Method and system for determining sample composition from spectral data

Abstract

Methods and systems for determining sample composition from spectral data. Methods and systems are disclosed for determining sample composition from spectral data acquired by a charged particle microscope system. Chemical elements in a sample are identified by processing the spectral data with a trained neural network (NN). If the identified chemical elements do not match a known elemental composition of the sample, the trained NN is retrained with the spectral data and the known elemental composition of the sample. The retrained NN can then be used to identify chemical elements within other samples.

Description

Technical Field

[0001] This specification generally relates to methods and systems for determining the composition of a sample from spectral data, and more specifically, to methods and systems for determining the composition of a sample based on spectral data acquired in response to irradiation of a sample with a beam of charged particles. Background Technology

[0002] Charged particle microscopy is a well-known and increasingly important technique for imaging microscopic objects. 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 in response to electron beam irradiation, energy-dispersive X-ray spectroscopy (EDS or EDX) can be used for elemental analysis or chemical characterization.

[0003] One method for determining chemical elements in a sample involves comparing each detected energy spectrum with a known spectrum of the chemical element. For example, a chemical element can be identified by comparing the peak positions in the detected spectrum with the known peak positions of each chemical element. However, the applicant recognizes that the identified chemical elements may include a large number of false positives and / or false negatives, especially when the detected spectrum is sparse or contains interfering peaks. Summary of the Invention

[0004] In one embodiment, a method includes: irradiating a first sample with a beam of charged particles; detecting a first type of emission from the first sample and forming one or more first spectra from the detected first type of emission; identifying one or more first chemical elements within the first sample by processing the first spectra with a trained neural network; detecting a second type of emission from the first sample and displaying a sample image generated based on the detected second type of emission; selecting one or more pixels in the sample image; displaying one or more chemical elements corresponding to the selected pixels; retraining the trained neural network with the spectrum corresponding to the selected pixels and the known element composition in response to the displayed chemical elements that differ from the known element composition; irradiating one or more locations of a second sample with a beam of charged particles; acquiring one or more second spectra by detecting the first type of emission from the second sample; and identifying one or more second chemical elements within the second sample by processing the second spectra with the retrained neural network. In this manner, chemical elements can be identified rapidly and with high accuracy.

[0005] It should be understood that the concepts introduced in a simplified form in the embodiments provided above are chosen for this purpose. They are not intended to identify key or essential features of the claimed subject matter, the scope of which is solely defined by the claims following the embodiments. Furthermore, the claimed subject matter is not limited to embodiments that address any shortcomings pointed out above or in any part of the invention. Attached Figure Description

[0006] Figure 1 Demonstrating a charged particle microscope.

[0007] Figure 2 This is a flowchart of an example method for determining the composition of a sample from spectral data.

[0008] Figure 3 This is a flowchart of an instance method for retraining a trained neural network for elemental analysis.

[0009] Figure 4 This displays an instance user interface for showing element analysis results and retraining the trained neural network.

[0010] Figure 5 illustrate Figure 2 The instance data stream of the methods in the code.

[0011] Figure 6 This is a flowchart of a method for generating an instance of a trained neural network used for elemental analysis.

[0012] Throughout the diagram, several views are represented by similar reference numbers, which refer to the corresponding parts. Detailed Implementation

[0013] The following description relates to methods and systems for determining the composition of a sample based on spectral data, such as energy spectra obtained by detecting emission in response to irradiation of the sample with a beam of charged particles. In US Patent 9048067, Owen discloses a method for identifying minerals by sequentially decomposing spectral data into elements in a mineral definition. Decomposition determines the proportion of chemical elements in an unknown mineral by fitting the known elemental spectra from a periodic table to the detected X-ray spectra, wherein the periodic table is obtained by selecting mineral definitions from a mineral database. The applicant recognizes that accurate compositional analysis requires accurate periodic tables. When the acquired X-ray spectra of an unknown sample are sparse (i.e., the total number of counts for each spectrum is low), periodic tables may be inaccurate. In sparse spectra, it is difficult to accurately locate peaks or distinguish peaks that are close to or interfering with each other. Increasing the number of X-ray counts (and thus, a denser spectrum) improves the accuracy of element identification. However, it can also significantly increase the total data acquisition time. Furthermore, due to factors including detector aging and different operating conditions, the detected X-ray spectra of the same mineral under different microscopes may differ, so the known elemental spectra may not be properly fitted to the detected X-ray spectra.

