Electrode three-dimensional reconstruction segmentation method and device and electronic equipment

Abstract

The application provides an electrode three-dimensional reconstruction segmentation method and device and electronic equipment. The method comprises: collecting multi-modal data; wherein the multi-modal data comprises morphological data and element data; constructing an implicit neural field model; wherein the input of the implicit neural field model is a spatial three-dimensional coordinate, and the output of the implicit neural field model is a probability vector of the spatial three-dimensional coordinate belonging to each phase; training the implicit neural field model based on a composite loss function of the multi-modal data; wherein the composite loss function comprises a data fidelity loss term and a physical prior loss term; and performing electrode three-dimensional reconstruction based on the trained implicit neural field model. In this way, the conductive agent, the binder and the pore phase with similar gray values in the composite electrode can be accurately segmented through multi-modal data fusion and physical prior constraints, so that the low-contrast image data obtained by using a focused ion beam-scanning electron microscope can be subjected to high-precision three-dimensional reconstruction.

Description

Technical Field

[0001] This invention relates to the fields of materials characterization and artificial intelligence fire extinguishing technology, and in particular to a method, apparatus and electronic device for three-dimensional reconstruction and segmentation of electrodes. Background Technology

[0002] In the development of electrochemical energy storage devices such as lithium-ion batteries and solid-state batteries, the microscopic three-dimensional structure of composite electrodes (typically containing active materials, conductive agents, binders, and pores) has a decisive influence on their electrochemical performance (such as rate performance and cycle life). The focused ion beam-scanning electron microscopy (FIB-SEM) dual-beam system is currently the primary means of acquiring three-dimensional data of electrode microstructures. It obtains a series of two-dimensional images through layer-by-layer cutting and imaging, and then reconstructs them into three-dimensional volumetric data.

[0003] To perform quantitative analysis on the reconstructed 3D volumetric data (such as calculating porosity, tortuosity, and relative connectivity), precise segmentation of each phase within the volumetric data is essential. Existing segmentation techniques mainly include: 1. Traditional image processing methods: such as thresholding methods based on grayscale values ​​(e.g., Otsu's method) and watershed algorithms. 2. Methods based on Convolutional Neural Networks (CNNs): such as deep learning networks like 3D U-Net (a 3D U-shaped network). These networks are trained on small, manually labeled datasets to learn how to distinguish different phases from the grayscale values ​​and texture features of FIB-SEM images.

[0004] However, the existing segmentation techniques described above have serious drawbacks when dealing with composite electrodes, especially lithium-ion battery electrodes. The conductive agent (such as carbon black), binder (such as polyvinylidene fluoride (PVDF)), and epoxy resin used to fill the pores during sample preparation all have very similar atomic numbers (Z values). This results in highly similar grayscale values ​​and extremely low contrast in backscattered electron (BSE) images. Traditional thresholding methods cannot find an effective threshold to separate these three phases due to their severely overlapping grayscale distributions. For CNN methods such as 3D U-Net, which rely primarily on local grayscale values ​​and texture information, the network struggles to accurately distinguish the boundaries of these three phases when the input information itself is highly ambiguous, often leading to misclassifications, such as misidentifying some binders as pores or conductive agents as binders.

[0005] This inaccurate segmentation leads to huge errors in subsequent quantitative analyses (such as volume fraction and specific surface area), which in turn causes the electrochemical simulation model based on this microstructure to be distorted and unable to accurately guide the design and optimization of electrodes. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide an electrode three-dimensional reconstruction segmentation method, apparatus and electronic device for high-precision three-dimensional reconstruction of low-contrast image data acquired using focused ion beam scanning electron microscopy (FIB-SEM).

[0007] In a first aspect, embodiments of the present invention provide a method for three-dimensional reconstruction and segmentation of electrodes. The method includes: acquiring multimodal data; wherein the multimodal data includes morphological data and elemental data; constructing an implicit neural field model; wherein the input of the implicit neural field model is three-dimensional spatial coordinates, and the output of the implicit neural field model is a probability vector of the three-dimensional spatial coordinates belonging to each phase; training the implicit neural field model based on a composite loss function of the multimodal data; wherein the composite loss function includes a data fidelity loss term and a physical prior loss term; and performing three-dimensional reconstruction of electrodes based on the trained implicit neural field model.

[0008] In optional embodiments of this application, the aforementioned data fidelity loss includes: morphological loss and elemental loss; the method further includes: reconstructing backscattered electron images based on the probability vectors and average gray values ​​of each phase, calculating morphological loss based on the reconstructed backscattered electron images and real backscattered electron data; reconstructing elemental maps of energy-dispersive X-ray spectra based on the probability vectors and elemental composition of each phase, and calculating elemental loss based on the reconstructed elemental maps and real energy-dispersive X-ray spectral data.

[0009] In optional embodiments of this application, the aforementioned physical prior loss term includes: global volume fraction constraint; the method further includes: calculating the volume fraction of each phase predicted by the implicit neural field model by integrating the probability vector of the entire volume or a large number of sampling points; obtaining the true volume fraction based on the slurry formulation during electrode preparation; and determining the global volume fraction constraint based on the volume fraction of each phase predicted by the implicit neural field model and the true volume fraction.

[0010] In an optional embodiment of this application, the above-mentioned step of reconstructing the electrode three-dimensionally based on the trained implicit neural field model includes: if the currently trained implicit neural field model converges, determining that the implicit neural field model training is complete; inputting multiple spatial three-dimensional coordinates within the three-dimensional volume of the electrode into the trained implicit neural field model, and outputting the phase probability of each point within the three-dimensional volume of the electrode; wherein, the phase probability of each point represents the probability vector of the point belonging to each phase; performing an argmax operation on each point within the three-dimensional volume of the electrode based on the phase probability of each point within the three-dimensional volume of the electrode to obtain three-dimensional segmentation data.

