Method and system for component analysis of spectral data

Abstract

Methods and systems for component analysis of spectral data. Emissions from a sample are acquired in the form of spectral data in response to irradiation by a charged particle beam. The spectral data are decomposed into abundance and spectral components by a machine learning estimator based on characteristics of a detector. An image showing compositional information of the sample is generated based on the abundance and the spectral components.

Description

Technical Field

[0001] This description generally relates to methods and systems for processing spectral data, and more specifically, to extracting sample composition information based on spectral data. Background Technology

[0002] Charged particle microscopy is a well-known and increasingly important technique for imaging microscopic objects, particularly those in the form of electron microscopy. The various types of emission from a sample in response to charged particle irradiation can provide information about the sample's structure and composition. For example, based on the energy spectrum of X-ray emission, energy-dispersive X-ray spectroscopy (EDS or EDX) can be used for elemental analysis or chemical characterization. Since different elements or chemicals have their characteristic energy spectra, the sample composition can be estimated by comparing the EDS spectrum with the characteristic energy spectrum.

[0003] EDS spectra can be generated by X-ray emission from multiple elements or chemicals. One method for decomposing EDS spectra is disclosed in US Patent 6584413B1 to Keenan et al., which uses multivariate spectral analysis to decompose spectral data into a concentration-intensity matrix and a spectral shape matrix. However, the inventors recognized that multivariate spectral analysis requires a full-spectrum image of all pixels, which can lead to high computational intensity and long data-to-image time. Summary of the Invention

[0004] In one embodiment, a method includes irradiating a sample with a beam of charged particles; acquiring a spectrum by detecting emission from the sample in response to the irradiation using a detector; and decomposing the acquired spectrum into abundance and multiple spectral components using a machine learning estimator based on characteristics of the detector. In this manner, an image representing the composition of the sample can be formed based on the output of the machine learning estimator and displayed during data acquisition. Further, instead of processing the full-spectrum image, each spectrum is decomposed as the spectral machine learning estimator receives it, reducing computational intensity.

[0005] It should be understood that the above overview is provided to introduce some concepts further described in the detailed description in a simplified form. It is not intended to identify key or essential features of the claimed subject matter, the scope of which is uniquely defined by the claims following the detailed description. Furthermore, the claimed subject matter is not limited to embodiments that address any shortcomings mentioned above or in any part of this disclosure. Attached Figure Description

[0006] Figure 1 The illustration shows a charged particle microscope.

[0007] Figure 2AThe diagram illustrates the data flow used to analyze spectral data using a machine learning estimator.

[0008] Figure 2B The diagram shows Figure 2A The data flow within the machine learning estimator.

[0009] Figure 3 The method for analyzing spectral data is shown.

[0010] Figure 4 The diagram illustrates the data generated during the analysis of spectral data.

[0011] Figure 5A and 5B This is an example component image.

[0012] Figure 6A and 6B This is an example phase diagram.

[0013] Figure 7 It is a method used to acquire data and analyze it using a charged particle microscope.

[0014] Throughout the various views in the accompanying drawings, similar reference numerals refer to the corresponding parts. Detailed Implementation

[0015] The following description relates to systems and methods for analyzing spectral data, such as spectral data from energy-dispersive X-ray spectroscopy (EDS or EDX). EDS spectra can be used as follows: Figure 1 Charged particle microscopy, such as charged particle microscopy, is used to acquire emissions from the sample in response to irradiation of the sample location by a charged particle beam, such as an electron beam. The X-ray detector detects these emissions as an energy spectrum. The energy spectrum displays the number of counts (intensities) of emissions at different energies. To obtain compositional information at the sample location, a machine learning estimator decomposes the energy spectrum into abundance and multiple spectral components. Figure 4 As shown, abundance is a 1D array or vector. Abundance values ​​are indexed by components. Each abundance value corresponds to one component and one of multiple spectral components. Each component can correspond to an element, a mixture of elements, or a chemical phase. The number of spectral components is the same as the number of components. The machine learning estimator can be updated based on the decomposition results. For example, the machine learning estimator can be updated based on one or both of the spectral components and the quantized spectral components. The quantization method involves extracting the spectrum as input and returning the concentration of the chemical element identified in the spectrum and the corrected or quantized spectrum based on known spectra such as the theoretical spectrum of the chemical element. The second energy spectrum obtained by irradiating a second sample location with a charged particle beam can be decomposed by the updated machine learning estimator.

