An artificial intelligence-based new material performance prediction method and system

Abstract

The application relates to the technical field of the cross of new material research and development and artificial intelligence, in particular to a new material performance prediction method and system based on artificial intelligence. The method first acquires the component structure and microstructure multi-source heterogeneous data of a target new material, completes cross-modal feature alignment and fusion through a pre-trained multi-modal feature encoder, generates material fusion feature representation, loads the fusion feature to a graph neural network model enhanced by physical knowledge, iteratively updates and generates material deep feature coding rich in physical semantics, and obtains multi-target performance initial prediction values in parallel through a multi-task performance predictor and captures model cognitive uncertainty. The application completes the reliability evaluation based on the uncertainty, triggers an active learning feedback mechanism to iteratively optimize the model for low confidence samples, effectively improves the accuracy and reliability of new material performance prediction, and can greatly shorten the research and development cycle and reduce the research and development trial and error cost.

Description

Technical Field

[0001] This invention relates to the field of new material research and development and artificial intelligence, specifically to a method and system for predicting the performance of new materials based on artificial intelligence. Background Technology

[0002] Traditional new materials research and development models heavily rely on the experience of researchers and trial-and-error experimental verification. From material composition design, structural optimization, and preparation process adjustment to final performance testing, dozens or even hundreds of experimental cycles are required, with a complete research and development cycle often lasting several to over a decade. This model not only requires significant investment in experimental equipment and raw material consumption but also struggles to comprehensively consider the coupled effects of multi-scale factors such as composition, structure, and microstructure on the final material performance. This can easily lead to deviations in research direction, resulting in a disproportionate relationship between research input and output.

[0003] Existing technologies have significant limitations in integrating and utilizing multi-source heterogeneous data. Most performance prediction methods only use single-component data or crystal structure data of materials as model inputs, failing to effectively integrate multi-dimensional and multi-scale material data such as composition, structure, and microstructure. This results in an inability to comprehensively characterize various key factors affecting the final performance of materials, leading to inherent deficiencies in the model's feature representation capabilities and hindering substantial breakthroughs in prediction accuracy. Furthermore, existing technologies have limited capabilities in standardizing unstructured material data such as text, spectra, and images, failing to achieve effective alignment and deep fusion of data from different modalities. This prevents the full utilization of massive amounts of publicly available material research data, resulting in a serious waste of data resources.

[0004] Existing technologies suffer from significant deficiencies in the physical interpretability and generalization ability of their models. Most current deep learning-based material property prediction models employ a purely data-driven black-box learning model. The model training and feature learning processes do not incorporate prior physical knowledge from materials science, making them prone to generating feature representations and prediction results that do not conform to the fundamental principles of materials physicochemistry. While these black-box models can achieve certain predictive results within the material systems covered by the training set, the reliability of the predictions drops drastically when faced with novel material systems outside the training set, resulting in severely insufficient generalization ability. Furthermore, the decision-making process of black-box models cannot be scientifically and rationally explained, making their predictions unreliable for materials researchers and failing to provide effective scientific support for new materials development decisions.

[0005] Existing technologies lack closed-loop design for iterative optimization of models. The training process of existing prediction models is mostly static; once initial training is complete, model parameters remain fixed, preventing autonomous iterative optimization during practical applications. When the prediction requirements for novel material systems lead to a decrease in model accuracy, manual sample selection, data labeling, and model retraining are required, making optimization cumbersome and complex. Furthermore, existing technologies cannot accurately select the most valuable samples for improving model performance, often requiring a large amount of newly labeled data to achieve even a small improvement. This results in low model iteration efficiency and an inability to adapt to the continuously changing and rapidly iterating application requirements in the development of new materials.

[0006] Based on the numerous shortcomings and industry pain points in the existing technologies, developing a new material performance prediction method and system that can achieve comprehensive fusion of multi-source heterogeneous data, deep integration of physical prior knowledge and data-driven learning, parallel prediction of multiple performance indicators, quantitative evaluation of the reliability of prediction results, and autonomous closed-loop iterative optimization capability has become a core technical problem that urgently needs to be solved in the field of new material research and development and artificial intelligence. Summary of the Invention

[0007] The purpose of this invention is to provide a new material performance prediction method and system based on artificial intelligence to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] A novel material performance prediction method based on artificial intelligence includes the following steps:

[0010] Step S1: Obtain initial data on the multi-source heterogeneity of the target new material; the initial data on multi-source heterogeneity includes composition data, structural data, and microstructure data;

[0011] Step S2: Input the multi-source heterogeneous initial data into the pre-trained multimodal feature encoder to perform cross-modal feature alignment and fusion, and generate a material fusion feature representation for characterizing the target new material;

[0012] Step S3: Load the material fusion feature representation as the initial node feature into the pre-constructed physical knowledge-enhanced graph neural network model; wherein, the graph neural network model uses atoms or atomic groups in the target new material as nodes and inter-atomic interactions as edges to construct a material structure topology graph, and introduces physical consistency constraints based on prior knowledge of the material gene library during the message passing process of the model.

[0013] Step S4: Iteratively update the material structure topology graph using a graph neural network model to aggregate local atomic environment information and global topological structure information, generating material deep feature encoding rich in physical semantics;

[0014] Step S5: Input the material deep feature encoding into the multi-task performance predictor to calculate the initial predicted values ​​of multiple target properties of the target new material in parallel, and at the same time capture the model cognitive uncertainty of the multi-task performance predictor when generating the initial predicted values.

[0015] Step S6: Evaluate the credibility of the initial prediction based on the model's cognitive uncertainty: If the credibility meets the preset conditions, the initial prediction is output as the final prediction result; if the credibility is lower than the preset threshold, the current sample data and its corresponding multi-source heterogeneous initial data are marked as high-value samples, triggering the active learning feedback mechanism, adding the high-value samples to the training set to iteratively optimize the multimodal feature encoder, graph neural network model and multi-task performance predictor, and re-execute steps S2 to S6 using the optimized model until the final prediction result that meets the credibility requirements is obtained.

[0016] An artificial intelligence-based new material performance prediction system, the system being used to execute an artificial intelligence-based new material performance prediction method, the system comprising:

[0017] The multi-source heterogeneous data acquisition module is used to acquire the initial multi-source heterogeneous data of the target new material; the initial multi-source heterogeneous data includes composition data, structural data and microstructure data;

[0018] The multimodal feature encoding module has a built-in pre-trained multimodal feature encoder, which is used to input multi-source heterogeneous initial data into the multimodal feature encoder, perform cross-modal feature alignment and fusion, and generate a material fusion feature representation for characterizing the target new material;

[0019] The physics-enhanced graph neural network module has a pre-built physics-enhanced graph neural network model built in, which is used to load the material fusion feature representation as the initial feature of the node into the graph neural network model. The graph neural network model uses atoms or atomic groups in the target new material as nodes and inter-atomic interactions as edges to construct a material structure topology graph. In the process of message passing in the model, physical consistency constraints based on prior knowledge of the material gene library are introduced.

[0020] The deep feature encoding generation module is used to iteratively update the material structure topology graph through a graph neural network model, so as to aggregate local atomic environment information and global topological structure information to generate material deep feature encoding rich in physical semantics.