[0011] In an optional embodiment of this application, the steps for acquiring multimodal data include: sequentially cutting and impregnating the sample, embedding it in resin, curing and polishing it to obtain the starting surface of the cut; performing ion beam cutting and electron beam imaging on the sample to obtain morphological data; and pausing the cutting process during the ion beam cutting of the sample and performing elemental surface scanning on the current cross-section using an energy-dispersive X-ray spectrometer to obtain elemental data.

[0012] In optional embodiments of this application, after the steps of ion beam cutting and electron beam imaging of the sample to obtain morphological data, the method further includes: image registration of the image sequence of the morphological data; after pausing the cutting and performing elemental surface scanning of the current cross section by an energy-dispersive X-ray spectrometer to obtain elemental data during the ion beam cutting of the sample, the method further includes: spatially aligning the energy-dispersive X-ray spectrum of the elemental data with the image sequence of the morphological data.

[0013] In an optional embodiment of this application, the implicit neural field model includes an input layer, a hidden layer, and an output layer; the input layer is used to encode the spatial three-dimensional coordinates of the input; the hidden layer is used to process the data output by the input layer through an activation function; wherein, the activation function includes: an activation function with a sine function or a ReLU activation function; the output layer is used to convert the data output by the output layer into probability vectors of each phase through a Softmax activation function.

[0014] In optional embodiments of this application, the composite loss function further includes: interface morphology loss; the method further includes: determining the interface morphology loss based on the gradient of the probability vector of each phase.

[0015] Secondly, embodiments of the present invention also provide an electrode three-dimensional reconstruction and segmentation device, the device comprising: a multimodal data acquisition module for acquiring multimodal data; wherein the multimodal data includes morphological data and elemental data; an implicit neural field model construction module for constructing an implicit neural field model; wherein the input of the implicit neural field model is spatial three-dimensional coordinates, and the output of the implicit neural field model is a probability vector of spatial three-dimensional coordinates belonging to each phase; an implicit neural field model training module for training the implicit neural field model based on a composite loss function of the multimodal data; wherein the composite loss function includes a data fidelity loss term and a physical prior loss term; and an electrode three-dimensional reconstruction module for performing electrode three-dimensional reconstruction based on the trained implicit neural field model.

[0016] Thirdly, embodiments of the present invention also provide an electronic device, including a processor and a memory, wherein the memory stores computer-executable instructions that can be executed by the processor, and the processor executes the computer-executable instructions to implement the above-described electrode three-dimensional reconstruction and segmentation method.

[0017] The embodiments of the present invention bring the following beneficial effects: This invention provides a method, apparatus, and electronic device for three-dimensional reconstruction and segmentation of electrodes. The method involves acquiring multimodal data, including morphological and elemental data; constructing an implicit neural field model, where the input to the implicit neural field model is three-dimensional spatial coordinates, and the output is a probability vector indicating whether the three-dimensional spatial coordinates belong to each phase; training the implicit neural field model based on a composite loss function of the multimodal data, where the composite loss function includes a data fidelity loss term and a physical prior loss term; and performing three-dimensional reconstruction of the electrode based on the trained implicit neural field model. This approach, through multimodal data fusion and physical prior constraints, can accurately segment conductive agents, binders, and porous phases with similar grayscale values ​​in composite electrodes, thereby enabling high-precision three-dimensional reconstruction of low-contrast image data acquired using focused ion beam scanning electron microscopy.

[0018] Other features and advantages of this disclosure will be set forth in the following description, or some features and advantages may be inferred from the description or determined without doubt, or may be learned by practicing the techniques described above.

[0019] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0020] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0021] Figure 1 A flowchart of an electrode three-dimensional reconstruction and segmentation method provided in an embodiment of the present invention; Figure 2 A flowchart of another electrode three-dimensional reconstruction and segmentation method provided in an embodiment of the present invention; Figure 3 A schematic diagram comparing an existing 3D segmentation process with the electrode 3D reconstruction segmentation method provided in this embodiment of the invention; Figure 4 This is a schematic diagram of the structure of an electrode three-dimensional reconstruction and segmentation device provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Currently, in order to perform quantitative analysis on the reconstructed three-dimensional volumetric data (such as calculating porosity, tortuosity, and relative connectivity), it is necessary to accurately segment the various phases in the volumetric data. Existing segmentation techniques, when processing composite electrodes, have the drawback of failing to separate the conductive agent, binder, and epoxy resin used to fill the pores during sample preparation. This leads to significant errors in subsequent quantitative analysis, resulting in distortion of the electrochemical simulation model based on this microstructure, and thus failing to accurately guide the design and optimization of the electrode.

[0024] Based on this, the present invention provides an electrode three-dimensional reconstruction segmentation method, device and electronic device, specifically providing an electrode three-dimensional reconstruction segmentation method based on physical prior and multimodal fusion, which can accurately segment conductive agent, binder and porous phase with similar gray values ​​in composite electrodes, so as to overcome the segmentation problem caused by low contrast and information ambiguity in the prior art. Its core lies in using an implicit neural representation model and designing a composite loss function to train the model.

[0025] To facilitate understanding of this embodiment, a detailed description of an electrode three-dimensional reconstruction and segmentation method disclosed in this embodiment of the invention will be provided first.

[0026] Example 1: This invention provides a method for three-dimensional reconstruction and segmentation of electrodes, see [link to relevant documentation]. Figure 1 The flowchart shown illustrates a three-dimensional reconstruction and segmentation method for electrodes, which may include the following steps: Step S102: Collect multimodal data; wherein, multimodal data includes: morphological data and element data.

[0027] In this embodiment, high-resolution morphological data can be obtained: for example, layer-by-layer high-resolution (e.g., 5nm / pixel) grayscale images acquired by the BSE or SE (Secondary Electron) detector of FIB-SEM.

