Method and device for processing multidimensional microscopy data for interface positioning between structurally homogeneous areas of a material sample

Abstract

Method and device for processing multidimensional microscopy data for interface positioning between structurally homogeneous areas of a material sample. This device is configured to implement, for the processing of at least one microscopy image of said sample forming an input data block, modules for: -normalization (20) of the input data block to obtain a normalized data block, the normalization including an adjustment of the contrast of the input data block; -segmentation (24) of the normalized data block allowing to determine a first data block representative of the first homogeneous area and a second data block representative of the second homogeneous area, such that the point-by-point sum of the values ​​of the first data block and the second data block is equal to the same value;-calculation (26) of a standard deviation data block between said first and second data blocks and thresholding of the standard deviation data block, to obtain the structural interface between the first homogeneous zone and the second homogeneous zone. Figure for the abstract: Figure 1;

Description

Title of the invention: Method and device for processing multidimensional microscopy data for interface positioning between structurally homogeneous areas of a material sample

[0001] The present invention relates to a method for processing multidimensional microscopy data for interface positioning between structurally homogeneous areas of a material sample, as well as an associated multidimensional microscopy data processing device and an associated computer program.

[0002] The invention lies in the field of processing multidimensional microscopy data, obtained by observing samples composed of one or more materials, for the analysis of chemical and physical properties of their structures.

[0003] The invention relates more particularly to automated metrology using specialized algorithms.

[0004] More particularly, the invention finds applications in the localization of chemical and / or structural interfaces in materials, for example applied in quality inspection in a production line of materials and devices.

[0005] By definition, an interface is the surface of discontinuity forming a common boundary between two distinct layers or two distinct zones. When the difference between the two zones is chemical in nature, then the boundary is called a chemical interface. For zones that differ in their structure, then the interface is said to be structural. For example, a reflection twin is a complex crystal, formed of simple crystals of the same species oriented differently and separated by a purely structural interface, called the twin plane.

[0006] As is well known, the interface plays a crucial role in the functioning of devices. The physics of the most advanced devices relies on interface engineering. It is therefore essential to measure and precisely position these interfaces using imaging techniques such as transmission electron microscopy.

[0007] It is useful to have efficient tools to carry out the inspection of materials, and to detect the presence of interfaces of a chemical or structural nature, in the calibration phase or in the technological optimization phase of various devices using materials.

[0008] The analysis of chemical and / or structural interfaces of materials by processing spectral (1D) and image (2D) data obtained by microscopy has been developed for this purpose. The proposed method is also valid in hyperspaces of dimension greater than 3.

[0009] Data sets of spectra and images, each representative of at least a portion of the observed sample, are obtained by any type of microscopy characterization that generates an N-dimensional data cube (or multidimensional data), where N is an integer greater than or equal to 2. This data cube contains droplets, also called spots, spikes, or blobs, that stand out against a homogeneous background (e.g., light droplets or spots on a dark homogeneous background). These droplets are representative of the chemical and structural characteristics of the observed sample material, for example, an alignment of atoms along the direction of observation. The data cube may also contain noise. The image contrast can be related to the chemistry and structure of the sample through simulations that include instrumental parameters.

[0010] For example, when observing crystals, atoms arranged in a regular pattern classically represent the crystal lattice. The simplest atomic model is a hard sphere model, which represents the particulate nature of an atom. The most powerful microscopes allow us to visualize atoms, which generally appear as droplets. In two dimensions, droplets are also called spots. In one dimension, droplets are points. Between the droplets are signals and interferences that arise from the essentially wave-like nature of matter.

[0011] Mathematically, a droplet is defined in this document as a simply connected component of a discrete topological space, in the sense of general topology. This means that any loop traced in a droplet can be reduced by homotopy to a point. Physically, a droplet is defined as the electrical signal produced by the pixels of a matrix detector following the impact of a particle (electron, photon, ion, fermion, etc.). The particle is always much smaller than the pixel, so the droplet can always be reduced to a point by homotopy, in accordance with the mathematical definition of a droplet. There is therefore a precise agreement between the physical and mathematical definitions of a droplet.

[0012] Images and spectra are obtained, for example, by high-resolution transmission electron microscopy (HRTEM and HRSTEM and 4D-STEM), X-ray fluorescence microscopy (HRSTEM-EDX), electron energy loss microscopy (HRSTEM-EELS and HRSTEM-VEELS), energy-filtered transmission electron microscopy (EF-HRTEM, EF-HRSTEM), atomic force microscopy (AFM), scanning tunneling microscopy (STM), atom probe tomography or electron tomography or future positron-based variants, for example. Other types Microscopies using charged particles are also concerned, such as scanning electron microscopy (SEM), focused ion beams (FIBs), dual beams (DB) which include beams of ions and electrons and / or positrons.

[0013] To obtain structural interfaces, the microscopy images obtained are atomic resolution images, i.e., images with a spatial resolution less than or equal to 2 nanometers. For chemical interfaces, this dimension constraint does not apply.

[0014] The method for determining structural interfaces presented below can also be applied to other more conventional techniques such as optical microscopy and all imaging variants more generally, for example in the space domain.

