Methods and systems for component analysis of spectral data

Abstract

To provide a method and a system for processing spectral data for extracting sample compositional information.SOLUTION: Responsive to irradiation of a charged particle beam 3, emission from a sample 6 is acquired in the form of spectral data by being detected by a detector 19. The spectral data is decomposed by a machine learning estimator to abundances and spectral components based on a character of the detector. Images showing compositional information of the sample 6 are generated based on the abundances and the spectral components.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] This specification 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 in the form of electron microscopy. Multiple types of emissions from a sample in response to charged particle irradiation can provide structural and compositional information about the sample. For example, energy-dispersive X-ray spectroscopy (EDS or EDX) can be used for elemental analysis or chemical characterization based on the energy spectrum of X-ray emissions. Since various elements or chemical substances have characteristic energy spectra, the sample composition may be estimated by comparing the EDS spectrum with the characteristic energy spectrum.

[0003] EDS spectra can result from X-ray emissions from multiple elements or chemicals, and one method for decomposing EDS spectra is disclosed in U.S. Patent No. 6584413B1 by Keenan et al., in which spectral data is decomposed into a concentration-intensity matrix and a spectral shape matrix using multivariate spectral analysis. However, the inventors recognize that multivariate spectral analysis requires a full spectral image of all pixels, which can result in high computational intensity and long data-to-image times. [Overview of the Initiative]

[0004] In one embodiment, the method includes irradiating a sample with a charged particle beam, acquiring a spectrum by detecting emissions from the sample in response to the irradiation via a detector, and decomposing the acquired spectrum into abundances and multiple spectral components using a machine learning estimator based on the characteristics of the detector. In this way, an image representing the sample composition is formed based on the output of the machine learning estimator and can be displayed during data acquisition. Furthermore, instead of processing the entire spectral image, each spectrum is decomposed as it is received by the machine learning estimator, reducing the computational intensity.

[0005] It should be understood that the above summary is provided in a simplified manner to introduce a selection of concepts that will be further described in the modes for carrying out the invention. It is not intended to identify the main or essential features of the claimed subject matter, and its scope is uniquely defined by the claims that follow the modes for carrying out the invention. Furthermore, the claimed subject matter is not limited to implementations that solve any of the defects mentioned above or in any part of this disclosure. [Brief explanation of the drawing]

[0006] [Figure 1] This shows a charged particle microscope. [Figure 2A] This document demonstrates a data flow for analyzing spectral data using a machine learning estimator. [Figure 2B] Figure 2A shows the data flow within the machine learning estimator. [Figure 3] This section describes a method for analyzing spectral data. [Figure 4] This shows the data generated during the analysis of spectral data. [Figure 5A] This is an example of a component map. [Figure 5B] This is an example of a component map. [Figure 6A] This is an example of a phase map. [Figure 6B] This is an example of a phase map. [Figure 7]This method involves acquiring and analyzing data using a charged particle microscope.

[0007] The same reference number refers to the corresponding part across several drawings within the same drawing. [Modes for carrying out the invention]

[0008] The following description concerns systems and methods for analyzing spectral data, such as spectral data from energy-dispersive X-ray spectroscopy (EDS or EDX). EDS spectra can be acquired using a charged particle microscope, such as the charged particle microscope shown in Figure 1. In response to irradiating a sample location with a charged particle beam, such as an electron beam, emissions from the sample, such as X-ray emissions, are detected via an X-ray detector. The X-ray detector outputs the detected emissions in the form of an energy spectrum. The energy spectrum shows the count (intensity) of the emissions at various energies. To obtain compositional information at the sample location, the energy spectrum is decomposed into abundances and multiple spectral components by a machine learning estimator. As shown in Figure 4, the abundances are a 1D array or vector. The abundance values ​​are indexed by component. Each abundance value corresponds to one component and one of multiple spectral components. Each component may 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 may be updated based on the decomposed results. For example, the machine learning estimator is updated based on either or both of the spectral components and the quantified spectral components. The quantification process involves taking a spectrum as input and returning a corrected or quantified spectrum based on known spectra, such as the concentrations of the chemical elements identified in the spectrum and the theoretical spectra of the chemical elements. A second energy spectrum obtained by irradiating a second sample location with a charged particle beam can be decomposed by the updated machine learning estimator.