[0021] The multi-task performance prediction module has a built-in multi-task performance predictor. It is used to input the material deep feature encoding into the multi-task performance predictor and calculate the initial predicted values ​​of multiple target properties of the target new material in parallel. At the same time, it captures the model cognitive uncertainty of the multi-task performance predictor when generating the initial predicted values.

[0022] The credibility assessment and active learning feedback module is used to assess the credibility of the initial prediction value based on the model's cognitive uncertainty. If the credibility meets the preset conditions, the initial prediction value is output as the final prediction result. If the credibility is lower than the preset threshold, the current sample data and its corresponding multi-source heterogeneous initial data are marked as high-value samples, triggering the active learning feedback mechanism. The high-value samples are added to the training set to iteratively optimize the multimodal feature encoder, graph neural network model, and multi-task performance predictor. The optimized model is then used to re-trigger the multimodal feature encoding module, the physics-enhanced graph neural network module, the deep feature encoding generation module, the multi-task performance prediction module, and the credibility assessment and active learning feedback module to perform their respective functions until the final prediction result that meets the credibility requirements is obtained.

[0023] The central control module is used to schedule and control the working processes of the multi-source heterogeneous data acquisition module, the multimodal feature encoding module, the physical knowledge-enhanced graph neural network module, the deep feature encoding generation module, the multi-task performance prediction module, and the credibility assessment and active learning feedback module.

[0024] As can be seen from the technical solutions provided by the present invention above, the beneficial effects of the new material performance prediction method and system based on artificial intelligence provided by the present invention are:

[0025] This invention simultaneously acquires and processes the compositional, structural, and microstructure data of the target new material. Through a standardized processing flow, unstructured text-based image data is converted into structured data in a unified format. Then, a multimodal feature encoder is used to complete cross-modal feature alignment and fusion, effectively eliminating the semantic gap between different types of data. This achieves full-dimensional information fusion of the material from atomic-level composition to microstructure morphology, providing a comprehensive and high-quality feature foundation for subsequent performance prediction and ensuring the accuracy of the prediction results from the data source.

[0026] This invention constructs a physics-enhanced graph neural network model. It uses atoms or atomic groups in a material as nodes and interatomic interactions as edges to build a topological graph that matches the material's physical structure. In the process of model message passing and feature aggregation, a physical consistency constraint based on prior knowledge of the material gene library is introduced. The model learning process is dynamically adjusted through a differentiable constraint function, which forces the model feature learning to always follow the physical and chemical laws of the material. This breaks the limitations of the traditional black box model, makes the generated feature codes have rich physical semantics, and makes the prediction results more in line with the basic principles of materials science. It significantly improves the scientific reliability and generalization ability of the model.

[0027] This invention achieves simultaneous parallel prediction of multiple performance indicators of a target new material through a multi-task performance predictor. It eliminates the need to build and train prediction models separately for each single performance indicator, greatly simplifying the performance prediction process and improving the efficiency of multi-dimensional performance prediction. At the same time, by leveraging a variable hierarchical model based on Bayesian inference, it accurately captures and quantifies the cognitive uncertainty of the model. Based on this, a complete reliability evaluation system for prediction results is constructed, solving the industry pain point that traditional deterministic prediction models cannot determine the reliability of results. It can intuitively present the confidence level of prediction results, providing a quantifiable reliability basis for scheme selection and decision-making in the process of new material research and development.

[0028] Based on the confidence assessment of the prediction results, this invention automatically marks low-confidence samples as high-value samples, triggering an active learning feedback mechanism. After sample labeling is completed through human-computer interaction, high-value training samples are expanded into the original training set, and the entire model is jointly fine-tuned and optimized, forming a complete working loop of prediction, evaluation, optimization, and re-prediction. This mechanism can continuously compensate for the generalization shortcomings of the model caused by insufficient training data or out-of-range sample distribution, continuously reduce the prediction uncertainty of the model for new material samples, and enable the model to autonomously complete performance iteration and capability upgrades in continuous new material research and development applications, maintaining a high level of prediction accuracy and adaptability in the long term.

[0029] This invention achieves accurate pre-prediction of material properties through artificial intelligence technology, enabling simulation evaluation and potential screening of material properties before experiments are conducted, significantly reducing unnecessary experimental verification steps and effectively avoiding ineffective R&D investment. At the same time, the model's multi-dimensional information fusion capability and continuous optimization characteristics can adapt to the R&D needs of various new material systems, further accelerating the entire R&D process of new materials from composition design and structural optimization to performance verification, significantly improving the efficiency and success rate of new material R&D, and providing efficient technical support for the rapid development of the new materials industry. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the steps in the method for predicting the properties of new materials based on artificial intelligence according to the present invention.

[0031] Figure 2 This is a schematic diagram of the structure of a novel material performance prediction system based on artificial intelligence according to the present invention. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0033] To better understand the above technical solutions, the following will provide a detailed explanation of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0034] like Figure 1-2 As shown, this embodiment of the invention provides a new material performance prediction method and system based on artificial intelligence, including:

[0035] The present invention provides a novel material performance prediction method based on artificial intelligence, comprising the following steps:

[0036] Step S1: Obtain initial data on the multi-source heterogeneity of the target new material; the initial data on multi-source heterogeneity includes composition data, structural data, and microstructure data;

[0037] Step S2: Input the multi-source heterogeneous initial data into the pre-trained multimodal feature encoder to perform cross-modal feature alignment and fusion, and generate a material fusion feature representation for characterizing the target new material;

[0038] Step S3: Load the material fusion feature representation as the initial node feature into the pre-constructed physical knowledge-enhanced graph neural network model; wherein, the graph neural network model uses atoms or atomic groups in the target new material as nodes and inter-atomic interactions as edges to construct a material structure topology graph, and introduces physical consistency constraints based on prior knowledge of the material gene library during the message passing process of the model.

[0039] Step S4: Iteratively update the material structure topology graph using a graph neural network model to aggregate local atomic environment information and global topological structure information, generating material deep feature encoding rich in physical semantics;

[0040] Step S5: Input the material deep feature encoding into the multi-task performance predictor to calculate the initial predicted values ​​of multiple target properties of the target new material in parallel, and at the same time capture the model cognitive uncertainty of the multi-task performance predictor when generating the initial predicted values.

[0041] Step S6: Evaluate the credibility of the initial prediction based on the model's cognitive uncertainty: If the credibility meets the preset conditions, the initial prediction is output as the final prediction result; if the credibility is lower than the preset threshold, the current sample data and its corresponding multi-source heterogeneous initial data are marked as high-value samples, triggering the active learning feedback mechanism, adding the high-value samples to the training set to iteratively optimize the multimodal feature encoder, graph neural network model and multi-task performance predictor, and re-execute steps S2 to S6 using the optimized model until the final prediction result that meets the credibility requirements is obtained.