[0015] In the prior art, methods for measuring interfaces in materials observed by microscopy are known. In particular, a technique called "Fast marching level sets" is described in patent application published under number US2021 / 0263430 A1. This technique is based on the "fast marching level sets" algorithm, which propagates a mobile and deformable interface from a chosen seed until the interface encounters a chosen discontinuity. This technique has been classically used for a very long time (Forcadel, N., Le Guyader, C. & Goût, C. Generalized fast marching method: applications to image segmentation. Numer Algor 48, 189-211 (2008)) to determine the position of interfaces. Patent application US2023 / 0196189A1 by Carl Zeiss proposes 3D metrology of semiconductors in a FIB / SEM using machine learning.A potential drawback of these techniques is that they do not allow for the universal differentiation of chemical and structural interfaces. Furthermore, these methods lack atomic spatial resolution and cannot eliminate the effect of roughness on layer thickness measurements, which is essential for ultrathin layers.

[0016] The evolution of the most advanced technological nodes of the ITRS (International Technology Roadmap for Semiconductors) generates the need to locate interfaces with finer resolution, in order to better understand and optimize semiconductor manufacturing processes, as well as performance and reliability.

[0017] To this end, the invention proposes, according to one aspect, a method for processing multidimensional microscopy data for positioning interfaces between structurally homogeneous areas of a material sample, comprising acquiring at least one microscopy image of said sample forming an input data pad, each image of said input data pad being representative of a portion of the observed sample, said input data pad being represented in an N-dimensional space, N being greater than or equal to two, each data point of said pad corresponding to a point in N-dimensional space. This process involves the following steps:

[0018] -normalization of the input data pad to obtain a normalized data pad, the normalization comprising an adjustment of the contrast of the input data pad between a predetermined minimum and maximum value;

[0019] - segmentation of the normalized data block allowing the determination of a first a data block representing a first homogeneous area and a second data block representing a second homogeneous area from the normalized data block, such that the point-by-point sum of the values ​​of the first data block and the second data block is equal to the same value;

[0020] - calculation of a standard deviation data block between said first and second blocks data;

[0021] - thresholding of the standard deviation data block, to obtain the structural interface between the first homogeneous zone and the second homogeneous zone.

[0022] Advantageously, the proposed method allows for the positioning of a structural interface with very high precision, down to the individual atom, from a single full-field image acquisition, thus enabling its application with high-speed acquisition technologies. Advantageously, the proposed method is fast and efficient in terms of computational resource consumption. Furthermore, another advantage of the proposed method is that it also allows for the precise measurement of the average thicknesses of the thinnest currently measurable layers, i.e., 1 atomic monolayer thickness, even in the corners. Advantageously, the method is even applicable to hyperspaces (with dimensions greater than 3).

[0023] According to other advantageous aspects of the invention, the method for processing multidimensional microscopy data for interface positioning between structurally homogeneous areas comprises one or more of the following features, taken individually or in all technically possible combinations.

[0024] The method further comprises a step of rectifying the structural interface, comprising a calculation of a transformation enabling the transformation of said structural interface into a rectified interface, said transformation being followed by a continuous elastic deformation of the normalized data block around the rectified interface.

[0025] The calculation of a transformation is a function of a plurality of points positioned on a curve in the structural interface.

[0026] The segmentation of the normalized data block further comprises a segmentation of the normalized data block into two structural classes, respectively a first structural class and a second structural class, each class corresponding to a homogeneous structure, said segmentation being a machine learning segmentation on a structural knowledge database, said machine learning segmentation providing the first data block representative of a probability of belonging to the first structural class and the second data block representative of a probability of belonging to the second structural class.

[0027] The thresholding of the standard deviation data block implements a threshold of predetermined value preferably between 0.00001% and 1% of the maximum intensity of the standard deviation data block.

[0028] The thresholding of the standard deviation data block implements a threshold of dynamically determined value, the threshold value being the smallest value allowing a continuous interface to be obtained.

[0029] The method further comprises a calculation of a structural profile associated with the rectified interface by averaging parallel to the rectified interface.

[0030] The method further includes a calculation of an average thickness of an area of ​​interest from the structural profile.

[0031] According to another aspect, the invention relates to a computer program comprising software instructions which, when executed by a computer, implement a method for processing multidimensional microscopy data for interface positioning between structurally homogeneous areas as defined above.

[0032] The invention also relates to a multidimensional microscopy data processing device for interface positioning between structurally homogeneous areas of a material sample, comprising acquiring at least one microscopy image of said sample forming an input data pad, each image of said input data pad being representative of a portion of the observed sample, said input data pad being represented in an N-dimensional space, N being greater than or equal to two, each data point of said pad corresponding to a point in the N-dimensional space. This device is configured to implement:

[0033] -a normalization module for the input data pad to obtain a normalized data pad, the normalization comprising an adjustment of the contrast of the input data pad between a predetermined minimum and maximum value;

[0034] - a normalized data block segmentation module allowing determine a first data block representative of a first homogeneous area and a second data block representative of a second homogeneous area from the normalized data block, such that the point-by-point sum of the values the first data block and the second data block are equal to the same value;

[0035] - a calculation module for a standard deviation data block between said first and second data blocks, and standard deviation data block thresholding, to obtain the structural interface between the first homogeneous zone and the second homogeneous zone.