[0028] This embodiment can also acquire sparse or low-resolution elemental data: for example, elemental surface distribution maps (such as carbon C, fluorine F, oxygen O, nickel Ni, cobalt Co, etc.) acquired by an energy dispersive X-ray spectroscopy (EDS) detector. This elemental data does not need to be acquired at high resolution layer by layer; it can be low-resolution or sparse data acquired only once every N layers (such as N=10 or 20).

[0029] It should be noted that in this embodiment, Raman spectroscopy (Raman Mapping, which can effectively distinguish different carbon materials and binders) or other signals that can provide phase identification can be used instead of EDS data.

[0030] Step S104: Construct an implicit neural field model; wherein, the input of the implicit neural field model is three-dimensional spatial coordinates, and the output of the implicit neural field model is the probability vector of the three-dimensional spatial coordinates belonging to each phase.

[0031] In this embodiment, a neural network can be constructed as an implicit neural field model, such as a multilayer perceptron (MLP). This network can receive a three-dimensional spatial coordinate (x, y, z) as input, and the output of the network can be a probability vector of the coordinate point (x, y, z) belonging to various phases i (such as active materials, conductive agents, binders, pores). (x,y,z).

[0032] It should be noted that the implicit neural field model in this embodiment can be an MLP or other variants, such as a fast neural representation based on hash grids (e.g., Instant-NGP).

[0033] Step S106: Train the implicit neural field model based on the composite loss function of multimodal data; wherein, the composite loss function includes: data fidelity loss term and physical prior loss term.

[0034] In this embodiment, a composite loss function can be designed. To train the neural field model, the loss function includes a data fidelity loss term. and physical prior loss term The following formula can be used to calculate it: , These are preset weight values.

[0035] In some embodiments, the data fidelity loss term includes: morphological loss and elemental loss; backscattered electron images can be reconstructed based on the probability vectors and average gray values ​​of each phase, and morphological loss can be calculated based on the reconstructed backscattered electron images and real backscattered electron data; elemental maps of energy-dispersive X-ray spectra can be reconstructed based on the probability vectors and elemental composition of each phase, and elemental loss can be calculated based on the reconstructed elemental maps and real energy-dispersive X-ray spectral data.

[0036] Data fidelity loss item in this embodiment This can include: morphological loss Through probability The BSE image is reconstructed using the average gray values ​​of each phase, and the L2 loss is calculated as the morphological loss by comparing it with the acquired high-resolution BSE ground truth (i.e., real BSE data). .

[0037] Data fidelity loss item in this embodiment It may also include: elemental loss Through probability EDS elemental maps were reconstructed using the known elemental composition of each phase (e.g., the F element of the binder PVDF), and the L2 loss was calculated as the elemental loss using the acquired sparse / low-resolution EDS ground truth (i.e., the true EDS data). Among them, element loss It can be calculated at coordinate points where EDS data is available.

[0038] In some embodiments, the physical prior loss term includes: a global volume fraction constraint; the volume fraction of each phase predicted by the implicit neural field model can be calculated by integrating the probability vector of the entire volume or a large number of sampling points; the true volume fraction is obtained based on the slurry formulation during electrode preparation; and the global volume fraction constraint is determined based on the volume fraction of each phase predicted by the implicit neural field model and the true volume fraction.

[0039] The physical prior loss term in this embodiment This can include global volume fraction constraints. During training, the phase probability is calculated over the entire volume (or a large batch of sampled points). Integrating the components, the volume fractions of each phase predicted by the model are calculated. Simultaneously, the known true volume fractions were obtained based on the slurry formulation used in electrode preparation (e.g., 90% AM, 5% binder, 5% conductive agent). .in, Defined as the difference between the model's predicted values ​​and the actual formula values ​​(e.g., L2 loss). ).

[0040] It should be noted here that the physical prior loss term in this embodiment In addition to global volume fraction constraints In addition, interface morphology constraints can also be included. For example, based on the physical fact that "binders tend to coat the surfaces of active materials and conductive agents," a loss term can be designed to encourage the binder phase to appear at the interface of other solid phases and to penalize the binder that "floats" in the pores.

[0041] Step S108: Perform three-dimensional reconstruction of the electrodes based on the trained implicit neural field model.

[0042] In this embodiment, the implicit neural field model can be either not part of the local workstation or cloud server, or it can be integrated into the control software of the FIB-SEM microscope. Electrode 3D reconstruction can be performed based on the trained implicit neural field model, thereby achieving near real-time 3D reconstruction and segmentation.

[0043] In some embodiments, if the currently trained implicit neural field model converges, it is determined that the implicit neural field model training is complete. An implicit neural field model is trained by taking multiple spatial three-dimensional coordinates within the three-dimensional volume of the electrode as input, and outputs the phase probability of each point within the three-dimensional volume of the electrode. The phase probability of each point represents the probability vector of that point belonging to each phase. Based on the phase probability of each point within the three-dimensional volume of the electrode, an argmax operation is performed on each point within the three-dimensional volume of the electrode to obtain three-dimensional segmentation data.

[0044] In this embodiment, the implicit neural field model training can be determined to be complete after the model training converges. By densely querying the three-dimensional coordinates (x, y, z) in space, the phase probability of each point within the entire three-dimensional volume can be obtained. By taking argmax( The three-dimensional segmentation data is obtained as the final high-precision segmentation result.