[0042] In this embodiment, the core function of step S1 is to retrieve and extract various types of raw data for the target new material from public material databases and internal material knowledge bases. Unstructured raw data is standardized and quantified using techniques such as natural language processing, graph analysis, and image preprocessing. This transforms text-based graph image data into structured component data and microstructure data that are uniformly formatted and interconnected. This provides high-quality, directly input-input-to-the-model basic data support for subsequent cross-modal feature alignment and fusion of the multimodal feature encoder. The detailed steps are as follows:

[0043] Step S1-1: Retrieval and extraction of raw data related to the target new material:

[0044] The system receives the identification information or basic physical property parameters of the target new material input from the user terminal, uses this information as search keywords, and calls the search interfaces of public material databases and internal material knowledge bases to perform precise matching and retrieval in the databases and knowledge bases. The search process covers three major categories of data: composition-related text records, structure-related spectral files, and microstructure-related image files, ensuring the comprehensiveness of the search results. The search results are deduplicated and filtered to remove invalid, incomplete, and irrelevant data, and extract original composition text, original structure spectral files, and original microstructure image files that are highly relevant to the target new material, providing raw data materials for subsequent data processing steps.

[0045] Step S1-2: Raw component text processing and component data generation:

[0046] The extracted raw component text is processed by natural language processing, which sequentially performs text cleaning, word segmentation, entity recognition, and information extraction. The text accurately identifies and extracts the element types, stoichiometry, and doping information. The extracted multi-dimensional component information is then vectorized to construct a fixed-dimensional structured component feature vector. Each dimension of this vector corresponds to the content percentage of a type of element, the type and content of doping elements, and other feature values. The converted structured component feature vector is the component data that can be directly used as input to the model.

[0047] Step S1-3: Parsing the original structure map file and generating structure data:

[0048] The original structural map files are professionally analyzed, and crystal structure parameters and space group information are extracted from the files using a map feature extraction algorithm. The crystal structure parameters include core physical parameters such as interatomic spacing, unit cell parameters, and crystal plane indices. Based on the extracted crystal structure parameters, an initial atomic structure model containing atomic occupancy and bonding relationships is reconstructed in three-dimensional space using an atomic structure reconstruction algorithm. This model can accurately characterize the atomic hierarchical structural features of the target new material, and the reconstructed initial atomic structure model is the structural data that can be directly used as model input.

[0049] Steps S1-4: Processing of raw microstructure image files and generation of microstructure data:

[0050] A complete image processing workflow is performed on the acquired raw microstructure image files. First, image preprocessing is performed, including image denoising, image enhancement, and image size normalization, to eliminate noise interference, improve image clarity, and standardize image size specifications. Then, semantic segmentation is performed on the preprocessed images. The image semantic segmentation model identifies key regions such as grain phase regions and defects in the images. Quantitative analysis is performed on each identified region to calculate and extract quantitative features such as grain size, phase distribution ratio, and the number and size of defects. The extracted quantitative features are then mapped to generate a quantitative microstructure feature map that can quantitatively reflect the microstructure characteristics of the target new material. The generated quantitative microstructure feature map is the microstructure data that can be directly used as input to the model.

[0051] Steps S1-5: Multi-type data alignment and encapsulation and multi-source heterogeneous initial data generation:

[0052] The component feature vectors generated in step S1-2, the initial atomic structure model generated in step S1-3, and the quantitative tissue feature map generated in step S1-4 are aligned. Using the unique identifier of the target new material as the association key, a one-to-one correspondence is established between the three types of data to ensure that the three types of data represent different dimensional features of the same target new material. The aligned three types of data are then encapsulated in a unified data structure according to the input data format requirements of the multimodal feature encoder. The unified and interconnected dataset after encapsulation is the final multi-source heterogeneous initial data, which will be directly input into the multimodal feature encoder for subsequent cross-modal feature alignment and fusion operations.

[0053] In this embodiment, the core function of step S2 is to extract modality-specific features from the multi-source heterogeneous initial data of the target new material. Through unified semantic space mapping and cross-modal attention mechanisms, precise semantic alignment of features at different modalities is achieved. Then, based on materials science principles, a contrastive learning loss function is constructed to complete the fusion and compression of multi-dimensional features. Finally, a material fusion feature representation that comprehensively characterizes the composition, structure, and microstructure of the target new material is generated, providing a high-quality input foundation for subsequent node feature loading and deep feature encoding of the graph neural network model. The detailed steps are as follows:

[0054] Step S2-1: Extraction of modal-specific features from multi-source heterogeneous data:

[0055] Compositional data, structural data, and microstructure data are input into the corresponding feature extraction branches of the multimodal feature encoder. Compositional data undergoes dimensionality mapping and sequence transformation through an embedding layer to generate a dimensionally unified compositional feature sequence. Structural data is extracted and encoded through a graph convolutional network to generate structural topological features that characterize the atomic hierarchical arrangement of the material. Microstructure data is extracted and filtered for morphological and textural features in the image through a convolutional neural network to generate tissue morphology features that characterize the differences in the microstructure of the material. The features output from all three branches undergo dimensionality standardization, providing a unified input basis for subsequent unified semantic space mapping.

[0056] Step S2-2: Unified semantic space mapping and cross-modal feature alignment:

[0057] Compositional feature sequences, structural topology features, and tissue morphology features are jointly mapped to a unified material feature semantic space to ensure that the three types of features have the same feature dimensions and semantic benchmarks. In this semantic space, the correlation weights between different modal features are calculated based on a cross-modal attention mechanism. The correlation weights are used to perform weighted interaction on the features of each modality to achieve feature alignment of different modal data at the semantic level.

[0058] The formula for calculating the cross-modal attention relevance weight is:

[0059] ,in, This is the correlation weight matrix between different modal features. The feature matrix of the source mode, The feature matrix of the target mode, To unify the feature dimensions in the semantic space, This is a normalization function used to convert the calculated correlation values ​​into weight coefficients in the range of 0 to 1;

[0060] After completing the relevance weight calculation, with By weighted fusion of features corresponding to different modalities, features representing the same material property in different modalities are aligned in semantic space, eliminating the semantic gap between data from different modalities and ensuring that multimodal features have a unified semantic expression logic.

[0061] Step S2-3: Contrastive learning constraints and multimodal feature fusion compression:

[0062] After semantic alignment, the multimodal features are concatenated and fused for compression. By introducing a contrastive learning loss function based on materials science principles, the fused features are constrained and optimized, ultimately generating a material fusion feature representation containing multi-dimensional information on composition, structure, and microstructure.

[0063] The formula for calculating the contrastive learning loss function is:

[0064] ,in, To compare the learning loss values, This represents the total number of material samples within the training batch. The preset interval threshold, For the first Similarity between different modal features of a material sample For the first The similarity between the characteristics of an individual material sample and other material samples within the same batch;

[0065] pass Backpropagation optimization narrows the representation distance of different modal features of the same material in the semantic space, while widening the feature distance between different materials, thus strengthening the exclusive representation ability of the fused features for the target new material. After optimization, the cascaded multimodal features are compressed in dimension through a fully connected layer to generate a material fusion feature representation with fixed dimension and complete information, which is then output to the subsequent graph neural network model for deep feature encoding.

