A material performance prediction method and electronic device

Abstract

The application discloses a material performance prediction method and an electronic device, and relates to the technical field of computers, and comprises the following steps: performing multi-channel voxelization processing on atomic structure data of a material to be predicted, and mapping an atomic point cloud to a multi-channel sparse tensor, so that the geometric morphology and accessibility information of a pore network are completely preserved; and subsequently, a pre-trained performance prediction model is used to simultaneously capture the correlation features between pore features of different scales and long-range structures, so that high-precision performance prediction of a porous carbon material is realized according to complete and accurate material feature information. The technical problem of insufficient performance prediction precision caused by incomplete and inaccurate material feature information is solved, and the technical effect of accurate material performance prediction is achieved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method for predicting material properties and an electronic device. Background Technology

[0002] Carbon-based materials, due to their diverse structures and excellent electrochemical properties, have become one of the most promising candidate anode materials for sodium-ion batteries. Accurate prediction of their performance is crucial for battery development. Currently, methods for predicting the performance of carbon-based materials, such as first-principles calculations and machine learning models, struggle to fully and accurately characterize the material's internal features from the atomic to the mesoscopic scale while maintaining computational efficiency. This results in insufficient accuracy in predicting the performance of porous carbon materials, hindering high-throughput performance screening and optimization design. Summary of the Invention

[0003] This application provides a material property prediction method and an electronic device to at least solve the problems in the related art.

[0004] This application provides a method for predicting material properties, including: Obtain atomic structure data of the material to be predicted; wherein, atomic structure data is used to characterize the spatial arrangement of atoms in the material to be predicted; Atomic structure data is processed into three-dimensional voxels to generate multi-channel three-dimensional tensors; where the data at each voxel position in the tensor is a multi-dimensional feature vector, used to characterize at least one material-based property of the local space represented by the corresponding voxel. Input a multi-channel three-dimensional tensor into a pre-trained performance prediction model and obtain the performance prediction results of the material to be predicted from the output of the performance prediction model. The performance prediction model is configured to extract multi-scale porosity features and long-range structural correlation features from multi-channel three-dimensional tensors, and perform performance prediction based on the multi-scale porosity features and long-range structural correlation features. The multi-scale porosity features are used to characterize the pore geometry and topological properties at multiple scales inside the material to be predicted; the long-range structural correlation features are used to characterize the cooperative relationship between structural units at different distances in the material to be predicted.

[0005] This application also provides a material property prediction device, comprising: The acquisition unit is used to acquire atomic structure data of the material to be predicted; wherein, the atomic structure data is used to characterize the spatial arrangement of atoms in the material to be predicted. The generation unit is used to perform three-dimensional voxelization processing on atomic structure data to generate a multi-channel three-dimensional tensor; wherein, the data of each voxel position in the tensor is a multi-dimensional feature vector, which is used to characterize at least one material-based property of the local space represented by the corresponding voxel. The prediction unit is used to input multi-channel three-dimensional tensors into a pre-trained performance prediction model and obtain the performance prediction results of the material to be predicted from the output of the performance prediction model. The performance prediction model is configured to extract multi-scale porosity features and long-range structural correlation features from multi-channel three-dimensional tensors, and perform performance prediction based on the multi-scale porosity features and long-range structural correlation features. The multi-scale porosity features are used to characterize the pore geometry and topological properties at multiple scales inside the material to be predicted; the long-range structural correlation features are used to characterize the cooperative relationship between structural units at different distances in the material to be predicted.

[0006] This application also provides an electronic device, which includes a computing module, an interconnect module, and a storage module integrated on at least one cabinet unit, wherein... The calculation module includes multiple servers for executing any of the above-mentioned material property prediction methods; The interconnect module is used to enable data conversion between multiple servers; The storage module is used to store atomic structure data, multi-channel three-dimensional tensors, and performance prediction results involved in the computation module.

[0007] This application also provides an electronic device, comprising: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described material property prediction methods.

[0008] This application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of any of the above-described material property prediction methods.

[0009] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described material property prediction methods.

[0010] This application utilizes multi-channel voxelization processing of the atomic structure data of the material to be predicted, mapping the atomic point cloud into a multi-channel sparse tensor to fully preserve the geometric morphology and accessibility information of the pore network. Subsequently, based on a pre-trained performance prediction model, it simultaneously captures pore features at different scales and long-range structural correlation features, enabling high-precision performance prediction of complex materials such as porous carbon based on complete and accurate material feature information. This solves the technical problem of insufficient performance prediction accuracy caused by incomplete and inaccurate material feature information representation, achieving the technical effect of accurate material performance prediction. Attached Figure Description

[0011] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 A schematic flowchart illustrating a material property prediction method provided in an embodiment of this application; Figure 2 A schematic diagram of a carbon material structure provided in an embodiment of this application; Figure 3 A schematic flowchart illustrating a material property prediction method provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0014] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0015] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0016] The specific application environment architecture or specific hardware architecture on which the material property prediction method depends is described here.

[0017] The embodiments of this application provide a material performance prediction method that can not only accurately predict the sodium storage performance of materials without long-range ordered or porous carbon, but also has the ability to handle various carbon structure types, capture the correlation between multi-scale pores and long-range structures, identify functional cavities, and operate efficiently in a high-performance hardware environment. This significantly improves the screening efficiency and design level of sodium-ion battery anode materials, and also has the potential to be extended to other secondary battery systems such as potassium-ion and magnesium-ion batteries. The following embodiments, combined with the execution flow of the material performance prediction method, provide a detailed description of the method.