[0009] A machine learning estimator decomposes the received spectrum and outputs abundances and spectral components based on the characteristics of the detector. Before output from the machine learning estimator, the spectral components may be adjusted based on the characteristics of the detector. In one example, the adjustment is based on the energy-dependent characteristics of the detector. Due to noise in the amplification chain and the detector material, a typical energy spectrum obtained by a detector, such as EDS data, will be smoother at higher energies compared to the energy spectrum at lower energies. In other words, the peaks in the energy spectrum at lower energies will be narrower than the peaks at higher energies. Through adjustment, the spectral components output from the machine learning estimator may simulate the characteristics of a typical energy spectrum.

[0010] A machine learning estimator may be initiated based on sample information and / or known spectral components. For example, known spectral components of the sample's potential composition may be used as initial spectral components. In another example, one or more sample locations may be selected by the microscope operator, and high-density spectra are obtained at the selected sample locations with long residence times. These high-density spectra may be used to initialize the estimator. As more spectra are analyzed, the output of the machine learning estimator may become more accurate, and compositions at more sample locations may be identified.

[0011] The output of the machine learning estimator, i.e., the abundance at each sample location and the spectral components of each component, may be displayed directly or further analyzed to extract more information. In one example, one or more component maps are generated based on the abundances. The component maps may 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, allowing for the acquisition of sparse spectra. In sparse spectra, one or more low-amplitude spectral peaks may be missing due to insufficient signal. Missing peaks may be present in the high-density spectra obtained with longer residence times. In other words, sparse spectra may be considered undersampled spectra from high-density spectra. In another example, as shown in Figures 5A-5B and 6A-6B, abundances from multiple sample locations can be further clustered or grouped based on the similarity of the abundance distributions to obtain one or more phase maps. Furthermore, high-density spectra may be reconstructed based on the abundances and spectral components generated from the sparse spectra. Quantified images showing the amounts of components in the sample may be generated based on the reconstructed high-density spectra.

[0012] Machine learning estimators for decomposing spectral data may be built or developed based on mathematical models that describe the multinomial distribution of the spectrum. For example, machine learning estimators may be built based on unsupervised learning methods that enable component / cluster identification and characterization. These methods may include non-negative matrix factorization (NMF), singular value decomposition (SVD), independent component analysis (ICA), latent Dirichlet allocation (LDA), and K-means.

[0013] In one example, a machine learning estimator is constructed based on a non-negative least squares method, where the data is modeled as a superposition of spectral components and additive noise, and reconstructed by minimizing the least squares criterion. To reduce the ambiguity of the solution, a Dirichlet before the abundance is assumed. Methods such as ADMM (Alternating Directional Multipliers) can be used to minimize the criterion.

[0014] In another example, spectral data is modeled using LDA. A machine learning estimator is built based on the stochastic variational Bayesian estimation of the LDA and the properties of the detector. As shown in Latent Dirichlet Allocation by David M. Blei, Andrew Y. Ng and Michael I. Jordan, incorporated herein by reference in J. Mach. Learn, Res. 3, March 2003, 993-1022; Online Learning of Latent Dirichlet Allocation in 2010 by Matthew D. Hoffman, David M. Blei and Francis Bach, in the Proceedings of the 23rd International Conference on Neural Information Processing Systems, Vol. 1 (NIPS'10), Learn. Res. 14, 1 (May 2013), 1303-1347, Learn. Res. 14, 1 (May 2013), 1303-1347, Curran Associates Inc., Red Hook, NY, USA, 856-864; and Probabilistic Variational Inference by Matthew D. Hoffman, David M. Blei and John Paisley, incorporated herein by reference, LDA is used for natural language processing of large document corpora. In natural language processing, LDA is used to determine the topic and the probability of a topic for each document based on the words in the document. In this specification, the LDA model is adapted to model spectral data. Specifically, the documents, words, and topics in natural language processing correspond to sample locations, spectral energy, and components of spectral data analysis, respectively. The energy of the energy spectrum has a multinomial distribution. The mean of the energy is a linear combination of k global basis vectors, where k is the number of components. Each value of the abundance is obtained from a Dirichlet distribution with a small concentration parameter to reduce the chemical complexity of each sample location.