[0016] The machine learning estimator decomposes the received spectrum and outputs abundance and spectral components based on detector characteristics. The spectral components can be adjusted based on detector characteristics before being output from the machine learning estimator. In one instance, this adjustment is based on the energy-related characteristics of the detector. Due to noise in the amplification chain and the detector material, typical energy spectra acquired by the detector, such as EDS data, are smoother at higher energies compared to lower-energy spectra. In other words, the peaks in the lower-energy spectrum are narrower than those in the higher-energy spectrum. By adjusting these components, the spectral components output from the machine learning estimator can simulate the characteristics of a typical energy spectrum.

[0017] Machine learning estimators can be launched based on sample information and / or known spectral components. For example, known spectral components of the possible compositions of a sample can be used as initial spectral components. In another instance, a microscope operator can select one or more sample locations and acquire a dense spectrum at the selected sample location with the longest dwell time. The dense spectrum can be used to initialize the estimator. As more spectra are analyzed, the machine learning estimator output can become more accurate and can identify the composition of more sample locations.

[0018] The machine learning estimator output, i.e., the abundance at the sample location and the spectral components of each component, can be directly displayed, or further analyzed to extract more information. In one instance, one or more component maps are generated based on the abundance. The component maps can be displayed on a display device and updated as more spectra are analyzed. To reduce data acquisition time, the residence time of the charged particle beam at each sample location can be relatively short, resulting in a sparse spectrum. In a sparse spectrum, one or more low-amplitude spectral peaks may be missing due to insufficient signal. The missing peaks may exist in the dense spectrum obtained with longer residence times. In other words, the sparse spectrum can be considered an undersampled spectrum derived from the dense spectrum. In another instance, the abundance from multiple sample locations can be further clustered or grouped based on the similarity of the abundance distribution to obtain one or more phase maps, such as... Figures 5A-5B as well as Figures 6A-6B As shown. Furthermore, the dense spectrum can be reconstructed based on the abundance and spectral components generated from the sparse spectrum. A quantized image showing the amount of composition in the sample can be generated based on the reconstructed dense spectrum.

[0019] Machine learning estimators for decomposing spectral data can be built or developed based on mathematical models that take into account the multinomial distribution of the spectrum. For example, machine learning estimators can be built based on unsupervised learning techniques capable of component / cluster identification and representation. These techniques can include nonnegative matrix factorization (NMF), singular value decomposition (SVD), independent component analysis (ICA), latent Dirichlet allocation (LDA), and K-means.

[0020] In one instance, the machine learning estimator is constructed based on nonnegative least squares. Here, the data is modeled as a superposition of spectral components plus additive noise and recovered by minimizing the least squares criterion. To reduce the ambiguity of the solution, a Dirichlet prior for abundance is assumed. To minimize the criterion, methods such as ADMM (Alternating Direction Method of Multipliers) can be used.

[0021] In another example, LDA is used to model the spectral data. The machine learning estimator is built based on stochastic variational Bayes inference of LDA and the characteristics of the detector. LDA has been used in natural language processing of large document corpora, as presented in the following: “Latent Dirichlet Allocation”, David M. Blei, Andrew Y. Ng, and Michael I. Jordan, *Journal of Machine Learning Research*, 3, March 2003, 993-1022; “Online Learning for Latent Dirichlet Allocation”, Matthew D. Hoffman, David M. Blei, and Francis Bach, 2010, *Proceedings of the 23rd International Conference on Neural Information Processing Systems*, Vol. 1 (NIPS'10), Curran Associates Inc., Red Hook, NY, USA, 856-864; and “Stochastic Variational Inference”. "Variational inference," Matthew D. Hoffman, David M. Blei, Chong Wang, and John Paisley, *Machine Learning Research*, 14, 1 (May 2013), 1303–1347, cited in this paper. In natural language processing, LDA is used to determine the topic and topic probability of each document based on the words within the document. Here, the LDA model is suitable for modeling spectral data. Specifically, in natural language processing, document, word, and topic correspond to sample location, energy, and components of the spectrum, respectively, in spectral data analysis. The energy of the spectrum has a multinomial distribution. The average energy is a linear combination of k global fundamental vectors, where k is the number of components. Each value of abundance is derived from a Dirichlet distribution with small concentration parameters to keep the chemical complexity low for each sample location.