[0066] In this embodiment, the core function of step S3 is to construct a material structure topology graph with physical entity meaning based on the structural data of the target new material. The material fusion feature representation generated in step S2 is loaded as the initial feature of the graph node. During the message passing and feature aggregation process of the graph neural network, physical consistency constraints based on prior knowledge of the material gene library are embedded to ensure that the model feature learning process conforms to the basic principles of material physicochemistry. Finally, intermediate feature codes with physical semantic information are generated, providing a feature foundation that conforms to physical laws for the subsequent generation of material deep feature codes. The detailed steps are as follows:

[0067] Step S3-1: Construction of Material Structure Topology Diagram:

[0068] Based on the structural data of the target new material, the atomic-level information of the material and the construction of its topology map are completed. First, the types of atoms, spatial coordinates, and bonding relationships between atoms in the structural data are analyzed. The material structure topology map is constructed based on the core information obtained from the analysis. In the material structure topology map, each atom or group of atoms is defined as a node in the graph, and the chemical bonds or physical interactions between atoms are defined as edges connecting the nodes. After the nodes and edges are defined, the correspondence between nodes and edges is verified to ensure that each edge connects two nodes with physical interactions, eliminating isolated nodes and invalid edges. Finally, a material structure topology map that perfectly matches the atomic structure of the material is generated.

[0069] Step S3-2: Initialize and load node and edge features:

[0070] The material fusion feature representation generated in step S2 is used as the initial feature of the node, completing the one-to-one correspondence loading of features and nodes. This ensures that the initial feature of each node can fully characterize the multi-dimensional information of the composition, structure, and microstructure of the corresponding atom or atomic group. At the same time, each edge in the material structure topology graph is assigned an initialized edge feature. The initialization of the edge feature is based on the distance and bonding type between the atoms represented by the two corresponding nodes. Different bonding types and atomic spacings correspond to different initial values ​​of the edge feature. After loading the node features and edge features, a graph structure data with complete physical entity meaning is formed, providing the input basis for subsequent message passing and feature aggregation in the graph neural network.

[0071] Step S3-3: Embedding Physical Consistency Constraints and Aggregating Graph Message Passing:

[0072] The initialized graph structure data is input into the graph neural network model, and multi-level message passing and feature aggregation operations are performed. During the node update process of each layer of the graph neural network, physical consistency constraints based on prior knowledge of the material gene library are introduced to achieve a deep integration of data-driven learning and physical prior knowledge. First, through the pre-constructed material gene library, the physical interaction potential energy curves and charge distribution laws between corresponding atoms are queried. The queried prior physical quantities are converted into differentiable constraint functions. The weight coefficients in the message passing process are dynamically adjusted by the constraint functions, so that the node features obtained by the model aggregation simultaneously conform to the data-driven laws and known physicochemical principles.

[0073] The calculation formula for message passing and updating in a graph neural network node is as follows:

[0074] ,in, For the first target node in layered network The updated hidden state It is a non-linear activation function. For the target node The set of neighboring nodes, For the first Neighbor nodes in a layered network To the target node Message passing weights, For the first Neighbor nodes in a layered network The hidden state, For the first The trainable weight matrix of a layer network, For the first Bias terms of the layer network;

[0075] The formula for calculating the adjustment of message passing weights by physical consistency constraints is as follows:

[0076] ,in, For the first Initial message passing weights in a layered network without physical constraints This is a differentiable constraint function constructed based on prior knowledge from a materials gene library. Atoms obtained from a materials gene library With atoms The physical potential energy between them Atoms obtained from a materials gene library With atoms The correlation value of charge distribution between them;

[0077] After updating the nodes of each network layer, the consistency between the node features and the prior physical knowledge is verified to ensure that the feature aggregation process always follows the basic laws of materials physics and chemistry, and to avoid the model from producing feature learning results that do not conform to physical principles.

[0078] Step S3-4: Global feature pooling and physical semantic feature generation:

[0079] After the graph neural network model completes message passing and node feature updates at a preset number of layers, a global pooling operation is performed on the node features of the entire graph. The global pooling operation aggregates local atomic environment information and global topological structure information, transforming scattered node-level features into a unified graph-level feature vector. The global pooling operation can be selected as global summation pooling or global average pooling based on the material type and task requirements. During the pooling process, the physical semantic information and structural topological information contained in the features are preserved. After the pooling operation is completed, a material deep feature code containing complete physical semantic information is generated. This code will be directly input into subsequent network layers to complete further feature refinement and performance prediction tasks.

[0080] In this embodiment, the core function of step S4 is to perform hierarchical iterative updates on the material structure topology graph with loaded initial features through a two-layer graph convolutional network. This process sequentially aggregates local coordination environment information and captures mid-range ordered structure information. Then, global pooling is used to aggregate node features at the graph level. Finally, physical property prediction branches and residual connections are used to refine and enhance the features, generating a material deep feature code that simultaneously covers local atomic environment information and global topology information, and fully conforms to the principles of material physicochemicals. This provides a highly discriminative and reliable feature input foundation for subsequent multi-task performance prediction. The detailed steps are as follows:

[0081] Step S4-1: Local feature aggregation and node hidden state update of the first convolutional layer:

[0082] The initial node features and edge features obtained from the initialization and loading in step S3 are input into the first graph convolutional layer of the graph neural network model. The first graph convolutional layer performs feature aggregation operation based on the message passing paradigm, using the chemical bond type and interatomic distance as control coefficients for the aggregation process. For each target node, the feature information of all its first-order neighbor nodes is aggregated to generate preliminary aggregated features that can fully characterize the local coordination environment information around the target node. After the preliminary aggregated features are generated, the preliminary aggregated features are corrected according to the physical consistency constraints introduced in step S3, and feature components that do not conform to the physical and chemical principles of materials are removed to ensure that the aggregated features are consistent with the known interatomic interaction laws, and finally the updated node hidden state of the first layer is obtained.

[0083] The formula for calculating the feature aggregation of the first convolutional layer is:

[0084] ,in, The target node output by the first convolutional layer The updated node is in a hidden state. It is a non-linear activation function. For the target node The set of first-order neighbor nodes, These are polymerization control coefficients set based on the type of interatomic chemical bonds and the interatomic distance. For the target node Corresponding neighbor nodes Initial characteristics of nodes, This is the trainable weight matrix of the first convolutional layer. This refers to the bias term of the convolutional layer in the first graph;

[0085] Step S4-2: Cluster-level feature aggregation and secondary update of node hidden states in the second convolutional layer:

[0086] The updated node hidden states and their corresponding edge features from the first layer are input into the second graph convolutional layer of the graph neural network model. The second graph convolutional layer introduces a hierarchical attention mechanism based on atomic clusters. First, based on the crystal structure characteristics and atomic bonding relationships of the material, atoms that are spatially close and have strong interactions are divided into different atomic clusters. Then, the influence weights of different atomic clusters on the target node are dynamically calculated. Based on the calculated influence weights, the updated node hidden states from the first layer are weighted and aggregated to capture the medium-order structure information of the material and generate cluster-level aggregated features after attention weighting. After the cluster-level aggregated features are generated, they are corrected again based on physical consistency constraints to ensure that the feature information is consistent with the prior physical knowledge in the material gene library, and finally the updated node hidden states from the second layer are obtained.