[0015] In some examples, as shown in FIG. 2B, the LDA-based machine learning estimator includes an LDA state, an E-step for generating abundances and updating state variables of the LDA state, an M-step for updating 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-dependent characteristics of the detected emissions. The adjusted component spectrum becomes smoother at higher energies compared to lower 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 components may be set as the current spectral components and fed back to the LDA state. Alternatively or additionally, the smoothed spectral components may be quantified based on known spectral components before being fed back to the LDA state.

[0016] As shown in FIG. 3, upon receiving each spectrum, the machine learning estimator outputs the corresponding abundances and updates the state variables of the E-step. After processing a batch of spectra, the spectral components are updated based on the state variables updated in the M-step. Thus, in each E-step, the state variables are updated based only on the last received spectrum and not on spectra previously acquired at other sample locations. Therefore, the computational intensity is reduced compared to multivariate spectral analysis.

[0017] In some examples, as shown in FIG. 7, the charged particle beam is configured to repeatedly scan a plurality of sample locations within the field of view (FOV). The charged particle beam is irradiated onto each sample location for a dwell time. When an X-ray signal is detected, the energy spectrum is transmitted to the machine learning estimator. An image formed based on the output of the machine learning estimator can be displayed in real time during the data acquisition process. Thus, the operator can adjust or terminate the scan based on the observed data quality.

[0018] Turning to FIG. 1, FIG. 1 is a very schematic depiction of an embodiment of a dual beam charged particle microscope (CPM) in which the present invention is implemented, and more specifically, shows an embodiment of a scanning electron microscope (SEM) system. The system axis is shown as axis 110. The microscope 100 includes a particle optical column 1 that creates a beam 3 of charged particles (in this case, an electron beam) propagating along the particle optical axis 101. The particle optical axis 101 may be aligned with the Z-axis of the system. The column 1 is attached to a vacuum chamber 5 that includes a sample holder 7 for holding / positioning the sample 6 and an associated actuator 8. The vacuum chamber 5 is evacuated using a vacuum pump (not shown). A vacuum port 9 that may be opened to introduce / remove items (components, samples) inside the vacuum chamber 5 is also shown. The microscope 100 may include a plurality of such ports 9 as needed.

[0019] The column 1 (in this 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, among other things, the deflection unit 15, the lenses 11, 13, and the detectors 19, 21, and for displaying the information collected from the detectors 19, 21 on a display unit 27.

[0020] Detectors 19 and 21 are selected from a variety of potential detector types that can be used to inspect different types of "induced" radiation emitted from the sample 6 in response to irradiation by the (incident) beam 3. Detector 19 may be a solid-state detector (such as a photodiode) used to detect cathodoluminescence emitted from the sample 6. Alternatively, it may be an X-ray detector such as a silicon drift detector (SDD) or a silicon lithium (Si(Li)) detector. Detector 21 may be an electron detector in the form of a solid-state photomultiplier tube (SSPM) or a vacuum photomultiplier tube (PMT), which can be used to detect backscattered and / or secondary electrons emitted from the sample 6. Those skilled in the art will understand that many different types of detectors can be selected in a configuration such as the one shown, including, for example, annular / segmented detectors. By scanning beam 3 over the sample 6, stimulated radiation, including, for example, X-rays, infrared / visible / ultraviolet rays, secondary electrons (SE), and / or backscattered electrons (BSE), is emitted from the sample 6. Since 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 makes it possible to use the signal from detector 21 (for example) to create a BSE image of (part of) the sample 6, which is essentially a map of the signal as a function of the scanning path position on the sample 6.

[0021] Signals from detectors 19 and 21 pass through control lines (buses) 25, are processed by the controller 26, and are displayed on the display unit 27. Such processing may include operations such as combining, integrating, subtracting, false coloring, edge enhancement, and other processes known to those skilled in the art. Furthermore, an automated recognition process (such as that used in particle analysis) may be included in such processing. The controller includes a processor and non-temporary memory for storing computer-readable instructions. The methods disclosed herein can be carried out by executing computer-readable instructions within the processor.

[0022] It should be noted that many improvements and alternatives to such settings are known to those skilled in the art, such as the use of a controlled environment within a microscope (a relatively large number of them), for example, maintaining a background pressure of several millibars (as used in environmental SEMs or low-pressure SEMs).