[0036] The invention will become clearer upon reading the following description, given solely by way of non-limiting example, and made with reference to the drawings in which:

[0037] [Fig-1] [Fig.1] is a block diagram of a detection and measurement system interfaces in a material sample comprising a multidimensional microscopy data processing device according to an embodiment;

[0038] [Fig.2] [Fig.2] is a synoptic diagram of the main steps of a treatment process for the localization of the interface between homogeneous areas of a sample according to an embodiment;

[0039] [Fig.3] [Fig.3] is an example of a two-dimensional input data tile;

[0040] [Fig.4] [Fig.4] illustrates the result obtained after the homogenization (left) and normalization (right) step on the data block of [Fig.3];

[0041] [Fig.5] [Fig.5] illustrates an example of segmentation and interface results;

[0042] [Fig.6] [Fig.6] illustrates an example of interfaces before and after rectification;

[0043] [Fig.7] [Fig.7] illustrates the positioning of structural interfaces on an image example digital;

[0044] [Fig.8] [Fig.8] illustrates a determination of average thicknesses, from the profile of intensities averaged over the entire height of a rectified image including the position of the interfaces.

[0045] Fig. 1 schematically illustrates a system 2 for detecting and measuring interfaces in a material sample from multidimensional data representative of the sample, acquired by a characterization machine 4.

[0046] Any microscopy technique suitable for the characterization of such a sample is applicable.

[0047] In one embodiment, the characterization machine 4 is a transmission electron microscope (TEM) which allows images to be acquired of a sample composed of one or more materials, including, for example, crystalline materials.

[0048] The electron microscope 4 allows the simultaneous acquisition of electron diffraction images, EELS (Electron Energy Loss Spectroscopy) spectra, EDX (Energy Dispersive X-Ray Analysis) spectra, and signals from different sensors (BF for Bright field, DF for Dark field, DPC). (for "Differential phase contrast" or any other suitable or customized sensor), for each point of the sample in scanning electron microscopy mode called STEM mode (for "Scanning Transmission Electron Microscope"). Another possible acquisition mode is TEM mode, which allows obtaining an overall image without having to scan the electron beam.

[0049] In other embodiments, the characterization machine 4 is a probe-type microscope (AFM, STM, KFM), an X-ray fluorescence imager, SAXS, SANS, XPS, based on the absorption of X-rays, neutrons, fermions, or alpha, beta, gamma, UV, visible; IR, millimeter, radio frequency, etc. Any imager where the contrast can be related to the actual position of the atoms via a physical simulation that integrates the instrumental parameters can be suitable (SIMS image and variants, RBS, thermoluminescence, X-ray diffraction image, electrons, neutrons, fermions, positrons, etc.).), electron tomography, X-ray tomography, neutron tomography, PEEM, KPEEM, SEM, FIB, dual beam, X-ray fluorescence, ICP-AES, Auger imaging, optical imaging, interferometry, roughness measurement, photoluminescence imaging, Raman spectrometry imaging, field-effect ion microscopy imaging, NMR imaging, any chemical and / or structural imaging technique generating data of dimension greater than or equal to 1, the data being related to a real position of the atoms, for example by a physical simulation that integrates the instrumental parameters. .

[0050] Each acquired spectrum or image is represented as a digital image, comprising points or pixels, each image being representative of at least a part of the observed sample.

[0051] In the case of structural interfaces, it is appropriate to use a characterization method which has sufficient spatial resolution to obtain this information, for example atomic resolution.

[0052] The set of acquired spectra and images forms an N-dimensional data cube, where N is a natural number greater than or equal to 2. Such a data cube is also called a "datacube".

[0053] In one embodiment, several spectra or images are acquired over time, showing an evolution of the sample during the analysis. In this embodiment, time is an additional dimension of the data frame.

[0054] In addition, another dimension of the data block is the microscope focus, electron energy, an angle (of the sample, electron collection, electron convergence), or any other setting parameter of the characterization machine that may vary in a controlled manner during the measurement.

[0055] Incomplete data may optionally be extrapolated from neighboring values ​​where appropriate, for example to correct any imperfections encountered during data acquisition. As a general rule, measurements at the atomic scale are very often marred by defects because simple acoustic or electronic noise can sometimes disrupt them.

[0056] In a multidimensional data block of a crystalline sample, the unit cell usually forms a regular network of droplets representative of the arrangement of atoms. Simulations based on theoretical knowledge of crystalline materials make it possible to predict what the expected data blocks will be for a given material and for a particular setting of the characterization machine.

[0057] A multidimensional data block of an observed sample is transmitted to a multidimensional data processing device 6 for the positioning of the chemical and / or structural interfaces of the sample.

[0058] For example, the transmission is carried out by a wired link or by a wireless link (optical, radio, or other).

[0059] The processing device 6 is, in one embodiment, a programmable electronic device, e.g. a computer.

[0060] The device 6 comprises a processor 8 (CPU and / or GPU) associated with electronic memories 10, of the RAM or ROM or SSD type, or any other type known to those skilled in the art. Optionally, the device 6 is connected to a communication network via a dedicated secure interface, for example via a communication unit 12.

[0061] Advantageously, the device 6 includes or is connected to a knowledge base 14. The knowledge base 14 stores existing data for each specific interface case. The knowledge base is populated with data from previously recorded data blocks and from interface areas that have been previously detected and validated.