[0018] Currently, sodium-ion batteries (SIBs) have broad application prospects in large-scale energy storage and electric vehicles due to the abundant and inexpensive availability of sodium resources and their high safety. Compared to lithium-ion batteries, sodium ions have a larger radius (1.02 Å, compared to 0.76 Å in lithium-ion batteries), making the anode material more prone to volume expansion and localized structural deformation during sodium storage. Therefore, developing anode materials with high capacity, low sodium intercalation potential, and excellent structural stability is key to improving the performance of sodium-ion batteries.

[0019] Carbon-based materials, due to their diverse structural types (such as hard carbon, soft carbon, graphite, and doped carbon), excellent electrical conductivity, and chemical stability, have become core candidates for sodium-ion battery anode materials. In carbon structures, sodium storage mechanisms mainly include surface adsorption, pore filling, and interlayer embedding. Different microstructures (such as pore size distribution, defect density, and degree of graphitization) significantly affect sodium storage capacity and kinetic performance. Hard carbon materials, as anode materials for sodium-ion batteries, exhibit excellent low-potential plateaus and high specific capacities, especially due to their lack of long-range ordered structures and abundant porous networks, and are considered strong contenders for commercialization.

[0020] Relevant methods for predicting material properties mainly include precise calculations based on first-principles calculations (DFT) and molecular dynamics (MD), traditional machine learning models based on sample features, and crystal structure modeling methods based on graph neural networks (GNN). Among these, first-principles and molecular dynamics methods can accurately calculate the electronic structure, ion migration barriers, and stability of materials at the atomic scale, with clear physical meaning, but they are computationally intensive and difficult to rapidly screen massive amounts of structures. Machine learning-based prediction methods rely on a large amount of material feature information (such as specific surface area, pore size distribution, coordination number, etc.), but the models are relatively simple, have poor interpretability, and suffer information loss in characterizing structural diversity and complex pore networks. In recent years, deep learning methods (such as GNNs) have been introduced into materials informatics to automatically extract local environmental features directly from graph structures composed of atoms and chemical bonds to adapt to different types of material structures. However, they have limited ability to capture long-range correlations, lack the ability to uniformly model multi-scale porosity features, and have insufficient prediction accuracy for amorphous and non-long-range ordered porous carbon structures. Moreover, long-range structural correlation information is often weakened in convolutional or graph networks, and the ability to identify functional cavities is also limited.

[0021] While related methods have advanced materials prediction to varying degrees, they still have significant limitations when dealing with the unique research subject of carbon materials for sodium-ion battery anodes. First, most methods extract features at a single spatial scale, failing to simultaneously capture pore features at different scales, which directly affect key performance parameters such as specific capacity and ion migration rate. Second, deep learning methods have limited ability to handle long-range structural correlations (such as intercellular pore connectivity and large-scale lattice defects), and their predictions of macropores or nonlocal structural features are not accurate enough. Furthermore, amorphous and porous structures without long-range order, such as hard carbon and doped carbon, are difficult to represent effectively within existing lattice parameter-driven modeling frameworks, limiting the generalization ability of the models.

[0022] Therefore, there is an urgent need for a predictive method that can accurately extract the geometric and topological features of carbon structures across multiple spatial scales and capture long-range correlation information to support high-throughput performance screening and optimized design of carbon anode materials for sodium-ion batteries.

[0023] The embodiments of this application provide a material property prediction method, and the method is described in detail below in conjunction with the execution flow of the material property prediction method.

[0024] Figure 1 A flowchart illustrating a material property prediction method provided in this application embodiment is shown, specifically including as follows: Figure 1 The following steps are shown: S101. Obtain the atomic structure data of the material to be predicted.

[0025] Among them, atomic structure data is used to characterize the spatial arrangement of atoms in the material to be predicted, and the atomic structure data is carbon material structure data.

[0026] Understandably, the material to be predicted is a porous material used for predicting anode materials in sodium-ion batteries, such as hard carbon and soft carbon in carbon materials, and porous materials such as metal-organic frameworks (MOFs) and covalent organic frameworks (COFs) in metallic materials. The following examples use a certain carbon material as an example for detailed explanation, and the atomic structure data is the carbon material structure data. The atomic structure data is used to characterize the spatial arrangement of atoms in the material to be predicted, and can be in common file formats such as CIF or XYZ. CIF (Crystal Information File) is a standard file format for describing crystal structures (such as atom types, coordinates, and unit cell parameters). XYZ is a simpler chemical file format, mainly containing information such as atom types and three-dimensional rectangular coordinates. The atomic structure data of carbon materials is the core information describing the three-dimensional spatial arrangement of its atoms, including atom types, coordinates, and bonding relationships. For example, the layered ordered coordinates of graphite (CIF file) and the disordered set of carbon atom positions in hard carbon.

[0027] S102. Perform three-dimensional voxelization on the atomic structure data to generate a multi-channel three-dimensional tensor.