[0023] Figure 2A shows the data flow for analyzing spectral data. Acquired spectra 201, such as EDS spectra obtained by the charged particle microscope in Figure 1, are sent to machine and learning estimators 203. Acquired spectra 201 may also be stored in the EDS data cube 202. The machine learning estimator 203 decomposes the received spectra 201 into abundances 207 and spectral components 208 based on known characteristics of the detected radiation 204. The machine learning estimator 203 may optionally be initialized with sample-specific information 205 and / or known spectral components 206. Sample-specific information 205 may be potential sample compositions, such as latent elements. The machine learning estimator 203 may be initialized with spectral components from potential sample compositions. Known spectral components 206 may be a library of spectral components. Spectral components 208 may be quantified in the quantification step 212 based on the theoretical spectra of chemical elements. The quantified spectral components 252 may be fed back to the machine learning estimator 203 to update state variables. The quantification step compares the input spectrum or spectral components with the theoretical spectra of multiple chemical elements and outputs the quantified spectrum or quantified spectral components, and the concentrations of the chemical elements within the components (or spectral components). In 212, the spectral components are quantified, and the elemental concentrations per component 251 and the quantified spectral components 252 are output. A quantified image 214 showing the elemental concentrations per sample location is generated based on the elemental concentrations 251 and abundances per component. Each pixel in the quantified image 214 may be calculated by taking the combined / sum of the concentrations for each spectral component, weighted equal to the abundance of the pixel. For example, if the first spectral component contains Fe at a concentration of 0.4 and oxygen at a concentration of 0.6, and the second spectral component contains Si at 0.33 and oxygen at 0.66, then the estimated concentrations at the sample location (or pixel) with abundances (0.5, 0.5) would be Fe…0.2, Si…0.165, O…0.63. The pixel values ​​of the quantified image 214 are determined based on the estimated density.The component map 219 may be generated based on the most recent abundances at multiple sampling locations, with each component map showing the spatial distribution of the amount of a particular component. Figures 5A and 5B are exemplary component maps of a sample. The gray values ​​of the pixels represent the abundance values ​​of the corresponding component.

[0024] The abundances 207 and spectral components 208 may be further analyzed to provide compositional information of the sample. In one example, abundances 207 may be clustered in 216 to generate one or more phase maps 218. The clustering algorithm 216 groups together pixels that have similar abundance vectors, i.e., abundance vectors containing spectral components in similar proportions, and assigns the same phase to each cluster found. The clustering algorithm may be K-means. The phase maps have binary contrast; that is, each pixel in the phase map can be either one of two numerical values, such as 1 or 0. Figures 6A and 6B are phase maps for the same FOV as Figures 5A and 5B. Figures 5A-5B show that region 501 of the sample contains different components. Due to the similarity of the abundance distribution of pixels in region 501, Figure 6B shows that the sample in region 501 belongs to the same phase. Thus, the component map shows the relative amount of one component at each sample location, and the phase map shows the presence of a phase at each sample location. The phase may correspond to a specific combination of components. In one example, each component map may correspond to a single element. The phase map corresponds to a mineral with a specific bonding of elements. In some examples, multiple phase maps may be combined into a single composite phase map by color-coding different phases.

[0025] The concentration 251 of the chemical element of the component associated with spectral component 208 may be combined with the abundance in 207 in the addition block 213 to obtain a quantified image 214. Alternatively, the quantified image 211 may be generated by first generating a combined spectrum by combining the abundance and spectral component in the addition block 209, and then quantifying the combined spectrum with the known spectra of the spectral component in the quantification block 210. Each pixel of the quantified image 211 is generated based on the concentration of the element output from the quantification block 210. Compared to the quantified image 214, the computational intensity for generating the quantified image 211 is higher due to the increased data size in the quantification block 210.

[0026] Figure 2B shows the data flow within the machine learning estimator 203. In one example, spectral data is modeled with LDA, and the machine learning estimator performs inference on the LDA model using a stochastic variational Bayes method. Essentially, a family of lower bounds of the log-likelihood is considered, indexed by a set of variational parameters. The variational parameters are selected by an optimization procedure that seeks to find the strictest latent lower bounds. The stochastic variational Bayes method for the LDA model includes E-steps and M-steps, as described in the 2010 online learning of latent Dirichlet allocation by Matthew D. Hoffman, David M. Blei and Francis Bach in the minutes of the 23rd International Conference on Neural Information Processing Systems, Vol. 1 (NIPS'10), Curran Associates Inc., Red Hook, NY, USA, 856-864. The machine learning estimator further includes a smoothing step (or regularization step) to augment the LDA model based on prior information about the spectral data. In the E-step, for each document, the optimized values ​​of the variational parameters are found. In step M, the lower bound of the result of the log-likelihood with respect to the model parameters α and β is maximized. This corresponds to finding the maximum likelihood estimate where sufficient statistics for each document are expected under the approximate posterior calculated in step E. The log-likelihood of the data with parameters α and β may be expressed as follows:

number

[0027] 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 M step output. The spectral components of the LDA state 221 may be started 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 abundances 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 the detector characteristics and outputs the smoothed spectral components from the machine learning estimator 203 as spectral components 208. In one example, the smoothed spectral component is fed back to the LDA state 221 as the current spectral component. Additionally or alternatively, the smoothed spectral component is quantified at 212 before being fed back to the LDA state 221 as the current spectral component.

[0028] Figure 3 shows a method 300 for decomposing spectral data with a machine learning estimator. The spectral data may be modeled with LDA. The decomposition is based on a stochastic variational Bayesian method applied to the LDA model and the energy-dependent properties of the detector. Hereinafter, step E is repeated multiple times until a sufficient number of spectra have been analyzed before proceeding to step M. Step E updates or generates the abundance of each input spectrum (corresponding to a sample location or pixel) and updates the state variables of the LDA state. Step M then estimates the spectral components based on the updated state variables. The estimated spectral components generated by step M are smoothed based on the properties of the EDS data. Abundances are specific to each sample location or pixel, while spectral components are global parameters that apply to all sample locations or pixels.

[0029] In step 302, the spectral components are initialized. In one example, the spectral components are initialized randomly. In another example, the spectral components are initialized based on known spectral components. The known spectral components may be a library of known spectral components of various elements or phases. Alternatively, the known spectral components may be determined based on the potential sample composition. In yet another example, the spectral components may be initialized based on high-density spectra measured from one or more sample locations.

[0030] At step 303, the E-step counter is set to zero.

[0031] In 304, the machine learning estimator receives a spectrum obtained from the sample location by the detector. The spectrum may be a sparse spectrum.

[0032] In step 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 the abundance converges. Additionally, the E-step counter is incremented by one.

[0033] In step 308, the E-step counter is compared to a predetermined threshold number. The threshold number may be determined experimentally. In one example, the threshold number is 200. If the E-step counter is less than the threshold number, step E is repeated in step 310 and a new spectrum is processed. Otherwise, method 300 proceeds to step M in step 312.

[0034] In 312, during the M step of spectral data decomposition, the spectral components are estimated or updated based on the updated state variables from 308.

[0035] In 314, the spectral components estimated from 312 are smoothed based on the energy-dependent characteristics of the EDS data, which arise from the energy-dependent characteristics of the detector. J indicates the number of energy bins identified by the detector. Each spectral component s∈R created by the M-step in 312 J Regarding this, the sum of the spectral components is 1. That is,

number

number

number

[0036] In step 316, the abundance and current spectral components of the analyzed spectrum may be output from the machine learning estimator for further analysis or image formation, as shown in Figure 2. In another example, the estimator outputs the abundance of each analyzed spectrum and only the spectral components in step 316, after the abundances are generated in step 306. In yet another example, after all spectra have been analyzed by the machine learning estimator, the estimator outputs the abundance and spectral components.

[0037] In step 320, method 300 checks whether all spectra have been analyzed. If the answer is "yes", method 300 terminates. Otherwise, in step 318, the current component spectrum may be quantified via quantification block 212 in Figure 2 and fed back to the LDA model to process the next spectrum in step E. Alternatively, the current spectral component is used to process the next spectrum in step E without being quantified.

[0038] In this way, the machine learning estimator processes the acquired spectra in batches. Within each batch, the E-step is repeatedly executed, and in each execution of the E-step, one spectrum is processed. After each execution of the E-step, the abundance of the sample location (or pixel) corresponding to the spectrum is generated. After each batch of spectra is processed, the estimator executes the M-step to estimate the spectral components related to all the spectra received by the estimator. Next, the estimated spectral components are smoothed, and as a result, the LDA model is enhanced by the prior information regarding the characteristics of the EDS data. The abundance and spectral components may be processed and displayed in real time during data acquisition in various forms as shown in FIG. 2A.