[0022] In some instances, such as Figure 2BAs shown, the LDA-based machine learning estimator includes an LDA state, an E-step for generating abundance and updating the state variables of the LDA state, an M-step for updating the spectral components based on the updated state variables, and a smoothing step for adjusting the updated spectral components based on detector characteristics. In one example, the adjustment is based on the energy-related characteristics of the detected emission. Compared to lower energies, the adjusted component spectrum is smoother at higher energies. Peaks at lower energies are narrower than peaks at higher energies. In other words, the rate of change of the adjusted component spectrum decreases with increasing energy. The adjusted or smoothed spectral component can be set as the current spectral component and fed back to the LDA state. Alternatively or additionally, the smoothed spectral component can be quantized based on a known spectral component before being fed back to the LDA state.

[0023] like Figure 3 As shown, upon receiving each spectrum, the machine learning estimator outputs the corresponding abundance and updates the state variable in the E-step. After processing a batch of spectra, the spectral components are updated based on the updated state variable from the M-step. Therefore, in each E-step, the state variable is updated only based on the most recently received spectrum, rather than previously acquired spectra located at other sample locations. Consequently, the computational intensity is reduced compared to multivariate spectral analysis.

[0024] In some instances, such as Figure 7 As shown, a charged particle beam is configured to repeatedly scan multiple sample locations within the field of view (FOV). The charged particle beam irradiates each sample location for a specific duration. If an X-ray signal is detected, the energy spectrum is sent to a machine learning estimator. An image formed based on the machine learning estimator output can be displayed in real time during the data acquisition process. Therefore, the operator can adjust or terminate the scan based on the observed data quality.

[0025] Turning Figure 1 , Figure 1 A highly schematic depiction of an embodiment of the dual-beam charged particle microscope (CPM) of the present invention is provided, and more specifically, an embodiment of a scanning electron microscopy (SEM) system is shown. The system axis is shown as axis 110. The microscope 100 includes a particle beam 1 that generates a beam of charged particles 3 (in this case, an electron beam) propagating along a particle beam axis 101. The particle beam axis 101 can be aligned with the Z-axis of the system. The beam 1 is mounted on a vacuum chamber 5, which includes a sample holder 7 for holding / placing a sample 6 and one or more associated actuators 8. The vacuum chamber 5 is emptied using a vacuum pump (not depicted). A vacuum port 9 is also depicted, which can be opened to introduce or remove articles (components, samples) into or from the interior of the vacuum chamber 5. The microscope 100 may include multiple such ports 9 as needed.

[0026] Column 1 (in the present case) includes an electron source 10 and an illuminator 2. This illuminator 2 includes lenses 11 and 13 for focusing the electron beam 3 onto the sample 6, and a deflection unit 15 (for performing beam steering / scanning of the beam 3). The microscope 100 further includes a controller / computer processing device 26 for controlling, in particular, the deflection unit 15, lenses 11 and 13 and detectors 19 and 21, and for displaying information collected from detectors 19 and 21 on a display unit 27.

[0027] Detectors 19 and 21 are selected from a variety of possible detector types that can be used to examine different types of "stimulated" radiation emitted from sample 6 in response to irradiation by (shock) beam 3. Detector 19 may be a solid-state detector (e.g., a photodiode) for detecting cathodic emission emitted from sample 6. For example, the detector may alternatively be an X-ray detector, such as a silicon drift detector (SDD) or a silicon-lithium (Si(Li)) detector. For example, detector 21 may be an electron detector in the form of a solid-state photomultiplier tube (SSPM) or a vacuum photomultiplier tube (PMT). This can be used to detect backscattered and / or secondary electrons emitted from sample 6. Those skilled in the art will understand that many different types of detectors can be selected in, for example, the setup depicted, including, for example, ring / segmented detectors. Stimulated radiation, including, for example, X-rays, infrared / visible / ultraviolet light, secondary electrons (SE), and / or backscattered electrons (BSE), is emitted from sample 6 by scanning beam 3 across sample 6. Because such stimulated emission is position-sensitive (due to the scanning motion), the information obtained from detectors 19 and 21 will also be position-dependent. This fact allows, for example, the signal from detector 21 to be used to generate a BSE image of sample 6 (partially), which is essentially a mapping of the signal that varies with the position of the scan path on sample 6.