[0028] In this tensor, the data at each voxel position is a multidimensional feature vector, used to characterize at least one material-based property of the local space represented by the corresponding voxel.

[0029] Understandably, based on the above S101, an efficient voxelization conversion process maps the atomic point cloud in the atomic structure data into a multi-channel three-dimensional coefficient tensor, generating a multi-channel three-dimensional tensor that fully preserves the geometric shape and accessibility information of the pore network. Each spatial location (voxel location) of the tensor has a set of multi-dimensional feature vectors used to identify various atomic properties.

[0030] Understandably, a voxel is a three-dimensional pixel unit used to transform a material structure into an equally spaced three-dimensional mesh. Each voxel contains multiple channels of structural and physical features.

[0031] Optionally, the atomic structure data is subjected to three-dimensional voxelization to generate a multi-channel three-dimensional tensor, including: The atomic structure data is processed into three-dimensional voxels at a first set spatial resolution, mapping the atomic point cloud into a multi-channel three-dimensional tensor. The multi-dimensional feature vectors include atomic density feature vectors, local coordination number feature vectors, polyhedral volume feature vectors, atomic property feature vectors, and / or set element functional cavity feature vectors. The local coordination number refers to the number of atoms in the neighborhood of the voxel side length at a second set spatial resolution. The atomic property feature vector is a two-dimensional vector composed of atomic radius and electron number. The set element functional cavity information is identified by a porosity preservation filter.

[0032] Understandably, the material to be predicted is subjected to three-dimensional voxelization at a first set spatial resolution, mapping the atomic point cloud into a multi-channel sparse tensor. For example, the first set spatial resolution is 1 Å (0.1 nanometers), with each voxel spatially divided at 1 Å intervals. Each voxel contains information such as atomic density, local coordination number (e.g., the number of atoms in a neighborhood with a radius of 5 Å), Voronoi volume, atomic type label (atomic property feature vector), and functional cavity label of the set element identified by a porosity-preserving filter. This preserves the geometric morphology, pore network, and accessibility features of the atomic structure in the three-dimensional tensor. The atomic type label can be a two-dimensional vector recording the atomic radius and the number of valence electrons, respectively. The set element can be Na. + In the cavity information, 0 indicates that it is not Na. + Functional cavity, 1 indicates Na + Functional cavities. This process not only preserves the spatial topological information of carbon materials, but also, by utilizing porosity-preserving filters, explicitly preserves functional cavity information in multi-channel three-dimensional tensors, effectively distinguishing between functional pores with sodium storage potential and irrelevant pores, thereby improving the modeling accuracy of subsequent models in terms of structure and performance.

[0033] Understandably, a porosity-preserving filter is an algorithmic module that identifies functional cavity regions during voxelization and can be implemented in conjunction with software.

[0034] Understandably, in addition to its application in predicting carbon materials for sodium-ion battery anodes, it can also be used to predict material properties in other energy storage systems such as potassium-ion and magnesium-ion batteries.

[0035] Optionally, before performing three-dimensional voxelization on the atomic structure data, the method further includes: The atomic structure data is analyzed to obtain the initial coordinates of the atoms. While maintaining the three-dimensional periodicity of the atoms, the initial coordinates are uniformly converted into atomic coordinates based on the Cartesian coordinate system. The atomic coordinates are then standardized to perform three-dimensional voxelization on the standardized atomic structure data.

[0036] Understandably, before performing 3D voxelization on the atomic structure data, a structure analysis module is constructed to unify and standardize the atomic coordinate system. Standardization involves converting the atomic coordinates in different types of structure files into Cartesian coordinates. Specifically, the atomic structure data is parsed to obtain the initial coordinates of the atoms; for example, a list of fractional coordinates of carbon atoms is read from a graphite CIF file. Then, the initial coordinates of the atoms in different types of structure files are uniformly converted into Cartesian coordinates. The data obtained after standardization is the Cartesian coordinate information of a 3D periodic atomic lattice. Subsequent 3D voxelization can then be performed on the standardized atomic structure data.

[0037] S103. Input the multi-channel three-dimensional tensor into the pre-trained performance prediction model and obtain the performance prediction results of the material to be predicted output by the performance prediction model.

[0038] The performance prediction model is configured to extract multi-scale porosity features and long-range structural correlation features from multi-channel three-dimensional tensors, and perform performance prediction based on the multi-scale porosity features and long-range structural correlation features. The multi-scale porosity features are used to characterize the pore geometry and topological properties at multiple scales inside the material to be predicted; the long-range structural correlation features are used to characterize the cooperative relationship between structural units at different distances in the material to be predicted.

[0039] Understandably, based on S102 above, multi-scale porosity features and long-range structural correlation features are extracted from the multi-channel three-dimensional tensor using the performance prediction model. The multi-scale porosity features characterize the geometric and topological properties of pores at multiple scales within the material to be predicted, while the long-range structural correlation features characterize the cooperative relationships between structural units at different distances within the material. Subsequently, the multi-scale porosity features and long-range structural correlation features are fused using a cross-scale feature fusion mechanism within the performance prediction model, achieving a unification of local and global information. At the output layer of the performance prediction model, a multi-task regression approach is used to predict multiple key performance indicators based on the fused feature information, thereby obtaining a comprehensive evaluation of various aspects of the material's performance in a single prediction and generating the performance prediction result.