[0014] To address the aforementioned problems, a trained neural network (NN) adapted to the system operating conditions of a specific microscope is used to perform elemental analysis. In one instance, after a first sample is loaded into the microscope, a beam of charged particles irradiates one or more locations on the first sample. A first type of emission, such as X-rays, in the form of an energy spectrum is acquired from each location, where each energy spectrum corresponds to a sample location. The trained NN receives each of the acquired energy spectra as input and outputs the chemical element identified in the spectrum and the probability of each identified chemical element. After processing the acquired energy spectra, a first table of chemical elements including those in the first sample can be generated. A second type of emission from the sample can be detected simultaneously or separately from one or more locations on the sample. A sample image can be generated based on the detected second type of emission. The sample image shows the structure of the sample. The sample image can be an optical image, a scanning electron microscopy (SEM) image, a backscattered electron (BSE) image, or a transmission electron microscopy (TEM) image. In some instances, the sample image also includes chemical element information determined from the trained NN. One or more pixels are selected in the sample image. The chemical element corresponding to the selected pixel is displayed. In some instances, the first sample is a reference sample with a known elemental composition. If the displayed chemical elements do not match the known elemental composition at the corresponding sample location, the neural network (NN) is retrained using the spectrum corresponding to the selected pixel. The retrained NN is then used to perform elemental analysis on spectral data acquired from a microscope of another sample. In this way, the retrained NN is adapted to microscope-specific parameters and operating conditions. For example, the retrained NN can be adapted to detector-induced energy shifts in the spectral data. Furthermore, the retrained NN is adapted to specific use cases. For instance, through retraining, the NN can be adapted to distinguish interference peaks specific to certain minerals, enabling more accurate identification of chemical elements.

[0015] In another example, after a first sample is loaded into a microscope, one or more locations on the sample are irradiated with a charged particle beam, such as an electron beam. A first type of emission, such as X-rays, from each location is detected and used to form an energy spectrum. A second type of emission, such as scattered electrons, is detected simultaneously or separately from the first type of emission. A sample image is generated based on the second type of emission. One or more pixels are selected from the sample image, for example, by an operator. In one example, the one or more spectra corresponding to the selected pixel are selected from the energy spectrum formed based on the first type of emission. In another example, the sample region corresponding to the selected pixel is irradiated with a charged particle beam, and the spectrum corresponding to the selected pixel is recollected. The recollected spectrum may be more compact than the first spectrum. The chemical element corresponding to the selected pixel is identified by processing the selected spectrum or the recollected spectrum with a trained neural network (NN). If the identified chemical element differs from the known elemental composition of the selected sample region, the trained NN is retrained with the selected spectrum and the known composition.

[0016] In another example, a sample image demonstrating the structure of the sample is displayed to the operator. The sample image can be any of the following: optical, SEM, BSE, or TEM images. One or more pixels in the sample image are selected automatically or by the operator. The sample region corresponding to the selected pixel is then irradiated with a beam of charged particles, and spectral data is collected by detecting emission of a first type, such as X-rays. The chemical element corresponding to the selected pixel is identified by processing the collected spectral data using a trained neural network (NN). If the identified chemical element differs from the known elemental composition of the selected sample region, the trained NN is retrained using the spectral data and the known composition.

[0017] A retrained neural network (NN) can be stored in the microscope's non-transitory memory, from which spectral data for retraining can be retrieved to process the spectral data acquired by the microscope. The retrained NN can replace the trained NN for future elemental analysis. In some instances, both the trained and retrained NNs are stored in a library. NNs in the library can serve as baseline trained NNs for further retraining in the same or different microscopes. The microscope can irradiate one or more locations of a second sample with a charged particle beam and acquire energy spectra from each of these locations. The retrained NN processes each of the acquired spectra and outputs a chemical element and its corresponding probability. Chemical elements with probabilities higher than a threshold probability can be identified as a second chemical element of the second sample. A composition map can be generated based on the second chemical element. The composition map shows the spatial distribution of chemical elements or composition in the sample. In one instance, the composition map is an elemental map showing the spatial distribution of the second chemical element. In another instance, based on the second chemical element, the acquired spectrum of the sample is further decomposed into multiple spectral components and / or abundances. Each component may correspond to a chemical element, a mixture of chemical elements, or a chemical phase. Each component corresponds to a spectral component and an abundance. A spectral component can be the spectrum of the component. Abundance can be the quantity of the component. In one instance, abundance is the ratio of the components at a given sample location. A composition plot can show the spatial distribution of the components in the sample.