[0028] Signals from detectors 19 and 21 are transmitted along control lines (buses) 25, processed by controller 26, and displayed on display unit 27. Such processing may include operations such as combination, integration, subtraction, pseudo-coloring, edge enhancement, and other processing known to those skilled in the art. Furthermore, automatic identification processes (e.g., for particle analysis) may be included in such processing. The controller includes a processor and non-transitory memory for storing computer-readable instructions. The methods disclosed herein can be implemented by executing computer-readable instructions in the processor. For example, the machine learning estimator disclosed herein can be implemented by executing computer-readable instructions in the processor of the controller.

[0029] It should be noted that many refinements and alternatives to such setups will be known to those skilled in the art, such as the use of controlled environments within (relatively large volume) microscopes 100, for example, maintaining a background pressure of a few mbar (as used in ambient SEM or low-pressure SEM).

[0030] Figure 2A The diagram illustrates the data flow used for analyzing spectral data. The acquired spectrum 201, as shown... Figure 1 The EDS spectrum acquired by charged particle microscopy is sent to the machine learning estimator 203. Alternatively, the acquired spectrum 201 can be stored in an EDS data cube 202. The machine learning estimator 203 decomposes the received spectrum 201 into abundance 207 and spectral components 208 based on known characteristics 204 of the detected emission. The machine learning estimator 203 can optionally be initialized with sample-specific information 205 and / or known spectral components 206. Sample-specific information 205 can be possible sample compositions, such as possible elements. The machine learning estimator 203 can be initialized with spectral components from possible sample compositions. Known spectral components 206 can be a library of spectral components. Spectral components 208 can be quantized at the quantization step 212 based on the theoretical spectra of chemical elements. The quantized spectral components 252 can be fed back to the machine learning estimator 203 to update state variables. The quantization step compares the input spectrum or spectral component with the theoretical spectra of multiple chemical elements and outputs the quantized spectrum or quantized spectral component along with the concentration of the chemical element in said component (or spectral component). In step 212, the spectral components are quantized, and the elemental concentration 251 of each component and the quantized spectral components 252 are output. A quantized image 214 displaying the elemental concentration at each sample location is generated based on the elemental concentration 251 and abundance of each component. Each pixel of the quantized image 214 can be calculated by using a combination / sum of the concentrations of each spectral component with a weight equal to the pixel abundance. For example, if the first spectral component contains Fe at a concentration of 0.4 and O at a concentration of 0.6, and the second spectral component contains Si at a concentration of 0.33 and O at a concentration of 0.66, then the derived concentrations at the sample location (or pixel) with abundances of (0.5, 0.5) will be Fe…0.2, Si…0.165, O…0.63. The pixel values ​​of the quantized image 214 are determined based on the derived concentrations. Component maps 219 can be generated based on the nearest abundance of multiple sampling locations, where each component map shows the spatial distribution of the amount of a specific component. Figure 5A and 5B This is an example component image of the sample. The grayscale value of a pixel represents the abundance of the corresponding component.

[0031] Further analysis of abundance 207 and spectral components 208 can be performed to provide compositional information about the sample. In one instance, abundance 207 can be clustered at 216 to generate one or more phase maps 218. The clustering algorithm 216 groups pixels with similar abundance vectors, i.e., pixels containing abundance vectors of spectral components in similar proportions, and assigns the same phase to each existing cluster. The clustering algorithm can be K-means. The phase map has binary contrast. That is, each pixel of the phase map can be one of two numbers, such as 1 or 0. Figure 6A and 6B Is with Figure 5A and 5B Phase diagram of the same FOV. Figures 5A-5B The sample region 501 is shown to contain different components. Due to the similarity of pixel abundance distribution in region 501, Figure 6B The sample at region 501 is shown to belong to the same phase. Therefore, the component map shows the relative quantity of a component at each sample location, while the phase map shows the presence of a phase at each sample location. A phase can correspond to a certain combination of components. In one instance, each component map corresponds to one element. The phase map corresponds to a mineral with a specific combination of elements. In some instances, multiple phase maps can be combined into a composite phase map by color-coding different phases.