[0040] The performance prediction models include multi-scale convolutional networks, dilated convolutional networks, feature fusion networks, and residual networks. Multi-scale convolutional networks consist of multiple independent convolutional channels with different aperture ranges and receptive field ranges based on different kernel sizes and strides.

[0041] Understandably, a multi-scale convolutional network refers to a multi-scale three-dimensional convolutional neural network. This network uses a multi-scale parallel convolutional branch setup, setting independent convolutional channels for different aperture ranges (such as nanopores, mesopores, and macropores). To ensure that each network has a different receptive field, settings are made for different scales in terms of aperture range, convolutional kernel size, and stride. For example, different receptive fields are set for nanopores (approximately 1 nm, convolutional kernel size: 3×3×3, stride: 1), mesopores (approximately 10 nm, convolutional kernel size: 10×10×10, stride: 3), and macropores (approximately 50 nm, convolutional kernel size: 30×30×30, stride: 10) to extract pore features at different spatial scales. A dilated convolutional network is introduced between convolutional kernels to expand the receptive field without increasing the number of parameters. The receptive field of the dilated convolutional network is larger than that of the multi-scale convolutional network. For example, the dilated convolutional network expands the receptive field to the maximum scale (approximately 100 nm, kernel size: 30×30×30, expansion coefficient: 30, stride: 30) compared to the multi-scale convolutional network, preserving long-range structural correlations and capturing long-range structural correlation features. Different convolutional kernels correspond to different voxel sizes of the input data, which can better balance the relationship between material information density and software performance. For performance prediction, a ResNet residual architecture (e.g., a network depth of 24 layers) can be used. The input of ResNet is the voxel tensor information of a three-dimensional periodic atomic lattice, and the output is the predicted sodium-ion battery anode performance parameters (capacity, sodium embedding potential, etc.), ensuring stable gradient transfer in deep structures. This network structure is innovative in structure and specifically optimized for sodium-ion battery carbon material tasks compared to existing convolutional networks. Optionally, a multi-channel three-dimensional tensor is input into a pre-trained performance prediction model, and the performance prediction results of the material to be predicted are obtained from the output of the performance prediction model, including: Multi-scale porosity features are extracted from multi-channel 3D tensors using a multi-scale convolutional network; long-range structural correlation features are extracted from multi-channel 3D tensors using a dilated convolutional network; multi-scale porosity features and long-range structural correlation features are fused using a feature fusion network to generate fused features; and performance prediction is performed based on the fused features using a residual network to output the performance prediction results of the material to be predicted.

[0042] Understandably, based on the aforementioned network structure, multi-scale porosity features are extracted from the multi-channel 3D tensor using a multi-scale convolutional network. Long-range structural correlation features are extracted from the multi-channel 3D tensor using a dilated convolutional network. Subsequently, a ResNet residual architecture is employed to ensure stable gradient propagation in the deep structure. A feature fusion network is then used to fuse the multi-scale porosity features and long-range structural correlation features, generating fused features. This feature fusion network can be a cross-scale feature fusion block (CFFB) to simultaneously preserve both local and global information. The final output layer employs a multi-task regression approach to predict the performance of the sodium-ion battery anode (such as capacity and sodium intercalation potential). For example, it can simultaneously predict multiple key performance indicators such as specific capacity, sodium intercalation potential, volume expansion rate, and structural stability score, thus obtaining a comprehensive evaluation of multiple aspects of the material's performance in a single prediction.

[0043] In one embodiment, when multiple key performance indicators are predicted, a standard value for each indicator is calculated. The corresponding values ​​of each indicator are then compared with the standard values ​​to calculate a comprehensive score. The comprehensive score can be the average of the percentages calculated for the four variables. For example, the value x corresponding to a key performance indicator is compared with the standard value x0. The specific capacity can be scored using formula (1), and the remaining parameters can be scored using formula (2). The comprehensive performance score is the average of the percentages calculated for the four variables, where x0 can be defined by the user.

[0044] (x x0) / x0×100% formula (1) x0 / (x x0)×100% formula (2) Standardization using the above formula can eliminate dimensions.

[0045] Optionally, multi-scale pore features can be extracted from multi-channel 3D tensors using multi-scale convolutional networks, including: Different pore features are extracted from a multi-channel 3D tensor through multiple independent convolution channels; the different pore features are then concatenated to obtain multi-scale pore features.

[0046] Understandably, multiple parallel convolutional neural networks are used to extract pore features at different scales from a multi-channel 3D tensor. Subsequently, these pore features at different scales are concatenated in a fusion layer to obtain multi-scale pore features. In other words, after extracting pore features from networks at each scale, the scale features can be concatenated first, and then the concatenated multi-scale pore features can be fused with long-range structure-related features. Alternatively, the pore features extracted from networks at each scale can be directly fused with long-range structure-related features; the specific feature fusion method is not limited.

[0047] Optionally, long-range structural correlation features can be extracted from multi-channel 3D tensors using dilated convolutional networks, including: The dilation coefficient of the dilated convolutional network is determined based on the structural characteristics of the multi-channel 3D tensor; long-range structural correlation features at different structural scales in the multi-channel 3D tensor are extracted based on the dilation coefficient.