[0018] In some instances, a first spectrum at multiple first locations of a sample can be obtained by scanning the sample with a charged particle beam. A trained neural network is used to identify the chemical elements of the sample based on the first spectrum. The first spectrum can then be decomposed into various components based on the identified chemical elements. Alternatively, the abundance of each component can be determined via the decomposition process. In some instances, after identifying the chemical elements in the sample, a second spectrum at multiple second locations of the sample is obtained by scanning the sample with a charged particle beam. The first spectrum may be more dense than the second spectrum; that is, the number of counts for each element in the first spectrum is higher than the number of counts for each element in the second spectrum. Alternatively, the second sample locations may differ from the first sample locations. For example, the second sample locations may cover a larger area and / or have higher spatial resolution than the first sample locations. The second spectrum can be decomposed based on the identified chemical elements to generate a composition map.

[0019] In some instances, after the trained neural network (NN) identifies the chemical elements of a reference sample, the identified chemical elements are displayed to the operator or user, for example, in the periodic table. The operator can input the known composition / chemical elements of the reference sample into the microscope to retrain the trained NN. In another instance, the sample image is displayed simultaneously with the identified chemical elements. The sample image may show the structure of the reference sample. Alternatively or additionally, the sample image may show the spatial distribution of the identified chemical elements in the reference sample. The sample image may include all scanned sample locations or a subset of sample locations. The operator can select one or more regions in the sample image and retrain the NN with an integrated spectrum and the known elemental composition of the selected region. A region may be a line, area, or point in the sample image. In another instance, the integrated spectrum is displayed together with the sample image and the identified chemical elements in, for example, the same window. The integrated spectrum is the sum of all spectra in the selected region. By selecting a region of the sample and displaying the integrated spectrum, the operator can select a region that accurately reflects the known elemental composition. For example, the operator can select a region without abrupt changes in the distribution of chemical elements, or a region that does not include sample boundaries. In another instance, the operator can select regions with overlapping peaks and retrain the NN to better identify or distinguish the peaks.

[0020] In some instances, after acquiring spectral data from a reference sample, the trained neural network (NN) can be automatically retrained using the known elemental composition of the reference sample. Regions in the sample image are automatically selected instead of the operator's selection. The trained NN can then be retrained based on the spectral data from the selected region. In some instances, a sample image demonstrating the sample structure can be displayed to the operator. The operator can select the location where spectral data is acquired in the sample image. The sample image can be an optical image, SEM image, BSE image, or TEM image.

[0021] In some instances, trained neural networks (NNs) are generated by training a primary NN with training data that includes simulated and / or experimental data. Simulated training data can be generated based on a variety of chemical elements and their known spectra. Simulated training data can include individual chemical elements and their spectra, as well as combinations of chemical elements and corresponding combined spectra. For example, the combination of chemical elements might be elements found in a mineral, and the combined spectrum might be the spectrum of the mineral. Experimental data can be generated by imaging a reference sample with a known composition using one or more microscopes. Different trained NNs can be generated for different micromanipulation conditions. Operating conditions can include one or more of the following: accelerating voltage, beam current, offset angle, working distance, and detector settings (e.g., bias voltage).

[0022] A neural network (NN) can be trained offline to identify many (e.g., 100) chemical elements or combinations thereof. However, a trained NN may be insensitive to specific combinations of chemical elements, especially when the peaks of individual chemical elements interfere with each other. Retraining a trained NN adjusts or adapts it to be more sensitive to the specific types of materials / minerals of interest to the operator, thus resulting in more accurate identification.

[0023] In some instances, charged particle beams include electron beams and ion beams. Energy spectroscopy includes X-ray spectroscopy and electron energy loss spectroscopy. In other instances, elemental analysis can be applied to analyze other types of spectral data, such as Raman spectroscopy.

[0024] Turning Figure 1 , Figure 1 A highly schematic depiction of an embodiment of a charged particle microscope (CPM) in which the present invention is implemented; more specifically, it illustrates an embodiment of a scanning electron microscope (SEM) system. The system axis is shown as axis 110. Microscope 100 includes a particle beam 1 that generates a beam of charged particles 3 (in this case, an electron beam) propagating along the 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 contains a sample holder 7 and associated actuators 8 for holding / positioning a sample 6. The vacuum chamber 5 is evacuated using a vacuum pump (not depicted). Vacuum ports 9 are also depicted, which can be open to introduce articles (components, samples) into / remove articles (components, samples) from the interior of the vacuum chamber 5. Microscope 100 may include multiple such ports 9 as needed.

[0025] Column 1 (in the context of this invention) includes an electron source 10 and an illuminator 2. The 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. The display unit may also function as an input unit for receiving operator input.