[0032] The chemical element concentration 251 of the component associated with spectral component 208 can be combined with the abundance at summation box 213 and 207 to obtain a quantized image 214. Alternatively, a quantized image 211 can be generated by first combining the abundance with the spectral component at summation box 209 to generate a combined spectrum, and then quantizing the combined spectrum with the known spectrum of the spectral component at quantization box 210. Each pixel of the quantized image 211 is generated based on the element concentration output from quantization box 210. The computational intensity used to generate the quantized image 211 is higher than that used for the quantized image 214 due to the increased data size at quantization box 210.

[0033] Figure 2BThe data flow within the machine learning estimator 203 is shown. In one instance, spectral data is modeled using LDA, and the machine learning estimator uses stochastic variational Bayes to perform inference on the LDA model. Essentially, a series of lower bounds on the log-likelihood, indexed by a set of variational parameters, are considered. The variational parameters are chosen by an optimizer that attempts to find the most stringent possible lower bound. The stochastic variational Bayes of the LDA model comprises E-steps and M-steps, as described below: “Online Learning of Latent Dirichlet Allocations” Matthew D. Hoffman, David M. Blei, and Francis Bach, 2010, Proceedings of the 23rd International Conference on Neural Information Processing Systems, Vol. 1 (NIPS'10), Coren Gail / Coren Associates, Redhook, NY, 856-864. The machine learning estimator further includes a smoothing (or regularization) step for enriching the LDA model based on prior information about the spectral data. In the E-step, for each document, there exists an optimized value for the variational parameters. In the M-step, the lower bound of the obtained log-likelihood with respect to the model parameters α and β is maximized. This corresponds to finding the maximum likelihood estimate with sufficient expected statistics for each document under the approximate posterior, calculated in the E-step. The log-sample likelihood of the data with parameters α and β can be expressed as... The corpus of the given document The machine learning estimator 203 includes an LDA state 221 for storing state variables and spectral components, an E-step and an M-step for performing inference, and a smoothing step 228 for smoothing the estimated spectral components 229 from the output of the M-step. The spectral components in the LDA state 221 can be initiated based on sample-specific information 205 or a library of known spectral components 206. The LDA state 221 outputs the current state variables and current spectral components to the E-step 222 and receives the updated state variables 224 generated by the E-step. The E-step generates the abundance 207 of the spectrum 201 based on the current spectral components. The E-step also outputs the updated state variables 226 to the M-step 223. The M-step estimates the spectral components based on the updated state variables and outputs the estimated spectral components 229 to the smoothing step 228. The smoothing step 228 smooths the estimated spectral components based on detector characteristics and outputs the smoothed spectral components outside the machine learning estimator 203 as spectral component 208. In one instance, the smoothed spectral component is fed back to LDA state 221 as the current spectral component. Alternatively, the smoothed spectral component can be quantized at 212 and then fed back to LDA state 221 as the current spectral component.

[0034] Figure 3Method 300 for decomposing spectral data using a machine learning estimator is shown. Spectral data can be modeled using LDA. The decomposition is based on a stochastic variational Bayesian applied to the LDA model and the energy-related properties of the detector. Here, the E step is repeated multiple times before proceeding to the M step until a sufficient number of spectra have been analyzed. The E step updates or generates the abundance (corresponding to a sample location or pixel) for each input spectrum and updates the state variables of the LDA state. The M step then estimates the spectral components based on the updated state variables. The estimated spectral components generated by the M step are smoothed based on the characteristics of the EDS data. The abundance is specific to each sample location or pixel, while the spectral components are global parameters applied to all sample locations or pixels.