[0062] In addition, optionally, the knowledge base 14 is fed by other measurements carried out on the sample, by the development recipes which allow estimating the nature and nominal dimensions expected, by computer-aided design (CAD) which defines the nominal dimensions of the devices through the masks made during the design (CAD) or the design of the devices, by information from the internet (big data, databases, data mining, etc.), through simulations, including predictive ones, through internal company reports and other databases, through scientific and technical knowledge published in literature and patents, through know-how from human or artificial intelligence-based scientific and technical communities, through reference samples that allow the extraction of physical parameter values ​​(e.g., average atomic number) at well-known abrupt reference interfaces.

[0063] Cross-referencing all this data makes it possible to estimate where the chemical and structural interfaces are expected.

[0064] In addition, the device 6 includes a human-machine interface 16 or AI-machine in the case of dedicated artificial intelligences with autonomous communication capabilities (conversational AI for example) including in particular a data display screen or a more advanced multidimensional communication system with a user or with an artificial intelligence.

[0065] Elements 8, 10, 12, 14 and 16 of device 6 are adapted to communicate via a communication bus 15 or any other state-of-the-art means of data exchange, for example optical or microwave.

[0066] The processor 8 (CPU or GPU) is configured to run a data processing module 18, to implement the multidimensional microscopy data processing method for interface positioning between chemically homogeneous areas of a material sample.

[0067] In one embodiment, module 18 is implemented as a computer program comprising software instructions which, when executed by a computer, implement a multidimensional microscopy data processing method for interface positioning between chemically and / or structurally homogeneous areas of a material sample, as described in more detail below.

[0068] Module 18 includes, in particular:

[0069] -a normalization module 20 of the input data pad to obtain a normalized data pad, the normalization comprising an adjustment of the contrast of the input data pad between a predetermined minimum value and a predetermined maximum value;

[0070] - an optional module 22 for applying edge-preserving filtering on the data from the normalized data block to obtain a homogeneous normalized data block;

[0071] - a module 24 for segmenting the normalized data block or, where applicable, normalized, homogenized data allowing the determination of a first data block representative of a first homogeneous area and a second data block representative of a second homogeneous area from the normalized data block, such that the point-by-point sum of the values ​​of the first data block and the second data block is equal to the same value;

[0072] - a module 26 for calculating a standard deviation data block between said first and second data blocks, and standard deviation data block thresholding, to obtain the structural interface between the first homogeneous zone and the second homogeneous zone;

[0073] - a rectification module 28, a calculation of a transformation allowing transform the structural interface into a rectified interface, the transformation being followed by a continuous elastic deformation of the normalized data pad around the rectified interface.

[0074] The multidimensional data processing software for positioning chemical interfaces is further capable of being stored, in the form of a computer program comprising software instructions, on a computer-readable medium, not shown. The computer-readable medium is, for example, a medium capable of storing electronic instructions and being coupled to a bus of a computer system. By way of example, the readable medium is an optical disc, a magneto-optical disc, a ROM, a RAM, any type of non-volatile memory (e.g., EPROM, EEPROM, FLASH, NVRAM, RRAM, PCRAM, 2D NAND, 3D NAND, SLC NAND, MLC NAND, TLC NAND, V-NAND, QLC), a magnetic card or an optical card, an SSD, and any variant known to those skilled in the art in the field of digital data storage.

[0075] In an alternative not shown, modules 20, 22, 24, 26, 28 are each implemented as a programmable logic component, such as an FPGA (Field Programmable Gate Array), a GPU (graphics processor) or a GPGPU (General-purpose processing on graphics processing), or as a dedicated integrated circuit, such as an ASIC (Application-Specific Integrated Circuit).

[0076] Fig. 2 is a synoptic diagram of the main steps of a multidimensional microscopy data processing method for positioning interfaces between chemically and / or structurally homogeneous areas of a material sample according to one embodiment.

[0077] The method includes a step 40 of acquiring an N-dimensional microscopy data block (of images and spectra), the data being representative of an observed crystalline material sample, N being an integer greater than or equal to 2. The data block is obtained by any characterization machine 4 as described above.

[0078] Each data point in the data block corresponds to a point in N-dimensional space, the point being computably related to the position of the atoms in the material,

[0079] In the particular case where N equals two, each multidimensional data block is a two-dimensional (2D) image. Each image data point (or image point, or pixel) has a numerical value, or intensity, represented by a chosen number of bits, for example 8, 16, 32 or 64.

[0080] The term image will be used hereafter to refer to 2D data tiles, it being understood that the methods described apply analogously to N-dimensional data tiles, with N greater than 2.

[0081] By way of example, a representative digital image 45 of an input data block with N=2 is shown in [Fig. 3], this image being the image of a material sample in which the drops (light spots) represent atomic columns of crystals. As can be seen, interface positioning is particularly difficult.

[0082] The method includes a step 42 of normalizing the input data pad to obtain a normalized data pad.