[0048] Understandably, the dilation coefficient of the dilated convolutional network is adjustable to adapt to the long-range information capture requirements under different structural sizes. Specifically, the dilation coefficient of the dilated convolutional network is dynamically determined based on the material porosity distribution and structural periodicity features extracted from the multi-channel 3D tensor, that is, the dilation coefficient is determined according to the actual structure between atoms. Subsequently, under this dilation coefficient, the dilated convolutional network extracts long-range structural correlation features spanning a large spatial range in the multi-channel 3D tensor.

[0049] Optionally, after outputting the performance prediction results of the material to be predicted, the method further includes: Obtain multi-objective constraints; where multi-objective constraints are a set of constraints containing multiple performance indicators and their corresponding threshold ranges; store the performance prediction results of the material to be predicted in the material database, and select candidate materials that meet the multi-objective constraints based on the performance prediction results of different materials stored in the material database.

[0050] Understandably, to achieve performance screening and optimized design, a post-processing screening module is integrated based on the performance prediction results. Users can automatically screen candidate materials that meet specific performance requirements according to set multi-objective constraints (such as specific capacity not less than 300 mAh / g, ion migration barrier less than 0.2 eV, etc.), and use the screening results for updating the material database, performance optimization, or input for new structural designs. In addition, high-throughput batch processing and rapid iteration are supported to efficiently complete the automated selection of candidate materials in the structural database.

[0051] Optionally, the performance prediction model can be trained using the following steps: Obtain a performance prediction model trained on a first material dataset and a second material dataset; wherein the first and second material datasets target different material types, and the material type of the material to be predicted corresponds to the material type targeted by the second material dataset; during the training of the performance prediction model based on the second material dataset, freeze some feature extraction modules in the performance prediction model, adjust the model parameters of the task execution module in the performance prediction model, wherein the task execution module is used to execute the material performance prediction task; and transfer the weights trained on other prediction tasks to the task execution module to obtain the trained performance prediction model.

[0052] Understandably, introducing transfer learning strategies during the training phase of performance prediction models is a strategy that uses model weights pre-trained on similar tasks as initialization to reduce training time and improve prediction accuracy. For example, this can be used for cross-material system and cross-task transfer on different sodium-ion battery datasets. Cross-material system transfer refers to transferring model parameters trained on a first material dataset to a model to be trained on a second material dataset, where the material to be predicted belongs to the material type corresponding to the second material dataset. For example, when transferring from a hard carbon dataset to a doped carbon or porous carbon dataset, the first few layers of the general feature extraction module of the model trained on the first material dataset are frozen, and only subsequent task-specific layers (such as material performance prediction tasks) are fine-tuned to achieve fast convergence and prevent overfitting. Cross-task transfer refers to the ability to transfer weights trained on an existing capacity prediction task to sodium embedding potential prediction or stability regression tasks.

[0053] Understandably, to improve model adaptability, a feature reweighting mechanism and adaptive learning rate adjustment strategy can be introduced during the transfer learning phase. This allows the model to automatically adjust feature importance based on the distribution of the target data, further enhancing its transfer generalization ability. For example, during the model training and prediction phases, weight parameters pre-trained on an existing sodium-ion battery database can be used as initialization to improve the model's convergence speed and prediction accuracy on complex structured data.

[0054] For example, Figure 2 This is a schematic diagram of a carbon material structure provided in an embodiment of this application. Figure 2 The diagram shows the three-dimensional atomic structure derived from CIF or XYZ files, followed by a three-dimensional tensor representation after voxelization at 1 Å resolution and multi-channel feature encoding. Each channel records the atomic density, local coordination number, Voronoi volume, and Na+. + Information about functional cavities, etc.

[0055] The method provided in this application proposes a three-dimensional structural representation method combining multi-channel voxelization and porosity-preserving filters to achieve explicit encoding of functional cavities, overcoming the limitations of statistical feature or graph structure representation. Secondly, in terms of model structure, a three-dimensional convolutional neural network architecture integrating multi-scale convolution and dilated convolution is constructed, and a cross-scale feature fusion module is introduced to achieve unified modeling of multi-scale porosity features and long-range correlation information. Thirdly, in terms of training and task design, a transfer learning and multi-task regression mechanism is combined to enable the model to achieve collaborative prediction of multiple performance indicators under small sample conditions. Finally, at the system implementation level, an all-in-one hardware and software co-optimization scheme for material prediction tasks is proposed to achieve deep coupling design of algorithm and hardware.

[0056] In terms of feasibility, based on technologies such as 3D convolutional neural networks and transfer learning, the input data consists of standardized material structure files (CIF, XYZ), with clear and easily accessible data sources. The GPU servers, high-speed storage, and CUDA software environment on which it relies provide a stable foundation for implementation. At the same time, the integrated deployment through an all-in-one machine reduces system complexity and demonstrates strong feasibility.

[0057] In terms of cost, the computational resource consumption and time costs are significantly reduced during material performance prediction and screening. Specifically, prediction based on a 3D convolutional neural network can complete the performance prediction of a single structure in a short time after model training, greatly improving screening efficiency. Regarding hardware resource utilization, optimization strategies such as sparse tensor representation, mixed-precision computation, and convolutional kernel fusion effectively reduce GPU memory usage and memory bandwidth pressure, significantly improving the utilization efficiency per unit of computing power. Simultaneously, the all-in-one deployment solution, through software and hardware co-optimization, also reduces the additional costs of deployment, operation, and network communication associated with traditional distributed clusters. At the R&D process level, it reduces reliance on high-cost computing resources (such as large-scale CPU clusters), shortens the material R&D cycle, and reduces trial-and-error costs and manpower investment.