[0035] At position 302, the spectral components are initialized. In one instance, the spectral components are randomly initialized. In another instance, the spectral components are initialized based on known spectral components. Known spectral components can be a library of known spectral components for various elements or phases. Alternatively, known spectral components can be determined based on possible sample compositions. In yet another instance, the spectral components can be initialized based on dense spectra measured at one or more sample locations.

[0036] At position 303, set the E-step counter to zero.

[0037] At position 304, the machine learning estimator receives the spectrum obtained by the detector from the sample location. The spectrum may be a sparse spectrum.

[0038] At position 306, the received spectrum is analyzed via the E-step. The E-step estimates the abundance of the spectrum and updates the state variables of the LDA model. For example, the abundance is calculated iteratively, and the state variables are updated after abundance aggregation. Additionally, the E-step counter is incremented by 1 or incremented by 1.

[0039] At 308, the E-step counter is compared with a predetermined threshold number. The threshold number can be determined experimentally. In one instance, the threshold number is 200. If the E-step counter is less than the threshold number, the E-step is repeated at 310 and the new spectrum is processed. Otherwise, method 300 proceeds to the M-step at 312.

[0040] At 312, during the M-step of spectral data decomposition, the spectral components are estimated or updated based on the updated state variables from 308.

[0041] At position 314, the estimated spectral components from position 312 are smoothed based on the energy correlation characteristics of the EDS data, which are generated by the energy correlation characteristics of the detector. This indicates the number of energy boxes distinguished by the detector. For each spectral component produced by the M-step at 312... The sum of the spectral components is 1. That is to say, Equation 1 The energy associated with the spectral components is split into... One energy tank. Use The energy bins are indexed; the larger the index, the higher the energy within the bin. The estimated spectral components can be sparse. The values ​​of the estimated spectral components may oscillate dramatically because there are no additional restrictions on the estimated spectral components besides Equation 1. The estimated spectral components are smoothed based on the detector characteristics, such that the smoothed spectral components have a smaller rate of change with energy at higher energies compared to the rate of change at lower energies. In other words, the smoothed spectral components become smoother with increasing energy, and the peaks at higher energies are wider than the peaks at lower energies. In one instance, the estimated spectral components can be obtained by solving for the smoothing: Equation (2) Where vector s represents the spectral components before smoothing, vector These are non-negative smoothed spectral components. The spectrum in the j-th box The value, It involves Fano noise and is a non-negative number that controls the intensity of smoothing, and the parameter The larger, The smoother and Leave The farther away. As adjacent boxes The rate of change of the smoothed spectral coefficients for each j is relatively small. (The coefficients are then...) Set as ,in The width of the box, and Involves the zero-energy peak width. Coefficient Its function is vector exist When the value is large, it has a small change / variation. The parameters in Equation 2 (such as...) , and The smoothed spectral components depend on the characteristics of the detector and may vary depending on the detector type and design. Therefore, compared to peaks at higher energies, the smoothed spectral components... The peaks at lower energies are narrower. This characteristic of the smoothed spectral components is similar to that of EDS data.

[0042] At position 316, the machine learning estimator can output the abundance and current spectral components of the analyzed spectrum for further analysis or to form an image as shown in Figure 2. In another instance, the estimator generates the abundance at position 306, outputs the abundance of each analyzed spectrum, and outputs only the spectral components at position 316. In yet another instance, after all spectra are analyzed by the machine learning estimator, the estimator outputs both the abundance and spectral components.

[0043] At 320, method 300 checks whether all spectra have been analyzed. If the answer is yes, method 300 exits. Otherwise, at 318, the current component spectrum can be quantized via quantization box 212 in Figure 2 and fed back to the LDA model for processing the next spectrum in step E. Alternatively, the current spectral component can be used to process the next spectrum in step E without quantization.

[0044] In this way, the machine learning estimator processes the acquired spectra in batches. Within each batch, the E-step is repeated, with each E-step processing one spectrum. After each E-step, an abundance corresponding to the sample location (or pixel) of the spectrum is generated. After processing the spectra of each batch, the estimator runs the M-step to estimate the spectral components associated with all spectra received by the estimator. The estimated spectral components are then smoothed to enrich the LDA model with prior information about the characteristics of the EDS data. Abundance and spectral components can be obtained during data acquisition as follows: Figure 2A The various formats shown are processed and displayed in real time.