[0026] Detectors 19 and 21 are selected from a variety of possible detector types that can be used to detect different types of "stimulated" radiation emitted from sample 6 in response to irradiation by (irradiation) beam 3. Detector 19 can be a solid-state detector (e.g., a photodiode) for detecting cathodic emission emitted from sample 6. For example, it can alternatively be an X-ray detector, such as a silicon drift detector (SDD) or a lithium silicon (Si(Li)) detector. For example, detector 21 can be an electron detector in the form of a solid-state photomultiplier (SSPM) or a evacuated 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 over sample 6. Because this type of stimulated radiation 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 (a portion thereof), which is essentially a graph of the signal varying with the position of the scan path on sample 6.

[0027] 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, false shading, edge enhancement, and other processes known to those skilled in the art. Additionally, automated identification processes (e.g., for particle analysis) may be included in such processing. The controller includes processor 28 and non-transitory memory 29 for storing computer-readable instructions. The methods disclosed herein can be implemented by executing computer-readable instructions stored in the non-transitory memory 29 within processor 28.

[0028] It should be noted that many improvements and alternatives to this type of setup are known to those skilled in the art, such as using a controlled environment within the microscope 100 (which has a relatively large volume), for example, maintaining a background pressure of several millibars (as used in ambient SEM or low-pressure SEM).

[0029] Figure 2 This demonstrates a method 200 for extracting sample composition from spectral data. In one example, the spectral data can be used... Figure 1 The EDS spectrum is acquired by the microscope 100. The trained neural network (NN) performs elemental analysis to identify the chemical elements in the sample. The trained NN can be adapted to the microscope 100 and specific sample types by retraining (or updating) the trained NN with measurements from samples with known elemental compositions.

[0030] At position 202, load the sample into the microscope. At position 204, set the system parameters. System parameters include operating conditions, such as one or more of the following: accelerating voltage, beam current, working distance, and detector settings. Detector settings may include the integration time at each sample location. System parameters may also include, for example, scanning parameters for the scan path and scan area.

[0031] In some instances, sample images showing the structure or dimensions of the sample can be displayed to the operator. The sample image can be an EM image. The operator can select a scan area within the sample image. The operator can also move the sample relative to the charged particle beam to move it into the current field of view (FOV) for easier selection of the scan area.

[0032] At position 206, one or more locations on the sample are scanned using a beam of charged particles. The scanned sample locations are determined at position 204. At each sample location, one or more types of charged particles emitted from the sample are acquired by one or more detectors. The acquired charged particles may include electrons and X-rays. An energy spectrum, such as an X-ray spectrum, is obtained for each sample location. The acquired charged particles may also include backscattered electrons or secondary electrons.

[0033] At 208, method 200 determines whether elemental analysis of the spectral data acquired at 206 needs to be performed. In one instance, if the current sample is a reference sample used to retrain the trained neural network, then elemental analysis is required. In another instance, if elemental analysis has already been performed on the sample, then elemental analysis is not required. In yet another instance, if the elemental composition of the sample is known, for example when retraining the trained neural network with a reference sample having a similar composition to the current sample, then elemental analysis is not required. If elemental analysis is not required, then method 200 moves to 216 and decomposes the spectral data based on the known elemental composition. If elemental analysis is required, then method 200 moves to 210 to perform elemental analysis on the spectral data.

[0034] At position 210, chemical elements within the sample are optionally identified using a trained neural network (NN). Based on the operating conditions set at position 204, a trained NN can be selected from a group of trained NNs, such as a baseline NN. The trained NN can be a convolutional NN or a network used for natural language processing. The trained NN analyzes one or more of the acquired spectra and outputs a table of chemical elements and their corresponding probabilities for each spectrum. Chemical elements identified at a specific sample location are those with a probability equal to or higher than a threshold probability. For example, the threshold probability is 0.75. In some instances, the threshold is any number between 0 and 1, including 1. In some instances, elemental analysis can be performed on a subset of the acquired spectra.

[0035] At 212, method 200 determines whether the neural network (NN) needs retraining. If the trained NN has not yet been retrained using spectral data acquired from the current microscope, then retraining may be necessary. If the NN has already been retrained using spectral data acquired from a microscope of a reference sample with similar composition, then retraining may not be necessary. If retraining is not required, method 200 proceeds to 216. If retraining is required, method 200 proceeds to 214 to retrain the trained NN. Details of the retraining process are shown in... Figure 3 middle.

[0036] After retraining the trained neural network at position 214, the retrained neural network can be saved. The retrained neural network can be used to analyze samples with similar composition or spectral data acquired under similar operating conditions. In some instances, the spectral data used to retrain the trained neural network can be further decomposed at position 216.