[0045] This invention provides a method for three-dimensional reconstruction and segmentation of electrodes, which involves acquiring multimodal data, including morphological and elemental data; constructing an implicit neural field model, where the input to the implicit neural field model is three-dimensional spatial coordinates, and the output is a probability vector indicating whether the three-dimensional spatial coordinates belong to each phase; training the implicit neural field model based on a composite loss function of the multimodal data, where the composite loss function includes a data fidelity loss term and a physical prior loss term; and performing three-dimensional reconstruction of the electrode based on the trained implicit neural field model. This method, through multimodal data fusion and physical prior constraints, can accurately segment conductive agents, binders, and porous phases with similar grayscale values ​​in composite electrodes, thereby enabling high-precision three-dimensional reconstruction of low-contrast image data acquired using focused ion beam scanning electron microscopy.

[0046] Example 2: This invention provides another method for three-dimensional reconstruction and segmentation of electrodes, implemented based on the above embodiments. The focus is on describing the specific implementation of the three-dimensional reconstruction and segmentation of electrodes. See also... Figure 2 The flowchart shown represents another electrode three-dimensional reconstruction and segmentation method, which may include the following steps: Step S202: Collect multimodal data; wherein, multimodal data includes: morphological data and element data.

[0047] In some embodiments, the sample may be sequentially cut and impregnated, embedded in resin, cured and polished to obtain the starting surface for cutting; the sample may be subjected to ion beam cutting and electron beam imaging to obtain morphological data; during the ion beam cutting of the sample, the cutting may be paused and the current cross-section may be scanned by an energy-dispersive X-ray spectrometer to obtain elemental data.

[0048] In this embodiment, multimodal data can be collected through steps A1-A3: Step A1, sample preparation, includes: 1. Cutting and impregnation: Cut the target area (e.g., 1×1mm) from the composite electrode sheet. 2 ), and perform a thorough cleaning (e.g., using DMC solvent) to remove residual electrolyte.

[0049] 2. Resin embedding: The sample is placed in a low-viscosity epoxy resin (such as EpoFix) and impregnated under vacuum to ensure that the resin fully penetrates all the pores of the electrode. This step is crucial because the resin filling (whose gray value is also similar to that of the C / Binder) is part of the problem that causes "similar gray values ​​in the three phases".

[0050] 3. Curing and polishing: After the resin has cured, the sample is mechanically ground and polished (e.g., using a diamond suspension) to obtain a smooth, scratch-free surface, which is the starting surface for FIB-SEM cutting.

[0051] Step A2, FIB-SEM imaging parameter settings: In this embodiment, a FIB-SEM dual-beam microscope can be used for data acquisition.

[0052] 1. Focused Ion Beam (FIB) Ion Beam Cutting: Using a gallium (Ga+) ion source or a xenon (Xe+) plasma source, a "trench" is cut into the sample surface with a certain beam current (e.g., 30kV, 10nA) to expose a fresh cross section.

[0053] 2. Scanning Electron Microscopy (SEM) for Electron Beam Imaging (Morphological Data): The cross-section is imaged using an electron beam (e.g., 2.0 kV, 1 nA). A backscattered electron (BSE) detector (such as an In-lens Duo or EsB detector) is preferred because it is sensitive to atomic number (Z value). Although the Z values ​​of the three phases are similar in this embodiment, BSE remains the best choice for providing the main morphology and subtle contrast. The image resolution is set to high resolution, for example, a pixel size of 5 nm × 5 nm.

[0054] 3. Layer-by-layer dicing: The FIB dices the material layer by layer at a set step size (e.g., 10 nm in the z-direction). For each layer diced, a BSE image is captured by SEM. This process is repeated to obtain a continuous sequence of two-dimensional BSE images. .

[0055] Step A3, EDS spectrum acquisition (elemental data): This embodiment allows for periodic pausing of the cutting process during FIB-SEM cutting, and the use of an energy-dispersive X-ray spectroscopy (EDS) detector coupled with SEM to perform elemental surface mapping on the current cross-section.

[0056] To balance time and resolution, a high-resolution EDS image (e.g., pixel size 5nm×5nm) can be acquired every N layers (e.g., N=20, i.e., every 200nm).

[0057] EDS spectra can be acquired at each layer, but at a lower resolution (e.g., by binning 10x10 to make the pixel size 50nm×50nm), or a very short acquisition time (Dwell Time) can be set.

[0058] Among them, the collected element channels It can include at least: carbon C (from conductive agents, binders, resins), fluorine F (a key signal, from PVDF binders), oxygen O (from active materials, some binders such as CMC (carboxymethyl cellulose sodium)), and characteristic elements of active materials (such as nickel Ni, cobalt Co, manganese Mn, phosphorus P, iron Fe, etc.).

[0059] In some embodiments, the image sequence of morphological data can be image registered; the energy dispersive X-ray spectrum of elemental data can be spatially aligned with the image sequence of morphological data.

[0060] This embodiment can also perform data preprocessing (alignment): due to slight drift during the FIB-SEM acquisition process, it is necessary to preprocess the BSE image sequence. Image registration is performed to eliminate inter-layer (x,y) shift. This will transfer sparse / low-resolution images... Atlas and high resolution The sequence is spatially aligned to ensure that the (x,y,z) coordinate system is consistent.

[0061] Step S204: Construct an implicit neural field model; wherein, the input of the implicit neural field model is three-dimensional spatial coordinates, and the output of the implicit neural field model is the probability vector of the three-dimensional spatial coordinates belonging to each phase.

[0062] In some embodiments, the implicit neural field model includes an input layer, a hidden layer, and an output layer; the input layer is used to encode the spatial three-dimensional coordinates of the input; the hidden layer is used to process the data output by the input layer through an activation function; wherein the activation function includes an activation function with a sine function or a ReLU activation function; the output layer is used to convert the data output by the output layer into probability vectors of each phase through a Softmax activation function.

[0063] In this embodiment, an implicit neural field model can be constructed, which is essentially a coordinate-based MLP.