[0045] Figure 4 Showing abundance Spectra obtained by derived free charged particle beam microscopy spectral components Abundance It is a vector with k values, where each abundance value corresponds to a component. Each spectral component is a vector representing the spectrum of that component. The spectral components can be normalized so that the sum of the vectors is one. The predicted spectrum can be formed by combining the abundance with the following spectral components. : Compared to the spectrum The signal-to-noise ratio of the predicted spectrum is improved, and the predicted spectrum can represent the most likely dense spectrum at the sample location.

[0046] Figure 7 A method 700 for imaging a sample using charged particle microscopy is shown. Emissions from the sample in response to charged particle irradiation are detected in spectral form by a detector. The acquired spectrum is decomposed into abundance and spectral components by a machine learning estimator, such as... Figure 3As shown, an image displaying the compositional information of a sample can be generated in real time based on the output of a machine learning estimator.

[0047] At position 702, the sample is loaded onto the charged particle microscope, and imaging parameters are set. Imaging parameters can include parameters for the charged particle beam and scanning parameters. For example, one or more fields of view (FOVs) can be selected, and the residence time of the charged particle beam at each sample location during FOV scanning can be determined. The FOV can be selected based on SEM or BSE imaging. The residence time can be determined based on the sample type. Furthermore, the sample type or possible sample composition can be set by the operator in the charged particle microscope to initialize the machine learning estimator.

[0048] At 704, the machine learning estimator can optionally be initialized based on one or more dense spectra collected from the sample. For example, a sample image, such as an SEM or BSE image, can be acquired at 705 and displayed to the operator. The operator can select one or more regions or sample locations on the sample image to collect dense spectra. The selected sample location can be a region with different sample compositions. At 706, a charged particle beam, such as an electron beam, is directed to the selected sample location, and the EDS spectrum is acquired. The longer dwell time at the selected sample location compared to the dwell time used to scan the FOV determined at 702 allows for the acquisition of dense spectra. The dense spectra can be used to initialize the estimator at step 302 of method 300.

[0049] At position 708, multiple sample locations are repeatedly scanned with a charged particle beam. If a spectrum is detected at a particular sample location, it is sent to a machine learning estimator. If no spectrum is detected, the charged particle beam moves to the next sample location without sending a signal to the machine learning estimator. In one instance, no spectrum is detected at a sample location when the total count output from the detector at that location is below a threshold count. Based on the received spectra, the machine learning estimator outputs the abundance of each spectrum in the analyzed spectrum and updates the spectral components after analyzing a batch of spectra.

[0050] At point 710, an image showing the structure and composition information of the sample can be displayed. For example, one or more of the component maps, spectral components, phase maps, and quantized images can be displayed based on the output of a machine learning estimator, as shown in Figure 2. Steps 708 and 710 can be performed in parallel to display the images during data acquisition.

[0051] At 712, method 700 checks if further scans are needed. In one instance, if the sample has already been scanned a predetermined number of times, no additional scans are required. In another instance, the operator can determine whether further scans are needed based on the image displayed at 710. If further scans are needed, method 700 continues scanning the sample with a charged particle beam. Otherwise, method 700 proceeds to 714. At 714, the decomposed data from the machine learning estimator output, along with the EDS data cube, is saved.

[0052] In this way, an image representing the composition of the sample can be formed, and the image can be displayed in real time based on the decomposed data. The technical impact of using LDA to model spectral data is that LDA modeling considers multiple distributions of the spectrum. The technical impact of smoothing spectral components based on known characteristics of the spectral data is that it adjusts the inference of LDA to generate spectral components with characteristics similar to the spectral data. The technical impact of outputting abundance and spectral components after processing a batch of received spectra is that an image of the constituent samples can be displayed during data acquisition. Furthermore, computational intensity is reduced.

Claims

1. A method comprising: Irradiate the sample with a beam of charged particles; The spectrum is obtained by detecting the emission from the sample in response to the irradiation using a detector; as well as Based on the characteristics of the detector, a machine learning estimator decomposes the acquired spectrum into abundance and multiple spectral components, wherein the characteristics of the detector include the energy-related characteristics of the detector.