[0087] The formula for calculating the hierarchical attention weights of atomic clusters is:

[0088] ,in, For the target node The corresponding number The influence weight of an individual atomic cluster For a leaky linear rectified activation function, For the trainable weight vector of the attention mechanism, The target node output by the first convolutional layer The hidden state of the node. For the first The aggregated feature vector of an atomic cluster This is a concatenation operation of feature vectors. For the target node The set of all atomic clusters within its spatial range;

[0089] The formula for updating node features in the second convolutional layer is:

[0090] ,in, The target node output by the convolutional layer of the second graph. The updated node is in a hidden state. This is the trainable weight matrix for the convolutional layer in the second graph. This refers to the bias term of the convolutional layer in the second figure;

[0091] Step S4-3: Global feature aggregation of pooling layer and generation of preliminary material depth feature encoding:

[0092] The updated hidden states of the nodes in the second layer are input into the pooling layer of the graph neural network model. The pooling layer performs a global pooling operation, which can be selected as either global summation pooling or global average pooling, depending on the material structure type and the requirements of the prediction task. The global pooling operation aggregates and transforms the hidden states of all nodes in the material structure topology graph, mapping the scattered node-level features into a fixed-dimensional graph-level feature vector. This feature vector can simultaneously represent the local atomic environment features and global topological structure features of the material. The graph-level feature vector generated after aggregation is the preliminary material deep feature encoding.

[0093] The formula for calculating the global average pooling operation is:

[0094] ,in, For preliminary material depth feature encoding, This represents the total number of nodes in the material structure topology diagram. The output of the second convolutional layer The hidden state of a node after its update;

[0095] The formula for calculating the global summation pooling operation is:

[0096] ;

[0097] Step S4-4: Enhancement of physical constraints in the feature refinement layer and generation of final material depth feature encoding:

[0098] The initial material deep feature encoding is input into the feature refinement layer of the graph neural network model. The feature refinement layer introduces a physical property prediction branch based on prior knowledge from the material gene library. Taking the initial material deep feature encoding as input, it predicts several fundamental physical quantities of the material in parallel. These fundamental physical quantities include physical parameters strongly correlated with the intrinsic properties of the material, such as lattice constant, forming energy, and elastic modulus. The predicted fundamental physical quantities are compared with the corresponding actual physical quantities of the material stored in the material gene library. The deviation between the predicted and actual values ​​is calculated, and residual connections are constructed based on the deviation results. The constructed residual connections are used to fine-tune and enhance the initial material deep feature encoding, strengthening the effective components related to the intrinsic physical properties of the material and suppressing invalid noise components. Finally, a material deep feature encoding rich in physical semantics is generated. This encoding is directly input into the multi-task performance predictor to complete the subsequent material performance prediction and uncertainty assessment tasks.

[0099] The formula for calculating the residual enhancement of the feature refining layer is:

[0100] ,in, Encode the final material deep features rich in physical semantics. It is a multilayer perceptron mapping network. This is the deviation vector between the predicted values ​​of the basic physical quantities output by the physical property prediction branch and the actual physical quantities in the material gene library.

[0101] In this embodiment, the core function of step S5 is to take the material deep feature encoding rich in physical semantics generated in step S4 as input, extract the common semantic features of multiple tasks through the shared feature encoding module of the multi-task performance predictor, and then use multiple parallel task branches that correspond one-to-one with the target performance to achieve synchronous prediction of multiple target performances of the new material. At the same time, the model cognitive uncertainty is quantitatively captured through variable hierarchical modeling based on Bayesian inference. Finally, an initial prediction set containing the initial prediction values ​​of each target performance and the corresponding uncertainty index is generated, providing the core judgment basis for subsequent credibility assessment and active learning feedback mechanism. The detailed steps are as follows:

[0102] Step S5-1: Shared feature encoding and multi-task intermediate feature representation generation:

[0103] The material depth feature code generated in step S4 is input into the shared feature encoding module in the multi-task performance predictor. The shared feature encoding module is composed of multiple fully connected layers stacked together. The fully connected layers perform feature dimension compression and nonlinear activation operations on the input material depth feature code, filter out redundant noise components in the features, extract common semantic features suitable for multiple performance prediction tasks, and finally generate an intermediate feature representation shared by multiple tasks.

[0104] The formula for calculating the feature transformation of the shared feature coding module is:

[0105] ,in, This represents the intermediate features shared by multiple tasks. Encode the material depth features rich in physical semantics generated in step 54. For the trainable weight matrix of the fully connected layer of the shared feature encoding module, For the bias terms of the fully connected layer of the shared feature encoding module, It is a non-linear activation function;

[0106] Step S5-2: Parallel task branch construction and Bayesian variational output processing:

[0107] The intermediate feature representations shared by multiple tasks are input into multiple parallel task branches that correspond one-to-one with the target performance. Each task branch consists of multiple stacked fully connected layers, used to complete the specific feature extraction and mapping for the corresponding performance prediction task. After the last fully connected layer of each task branch, a variable layering based on Bayesian inference is introduced. The variable layering breaks the fixed weight mapping mode of traditional deterministic networks by applying random sampling operations to the network weights during the forward propagation process, so that the output of each task branch presents a probability distribution form, providing a data foundation for the subsequent quantification of cognitive uncertainty.

[0108] The formula for calculating the weighted sampling of variable stratification is:

[0109] ,in, For the first The network weights are obtained by hierarchical sampling of each task branch. For the first The mean parameter of the hierarchical weights of each task branch. For random sampling variables that follow a standard normal distribution, For the first The standard deviation parameter of the hierarchical weights of each task branch;

[0110] The formula for calculating the probability distribution output of task branches is:

[0111] ,in, For the first The probability distribution of the hierarchical output of each task branch. For the first Bias terms for each task branch;

[0112] Step S5-3: Initial Prediction Calculation and Cognitive Uncertainty Quantification:

[0113] In each task branch, statistical calculations are performed on the probability distribution results of the variable stratification output; the expected value of the probability distribution is used as the initial prediction value of the target performance corresponding to that task branch, and the variance of the probability distribution is used as a measure of the model's cognitive uncertainty of the prediction result; cognitive uncertainty reflects the prediction confidence caused by insufficient training data or out-of-sample distribution. The larger the variance value, the lower the confidence of the model's prediction result of the performance, and vice versa.

[0114] The formula for calculating the initial predicted value is:

[0115] ,in, For the first Each task branch corresponds to an initial predicted value for the target performance. Calculate the operator for mathematical expectation;

[0116] The formula for calculating cognitive uncertainty is:

[0117] ,in, For the first The cognitive uncertainty measure of the prediction results for each task branch Variance calculation operator;

[0118] Step S5-4: Aggregation of multi-task prediction results and generation of initial prediction set:

[0119] The initial predicted values ​​output from all task branches and their corresponding cognitive uncertainty measures are collected and aligned. Using the unique identifier of the target new material as the association key, a one-to-one mapping relationship is established between each initial predicted value and its corresponding cognitive uncertainty measure, ensuring that all prediction results correspond to different target performances of the same target new material. After collection and alignment, an initial prediction set containing multiple target performance initial prediction results and corresponding uncertainty indicators is formed. The initial prediction set is then output to the subsequent credibility assessment module for subsequent credibility judgment and active learning triggering based on uncertainty of the initial predicted values.