[0083] The normalization step 42 includes, in particular, an adjustment (or normalization) of the contrast of the data block between a minimum value Imin and a maximum value Imax chosen from the range of possible values, the minimum value Imin being, by convention, strictly less than the maximum value Imax. For example, for a data block containing 16-bit encoded values, the maximum contrast is achieved for Imin=0 and Imax=216-1. A person skilled in the art can easily calculate the maximum contrast range for each type of image. To perform such an adjustment, a scaling operation is applied to the numerical values ​​(or intensities) of each point in the data block.

[0084] Preferably, a continuous background subtraction is also applied. For example, the continuous background subtraction involves subtracting the data block obtained after contrast adjustment from a corresponding block obtained by smoothing, for example by applying a Gaussian blur. This allows for the subtraction of slow contrast fluctuations.

[0085] Alternatively, any other continuous blur subtraction method known to the person skilled in the art is applicable, in particular among the methods described in the article "Traditional and recent approaches for background subtraction" by T. Bouwmans, published in Computer Science Review, in May 2014.

[0086] Optionally, the method also includes dimensional normalization, with the median distance between drops being brought down to a range of 2 to 12 pixels by subsampling, to accelerate processing.

[0087] The process then includes a step 43 of segmenting the normalized data block to determine an interface zone representative of a chemical interface or a structural interface.

[0088] The segmentation step 43 is specialized into two distinct processing routes, one of the processing routes being dedicated to chemical interfaces, and the other processing route being dedicated to structural interfaces.

[0089] To determine the chemical interface, in the multidimensional data processing method for the interface position between chemically homogeneous zones, the segmentation step 43 includes a substep 44 of applying an edge-preserving filter to the data of the normalized data block to obtain a homogenized normalized data block.

[0090] In other words, filtering 44 is a chemical homogenization filter. Advantageously, edge-preserving filtering preserves the interfaces between chemically homogeneous areas of the observed sample.

[0091] In one embodiment, filtering 44 consists of removing details that do not fall within the definition of the chemical interface, for example, digital noise, imager defects, possible measurement artifacts, etc., while preserving the actual position of the chemical interfaces, for example with a bilateral filter, or possibly an adaptive median filter, or isotropic scattering, or anisotropic scattering, or a Kuwahara filter, or BEEPS. For example, image 47 in [Fig. 4] was obtained with an isotropic anomalous scattering filter with 500 steps in an increment of 0.125, an anomalous scattering parameter Q=1 and a generalized scattering coefficient of 1 using the algorithm published in "Anomalous diffusion process applied to magnetic resonance image enhancement".

[0092] AC da S Senra Filhol, CE Garrido Salmon2 and LO Murta Junior 1 Published 26 February 2015 • © 2015 Institute of Physics and Engineering in Medicine; Physics in Medicine & Biology, Volume 60, Number 6; Citation: AC da S Senra Filho et al 2015 Phys. Med. Biol. 60 2355; DOI 10.1088 / 0031-9155 / 60 / 6 / 2355. These filtering techniques are published in the literature and are generally generalizable to work in hyperspaces of dimension greater than 3.

[0093] Preferably, step 44 also includes, after the application of edge-preserving filtering, a second contrast adjustment, so as to maximize the contrast to better distinguish the two chemically homogeneous layers. This step typically involves saturating the gray levels on either side of the interface, white on one side, for example at the bottom, and black on the other side of the interface, for example at the top. Through affine transformation of the intensities, all values ​​above the maximum become white, and values ​​below the minimum become black.

[0094] Figure 4 illustrates by way of example two digital images, respectively image 47 corresponding to the result obtained after applying edge-preserving filtering to digital image 45, and a digital image 49 obtained after applying contrast adjustment to digital image 45. In other words, image 49 is the normalized and homogenized data block obtained after filtering 44.

[0095] The resulting homogenized, normalized dataset is then optionally segmented (step 46) using chemical machine learning. This machine learning uses at least one training point, populated from knowledge base 14, which contains chemical knowledge. For example, the material growth recipe allows for a nominal estimation of the location of each layer of the material, each layer being chemically homogeneous. This information makes it possible to find at least one region corresponding to a known material.

[0096] For example, in the case of [Fig. 3], the material growth was carried out to obtain two distinct layers, which allows for an initial localization of each layer. Therefore, to launch machine learning on the two classes, it is sufficient to delimit at least one region in the upper zone for the first class, and at least one other region in the lower zone to define the second class.

[0097] Next, a classifier compatible with binary classes and optimized for chemical segmentation should be applied, for example, a Bayesian type (Bayesnet, Naive Bayes, Naive Bayes Updatable, etc.), a meta type (AdaBoostML, AttributeSelectedClassifier, Random Subspace, ClassificationViaClustering, MultiClassClassifier, RandomCommittee, Bagging, Filtered Classifier, LogitBoost, MulticlassClassiferUpdatable, Threshold Selector, etc.), a rules type (JRep, OneR, Part, etc.), or a tree type (Decision Stump, ForestPA, J48, J48Graft, Random Forest, Random Tree, REP Tree, fast random forest, etc.). Then, in terms of the associated convolution filter family, it is advantageous, for example, to select the Gaussian blur, Mean, Max, Entropy, Variance, Minimum, Median, Bilateral, Structure, and Neighbors filters.

[0098] We ultimately obtain two data sets representing the probability of belonging to each of the classes: a first data set representing the probability of belonging to the first class and a second data set representing the probability of belonging to the second class. The sum of the probabilities is 1, therefore the pointwise sum of the values ​​in the first and second data sets is equal to one.