2. The method of claim 1, wherein the machine learning estimator is developed based on a mathematical model that takes into account the multinomial distribution of the spectrum.

3. The method according to any one of claims 1-2, further comprising: The machine learning estimator is updated based on the multiple spectral components; The second spectrum is obtained by detecting the emission from the sample using the detector. as well as The acquired second spectrum is decomposed using an updated machine learning estimator based on the characteristics of the detector.

4. The method according to any one of claims 1-2, wherein decomposing the acquired spectrum based on the characteristics of the detector includes adjusting each of the plurality of spectral components based on the energy correlation characteristics of the detector.

5. The method of claim 4, wherein the energy correlation characteristic of the detector comprises that the emission peak detected at lower energies is narrower than the emission peak detected at higher energies.

6. The method of claim 1, wherein the abundance is a 1D array, and each value of the abundance corresponds to a component and one of the plurality of spectral components.

7. The method of claim 1, wherein the acquired spectrum is modeled using a latent Dirichlet allocation, and the machine learning estimator is constructed based on stochastic variational Bayesian inference.

8. The method according to any one of claims 1-2 and 6-7, further comprising initializing the machine learning estimator with a plurality of known spectral components.

9. The method of claim 8, further comprising: Select one or more sample locations; The selected location is irradiated with the charged particle beam and one or more dense spectra are obtained, wherein the dwell time for obtaining the dense spectrum is longer than the dwell time for obtaining the spectrum. as well as The known spectral components are determined based on the one or more dense spectra.

10. The method of claim 8, further comprising determining the known spectral components based on the sample composition.

11. The method of claim 1, further comprising: Theoretical spectra based on chemical elements are used to quantify one or more of the multiple spectral components. The machine learning estimator is updated based on the quantized spectral components; The second spectrum is obtained by detecting the emission from the sample; as well as The second spectrum is obtained using the updated machine learning estimator and the feature decomposition of the detector.

12. A method comprising: Multiple sample locations are repeatedly scanned using a charged particle beam; Multiple spectra are acquired in response to the detection of emission from the plurality of sample locations by a detector; as well as Based on the characteristics of the detector, a machine learning estimator decomposes the acquired spectrum into abundance and multiple spectral components, wherein each of the acquired multiple spectra corresponds to one of the abundances, and wherein the characteristics of the detector include the energy-related characteristics of the detector.

13. The method of claim 12, wherein the acquired plurality of spectra are modeled using Latent Dirichlet Allocation (LDA), and the method further comprises: Update the state variables of the LDA model with each of the multiple acquired spectra; as well as The plurality of spectral components are updated after the state variable has been updated for a predetermined number of times.

14. The method according to any one of claims 12-13, wherein the method further comprises: One or more component maps of the sample are generated, wherein each component map is generated based on the value corresponding to the component in the abundance.

15. The method according to any one of claims 12-13, further comprising clustering the abundance into one or more phases; and generating one or more phase maps based on the clustered abundance.

16. The method of any one of claims 12-13, wherein the plurality of spectral components are smoothed based on the characteristics of the detector, and the method further comprises updating the machine learning estimator based on the plurality of spectral components.

17. A charged particle microscope, comprising: Source, the source being used to generate a beam of charged particles; A scanner used to scan the beam of charged particles over a sample; A detector for detecting emissions from the sample in response to irradiation of the sample by the charged particle beam; The controller includes a non-transitory memory for storing computer-readable instructions that, when executed, cause the controller to: The sample was irradiated with the charged particle beam; Based on the emission spectrum detected from the sample; and Based on the characteristics of the detector, a machine learning estimator decomposes the acquired spectrum into abundance and multiple spectral components, wherein the characteristics of the detector include the energy-related characteristics of the detector.

18. The charged particle microscope of claim 17, wherein the source is an electron source and the detector is an X-ray detector.

19. The charged particle microscope according to any one of claims 17-18, wherein the machine learning estimator is based on stochastic variational Bayesian inference for latent Dirichlet allocation (LDA), and the controller is further configured to: The machine learning estimator is updated based on the multiple spectral components; A second spectrum is obtained by detecting the emission from the sample; and The acquired second spectrum is decomposed using an updated machine learning estimator.

Images