[0064] 1. Input Layer: This layer receives a three-dimensional spatial coordinate p=(x,y,z) and performs positional encoding. Because MLPs suffer from "spectral bias" when learning high-frequency functions, directly inputting (x,y,z) results in a very smooth output image that fails to capture the fine boundaries of the electrodes. Therefore, we first use a fixed encoding function... low-dimensional coordinates Mapped to a high-dimensional feature space, the following formula is used for calculation: ; Here, L is a hyperparameter (e.g., L=10) that determines the dimension of the encoding. It becomes the actual input to the MLP.

[0065] 2. Hidden Layers: This MLP consists of a series of fully connected layers. For example, a typical structure could be: 8 hidden layers, each with 256 neurons.

[0066] In this embodiment, the hidden layer can be configured with an activation function: 1. Preferably, a SIREN activation function can be used: an activation function with a sine function can be used. The SIREN activation function has proven to be very suitable as an implicit representation because it represents the derivative of the signal well, thus producing clear, sharp boundaries. 2. The ReLU activation function is also an option: The ReLU activation function, combined with the positional encoding described above, can also achieve good results.

[0067] 3. Output Layer: The last layer is a fully connected layer with an output dimension of N, representing the number of phases (e.g., N=4, indicating i=1 active material, i=2 conductive agent, i=3 binder, i=4 pores / resin). This N-dimensional vector is converted into a set of probability values ​​through a Softmax activation function. ,in . That is, representing the coordinate point Belongs to the phase The probability of.

[0068] Step S206: Train the implicit neural field model based on the composite loss function of multimodal data; wherein, the composite loss function includes: data fidelity loss term, physical prior loss term and interface morphology loss.

[0069] In this embodiment, a carefully designed composite loss function can be minimized. To train MLP parameters: ;in, These are the weight hyperparameters for each term, used to balance the contributions of different loss terms. During training, a batch of coordinate points can be randomly sampled from the collected data volume. (For example (points).

[0070] (1) Morphological loss (Data fidelity): Used to ensure that the segmentation results of the model are morphologically consistent with high-resolution BSE images. Consistent.

[0071] This embodiment can obtain sampling points. exist Ground Truth grayscale values Define a learnable (or pre-defined) parameter, i.e., per phase Average gray value The model is based on the output probability. Reconstruct the grayscale value of that point: This loss (e.g., L2 loss) is calculated as follows: (All points in Batch) (Calculate the mean).

[0072] (2) Element loss (Chemical Anchoring): Utilizing EDS signals to break down grayscale ambiguity. This is the decisive signal for distinguishing between binders (containing F) and conductive agents / pores (not containing F).

[0073] This embodiment addresses the sampling points. Check if it exists in In the data (i.e., on the sparse acquisition layer, or within a low-resolution grid). If present, this loss is calculated. Taking the F element in the binder (PVDF) as an example, its true EDS signal value is obtained. (Normalized). Reconstruct the F element signal at this point (assuming only the binder phase). Containing F, its F signal is ): This loss (e.g., L2 loss) is calculated as follows: This loss term forces the model to identify regions with strong F element signals as binders. Even if the BSE gray value of this area is exactly the same as the surrounding conductive agent / pores.

[0074] (3) Physical prior loss (Global Constraints): Use known macroscopic physical facts (slurry formulation) to constrain the global behavior of the model and prevent it from "guessing" in areas with insufficient information.

[0075] The prior knowledge can be: the slurry formulation for electrode preparation (e.g., NCM 92wt%, conductive agent 4wt%, binder 4wt%) and the density of each component (e.g., NCM). Conductive agent Adhesive ), calculate the true volume fraction of each solid phase. .

[0076] This embodiment can perform model prediction in the following way: during training, the probability of all points in a very large sampling batch (or the entire volume) is calculated. Calculate the mean (i.e., Monte Carlo integral) to obtain the current volume fraction predicted by the model. . This loss (e.g., L1 or L2 loss) is calculated as follows: (Sum only for the solid phase). This loss term "forces" the model to assign a total volume of each phase that conforms to the actual physical composition on a global scale.

[0077] (4) Interface morphology loss (Advanced Physics Priors): Introduce more refined physical common sense, such as "adhesives are used for bonding and should tend to be distributed on the surface of active materials / conductive agents".

[0078] In some embodiments, the interface morphology loss can also be determined based on the gradient of the probability vector of each phase.

[0079] In this embodiment, a loss term can be designed, which is calculated using a probability field. gradient To analyze the interface. For example, penalize those that "float" in the pores ( In, not with any solid phase ( or ) contact adhesive voxel ( This can be achieved through analysis. and It is achieved through the product of gradients.

[0080] Step S208: Perform three-dimensional reconstruction of the electrodes based on the trained implicit neural field model.

[0081] This embodiment assumes that the implicit neural field model has converged (e.g., (No longer decreasing), the MLP parameters of the implicit neural field model are fixed. At this point, the MLP has become a continuous function. It learned the structure of the entire three-dimensional volume.

[0082] In this embodiment, a high-resolution three-dimensional mesh coordinate system (e.g., 1000×1000×500) can be generated. The coordinates of each point are input in parallel into the trained MLP to obtain the phase probability of each point. .

[0083] This embodiment can be applied to each point. Execute a operate: The final result The matrix is ​​a high-resolution, high-precision three-dimensional segmented data (Label Field), which can be directly used for subsequent quantitative analysis (porosity, tortuosity, three-phase interface, etc.).

[0084] Recognizing that single BSE grayscale information is physically insufficient to distinguish conductive agents, binders, and pores, the method provided in this embodiment of the invention introduces two types of "additional information" to jointly constrain an implicit neural field model: chemical (elemental) specific information from EDS, and macroscopic (global) physical constraints from electrode manufacturing formulations.