[0058] Based on the above embodiments, Figure 3 A flowchart illustrating a material property prediction method provided in this application embodiment is shown, specifically including as follows: Figure 3 The following steps are shown: (1) Input atomic structure data; (2) Voxelization of structure data; (3) Extraction of multi-channel coefficient tensor features; (4) Prediction of transfer learning performance; (5) Screening of material properties; (6) Output of prediction and screening results.

[0059] Understandably, (1) to (6) above intuitively illustrate the complete path from atomic structure to performance prediction. For specific implementation steps, please refer to the above embodiments, which will not be repeated here.

[0060] Based on the above embodiments, Figure 4 This application provides a schematic diagram of the structure of an electronic device, which includes a computing module, an interconnection module, and a storage module integrated on at least one cabinet. The computation module includes multiple servers for executing any of the aforementioned material property prediction methods; the interconnection module is used to realize data conversion between multiple servers; and the storage module is used to store atomic structure data, multi-channel three-dimensional tensors, and performance prediction results involved in the computation module.

[0061] Understandably, the electronic device can be understood as an all-in-one machine adapted to high-performance computing environments, suitable for various user scenarios. The electronic device adopts an AI cluster layout, with multiple cabinets integrating multiple servers and a high-speed interconnect switch (interconnect module), and equipped with a storage system (storage module) scalable to PB levels, capable of meeting the storage and retrieval needs of large-scale structured data. The servers include computing modules. In one embodiment, the all-in-one machine uses two 19U racks to integrate eight 4U GPU servers. Each 4U server is configured with a dual-socket high-end server (e.g., with 32 performance cores, a base frequency of 2.8GHz, a maximum turbo frequency of 4.1GHz, and a TDP of 300W) and a processor (e.g., 32 cores per socket), equipped with eight GPUs (e.g., 80GB of VRAM). Nodes are interconnected through a high-performance data center network (e.g., 400Gb / s bandwidth, nanosecond-level latency), with a total VRAM of 5.12 TB and storage using a 6×102.4 TB NVMe array.

[0062] Understandably, the software architecture of the electronic device is based on a computing module. This architecture includes a front-end graphical interface for uploading and configuring filtering conditions for atomic structure files, and a back-end module for generating structural tensors, multi-scale convolutional prediction, and performance filtering. This enables fully automated processing from data input, model training, prediction to result filtering. The specific implementation process is detailed in the above embodiments and will not be elaborated upon here. In one embodiment, the software layer establishes a distributed training environment based on the development platform and cross-node communication standards. Combined with the inference engine and the aforementioned sparse tensor storage module, this further improves the model's throughput in batch prediction scenarios. Through the synergistic optimization of this hardware and software architecture, the electronic device can effectively reduce the bandwidth pressure on GPU memory and RAM through memory access optimization strategies such as data block loading, sparse tensor storage, mixed-precision computation, and convolutional kernel fusion. This highly integrated, high-bandwidth, and low-latency integrated hardware architecture design not only improves data transmission and training efficiency but also ensures the model's real-time prediction capability on high-resolution three-dimensional voxel data.

[0063] The electronic device disclosed herein, in terms of software, achieves a unified characterization of multi-scale porosity and long-range correlation information through multi-channel voxelized structural representation, a 3D deep network integrating multi-scale convolution and dilated convolution, and explicit modeling of functional cavities, thus achieving an effective balance between prediction accuracy and computational efficiency. Simultaneously, in terms of hardware, through integrated deployment and memory access optimization design, the model training, inference, and data processing processes are highly integrated, reducing system complexity and deployment costs while ensuring computational efficiency. This avoids, to some extent, the negative impacts of related distributed solutions in terms of communication overhead and environmental dependence, effectively improving system security and portability. Furthermore, through the deep coupling design of hardware and algorithms, it achieves the performance of traditional clusters within a relatively small space.

[0064] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0065] Embodiments of this application also provide a material property prediction device, which includes an acquisition unit, a generation unit, and a prediction unit, wherein: The acquisition unit is used to acquire atomic structure data of the material to be predicted; wherein, the atomic structure data is used to characterize the spatial arrangement of atoms in the material to be predicted. The generation unit is used to perform three-dimensional voxelization processing on atomic structure data to generate a multi-channel three-dimensional tensor; wherein, the data of each voxel position in the tensor is a multi-dimensional feature vector, which is used to characterize at least one material-based property of the local space represented by the corresponding voxel. The prediction unit is used to input multi-channel three-dimensional tensors into a pre-trained performance prediction model and obtain the performance prediction results of the material to be predicted from the output of the performance prediction model. The performance prediction model is configured to extract multi-scale porosity features and long-range structural correlation features from multi-channel three-dimensional tensors, and perform performance prediction based on the multi-scale porosity features and long-range structural correlation features. The multi-scale porosity features are used to characterize the pore geometry and topological properties at multiple scales inside the material to be predicted; the long-range structural correlation features are used to characterize the cooperative relationship between structural units at different distances in the material to be predicted.