[0037] At positions 216 and 206, the spectral data are decomposed based on elemental information. This elemental information may be known or predetermined and is loaded into the controller based on the sample type. Alternatively, the elemental information may be a chemical element identified at position 210. In one instance, each of the acquired spectra is decomposed into one or more spectral components and / or the abundance of each component. Each component corresponds to a spectral element and an abundance. Each component may be an element, a mixture of elements, or a chemical phase. The components may include various combinations of known or identified elements. One method of decomposing spectral data is illustrated in U.S. Application 17 / 166,885, filed February 3, 2021, by Petr Hlavenka et al., and is incorporated herein by reference in its entirety.

[0038] At position 218, one or more composition images are generated. Composition images include elemental maps showing the distribution of the identified chemical elements in the sample. Composition images also include images generated based on the spectral components and / or abundances obtained at position 216. In the composition images, compositional data can be overlaid on the structural images of the sample. Structural images can be generated based on scattered charged particles, such as scattered electrons, obtained at position 206.

[0039] At 220, method 200 determines whether additional scans of the sample are needed. Additional scans may be required to obtain more spectral data for retraining. In some instances, multiple first sample locations may be scanned in the first scan to determine elemental composition and / or retrain the trained neural network. The sample may be rescanned in the second scan, where multiple second sample locations may be scanned to generate a compositional image. The system parameters in the second scan may differ from those in the first scan. In one instance, the second scan may have a higher spatial resolution than the first scan. In another instance, multiple first sample locations may differ from second sample locations. The number of sample locations in the second scan may be greater than the number of sample locations in the first scan. In yet another instance, the number of counts for each spectrum used in the second scan may be less than that in the first scan. If additional scans are required, method 200 proceeds to 204 to reset the system parameters. Otherwise, method 200 terminates.

[0040] In this manner, elemental analysis is performed using a trained neural network (NN). Based on measurements of one or more reference samples with known elemental compositions using a microscope, the trained NN can be retrained, specifically for the microscope, operating conditions, and sample type. Once retrained, the NN can be stored and used for elemental analysis of other similar samples. Based on the identified chemical elements from the elemental analysis, the spectral data can be decomposed to obtain the spectral components and abundances at each sample location, which can be used to generate additional compositional images to provide a deeper insight into the sample composition.

[0041] Figure 3 A method 300 for retraining a trained neural network is demonstrated. The trained neural network can be retrained using an integrated spectrum generated from the acquired spectral data and a known element composition corresponding to the integrated spectrum. Method 300 allows interaction with an operator via a user interface to receive the known elements.

[0042] At position 302, a sample image demonstrating the sample structure is displayed. This can be based on... Figure 2 Scattered electrons acquired during the scan at position 206 produce an image of the sample. Alternatively, with Figure 2 The sample image is acquired separately by scanning at position 206. The sample image can be an optical image, SEM image, BSE image, or TEM image. In one example, if at... Figure 2 If elemental analysis is performed in 210, the sample image can be an elemental map showing the spatial distribution of the identified chemical elements in the sample. The elemental map can be generated based on both the spatially resolved scattered electrons and the identified chemical elements. For example, the signal from the scattered electrons can be displayed in grayscale to show the structural information of the sample. The identified elements can be color-coded and overlaid on the grayscale image.

[0043] In some instances, Figure 2The 210 identified chemical elements are displayed along with the sample image. In one instance, the identified chemical elements may be displayed in the periodic table. In another instance, the identified chemical elements are displayed as a list.

[0044] At 304, select one or more regions in the sample image. The operator can select regions after observing the elemental map. Alternatively, regions can be selected automatically. For example, the central region of the sample can be automatically selected. The selected region contains one or more pixels of the sample image. In one instance, the selected region may be a region reflecting the known composition of the sample. In another instance, the selected region corresponds to a characteristic spectrum, for example, one with overlapping peaks.

[0045] At 308, the charged particle beam is optionally directed to a sample region (or pixel) in the sample image corresponding to a selected region to collect additional spectral data, which is then analyzed using a trained neural network to identify chemical elements in the selected region. System parameters can be adjusted to collect the additional spectral data. In one example, the detector integration time is longer compared to the integration time at 206, resulting in a denser spectrum. In another example, the scanning sequence is adjusted.

[0046] At position 310, the chemical element identified in the selected region is displayed at position 310. Chemical elements can be identified at either position 210 or 310. In one example, the chemical element is displayed by updating the identified chemical element at position 306. In another example, the chemical element identified at position 310 can be displayed in the periodic table or as a list. One or more of the integrated spectrum, the identified chemical element, and the sample image can be selectively displayed individually or together on the display unit.

[0047] At position 312, an integrated spectrum is generated based on the acquired spectra within the selected region. For example, one or more spectra acquired from sample locations within the selected region are summed or averaged to produce the integrated spectrum. In some instances, the integrated spectrum is displayed to the operator.