[0085] See Figure 3 This diagram illustrates a comparison between an existing 3D segmentation process and the electrode 3D reconstruction segmentation method provided in this embodiment. Wherein, Figure 3 The left side shows the segmentation effect of the existing 3D segmentation process (3D U-Net) on FIB-SEM data. Figure 3 The right side shows the segmentation effect of the electrode 3D reconstruction segmentation method provided in this embodiment on FIB-SEM data. For example... Figure 3 As shown, the segmentation effect of the method provided in this embodiment of the invention is compared with that of the existing 3D segmentation process, which further demonstrates that the segmentation accuracy of the method provided in this embodiment of the invention is far higher than that of the existing 3D U-Net technology.

[0086] The method provided in the embodiments of the present invention has the following main advantages: 1. Solving grayscale ambiguity: Traditional methods rely solely on BSE grayscale values. This invention introduces EDS elemental signals (such as fluorine F in binders). Even if the EDS data is sparse or low-resolution, it provides a "chemical anchor" for the neural field, enabling it to clearly distinguish the binder phase, thereby breaking the grayscale ambiguity among the conductive agent, binder, and pores.

[0087] 2. Ensure overall accuracy: Innovatively introduces physical priors based on electrode slurry formulation. This constraint "forces" the model to assign total volumes of each phase on a global scale that conform to the actual physical composition. This greatly corrects potential misjudgments by the model in local gray-scale blurred regions, ensuring the accuracy of the final volume fraction statistics.

[0088] 3. High resolution and continuity: Based on the continuous function representation of implicit neural fields, segmentation results of arbitrary resolution can be generated during the inference stage, avoiding the voxel jagged effect of traditional CNNs, which is more beneficial for subsequent morphological analysis such as tortuosity and specific surface area.

[0089] In summary, the method provided in this embodiment can fundamentally solve the problem of segmenting the low-contrast phase of composite electrodes through multimodal data fusion and physical prior constraints, and its segmentation accuracy is far higher than that of existing technologies such as 3D U-Net.

[0090] Example 3: This invention provides another method for three-dimensional reconstruction and segmentation of electrodes, which is implemented based on the above embodiments. The specific examples of the three-dimensional reconstruction and segmentation method of electrodes are described in detail.

[0091] Example 1, NCM8119 (a high-nickel ternary lithium battery) positive electrode (PVDF binder): Material: NCM811 positive electrode sheet. Slurry formulation (solid content): NCM811 (density) 92wt%, conductive agent SuperP ( 4wt%, adhesive PVDF ( 4wt%. The target volume fraction was calculated to be... Approximately: AM (Active Material) 82.0%, Carbon (Conductive Carbon Black) 9.4%, Binder (Polymer Binder) 9.6%.

[0092] Data acquisition: BSE (5nm / pixel, z=10nm), EDS (high resolution, acquired every 20 layers, focusing on fluorine (F), carbon (C), oxygen (O), nickel (Ni), cobalt (Co), and manganese (Mn) channels).

[0093] Model: 8-layer MLP, 256 neurons / layer, SIREN activation function. Location encoding. .

[0094] Training: Loss weights are set to (BSE), (EDS-F channel) (Volume). Early stages of training It can be set to 0.5 to accelerate global convergence, and then reduced to 0.1 later to optimize details.

[0095] Result: As Figure 3 As shown, conductive agents, binders, and pores, which were almost indistinguishable in grayscale value in BSE, were clearly separated. Regions with strong F element signals were accurately identified as binders, even though their grayscale values ​​were the same as the porous resin. The final volume fractions were 82.2% AM, 9.3% Carbon, and 9.5% Binder, which highly matched the formulation.

[0096] Example 2, Graphite Anode (SBR / CMC Binder): Material: Graphite anode sheet. Slurry formulation (solid content): Graphite ( The composition is 90 wt% (SBR / CMC), 5 wt% (conductive agent), and 5 wt% (binder / CMC). Challenge: In this case, both the binder (SBR / CMC) and the conductive agent (Carbon) are carbon-based materials, and the C channels of EDS cannot distinguish between them. Solution: Instead, the O (oxygen) element in CMC is used as a chemical anchor. Under these conditions of weak chemical signals, Physical constraints become crucial. They globally ensure that the model does not misclassify excessive conductive agents as adhesives.

[0097] Data acquisition: BSE (5nm / pixel, z=10nm), EDS (low resolution, acquisition of each layer, with a focus on carbon C and oxygen O channels).

[0098] Model: 6-layer MLP, 128 neurons / layer, ReLU activation function with positional encoding.

[0099] Training: Loss weights are set to (BSE), (EDS-O channel) (Volume) Results: Although the chemical discrimination was lower than that of Example 1, in Despite the strong constraints, the model still successfully separated the three phases, and its accuracy was significantly better than any method that relied solely on BSE.

[0100] Example 3, System Deployment and Performance: Hardware: The method provided in this embodiment can be implemented on a deep learning workstation.

[0101] Software: Python language, based on the PyTorch deep learning framework. MLP models are built using torch.nn.Module (the basic class for implementing neural network structures in PyTorch), and positional encoding and composite loss functions are defined as callable modules.

[0102] Training: For a voxels ( The dataset contains 50,000 data points, and the total number of training iterations is 50,000. Each batch (Batch Size) samples... One point. The total training time was approximately 4 hours.

[0103] Reasoning: In the reasoning stage, Each coordinate point is fed into the trained MLP in batches, taking approximately 2 minutes in total to generate complete, high-resolution (5nm) 3D segmentation data. Conclusion: The method of this invention is not only revolutionary in accuracy but also entirely feasible in terms of computational efficiency.

[0104] Example 4 compares the existing 3DU-Net method with the method provided in this embodiment: Existing 3DU-Net method: Employs the state-of-the-art 3DU-Net segmentation model in the field. Training: Uses only the BSE data collected in step S100. Since there are no ground truth labels, two methods can be tried: (1) Unsupervised clustering: It is impossible to separate three phases with similar gray levels. (12) Manual annotation: It took researchers several weeks to manually annotate 20 layers (1000x1000) of training data.