[0099] Thus, for N=2, each image obtained is the photographic negative of the other.

[0100] More generally, in dimension N, the photographic negative of a first data block is a second data block, such that the sum of the values, point by point, of the first data block and the second data block is equal to the same value.

[0101] To gain statistical precision, it is advantageous to create a data block with all the classification results for each classifier and to take the median of this data block to obtain the most probable result by This technique combines all the results. While it maximizes accuracy, it obviously requires more time since it involves reviewing several machine learning methods. The statistical analysis performed on the dataset of classification results allows us to assess the dispersion of the results and the error bar for interface positioning. If there are significant differences between all the machine learning methods, then the median of all the results is considered the closest to the truth.

[0102] With regard to the measurement of structural interfaces, in the multidimensional data processing method for interface positioning between structurally homogeneous areas, the segmentation step 43 includes a dedicated structure-based machine learning segmentation 48, optimized for the analysis of structural interfaces from the normalized data block.

[0103] In this case, the training data consists of regions whose structure is known. Machine learning uses the knowledge base 14. For example, the material growth recipe allows for a nominal estimation of where each layer is expected. This information makes it possible to find at least one point that corresponds to a layer of known and structurally homogeneous material.

[0104] Segmentation 48 then allows the points of the normalized data block to be classified into two classes, a first class corresponding to the area above the interface and a second class corresponding to the area below the interface, each of the classes being representative of a structurally homogeneous layer.

[0105] For example, in the case of [Fig. 8], material growth has generated growth twins, which allows us to locate each region around the twins. Therefore, to run machine learning on the two classes, it is sufficient to delimit at least one region in the area above the twin for the first class, and at least one other block in the area below the twin to define the second class.

[0106] In the case of image 67 of [Fig.7], we observe that there are 2 classes of structure, one with quasi-horizontal droplet doublets, and the other with quasi-vertical droplet doublets, so it is sufficient to define 2 classes and to position at least one supervised learning box per class to feed the machine learning.

[0107] In other words, the segmentation step 48 makes it possible to obtain a first block of data representative of a probability of belonging to the first class (for example the class with almost horizontal drops) and a second block of data representative of a probability of belonging to the second class (for example, the class with almost vertical drops).

[0108] The sum of the probabilities is 1, therefore the pointwise sum of the values ​​of the first data block and the second data block is equal to one.

[0109] Thus, for N=2, each image obtained is the photographic negative of the other.

[0110] More generally, in dimension N, the photographic negative of a first data block is a second data block, such that the sum of the values, point by point, of the first data block and the second data block is equal to the same value.

[0111] For segmentation 48, the classifier used is specialized in structure analysis, for example of Meta type (AdaBoostMl, Bagging, LogitBoost, MultiClass Classifier, MultiClass Classifier Updatable, Random Committee, Random SubSpace, Threshold Selector, etc.), of Rules type (JRip, PART, etc.), of Tree type (Decision Stump, J48, J48 graft, Random Forest, Random Tree, REP Tree, Fast Random Forest, etc.). The most efficient convolution filters that can be used are of the Hessian, Gabor, Kuwahara, derivative, structure, neighbors, etc. type. Testing on each known particular case allows progress in the art of learning as quickly as possible, through the knowledge base 14 which includes not only the results of structural interface positioning, but also which machine learning method gave the best result, as quickly as possible, for structural classification.Injecting the results and effectiveness of structural interface localization methods into the knowledge base is therefore a way to dynamically optimize methodological variants for each particular case.

[0112] According to another variant, the 48 machine learning segmentation methods implement a neural network, trained by supervised learning, for example a convolutional neural network CNN (for “Convolutional Neural Networks”) or a “deep learning” type learning.

[0113] To improve statistical accuracy, it is advantageous to create a dataset containing all the classification results for each classifier and to take the median of this dataset to obtain the most probable result by combining all the results. With this technique, the accuracy obtained is maximized, but obviously requires more time since it is necessary to review several machine learning methods. The statistical analysis performed on the dataset of classification results allows for the evaluation of the dispersion of the results and the error bar for the positioning of the interfaces. If there are large differences between all the machine learning methods, then the median of all the results is considered to be closest to the truth.

[0114] The method then includes an interface calculation step 50, which implements the calculation of a point-to-point standard deviation block, between the first data block and the second data block which is its photographic negative, these data blocks being obtained by the segmentation step 43.

[0115] In the case of localizing interfaces between chemically homogeneous zones, the standard deviation is therefore calculated from the result of the segmentation by chemical machine learning 46 or from the block made up of the homogenized normalized data block obtained in step 44, which forms the first data block, and its photographic negative which forms the second data block.

[0116] In the case of localizing interfaces between structurally homogeneous areas, the standard deviation is therefore calculated from the result of the structural machine learning segmentation 48, between the first block of data containing the probability of belonging to the first class and the second block of data containing the probability of belonging to the first class.

[0117] The standard deviation is a block of data whose intensity is calculated point by point using the classic formula for standard deviation. The standard deviation of two images is therefore one image of the same size as the other two.