[0066] Optionally, the generating unit is used for: The atomic structure data is processed into three-dimensional voxels at a first set spatial resolution, and the atomic point cloud is mapped into a multi-channel three-dimensional tensor. Among them, the multidimensional feature vector includes atomic density feature vector, local coordination number feature vector, polyhedral volume feature vector, atomic property feature vector and / or set element functional cavity feature vector. The local coordination number refers to the number of atoms in the neighborhood with a second set spatial resolution of voxel side length. The atomic property feature vector is a two-dimensional vector composed of atomic radius and number of electrons. The set element functional cavity information is identified by the porosity preservation filter.

[0067] Optionally, the material property prediction device is also used for: Analyze atomic structure data to obtain the initial coordinates of the atoms; While maintaining the three-dimensional periodicity of atoms, the initial coordinates are uniformly converted into atomic coordinates based on the Cartesian coordinate system; The atomic coordinates are standardized in order to perform three-dimensional voxelization on the standardized atomic structure data.

[0068] The performance prediction models include multi-scale convolutional networks, dilated convolutional networks, feature fusion networks, and residual networks.

[0069] Optionally, the prediction unit is used for: Multi-scale pore features are extracted from multi-channel 3D tensors using a multi-scale convolutional network. Long-range structural correlation features are extracted from multi-channel 3D tensors using dilated convolutional networks; A feature fusion network is used to fuse multi-scale porosity features and long-range structural correlation features to generate fused features. Performance prediction is performed based on fusion features using residual networks, and the performance prediction results of the material to be predicted are output.

[0070] The multi-scale convolutional network contains multiple independent convolutional channels with different aperture ranges and receptive field ranges based on different kernel sizes and strides.

[0071] Optionally, the prediction unit is used for: Different pore features are extracted from multi-channel 3D tensors by using multiple independent convolution channels; By stitching together different pore features, multi-scale pore features are obtained.

[0072] Among them, the receptive field of dilated convolutional networks is larger than that of multi-scale convolutional networks.

[0073] Optionally, the prediction unit is also used for: The dilation coefficient of the dilated convolutional network is determined based on the structural characteristics of the multi-channel three-dimensional tensor. Long-range structural correlation features at different structural scales in multi-channel 3D tensors are extracted based on the expansion coefficient.

[0074] Optionally, the material property prediction device is also used for: Obtain multi-objective constraints; where multi-objective constraints are a set of constraints that include multiple performance indicators and their corresponding threshold ranges; The performance prediction results of the material to be predicted are stored in the material database, and candidate materials that meet the multi-objective constraints are selected based on the performance prediction results of different materials stored in the material database.

[0075] Optionally, the material property prediction device is also used for: Obtain the performance prediction model trained based on the first material dataset and the second material dataset; wherein the first material dataset and the second material dataset target different material types, and the material type of the material to be predicted corresponds to the material type targeted by the second material dataset; During the training of the performance prediction model based on the second material dataset, some feature extraction modules in the performance prediction model are frozen, and the model parameters of the task execution module in the performance prediction model are adjusted. The task execution module is used to perform material performance prediction tasks. In addition, the weights trained on other prediction tasks are transferred to the task execution module to obtain the trained performance prediction model.

[0076] Among them, the atomic structure data are carbon material structure data, and the performance prediction results are performance prediction results of carbon materials for sodium-ion battery anodes, including at least one of the following performance indicators: specific capacity, sodium intercalation potential, volume expansion rate, and structural stability.

[0077] For a description of the features in the embodiment corresponding to the material property prediction device, please refer to the relevant description of the embodiment corresponding to the material property prediction method, which will not be repeated here.

[0078] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above-described material property prediction method embodiments.

[0079] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above-described material property prediction method embodiments when run.

[0080] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0081] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above-described material property prediction method embodiments.

[0082] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps in any of the above-described material property prediction method embodiments.

[0083] Any of the components, modules, units, parts, methods, and operations described herein can be implemented using software, firmware, hardware (e.g., fixed logic circuitry), manual processing, or any combination thereof. Alternatively or additionally, any functionality described herein can be executed at least in part by one or more hardware logic components, such as, but not limited to, a central processing unit (CPU), a field-programmable gate array (FPGA), an application-specific integrated circuit (ASIC), an application-specific standard product (ASSP), a system-on-a-chip (SoC), a complex programmable logic device (CPLD), a microprocessor (MCU), etc. The terms "system," "computing device," or "apparatus" as used herein encompass various means, devices, and machines for processing data, including, for example, one or more programmable processors, computers, SoCs, or combinations thereof. The apparatus may also include code that creates an execution environment for the computer program in question, such as code constituting processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or one or more combinations thereof. The aforementioned computer program (also known as a program, software, software application, app, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and can be deployed in any form, including as a standalone program or as a module, component, subroutine, object, or other unit suitable for a computing environment.