[0105] Result: As Figure 3 As shown, even with expensive manual annotation, 3DU-Net still generates numerous errors at the boundaries of conductive agents, binders, and pores, resulting in unclear three-phase mixing. For example, on the dataset in Example 1, 3DU-Net predicts a binder volume fraction of only 1.5% (while the actual formulation is 9.6%), misclassifying the vast majority of binders as pores or conductive agents, leading to completely distorted analysis results.

[0106] Example 4: Corresponding to the above method embodiments, this invention provides an electrode three-dimensional reconstruction and segmentation device, see [link to relevant documentation]. Figure 4 The diagram shows a structural schematic of an electrode three-dimensional reconstruction and segmentation device, which includes: The multimodal data acquisition module 41 is used to acquire multimodal data, which includes morphological data and element data. Implicit neural field model construction module 42 is used to construct implicit neural field models; wherein, the input of the implicit neural field model is three-dimensional spatial coordinates, and the output of the implicit neural field model is the probability vector of the three-dimensional spatial coordinates belonging to each phase; Implicit neural field model training module 43 is used to train implicit neural field models based on composite loss functions of multimodal data; wherein, the composite loss function includes: data fidelity loss term and physical prior loss term; Electrode 3D Reconstruction Module 44 is used for electrode 3D reconstruction based on the trained implicit neural field model.

[0107] This invention provides an electrode 3D reconstruction and segmentation device that acquires multimodal data, including morphological and elemental data. An implicit neural field model is constructed, where the input is spatial 3D coordinates and the output is a probability vector indicating whether the spatial 3D coordinates belong to each phase. The implicit neural field model is trained based on a composite loss function of the multimodal data, including a data fidelity loss term and a physical prior loss term. Electrode 3D reconstruction is performed based on the trained implicit neural field model. This method, through multimodal data fusion and physical prior constraints, can accurately segment conductive agents, binders, and porous phases with similar grayscale values ​​in composite electrodes, thereby enabling high-precision 3D reconstruction of low-contrast image data acquired using focused ion beam scanning electron microscopy.

[0108] The aforementioned data fidelity loss terms include: morphological loss and elemental loss; the aforementioned implicit neural field model training module is also used to reconstruct backscattered electron images based on the probability vectors and average gray values ​​of each phase, calculate the morphological loss based on the reconstructed backscattered electron images and real backscattered electron data; reconstruct the elemental map of energy-dispersive X-ray spectra based on the probability vectors and elemental composition of each phase, and calculate the elemental loss based on the reconstructed elemental map and real energy-dispersive X-ray spectral data.

[0109] The aforementioned physical prior loss term includes: global volume fraction constraint; the aforementioned implicit neural field model training module is also used to calculate the volume fraction of each phase predicted by the implicit neural field model by integrating the probability vector of the entire volume or a large number of sampling points; obtain the true volume fraction based on the slurry formulation during electrode preparation; and determine the global volume fraction constraint based on the volume fraction of each phase predicted by the implicit neural field model and the true volume fraction.

[0110] The aforementioned electrode 3D reconstruction module is used to determine that the implicit neural field model training is complete if the currently trained implicit neural field model converges; it takes multiple spatial 3D coordinates within the electrode's 3D volume as input to the trained implicit neural field model and outputs the phase probability of each point within the electrode's 3D volume; wherein, the phase probability of each point represents the probability vector of that point belonging to each phase; based on the phase probability of each point within the electrode's 3D volume, it performs an argmax operation on each point within the electrode's 3D volume to obtain 3D segmentation volume data.

[0111] The aforementioned multimodal data acquisition module is used to sequentially cut and impregnate the sample, embed it in resin, cure and polish it to obtain the starting surface of the cut; to perform ion beam cutting and electron beam imaging on the sample to obtain morphological data; during the ion beam cutting process, the cutting is paused and the current cross-section is scanned by an energy-dispersive X-ray spectrometer to obtain elemental data.

[0112] The aforementioned multimodal data acquisition module is also used to perform image registration on the image sequence of morphological data; the aforementioned multimodal data acquisition module is also used to spatially align the energy dispersive X-ray spectra of elemental data with the image sequence of morphological data.

[0113] The aforementioned implicit neural field model includes an input layer, a hidden layer, and an output layer. The input layer is used to encode the spatial three-dimensional coordinates of the input. The hidden layer is used to process the data output by the input layer through activation functions. The activation functions include: activation functions with sine functions or ReLU activation functions. The output layer is used to convert the data output by the output layer into probability vectors of each phase through a softmax activation function.

[0114] The aforementioned composite loss function also includes: interface morphology loss; the aforementioned implicit neural field model training module is also used to determine the interface morphology loss based on the gradient of the probability vector of each phase.

[0115] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electrode three-dimensional reconstruction and segmentation device described above can be referred to the corresponding process in the aforementioned embodiments of the electrode three-dimensional reconstruction and segmentation method, and will not be repeated here.

[0116] Example 4: This invention also provides an electronic device for running the above-described electrode three-dimensional reconstruction and segmentation method; see [link to previous document]. Figure 5 The diagram shows the structure of an electronic device, which includes a memory 100 and a processor 101. The memory 100 is used to store one or more computer instructions, which are executed by the processor 101 to implement the above-mentioned three-dimensional reconstruction and segmentation method for electrodes.

[0117] Furthermore, Figure 5 The electronic device shown also includes a bus 102 and a communication interface 103, with the processor 101, the communication interface 103 and the memory 100 connected via the bus 102.

[0118] The memory 100 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device. Communication between this system network element and at least one other network element is achieved through at least one communication interface 103 (which can be wired or wireless), such as the Internet, wide area network, local area network, metropolitan area network, etc. The bus 102 may be an ISA bus, PCI bus, or EISA bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 5 The symbol is represented by a single double-headed arrow, but this does not mean that there is only one bus or one type of bus.