[0048] At 314, the identified chemical element in the selected region is compared with the known elemental composition in the selected region. The comparison can be performed via the microscope's controller or by the operator. If the identified chemical element does not match the known element, then method 300 proceeds to 316 to receive the known elemental composition from user input. Otherwise, method 300 proceeds to 322.

[0049] At point 316, the known elemental composition from the selected area is received via user input. For example, at point 310, the operator can select or deselect the displayed chemical elements.

[0050] At point 318, the trained neural network (NN) is retrained using integrated spectra and the received known composition information. In some instances, the retrained NN can be validated using spectral data not used for retraining. In one instance, the trained NN is replaced by and stored with the retrained NN. In some instances, one or more retrained NNs are stored as baseline NNs. The baseline NN can be used as a starting point for retraining, for example, for different samples or different microscope systems.

[0051] At position 320, the acquired spectral data can optionally be reprocessed with the retrained neural network to produce an updated elemental map. The updated elemental map can be used to evaluate the accuracy of the retrained neural network.

[0052] At 322, method 300 checks whether additional retraining of the neural network (NN) with the current spectral data is required. For example, the operator can determine whether to retrain the NN based on the updated elemental map. If retraining is required, method 300 moves to 306, and the NN can be retrained using spectral data from different regions of the sample. If no additional retraining is required or if additional spectral data is needed for retraining, method 300 terminates.

[0053] By receiving chemical elements selected by the operator, new training data that has not yet been used to train the NN offline can be obtained to fine-tune the NN to be more sensitive to specific user cases and / or specific compositions of interest.

[0054] Figure 4 This example demonstrates a user interface for displaying results from elemental analysis and receiving user input for retraining a trained neural network. An elemental diagram 402, a periodic table 403 showing the identified chemical elements, and an integrated spectrum 404 are simultaneously displayed in a window 401 on the display unit. The elemental diagram displays the identified elements in color-coded overlays on an electron micrograph of the sample. The identified elements Fe, Zn, and Sn are highlighted in the periodic table. The integrated spectrum 404 can be generated based on all spectral data received in the sample region displayed in the elemental diagram 402. Peaks in the integrated spectrum corresponding to the identified elements can be indicated in the integrated spectrum.

[0055] The operator can select one or more regions in element map 402. A region can be a line, area, or point in the sample image. During or after selection, the periodic table 403 and integrated spectrum 404 are updated with the identified elements and the updated integrated spectrum of the selected regions. If the identified elements differ from the known elemental composition, the operator can input the known elemental composition by selecting and / or deselecting elements in periodic table 403. The operator can then click the "Learn" button 405 in the window to retrain the NN with the input information.

[0056] Figure 5 This describes an instance data stream for extracting sample composition from spectral data using a trained neural network (NN). The acquired spectrum 501 is processed by the trained NN 502. For each processed spectrum, the trained NN generates a chemical element table 504 and a probability 505 for each chemical element. An instance list of chemical elements and probabilities is shown in 503. The list may contain all chemical elements in the periodic table. Probability values ​​range from 0 to 1. Chemical elements 506 in the sample contain all chemical elements identified from spectrum 501. For example, chemical element 506 includes chemical element 504 with a probability greater than a threshold probability. Based on the identified chemical element 506, each spectrum of spectrum 501 can then be decomposed into spectral components and / or abundances at 508. An instance of components and abundances at a sample location is shown in 509. In this instance, there are two components at the sample location. Each component is a combination of individual chemical elements. The first component contains 0.4% Fe and 0.6% oxygen, and the second component contains 0.33% Si and 0.66% oxygen (O). The abundance of each component is 0.5. That is, half of the composition at the sample location is the first component and the other half is the second component. An elemental map 511 can be generated based on the identified chemical elements 506 at each sample location. A compositional image 510 can be generated based on the spectral components and / or the abundance at the scanned sample location.

[0057] Figure 6 This demonstrates a method 600 for generating a trained neural network offline. The trained neural network can then be retrained online, such as... Figure 2 The method is shown in 200.

[0058] At position 602, various operating conditions are defined. Operating conditions may include one or more of the following: accelerating voltage, beam current, operating distance, and detector settings. A trained neural network can be generated for each operating condition.

[0059] At position 604, a training dataset is determined based on the operating conditions. The training dataset includes training spectra and their corresponding elemental compositions. The training spectra may include simulated data 606 and / or experimental data 608 obtained from one or more microscopes. The training spectra may include characteristic spectra of individual chemical elements and / or spectra of combinations of chemical elements.