[0039] FIG. 4 shows the abundance θ p and the spectral components s1 - s derived from the spectrum φ p acquired by the charged particle microscope. The abundance θ k is a vector having k values, and each value of the abundance corresponds to one component. Each spectral component is a vector representing the spectrum of the component. The spectral components may be normalized such that the sum of the vectors is 1. The predicted spectrum p may be formed by combining the abundance and the spectral components.

Number

Number

[0040] ​Figure 7 shows a method 700 for imaging a sample using a charged particle microscope. Emission from the sample in response to charged particle irradiation is detected by a detector in the form of a spectrum. The acquired spectrum is decomposed into abundances and spectral components by a machine learning estimator, as shown in Figure 3. Images showing the compositional information of the sample can be displayed in real time based on the output of the machine learning estimator.

[0041] In step 702, the sample is loaded into the charged particle microscope and imaging parameters are set. These imaging parameters may include parameters for the charged particle beam and scanning parameters. For example, one or more FOVs may be selected, and the residence time of the charged particle beam at each sample location as the FOV is scanned may be determined. The FOV may be selected based on SEM or BSE imaging. The residence time may be determined based on the sample type. Furthermore, the sample type or potential sample composition may be set by the operator in the charged particle microscope to initialize a machine learning estimator.

[0042] In 704, the machine learning estimator may optionally be initialized based on one or more high-density spectra collected from the sample. For example, a sample image, such as an SEM or BSE image, may be acquired in 705 and displayed to the operator. The operator may select one or more regions or sample locations on the sample image to collect high-density spectra. The selected sample locations may be regions having different sample compositions. In 706, a charged particle beam, such as an electron beam, is directed to the selected sample locations and an EDS spectrum is acquired. Since the residence time at the selected sample locations is longer than the residence time for scanning the FOV determined in 702, a high-density spectrum is acquired as a result. The high-density spectrum may be used to initialize the estimator in step 302 of method 300.

[0043] In 708, multiple sample locations are repeatedly scanned with a charged particle beam. If a spectrum is detected at a particular sample location, the spectrum 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. For example, if the total count from the detector output at a sample location is lower than a threshold count, the spectrum is not detected at that location. Based on the received spectra, the machine learning estimator outputs the abundance of each spectrum analyzed and updates the spectral components after analyzing a batch of spectra.

[0044] In step 710, an image showing the structure and composition information of the sample may be displayed. For example, as shown in Figure 2, one or more of the following may be displayed based on the output of the machine learning estimator: a component map, spectral components, a phase map, and a quantified image. Steps 708 and 710 may be performed in parallel so that the image is displayed while data acquisition is in progress.

[0045] In step 712, method 700 checks whether more scans are needed. In one example, if the sample has been scanned a predetermined number of times, no additional scans are needed. In another example, the operator may determine whether more scans are needed based on the image displayed in step 710. If further scans are needed, method 700 continues scanning the sample with the charged particle beam. Otherwise, method 700 proceeds to step 714. In step 714, the decomposed data from the output of the machine learning estimator and the EDS data cube are saved.

[0046] In this way, an image representing the sample composition may be formed and displayed in real time based on the decomposed data. The technical effect of modeling spectral data using LDA is that the LDA model describes the multinomial distribution of the spectrum. The technical effect of smoothing spectral components based on known properties of the spectral data is to adjust the LDA inference to produce spectral components that resemble the properties of the spectral data. The technical effect of outputting abundances and spectral components after processing a batch of received spectra is that a compositional sample image may be displayed during data acquisition. Furthermore, the computational intensity is reduced.

[0047] In one example, the method includes repeatedly scanning multiple sample locations of a sample with a charged particle beam, acquiring multiple spectra in response to detecting radiation from the multiple sample locations via a detector, and decomposing the acquired spectra into abundances and multiple spectral components using a machine learning estimator based on the characteristics of the detector, with each acquired spectrum corresponding to one of the abundances. A first example of the method, in which the acquired spectra are modeled using Latent Dirichlet Allocation (LDA), further includes updating a state variable of the LDA model for each of the acquired spectra, and updating the multiple spectral components after updating the state variable a predetermined number of times. A second example of the method optionally includes the first example and further includes generating one or more component maps of the sample, with each component map being generated based on the abundance values ​​corresponding to the components. A third example of the method optionally includes one or more of the first and second examples and further includes clustering the abundances into one or more phases, and generating one or more phase maps based on the clustered abundances. A fourth example of the method optionally includes one or more of the first to third examples, and further includes generating concentrations of the chemical elements corresponding to each spectral component of the multiple spectral components by quantifying the multiple spectral components based on the theoretical spectra of the chemical elements, and determining the concentrations of the chemical elements at each of the multiple sample locations by combining the concentrations of the chemical elements of the multiple spectral components and the abundances at the sample locations. A fifth example of the method optionally includes one or more of the first to fourth examples, and further includes smoothing the multiple spectral components based on the characteristics of the detector, and the method further includes updating a machine learning estimator based on the multiple spectral components.