[0118] Advantageously, the standard deviation between a first data frame and a second data frame, which is its photographic negative (also called its inverse), always yields minimum values ​​near the interface and maximum values ​​around it. This is because near the interface, the gray levels are closest to a mean gray level that remains unchanged due to contrast inversion. Conversely, far from the interfaces, white is transformed into black by contrast inversion, so the standard deviation between these two extreme values ​​is maximal because, by definition, the standard deviation is the root mean square of the deviations from the mean. Near the interface, the normalized gray levels are closest to a mean gray, so the standard deviation between the image and its inverse is minimal.

[0119] Basic thresholding is then performed in thresholding step 52, with the thresholding applied to the standard deviation tile. For example, the threshold value S is chosen between 0.00001% and 1% of the maximum intensity of the segmented tile, with all points with a value less than the threshold value S being considered part of the interface, or, in other words, belonging to the interface. Points with a value greater than the threshold value S are points belonging to one of the homogeneous zones. The smaller the threshold S, the finer the interface. The interface is the visible limit of the boundary between the two zones as S approaches zero.

[0120] In a particular application case illustrated, the threshold value is between 5 and 20.

[0121] In one embodiment, the value of the threshold S is dynamically adjusted to the smallest value allowing a continuous interface to be obtained.

[0122] As illustrated by way of example in [Fig. 5], the digital image 49 corresponds to the homogenized normalized data block, the digital image 51 is the photographic negative (or inverse) of the digital image 49, and the digital image 53 includes the result of the thresholding 52 of the standard deviation allowing to visualize precisely a first homogeneous zone 55, a second homogeneous zone 57 and the interface 65 which converges towards a line when S tends towards zero.

[0123] An interface is therefore a 2D line.

[0124] In 3D, the interface is a surface (for example, the soap bubble is the interface between the inside and the outside of the bubble). Here all the classic notions of interface apply.

[0125] The first homogeneous zone 55 corresponds to a first homogeneous layer of the material sample, and the second homogeneous zone 57 corresponds to a second homogeneous layer of the material, and the boundary between these two regions is the interface. When there are three homogeneous regions, then two interfaces are obtained as illustrated in [Fig. 6].

[0126] By way of example, [Fig.6] illustrates a digital image 61 of a thin layer on GaN, on which a first interface 63 (lower interface) and a second interface 62 (upper interface) obtained after the thresholding step 52.

[0127] Advantageously, these chemical and structural segmentation methods are fast, modular, and easy to integrate into more complex programs, and allow for good accuracy results thanks to the prior normalization processing of the input data set. These methods also work in hyperspaces of any dimension, by replacing the curvilinear boundary with its hypersurface equivalent.

[0128] Of course, what has been described above for two layers applies analogously to a larger number of layers, the proposed method allowing the determination of interfaces between each pair of superimposed layers.

[0129] The method further includes an interface rectification step 54.

[0130] In the method of positioning the interface between structurally homogeneous zones; the structural interface obtained previously is rectified.

[0131] Rectification 54 involves calculating a transformation to transform an interface into a straight line segment, and more generally into a rectified interface. The data block is then reorganized with respect to this imposed deformation, as if the data block represented a continuous, deformable, and elastic material bonded to this rectified interface (or reference interface).

[0132] For this, it is proposed to fit a polynomial curve, also called a spline, from a set of adjustable points positioned on the interface and to use for example the algorithm developed by Eva Kocsis and co-authors: “Averaging of Flexible Fibrous Macromolecules: The Clathrin Triskelion Has an Elastic Proximal Segment” Kocsis, E., Trus, BL, Steer, CJ, Bisher, ME and Steven, AC (1991) J. Struct. Biol. 107,6-14.

[0133] An example is illustrated in [Fig. 6]. In the case of a very thin oxide layer on GaN (left part of [Fig. 6]), the result obtained after rectification is shown on the right part of [Fig. 6]: the digital image 61' is obtained after rectification of the interface 63, which is transformed into the rectified interface 63'. By zooming in on the left side of [Fig. 6], the points used to obtain the spline can be seen at 63.

[0134] It is observed that at the end of the rectification step 54, the interface 63 becomes indeed rectilinear (rectified interface 63') and the entire image around this interface is reorganized with respect to this now straight line.

[0135] The advantage of the 54 rectification is the elimination of the interface roughness effect.

[0136] This effect is achieved in order to eliminate topological irregularities. Therefore, with the rectification process, it is now possible to eliminate these irregularities and average the thickness parallel to the interface over the entire region of interest to obtain the precise value of the average thickness without the effect of roughness (see the right-hand side of [Fig. 8]). Indeed, roughness strongly impacts the average thickness measurement result because interface oscillations led to an artificial overthickness in the absence of rectification.

[0137] Furthermore, this same technique advantageously allows for the measurement of average thicknesses in corners. For this, it is sufficient to define the spline on either side of the corner.

[0138] This rectification process makes it possible to smooth out all the irregularities of the interface, while preserving the structural information contained in the data block.

[0139] The rectification can also be carried out on interface 62 (upper interface in [Fig.6]) or on any interface, regardless of its initial form.

[0140] The interface rectification treatment 54 ultimately makes it possible to obtain the chemical or structural profile possibly corrected for topology effects, for example for complex geometries such as corners.

[0141] Structural profiles allow the knowledge base 14 to be fed in order to improve the automated recognition of structural interfaces following further learning.