[0120] In this embodiment, the core function of step S6 is to perform a quantitative evaluation of the reliability of the prediction results based on the model's cognitive uncertainty output by the multi-task performance predictor. Prediction results that meet the preset reliability conditions are directly output as the final value. Samples that do not meet the reliability criteria are marked as high-value samples, triggering an active learning feedback mechanism to complete sample labeling and model iterative optimization. The prediction process is repeated until a final prediction result that meets the reliability requirements is output, ensuring the continuous improvement of the reliability and generalization ability of the model's prediction results. The detailed steps are as follows:

[0121] Step S6-1: Initial prediction set processing and comprehensive credibility index calculation:

[0122] The system receives the initial prediction set output by the multi-task performance predictor. The initial prediction set contains the initial prediction values ​​of multiple target performances of the target new material and the model cognitive uncertainty measure corresponding to each initial prediction value. The model cognitive uncertainty measure of each target performance is normalized to eliminate the difference in dimensions. Then, it is weighted and fused according to the preset weight coefficients to calculate the comprehensive credibility index used to characterize the overall credibility of the current prediction result.

[0123] The total number of target performance metrics is denoted as , No. The cognitive uncertainty measure of the model corresponding to each target performance is denoted as . The maximum value of all target performance cognitive uncertainty measures is denoted as The minimum value is denoted as ;No. The formula for calculating the normalized value of the cognitive uncertainty of a target performance is:

[0124] ,in, For the first The normalized value of the cognitive uncertainty of the target performance. For the first The model's cognitive uncertainty measure corresponding to each target performance. The maximum value of the cognitive uncertainty measure for all target performance. The minimum value for the cognitive uncertainty measure of all target performance;

[0125] The preset weighting coefficients for each target performance are denoted as follows: All weight coefficients satisfy The weighting coefficients are pre-set based on the importance of each target's performance; the formula for calculating the comprehensive credibility index is:

[0126] ,in, This serves as a comprehensive reliability index for the current prediction results. For the first The weighting coefficients corresponding to each target performance. For the first The normalized value of the cognitive uncertainty of the target performance. The total number of target performance metrics;

[0127] The calculated comprehensive credibility index is associated with and stored with the unique identifier of the corresponding target new material, providing core data support for subsequent credibility judgment;

[0128] Step S6-2: Confidence threshold determination and output of confidence prediction results:

[0129] Load the preset credibility acceptance threshold, which is preset according to the accuracy requirements of the target new material research and development and the reliability requirements of the application scenario; compare the calculated comprehensive credibility index with the credibility acceptance threshold; if the comprehensive credibility index is greater than or equal to the credibility acceptance threshold, the current prediction result is determined to be credible, and all the initial prediction values ​​in the initial prediction set are directly output as the final prediction result, ending the prediction process.

[0130] Step S6-3: Screening and labeling of low-confidence prediction items:

[0131] If the overall confidence index is lower than the confidence acceptance threshold, the model cognitive uncertainty measure corresponding to each target performance is further iterated, and a preset single-task uncertainty threshold is loaded. The single-task uncertainty threshold is preset according to the accuracy requirements of the single performance prediction. The model cognitive uncertainty measure of each target performance is compared with the single-task uncertainty threshold one by one. Specific performances whose uncertainty measure exceeds the single-task uncertainty threshold are selected. The specific performance and its corresponding initial prediction value are marked as low-confidence prediction items. The position of the low-confidence prediction items and the corresponding uncertainty value are recorded to provide a basis for the subsequent generation of high-value samples.

[0132] Step S6-4: Encapsulation of high-value sample records and triggering of active learning:

[0133] In response to the labeling results of low-confidence predictions, the identification information of the target new material being processed, the initial data of multi-source heterogeneity, the initial prediction value, and the low-confidence predictions are encapsulated to generate a complete high-value sample record. This high-value sample record is written into the database to be labeled. While completing the data writing operation, the active learning feedback mechanism is triggered to start the subsequent sample labeling and model optimization process.

[0134] Step S6-5: High-value sample push and labeling confirmation:

[0135] In the active learning feedback mechanism, high-value sample records in the database to be labeled are pushed to the user terminal or expert review platform. The target new material information, multi-source heterogeneous initial data, initial prediction results and low-confidence prediction items are displayed through the human-computer interaction interface. The annotation and confirmation of the actual performance values ​​corresponding to the low-confidence prediction items by users or material experts are received. The confirmed actual performance values ​​are used as supervision labels and combined with the encapsulated multi-source heterogeneous initial data to form a labeled high-value training sample.

[0136] Step S6-6: Training set augmentation and joint model optimization:

[0137] High-value training samples with annotations are added to the original model training set to expand the scale of training data and improve the data distribution coverage of the training set. Based on the expanded training set, the multimodal feature encoder, graph neural network model and multi-task performance predictor are jointly fine-tuned and the parameters are optimized. The optimization process aims to reduce the model's cognitive uncertainty about the high-value training samples with annotations. The model parameters are iteratively updated through the backpropagation algorithm to enhance the model's feature representation and prediction capabilities for out-of-distribution samples and sparse data scenarios.

[0138] Steps S6-7: The model reloading and prediction process is executed cyclically.

[0139] After optimizing the model parameters, reload the initial multi-source heterogeneous data of the current target new material, jump to step S2, and use the optimized model to execute the entire process from feature encoding to performance prediction again; recalculate the comprehensive credibility index for the newly generated prediction results until the newly calculated comprehensive credibility index meets the credibility acceptance threshold, output the final prediction result that meets the credibility requirements, and end the current prediction cycle.

[0140] An AI-based new material performance prediction system is disclosed. This system executes an AI-based new material performance prediction method and consists of seven modules: a multi-source heterogeneous data acquisition module, a multi-modal feature encoding module, a physics-enhanced graph neural network module, a deep feature encoding generation module, a multi-task performance prediction module, a reliability assessment and active learning feedback module, and a central control module. Each module possesses independent core functions and operational logic, while forming a collaborative working system through data transmission and process scheduling. This system fully executes the AI-based new material performance prediction method, achieving accurate prediction of the multi-dimensional performance of target new materials. Furthermore, it leverages an active learning mechanism to iteratively optimize the model, continuously improving the reliability and generalization ability of the prediction results. The following is a detailed description of each module:

[0141] Multi-source heterogeneous data acquisition module:

[0142] The multi-source heterogeneous data acquisition module is the foundation of the entire prediction system's data source. Its core function is to acquire initial multi-source heterogeneous data of the target new material. This data includes three core categories: compositional data, structural data, and microstructure data. The module's operation revolves around relevant data retrieval, unstructured data processing, and structured data encapsulation for the target new material. First, it receives the target new material's identification information or basic physical property parameters input from the user. Using this information as the retrieval basis, it accurately retrieves and extracts original compositional text, original structural atlas files, and original microstructure image files related to the target new material from public material databases and internal material knowledge bases. The extracted raw data is then categorized and standardized. The process involves several steps: First, natural language processing is performed on the original composition text to extract element type, stoichiometry, and doping information, which is then converted into structured composition feature vectors. Next, professional analysis is conducted on the original structure map files to extract crystal structure parameters and space group information, and to reconstruct the initial atomic structure model. Then, preprocessing and semantic segmentation are performed on the original microstructure image files to identify and quantify grain size, phase distribution, and defect features, generating quantitative microstructure feature maps. Finally, the processed composition feature vectors, initial atomic structure models, and quantitative microstructure feature maps are aligned and format-encapsulated to generate unified and interconnected multi-source heterogeneous initial data, providing standardized input data for subsequent feature processing modules.