[0119] Processor 101 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of processor 101 or by instructions in software form. Processor 101 can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this invention can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a readily available storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 100, and processor 101 reads information from memory 100 and, in conjunction with its hardware, completes the steps of the method described in the foregoing embodiments.

[0120] This invention also provides a computer-readable storage medium storing computer-executable instructions. When these computer-executable instructions are called and executed by a processor, they cause the processor to implement the above-described electrode three-dimensional reconstruction and segmentation method. For specific implementation details, please refer to the method embodiments, which will not be repeated here.

[0121] The computer program products of the electrode three-dimensional reconstruction and segmentation method, apparatus and electronic device provided in the embodiments of the present invention include a computer-readable storage medium storing program code. The instructions included in the program code can be used to execute the methods in the preceding method embodiments. For specific implementation, please refer to the method embodiments, which will not be repeated here.

[0122] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the system and / or device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0123] Furthermore, in the description of the embodiments of the present invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in the present invention based on the specific circumstances.

[0124] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0125] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0126] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for three-dimensional reconstruction and segmentation of electrodes, characterized in that, The method includes: Collect multimodal data; wherein, the multimodal data includes: morphological data and element data; Construct an implicit neural field model; wherein the input of the implicit neural field model is three-dimensional spatial coordinates, and the output of the implicit neural field model is the probability vector of the three-dimensional spatial coordinates belonging to each phase; The implicit neural field model is trained based on a composite loss function of the multimodal data; wherein the composite loss function includes: a data fidelity loss term and a physical prior loss term; Electrode 3D reconstruction is performed based on the trained implicit neural field model.

2. The method according to claim 1, characterized in that, The data fidelity loss term includes: morphological loss and elemental loss; the method further includes: Backscattered electron images are reconstructed based on the probability vectors and average gray values ​​of each phase, and the morphological loss is calculated based on the reconstructed backscattered electron images and real backscattered electron data. The elemental map of the energy-dispersive X-ray spectrum is reconstructed based on the probability vectors of each phase and the elemental composition, and the elemental loss is calculated based on the reconstructed elemental map and the actual energy-dispersive X-ray spectral data.

3. The method according to claim 1, characterized in that, The physical prior loss term includes: global volume fraction constraint; the method further includes: The volume fraction of each phase predicted by the implicit neural field model is calculated by integrating the probability vector of the entire volume or a large number of sampling points. The true volume fraction is obtained based on the slurry formulation during electrode preparation; The global volume fraction constraint is determined based on the volume fractions of each phase predicted by the implicit neural field model and the actual volume fractions.

4. The method according to claim 1, characterized in that, The steps for three-dimensional reconstruction of electrodes based on the trained implicit neural field model include: If the currently trained implicit neural field model converges, it is determined that the training of the implicit neural field model is complete. The implicit neural field model, trained by inputting multiple three-dimensional coordinates within the three-dimensional volume of the electrode, outputs the phase probability of each point within the three-dimensional volume of the electrode; wherein, the phase probability of each point represents the probability vector of that point belonging to each phase. Based on the phase probability of each point within the three-dimensional volume of the electrode, an argmax operation is performed on each point within the three-dimensional volume of the electrode to obtain three-dimensional segmentation data.

5. The method according to claim 1, characterized in that, The steps for collecting multimodal data include: The sample was sequentially cut and impregnated, embedded in resin, cured and polished to obtain the starting surface for cutting; The sample was subjected to ion beam cutting and electron beam imaging to obtain the morphological data. During the ion beam cutting of the sample, the cutting is paused and the current cross-section is scanned using an energy-dispersive X-ray spectrometer to obtain the elemental data.

6. The method according to claim 5, characterized in that, After the step of performing ion beam cutting and electron beam imaging on the sample to obtain the morphological data, the method further includes: performing image registration on the image sequence of the morphological data; After pausing the cutting process during ion beam cutting of the sample and performing elemental surface scanning of the current cross section using an energy-dispersive X-ray spectrometer to obtain the elemental data, the method further includes: spatially aligning the energy-dispersive X-ray spectrum of the elemental data with the image sequence of the morphological data.

7. The method according to claim 1, characterized in that, The implicit neural field model includes: an input layer, a hidden layer, and an output layer; The input layer is used to encode the position of the input three-dimensional spatial coordinates; The hidden layer is used to process the data output by the input layer through an activation function; wherein the activation function includes: an activation function with a sine function or a ReLU activation function; The output layer is used to convert the data output by the output layer into probability vectors for each phase through the Softmax activation function.

8. The method according to claim 1, characterized in that, The composite loss function further includes: interface morphology loss; the method further includes: The interface morphology loss is determined based on the gradient of the probability vector of each phase.

9. A three-dimensional reconstruction and segmentation device for electrodes, characterized in that, The device includes: A multimodal data acquisition module is used to acquire multimodal data; wherein, the multimodal data includes: morphological data and element data; An implicit neural field model construction module is used to construct an implicit neural field model; wherein, the input of the implicit neural field model is a three-dimensional spatial coordinate, and the output of the implicit neural field model is a probability vector of the three-dimensional spatial coordinates belonging to each phase; An implicit neural field model training module is used to train the implicit neural field model based on a composite loss function of the multimodal data; wherein the composite loss function includes: a data fidelity loss term and a physical prior loss term; The electrode 3D reconstruction module is used to perform electrode 3D reconstruction based on the trained implicit neural field model.

10. An electronic device, characterized in that, It includes a processor and a memory, the memory storing computer-executable instructions that can be executed by the processor, the processor executing the computer-executable instructions to implement the electrode three-dimensional reconstruction and segmentation method according to any one of claims 1 to 8.

Images