[0142] The process then includes, after rectification 54, a step 56 of calculating the profiles of the layers, for example the average thicknesses of the layers, from the average of the intensities of the block parallel to the rectified interface ([Fig.8], right part).

[0143] In one embodiment, at step 56 of calculating the average intensity profiles, an average thickness of the area of ​​interest, located between 2 interfaces, is calculated.

[0144] For the interfaces are chemical, then we obtain the chemical thickness.

[0145] For the interfaces are structural, then we obtain the structural thickness.

[0146] It is possible to combine the types of interfaces to obtain other useful thicknesses, for example structural thicknesses for a given chemistry.

[0147] As shown in [Fig. 8], by averaging the height of the image on the left side of [Fig. 8], we obtain the intensity profile which allows us to measure the thickness between the two interfaces represented by light lines. The thickness is the distance between the two peaks on the right side of [Fig. 8], which represents the intensity averaged over the height of the image.

[0148] Advantageously, the proposed method allows the average thicknesses of ultrathin layers (up to 1 monolayer) to be measured, with correction of roughness effects, even in corners.

[0149] Preferably, the results of rectification step 54 are recorded in knowledge base 14 with a view to dynamically improving the knowledge base.

[0150] Preferably, the parameters of the chemical segmentation method by machine learning, and the parameters of the structural segmentation method by machine learning are stored, for example in the electronic memories of the programmable electronic device 6.

[0151] Of course, the determination of chemical interfaces on the one hand, and structural interfaces on the other, is applicable to the same sample of material observed and from the same dataset. Advantageously, the same knowledge base 14 is shared.

Claims

Demands

1. A method for processing multidimensional microscopy data for interface positioning between structurally homogeneous areas of a material sample, comprising acquiring at least one microscopy image of said sample forming an input data pad, the image or images of said input data pad being representative of a part of the observed sample, said input data pad being represented in an N-dimensional space, N being greater than or equal to two, each data point of said pad corresponding to a point in the N-dimensional space, the method being characterized in that it comprises steps of: -normalization (42) of the input data pad to obtain a normalized data pad, the normalization comprising adjusting the contrast of the input data pad between a predetermined minimum and maximum value;- segmentation (43) of the normalized data block allowing to determine a first data block representative of a first homogeneous zone and a second data block representative of a second homogeneous zone from the normalized data block, such that the sum point by point of the values ​​of the first data block and the second data block is equal to the same value; - calculation (50) of a data block of standard deviation between said first and second data blocks; - thresholding (52) of the data block of standard deviation, to obtain the structural interface between the first homogeneous zone and the second homogeneous zone.

2. A method according to claim 1, further comprising a rectification step (54) of the structural interface, comprising a calculation of a transformation enabling the transformation of said structural interface into a rectified interface, said transformation being followed by a continuous elastic deformation of the normalized data block around the rectified interface.

3. A method according to claim 2, wherein the calculation of a transformation is a function of a plurality of points positioned on a curve in the structural interface.

4. A method according to any one of claims 1 or 2, wherein the segmentation (43) of the normalized data block further comprises a segmentation (48) of the normalized data block into two structural classes, respectively a first structural class and a second structural class, each class corresponding to a homogeneous structure, said segmentation being a machine learning segmentation on a structural knowledge database, said machine learning segmentation providing the first data block representative of a probability of belonging to the first structural class and the second data block representative of a probability of belonging to the second structural class.

5. A method according to any one of claims 1 to 4, wherein said thresholding (52) of the standard deviation data block implements a threshold of predetermined value preferably between 0.00001% and 1% of the maximum intensity of the standard deviation data block.

6. A method according to any one of claims 1 to 4, wherein said thresholding (52) of the standard deviation data block implements a threshold of dynamically determined value, the threshold value being the smallest value enabling a continuous interface.

7. A method according to any one of claims 1 to 6, further comprising a calculation (56) of a structural profile associated with the rectified interface by averaging parallel to the rectified interface.

8. A method according to claim 7, further comprising a calculation of an average thickness of an area of ​​interest from the structural profile.

9. A computer program comprising software instructions which, when executed by a programmable electronic device, implement a method for processing multidimensional microscopy data for interface positioning between structurally homogeneous areas in accordance with claims 1 to 8.

10. A multidimensional microscopy data processing device for interface positioning between structurally homogeneous areas of a material sample, implementing the acquisition of at least one microscopy image of said sample forming an input data pad, the image or images of said input data pad being representative of a portion of the observed sample, said input data block being represented in an N-dimensional space, N being greater than or equal to two, each data point of said block corresponding to a point in the N-dimensional space, the device being characterized in that it is configured to implement: -a normalization module (20) of the input data pad to obtain a normalized data pad, the normalization comprising an adjustment of the contrast of the input data pad between a predetermined minimum and maximum value; - a segmentation module (24) of the normalized data block allowing to determine a first data block representative of a first homogeneous area and a second data block representative of a second homogeneous area from the normalized data block, such that the sum point by point of the values ​​of the first data block and the second data block is equal to the same value; - a module (26) for calculating a standard deviation data block between said first and second data blocks and for thresholding (26) the standard deviation data block, to obtain the structural interface between the first homogeneous zone and the second homogeneous zone.

Images