[0084] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0085] The above provides a detailed description of a material performance prediction method provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are merely for the purpose of helping to understand the method and core ideas of this application. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A method for predicting material properties, characterized in that, include: Acquire atomic structure data of the material to be predicted; wherein the atomic structure data is used to characterize the spatial arrangement of atoms in the material to be predicted; The atomic structure data is subjected to three-dimensional voxelization to generate a multi-channel three-dimensional tensor; wherein, the data of each voxel position in the tensor is a multi-dimensional feature vector, which is used to characterize at least one material-based property of the local space represented by the corresponding voxel. The multi-channel three-dimensional tensor is input into a pre-trained performance prediction model, and the performance prediction results of the material to be predicted are obtained from the output of the performance prediction model. The performance prediction model is configured to extract multi-scale porosity features and long-range structural correlation features from the multi-channel three-dimensional tensor, and perform performance prediction based on the multi-scale porosity features and the long-range structural correlation features. The multi-scale porosity features are used to characterize the pore geometry and topological properties at multiple scales within the material to be predicted; the long-range structural correlation features are used to characterize the cooperative relationship between structural units at different distances in the material to be predicted.

2. The method according to claim 1, characterized in that, The step of performing three-dimensional voxelization on the atomic structure data to generate a multi-channel three-dimensional tensor includes: The atomic structure data is subjected to three-dimensional voxelization processing at a first set spatial resolution, and the atomic point cloud is mapped into a multi-channel three-dimensional tensor. The multidimensional feature vector includes an atomic density feature vector, a local coordination number feature vector, a polyhedral volume feature vector, an atomic property feature vector, and / or a set element functional cavity feature vector. The local coordination number refers to the number of atoms in a neighborhood with a second set spatial resolution equal to the voxel side length. The atomic property feature vector is a two-dimensional vector composed of atomic radius and electron number. The set element functional cavity information is identified by a porosity preservation filter.

3. The method according to claim 1 or 2, characterized in that, Before performing three-dimensional voxelization on the atomic structure data, the method further includes: The atomic structure data is analyzed to obtain the initial coordinates of the atoms; While maintaining the three-dimensional periodicity of the atoms, the initial coordinates are uniformly converted into atomic coordinates based on the Cartesian coordinate system; The atomic coordinates are standardized to perform three-dimensional voxelization on the standardized atomic structure data.

4. The method according to claim 1, characterized in that, The performance prediction model includes a multi-scale convolutional network, a dilated convolutional network, a feature fusion network, and a residual network. The step of inputting the multi-channel three-dimensional tensor into the pre-trained performance prediction model and obtaining the performance prediction results of the material to be predicted output by the performance prediction model includes: Multi-scale pore features are extracted from the multi-channel three-dimensional tensor using the multi-scale convolutional network. Long-range structural correlation features are extracted from the multi-channel 3D tensor using the dilated convolutional network. The feature fusion network is used to fuse the multi-scale porosity features and the long-range structural correlation features to generate fused features. The residual network performs performance prediction based on the fusion features and outputs the performance prediction results of the material to be predicted.

5. The method according to claim 4, characterized in that, The multi-scale convolutional network comprises multiple independent convolutional channels with different aperture ranges and receptive field ranges based on different kernel sizes and strides. Extracting multi-scale pore features from the multi-channel 3D tensor using the multi-scale convolutional network includes: Different pore features are extracted from the multi-channel 3D tensor through the multiple independent convolution channels; By stitching together the different pore features, multi-scale pore features are obtained.

6. The method according to claim 4, characterized in that, The receptive field of the dilated convolutional network is larger than that of the multi-scale convolutional network. Extracting long-range structural correlation features from the multi-channel 3D tensor using the dilated convolutional network includes: The dilation coefficient of the dilated convolutional network is determined based on the structural characteristics of the multi-channel three-dimensional tensor. Based on the expansion coefficient, long-range structural correlation features at different structural scales in the multi-channel three-dimensional tensor are extracted.

7. The method according to claim 1, characterized in that, After outputting the performance prediction results of the material to be predicted, the method further includes: Obtain multi-objective constraints; wherein, the multi-objective constraints are a set of constraints that include multiple performance indicators and their corresponding threshold ranges; The performance prediction results of the material to be predicted are stored in the material database, and candidate materials that meet the multi-objective constraints are selected based on the performance prediction results of different materials stored in the material database.

8. The method according to claim 1, characterized in that, The performance prediction model is trained through the following steps: Obtain a performance prediction model trained based on a first material dataset and a second material dataset; wherein the first material dataset and the second material dataset target different material types, and the material type of the material to be predicted corresponds to the material type targeted by the second material dataset; During the training of the performance prediction model based on the second material dataset, some feature extraction modules in the performance prediction model are frozen, and the model parameters of the task execution module in the performance prediction model are adjusted, wherein the task execution module is used to perform material performance prediction tasks; and the weights trained on other prediction tasks are transferred to the task execution module to obtain a trained performance prediction model.

9. The method according to claim 1, characterized in that, The atomic structure data is carbon material structure data, and the performance prediction results are performance prediction results for sodium-ion battery anode carbon materials, including at least one performance index among specific capacity, sodium intercalation potential, volume expansion rate, and structural stability.

10. An electronic device, characterized in that, The electronic device includes a computing module, an interconnection module, and a storage module integrated on at least one cabinet unit, wherein... The computing module includes multiple servers for executing the material property prediction method as described in any one of claims 1-9; The interconnection module is used to realize data conversion between the multiple servers; The storage module is used to store the atomic structure data, multi-channel three-dimensional tensors, and performance prediction results involved in the computing module.

Images