Claims

1. Irradiating a sample with a charged particle beam, By energy-dispersive X-ray spectroscopy, a spectrum is obtained by detecting the emissions from the sample in response to the irradiation via a detector. A method comprising processing the acquired spectrum with a machine learning estimator based on the characteristics of the detector, decomposing it into a plurality of spectral components, and generating the abundance of sample locations corresponding to the spectrum.

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

3. Updating the machine learning estimator based on the aforementioned plurality of spectral components, A second spectrum is obtained by detecting the emission from the sample via the detector. The method according to claim 1 or 2, further comprising decomposing the acquired second spectrum with the updated machine learning estimator based on the characteristics of the detector.

4. The method according to any one of claims 1 to 3, wherein the characteristics of the detector include the energy-dependent characteristics of the detector, and the decomposition of the acquired spectrum based on the characteristics of the detector includes adjusting each of the plurality of spectral components based on the energy-dependent characteristics of the detector.

5. The method according to claim 4, wherein the energy-dependent characteristics of the detector include the peak of the detected emission at lower energies being narrower than the peak of the detected emission at higher energies.

6. The method according to any one of claims 1 to 5, wherein the abundance is a 1D array, and each value of the abundance corresponds to a component and a spectral component among the plurality of spectral components.

7. The method according to any one of claims 1 to 6, wherein the acquired spectrum is modeled using a latent Dirichlet allocation method, and the machine learning estimator is constructed based on stochastic variational Bayesian estimation.

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

9. Select one or more sample locations, Irradiating the charged particle beam to the selected location, and acquiring one or more high-density spectra for a residence time longer than the residence time for acquiring the spectrum, wherein the high-density spectrum is denser than the spectrum obtained. The method according to claim 8, further comprising determining the known spectral components based on the one or more high-density spectra.

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

11. Based on the theoretical spectrum of a chemical element, one or more of the aforementioned spectral components are quantified. The machine learning estimator is updated based on the quantified spectral components, A second spectrum is obtained by detecting the emission from the sample. The method according to any one of claims 1 to 10, further comprising decomposing the acquired second spectrum with the updated machine learning estimator and the characteristics of the detector.

12. The method according to claim 11, further comprising generating one or more component maps of the sample, wherein each component map is generated based on the abundance values ​​corresponding to the components.

13. It is a charged particle microscope, A source for generating a charged particle beam, A scanner for scanning the charged particle beam over a sample, A detector for detecting emissions from the sample in response to the charged particle beam irradiated onto the sample, A controller, which, when executed, the controller, Irradiating the sample with the charged particle beam, Based on the detected emissions from the sample, an energy-dispersive X-ray spectrum is obtained, A charged particle microscope comprising: a controller including a non-temporary memory for storing computer-readable instructions for processing the acquired spectrum using a machine learning estimator based on the characteristics of the detector, by decomposing it into a plurality of spectral components and generating the abundance of sample locations corresponding to the spectrum; and

14. The charged particle microscope according to claim 13, wherein the source is an electron source and the detector is an X-ray detector.

15. The machine learning estimator is based on stochastic variational Bayesian estimation for Latent Dirichlet Allocation (LDA), and the controller is based on this. Based on the aforementioned plurality of spectral components, the machine learning estimator is updated. By detecting the emission from the sample, a second spectrum is obtained. The charged particle microscope according to claim 13 or 14, wherein the updated machine learning estimator is further configured to decompose the acquired second spectrum.

16. The charged particle microscope according to any one of claims 13 to 15, wherein the controller is further configured to generate one or more component maps of the sample, each component map being generated based on the abundance values ​​corresponding to the components.

Images