[0060] At 610, the training data is augmented by introducing noise, nonlinearity, and adjusted gain. In one instance, noise can be introduced into the training data by shifting one or more peaks within a specific energy range of their positions. The energy range can be less than 100 eV. In one instance, the energy range is less than 50 eV. In another instance, noise can be added to the amplitude of the training spectrum.

[0061] At position 612, a primary neural network is trained using the training dataset to produce a trained neural network. In some instances, a trained neural network is produced for each operational condition specified at position 602.

[0062] At position 614, the validation data includes validation spectra, and corresponding elemental compositions are generated. The trained neural network (NN) is validated using the validation data. The validation spectra differ from the training spectra. For example, validation spectra may include experimental spectra obtained from different microscopes. Elemental analysis is performed on the validation spectra using the trained NN. The performance of the trained NN is determined by comparing the elemental compositions in the validation data with the chemical elements identified by the trained NN.

[0063] At position 616, if the performance of the trained NN is acceptable, then save the trained NN at position 620. Otherwise, generate an additional training dataset at position 618 to continue training the NN.

[0064] The advantages of using neural networks (NNs) for elemental analysis include: NNs can perform elemental analysis on spectral data, especially sparse spectral data, with high speed and accuracy. The advantage of retraining a trained NN by measuring a reference sample is that an offline-trained NN can be adapted to specific use cases, systems, and materials. The advantage of selecting a region of a reference sample for retraining is that it allows selection of sample regions that reflect the known elemental composition. The advantage of performing elemental analysis before decomposing the spectral data is that it allows for rapid and accurate identification of sample components.

[0065] In the representation, the non-transitory computer-readable medium contains instructions that, when executed by a processor of the controller, cause the controller to perform the following operations: irradiate one or more locations of a first sample with a beam of charged particles; acquire one or more first spectra by detecting a first emission from the first sample in response to irradiation of the first sample; determine one or more first chemical elements within the first sample by processing the first spectra with a trained neural network stored in the computer-readable medium; display one or more first chemical elements; receive a known chemical elemental composition of the first sample; retrain the trained neural network with one or more of the first spectra and the known elemental composition of the first sample; irradiate one or more locations of a second sample with a beam of charged particles; acquire one or more second spectra by detecting a first emission from the second sample in response to irradiation of the second sample; determine one or more second chemical elements within the second sample by processing the second spectra with the retrained neural network; and generate a compositional map of the second sample based on the second chemical elements.

Claims

1. A method for determining the composition of a sample, comprising: The first sample was irradiated with a beam of charged particles. Detect a first type of emission from the first sample, and form one or more first spectra based on the detected first type of emission; One or more first chemical elements in the first sample are identified by processing the first spectrum with a trained neural network. Detect a second type of emission from the first sample and display a sample image generated based on the detected second type of emission; Select one or more pixels from the sample image; Displays one or more chemical elements corresponding to the selected pixel; In response to chemical elements that are different from known elemental compositions, the trained neural network is retrained with the spectrum corresponding to the selected pixel and the known elemental composition. Irradiate one or more locations of the second sample with the charged particle beam; One or more second spectra are obtained by detecting the first type of emission from the second sample; and One or more second chemical elements in the second sample are identified by processing the second spectrum with a retrained neural network.

2. The method of claim 1, wherein identifying one or more first chemical elements in the first sample by processing the first spectrum with a trained neural network comprises: One or more of the first spectra are input into the trained neural network; The trained neural network outputs one or more chemical elements and probabilities, where each chemical element corresponds to a probability; and The chemical element whose probability is greater than a threshold probability is selected as the first chemical element.

3. The method according to any one of claims 1 to 2, further comprising generating an elemental map showing the spatial distribution of the second chemical element in the second sample.

4. The method according to any one of claims 1 to 2, further comprising decomposing each of the second spectrum into one or more spectral components and abundances based on the second chemical element, wherein each component corresponds to an abundance.

5. The method of claim 4, further comprising generating a compositional map of the second sample based on the spectral components and / or the abundance.

6. The method of claim 5, wherein each spectral component is a spectrum comprising one or more of the second chemical elements.

7. The method according to claim 1, wherein, The sample image is further displayed based on the first chemical element, wherein the sample image shows the spatial distribution of the first chemical element.

8. The method of claim 1, further comprising generating the trained neural network by training the neural network with training spectra of a plurality of chemical elements.

9. The method of claim 8, wherein the training spectrum comprises simulated spectrum and / or experimental spectrum.

10. The method of claim 1, further comprising receiving the known elemental composition of the first sample via user input.