[0143] Multimodal feature encoding module:

[0144] The multimodal feature encoding module is the core module for feature extraction and fusion in the system. It contains a pre-trained multimodal feature encoder. Its core function is to input multi-source heterogeneous initial data into the encoder, complete cross-modal feature alignment and fusion, and generate a material fusion feature representation capable of characterizing the target new material. The module operates according to the logic of branched feature extraction, cross-modal feature alignment, and multimodal feature fusion. First, component data, structural data, and microstructure data are input into the corresponding feature extraction branches of the multimodal feature encoder. Component data is converted into component feature sequences through an embedding layer. Structural data is used to extract the topological connections between atoms through a graph convolutional network to generate structural topological features. Microstructure data is extracted through a convolutional neural network. Image morphology and texture features are used to generate tissue morphology features, and all three types of features undergo dimensionality standardization. Subsequently, these three types of features are mapped to a unified material feature semantic space. Based on a cross-modal attention mechanism, correlation weights between different modal features are calculated. Weighted interactions are then applied to each modal feature according to these weights to achieve semantic feature alignment of different modal data. Finally, the aligned multimodal features are concatenated and fused, and a contrastive learning loss function based on materials science principles is introduced to shorten the representation distance between different modal features of the same material and widen the feature distance between different materials. Ultimately, a material fusion feature representation containing multi-dimensional information on composition, structure, and microstructure is generated, providing a feature foundation for subsequent graph neural network processing.

[0145] Physics-enhanced graph neural network module:

[0146] The physics-enhanced graph neural network module is the core module for physics knowledge fusion and graph structure feature processing in the system. The module contains a pre-built physics-enhanced graph neural network model. Its core function is to load the material fusion feature representation as the initial node feature into the model, construct a material structure topology graph, and introduce physical consistency constraints to complete message passing and feature aggregation. The module's workflow revolves around topology graph construction, feature loading, physical constraint embedding, and preliminary feature aggregation. First, based on the structural data of the target new material, it analyzes the spatial coordinates of atomic types and the bonding relationships between atoms. Using atoms or atomic groups as nodes and chemical bonds or physical interactions between atoms as edges, it constructs a material structure topology graph that matches the atomic structure of the material. Then, it loads the material fusion feature representation as the initial node feature of the topology graph, and simultaneously... Based on interatomic distances and bonding types, initial edge features are assigned to the edges of the topological graph, forming graph structure data with physical entity meaning. After inputting the graph structure data into the graph neural network model, physical consistency constraints based on prior knowledge of the material gene library are introduced during the node update process of each layer of the network. By querying the physical interaction potential energy curves and charge distribution patterns in the material gene library, the prior physical quantities are converted into differentiable constraint functions, and the weight coefficients of the message passing process are dynamically adjusted so that the node features aggregated by the model simultaneously conform to data-driven laws and physicochemical principles. Finally, after completing multi-layer message passing, a global pooling operation is performed on the node features of the entire graph to aggregate local atomic environment information and global topological structure information, generating feature codes containing physical semantic information, providing a foundation for the generation of deep feature codes.

[0147] Deep feature encoding generation module:

[0148] The deep feature encoding generation module is the core module for deep feature extraction and refinement in the system. Its core function is to iteratively update the material structure topology graph using a physically-enhanced graph neural network model, aggregating local atomic environment information and global topological structure information to generate material deep feature codes rich in physical semantics. The module completes iterative optimization and refinement of features through collaborative processing of a multi-level network structure. First, the graph structure data, including the initial features of loaded nodes and initialized edge features, is input into the first graph convolutional layer of the model. Using the type of inter-atomic chemical bonds and the inter-atomic distance as control coefficients, the first-order neighbor node feature information of the target node is aggregated. After correction by physical consistency constraints, the updated node hidden state of the first layer is obtained. Subsequently, the first-layer node hidden state and edge features are input into the second graph convolutional layer, introducing atomic cluster-based... A hierarchical attention mechanism dynamically calculates the influence weights of different atomic clusters on the target node. After weighted aggregation and correction by physical consistency constraints, the updated node hidden states of the second layer are obtained. The node hidden states of the second layer are input into the pooling layer. Through global summation pooling or global average pooling operations, all node hidden states are aggregated into a graph-level feature vector of fixed dimension to generate preliminary material deep feature encoding. Finally, the preliminary feature encoding is input into the feature refinement layer, which introduces a physical property prediction branch based on prior knowledge of the material gene library. The basic physical quantities of the material are predicted in parallel. The predicted values ​​are compared with the true values ​​in the material gene library. Residual connections are constructed based on the deviation to fine-tune and enhance the preliminary feature encoding. Finally, a material deep feature encoding rich in physical semantics is generated, providing a highly discriminative feature input for performance prediction.

[0149] Multi-task performance prediction module:

[0150] The multi-task performance prediction module is the core module for performance prediction and uncertainty capture in the system. It contains a pre-built multi-task performance predictor. Its core function is to input the material's deep feature encoding into the predictor, calculate multiple initial prediction values ​​for the target new material's performance in parallel, and simultaneously capture the cognitive uncertainty when the model generates these prediction values. The module operates according to the logic of shared feature extraction, parallel branch prediction, uncertainty quantification, and result aggregation. First, the material's deep feature encoding is input into the predictor's shared feature encoding module. Through a fully connected layer, feature dimension compression and nonlinear activation are performed to extract common semantic features applicable to multiple performance prediction tasks, generating intermediate feature representations shared by multiple tasks. Subsequently, these intermediate feature representations are input into the modules corresponding one-to-one with the target performance. Multiple parallel task branches are constructed, each consisting of multiple stacked fully connected layers. A variable layering based on Bayesian inference is introduced after the last fully connected layer. By applying random sampling to the network weights, the outputs of each task branch are made to present a probability distribution. In each task branch, the probability distribution of the variable layering output is statistically calculated. The expected value of the probability distribution is used as the initial prediction value of the corresponding target performance, and the variance of the probability distribution is used as a measure of the model's cognitive uncertainty of the prediction result. Finally, the initial prediction values ​​of all task branch outputs and their corresponding cognitive uncertainties are collected and aligned to establish a one-to-one mapping relationship, forming an initial prediction set containing the initial prediction results of multiple target performances and corresponding uncertainty indicators, which is then output to the subsequent credibility evaluation module.

