the present invention retains 3 dimensions, generally less than 2% of the information is lost, and if 2 dimensions are retained, generally less than 8% of the information is lost. It can be considered that the compressed 2D or 3D data can represent the original data.
[0031] 3. K-means clustering algorithm:
[0032] The K-means clustering algorithm is an unsupervised algorithm that does not require user intervention and can automatically classify similar samples into one category. The goal of the algorithm is to cluster n objects into specified k clusters based on the similarity between objects, and each object belongs to and only belongs to one cluster with the smallest distance to the cluster center.
[0033] First, k cluster centers need to be initialized. The initial cluster centers in the clustering algorithm are randomly selected. The algorithm will automatically iterate to the real k centers. Then, the Euclidean distance from each object (the compressed test point data in step 2) to each cluster center is calculated, and the distance from each object to each cluster center is compared in turn. The object is assigned to the cluster closest to the cluster center. After clustering, k cluster centers are finally obtained. 1 Then calculate k 1 The center of each cluster is called the new k 1 Cluster centers. Repeat the above process and the final result can converge. The k 1 The k cluster centers are used as the basic elements of new complex materials. 1 The test points corresponding to the cluster centers are used as primitives to form the EMA model. The EMA model is used to obtain the dispersion performance parameters of the complex material to describe the basic elements of the complex material.
[0034] Embodiment: A set of ellipsometry spectral measurement data is compressed using principal component analysis. Figure 1 It is the distribution of data compressed to 2 dimensions. Through the K-means clustering algorithm, different data points are clustered into 3 different clusters. The center of the cluster is represented by a large black dot. The sample points corresponding to the centers of these three clusters can be used as primitives to form an EMA model, thereby automatically obtaining an EMA model that describes the sample to be measured. The primitives in the EMA model can be used to calculate the refractive index and absorptivity of the material.
[0035] Figure 2 , 3 The thin lines represent the refractive index n and absorptivity k curves of each sample point. The thick lines are the refractive index n and absorptivity k curves corresponding to the EMA model primitives generated by the method of the present invention. It can be seen that the refractive index (absorptivity) curves of the primitives calculated by the method of the present invention are three completely different curves. It is precisely because these three refractive index (absorptivity) curves are completely different that they can be used as basic elements to describe the complex material and are very suitable as primitives of the EMA model.
[0036] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.
Claims
1. A method for constructing a dispersion model of non-standard complex materials using machine learning, characterized in that: The specific steps include: 1) Data collection of unknown samples: According to the structure and unevenness of the unknown samples, the measurement range and the location and number of the sample points are limited, and all the sample points are measured to obtain the ellipsometric spectra or Fourier coefficients corresponding to all the sample points; 2) Use the principal component analysis algorithm to compress the dimension of the data of each sampling point to obtain compressed dimensional data; 3) Using the K-means clustering algorithm to perform clustering calculation on the compressed dimensional data of each sample point in step 2), the k1 cluster centers obtained by the calculation result are used as the basic elements of the unknown sample to be tested, and the sample points corresponding to the k1 cluster centers are used as the primitives that compose the EMA model. The EMA model is used to obtain the dispersion performance parameters of the unknown sample material to be tested, which are used to describe the basic elements of the unknown sample material to be tested.
2. The method for constructing a non-standard complex material dispersion model using machine learning according to claim 1, characterized in that: In the step 1), all the sample points are sample points at different positions on a single unknown sample to be tested, or sample points at different positions on multiple samples.
3. The method for constructing a non-standard complex material dispersion model using machine learning according to claim 1 or 2, characterized in that: The measurement in step 1) can be replaced by model parameters obtained by fitting a certain material model at each sample point.
4. The method for constructing a non-standard complex material dispersion model using machine learning according to claim 3, characterized in that: The material model is an optical response model of the material, and the optical response model includes but is not limited to a Cauchy model, a Lorentz model, and a simple harmonic oscillator model.
5. The method for constructing a non-standard complex material dispersion model using machine learning according to claim 1, characterized in that: The dimensional data after compression in step 2) is 2-dimensional or 3-dimensional data that can represent the original data.
6. The method for constructing a non-standard complex material dispersion model using machine learning according to claim 1, characterized in that: The specific implementation steps of step 3) are as follows: first, randomly select k cluster centers; then, by calculating the Euclidean distance from each compressed sample point data in step 2) to each cluster center, compare the distance from each sample point data to each cluster center in turn, and distribute the sample point data to the cluster closest to the cluster center, and finally obtain k1 clusters after clustering; calculate the center of each cluster of the k1 clusters, which is called the new k1 cluster centers; repeat the above process until the result converges; finally, the k1 cluster centers obtained by the calculation result are used as the basic elements of the unknown sample to be tested, and the sample points corresponding to the k1 cluster centers are used as the primitives of the EMA model, and the dispersion performance parameters of the unknown sample to be tested are obtained by using the EMA model to describe the basic elements of the unknown sample to be tested.