11. The method of claim 10, wherein retraining the trained neural network using the spectrum corresponding to the selected pixel and the known element composition comprises: An integrated spectrum is generated from the spectrum corresponding to the selected pixel, and the trained neural network is retrained using the integrated spectrum and the known elements.

12. The method of claim 11, further comprising simultaneously displaying the integrated spectrum with one or more chemical elements of the first chemical element corresponding to the selected pixel.

13. The method of claim 11, wherein displaying the first chemical element on a display comprises displaying the first chemical element in a periodic table, and wherein the user input comprises selecting or deselecting the first chemical element displayed in the periodic table.

14. The method of claim 1, further comprising: After selecting one or more pixels in the sample image, the charged particle beam is guided to one or more sample regions corresponding to the selected pixels, and one or more spectra of the sample regions are obtained; and The one or more chemical elements corresponding to the selected pixels are identified by processing the spectrum obtained using the trained neural network.

15. A charged particle microscope system comprising: A sample holder, used to position the sample; A charged particle source for irradiating the sample with a beam of charged particles; A first detector is used to detect a first type of emission from the sample in response to irradiation by a beam of charged particles; A second detector is used to detect a second type of emission from the sample in response to irradiation by the charged particle beam; and A controller includes a non-transitory computer-readable medium storing instructions that, when executed by the controller's processor, cause the controller to perform the following operations: The first sample was irradiated with the charged particle beam; Detect the first type of emission from the first sample, and form one or more first spectra based on the detected first type of emission; Detect the second type of emission from the first sample and display a sample image generated based on the detected second type of emission; Receive a selection of one or more pixels from the sample image; Obtain one or more spectra corresponding to the selected pixel; One or more first chemical elements in the first sample are identified by using the spectrum obtained by processing with a trained neural network; In response to a first chemical element with a different elemental composition than the known elemental composition, the trained neural network is retrained using a selected spectrum and the known elemental composition; and The retrained neural network is stored in the non-transitory computer-readable medium.

16. The charged particle microscope system of claim 15, wherein the non-transitory computer-readable medium includes further instructions that, when executed by the processor of the controller, cause the controller to perform the following operations: Irradiate one or more locations of the second sample with the charged particle beam; One or more second spectra are obtained by detecting the first type of emission from the second sample; and One or more second chemical elements in the second sample are identified by processing the second spectrum with a retrained neural network.

17. The charged particle microscope system of claim 16, wherein the non-transitory computer-readable medium includes further instructions that, when executed by the processor of the controller, cause the controller to: detect the second type of emission from the second sample; and, based on the detected second type of emission, generate a composition map showing the distribution of the second chemical element in the second sample.

18. The charged particle microscope system of claim 17, wherein generating the composition map of the second sample based on the second chemical element comprises: Based on the second chemical element, the second spectrum is decomposed into one or more spectral components and the abundance of each spectral component. The composition map of the second sample is generated based on the spectral components and / or the abundance.

19. The charged particle microscope system according to any one of claims 15 to 18, wherein the charged particle beam is an electron beam, the first type of emission is X-rays, and the second type of emission is scattered electrons.

20. The charged particle microscope system of claim 15, wherein the trained neural network receives spectra and outputs one or more chemical elements and the probability of each chemical element.

21. The charged particle microscope system of claim 20, wherein determining one or more first chemical elements in the first sample by processing the first spectrum with a trained neural network includes selecting the chemical element output from the trained neural network as the first chemical element with a probability higher than a threshold probability.

22. The charged particle microscope system of claim 15, wherein obtaining one or more spectra corresponding to the selected pixel comprises selecting the one or more spectra from the first spectrum corresponding to the selected pixel.

23. The charged particle microscope system of claim 15, wherein obtaining one or more spectra corresponding to the selected pixel comprises: The charged particle beam is directed to the sample region corresponding to the selected pixel; And to obtain the one or more spectra by detecting the first type of emission from the sample region.

24. A non-transitory computer-readable medium comprising instructions that, when executed by a processor, cause a computing device to perform the following operations: Access to first detector data corresponding to a first type of emission from a sample, wherein the first detector data is obtained in response to irradiating the sample with a beam of charged particles; This causes the display of a sample image generated based on the data from the first detector; Receive a selection of one or more pixels from the sample image; Accessing second detector data corresponding to a second type of emission from one or more sample regions corresponding to a selected pixel, wherein the second detector data is obtained in response to irradiating the sample regions with the charged particle beam; One or more chemical elements in the sample are identified by processing the data from the second detector using a trained neural network. This causes the identified chemical element to be displayed; In response to a displayed chemical element that differs from the known elemental composition, the trained neural network is retrained using the second detector data and the known elemental composition. and The retrained network is stored in the non-transitory computer-readable medium.

Images