[0151] Credibility assessment and active learning feedback module:

[0152] The credibility assessment and active learning feedback module is the core module for the system's prediction result verification and model iterative optimization. Its core function is to assess the credibility of initial predictions based on model cognitive uncertainty, output the final prediction value based on the assessment result, or trigger the active learning feedback mechanism to complete model iterative optimization and re-execute the prediction process. The module's workflow encompasses six stages: credibility calculation, threshold judgment, low-confidence term screening, high-value sample processing, model optimization, and process loop. First, it receives the initial prediction set output from the multi-task performance prediction module, normalizes the model cognitive uncertainty measure for each target performance, and calculates a comprehensive credibility index representing the overall credibility of the prediction result based on preset weight coefficients. The comprehensive credibility index is compared with a preset credibility acceptance threshold. If the index is greater than or equal to the threshold, the prediction result is deemed credible, the initial prediction value is directly output as the final prediction result, and the prediction process ends. If the index is lower than the threshold, the cognitive uncertainty measure of each target performance is iterated, and items exceeding the single-task threshold are selected. For specific performance characteristics at a deterministic threshold, these characteristics and their corresponding initial predicted values ​​are marked as low-confidence prediction terms. Based on the marking results of these low-confidence prediction terms, the identification information of the target new material, multi-source heterogeneous initial data, initial predicted values, and low-confidence prediction terms are encapsulated to generate high-value sample records and write them into the unlabeled database, triggering an active learning feedback mechanism. The high-value sample records in the unlabeled database are pushed to the user end or expert review platform to receive the confirmed real performance values, which are then combined with the multi-source heterogeneous initial data to form labeled high-value training samples. These samples are added to the original model training set, and the multimodal feature encoder, the physics-enhanced graph neural network model, and the multi-task performance predictor are jointly fine-tuned and their parameters optimized based on the expanded training set to reduce the model's cognitive uncertainty regarding these samples. After model optimization, the multi-source heterogeneous initial data of the target new material is reloaded, triggering the multimodal feature encoding module and subsequent modules to re-execute the prediction process until the newly calculated comprehensive confidence index meets the threshold requirements, at which point the final prediction result is output.

[0153] Central control module:

[0154] The central control module is the core of the entire prediction system's scheduling and management. Its core function is to comprehensively schedule and control the workflow of the multi-source heterogeneous data acquisition module, the multi-modal feature encoding module, the physics-enhanced graph neural network module, the deep feature encoding generation module, the multi-task performance prediction module, and the credibility assessment and active learning feedback module. This module is responsible for the time-series control of the entire process, triggering each module to execute its corresponding function sequentially according to the step logic of the AI-based new material performance prediction method. This ensures the orderly transmission and processing of data between modules, achieving a streamlined process from acquiring initial multi-source heterogeneous data to outputting the final prediction results. The module also has real-time... The monitoring capabilities enable dynamic monitoring of the operational status of each module, timely feedback of any anomalies during module operation, and ensure the overall stable operation of the system. Upon triggering the active learning feedback mechanism, the module coordinates the work of various model optimization modules, controlling the timing and rhythm of parameter updates to ensure efficient joint fine-tuning of the multimodal feature encoder, the physics-enhanced graph neural network model, and the multi-task performance predictor. Furthermore, the module is responsible for uniformly managing various preset parameters within the system, including confidence thresholds, weight coefficients, and training parameters, providing unified parameter support for the operation of each module and achieving automated, standardized, and efficient management of the entire prediction system.

[0155] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A novel material performance prediction method based on artificial intelligence, characterized in that: Includes the following steps: Step S1: Obtain initial data on the multi-source heterogeneity of the target new material; the initial data on multi-source heterogeneity includes composition data, structural data, and microstructure data; Step S2: Input the multi-source heterogeneous initial data into the pre-trained multimodal feature encoder to perform cross-modal feature alignment and fusion, and generate a material fusion feature representation for characterizing the target new material; Step S3: Load the material fusion feature representation as the initial node feature into the pre-constructed physical knowledge-enhanced graph neural network model; wherein, the graph neural network model uses atoms or atomic groups in the target new material as nodes and inter-atomic interactions as edges to construct a material structure topology graph, and introduces physical consistency constraints based on prior knowledge of the material gene library during the message passing process of the model. Step S4: Iteratively update the material structure topology graph using a graph neural network model to aggregate local atomic environment information and global topological structure information, generating material deep feature encoding rich in physical semantics; Step S5: Input the material deep feature encoding into the multi-task performance predictor to calculate the initial predicted values ​​of multiple target properties of the target new material in parallel, and at the same time capture the model cognitive uncertainty of the multi-task performance predictor when generating the initial predicted values. Step S6: Evaluate the credibility of the initial prediction based on the model's cognitive uncertainty: If the credibility meets the preset conditions, the initial prediction is output as the final prediction result; if the credibility is lower than the preset threshold, the current sample data and its corresponding multi-source heterogeneous initial data are marked as high-value samples, triggering the active learning feedback mechanism, adding the high-value samples to the training set to iteratively optimize the multimodal feature encoder, graph neural network model and multi-task performance predictor, and re-execute steps S2 to S6 using the optimized model until the final prediction result that meets the credibility requirements is obtained.

2. The method for predicting the properties of new materials based on artificial intelligence according to claim 1, characterized in that: Obtain initial data on the multi-source heterogeneity of the target new material; Initial data for multi-source heterogeneity includes compositional data, structural data, and microstructure data, including: Step S1-1: Receive the identification information or basic physical property parameters of the target new material input by the user terminal, and retrieve and extract the original composition text, original structural map file and original microstructure image file related to the target new material from the public material database and internal material knowledge base based on the identification information or basic physical property parameters. Step S1-2: Perform natural language processing on the extracted original component text to identify and extract the element types, stoichiometry and doping information, and convert them into structured component feature vectors as component data; Step S1-3: Analyze the obtained original structure map file, extract crystal structure parameters and space group information, and reconstruct an initial atomic structure model containing atomic occupancy and bonding relationships based on the crystal structure parameters, as structural data; Steps S1-4: Perform image preprocessing and semantic segmentation on the acquired original microstructure image files, identify and quantify the grain size, phase distribution and defect features in the images, and generate a quantitative microstructure feature map to characterize the microstructure as microstructure data. Steps S1-5: Align and encapsulate the component feature vectors, initial atomic structure models, and quantitative tissue feature maps to generate multi-source heterogeneous initial data with a unified format and mutual correlation, which will then be input into the multimodal feature encoder.

3. The method for predicting the properties of new materials based on artificial intelligence according to claim 1, characterized in that: Multi-source heterogeneous initial data is input into a pre-trained multimodal feature encoder for cross-modal feature alignment and fusion, generating a material fusion feature representation for characterizing the target novel material, including: Step S2-1: Input the component data, structural data, and microstructure data into the corresponding feature extraction branches of the multimodal feature encoder, respectively; wherein, the component data is converted into a component feature sequence through the embedding layer, the structural data is extracted through a graph convolutional network to extract the topological connection relationship between atoms to generate structural topological features, and the microstructure data is extracted through a convolutional neural network to extract the morphological texture features in the image to generate tissue morphological features. Step S2-2: Map the component feature sequence, structural topology features, and tissue morphology features to a unified material feature semantic space. In this semantic space, calculate the correlation weights between different modal features based on a cross-modal attention mechanism. Use the correlation weights to perform weighted interaction on the features of each modality to achieve feature alignment of different modal data at the semantic level. Step S2-3: Perform feature concatenation and fusion compression on the aligned multimodal features. By introducing a contrastive learning loss function based on materials science principles, the representation distance of different modal features of the same material in the semantic space is narrowed, while the feature distance between different materials is widened, generating a material fusion feature representation containing multi-dimensional information on composition, structure and microstructure.

4. The method for predicting the properties of new materials based on artificial intelligence according to claim 1, characterized in that: Material fusion feature representations are loaded as initial node features into a pre-constructed, physically-enhanced graph neural network model. The graph neural network model uses atoms or atomic groups in the target new material as nodes and interatomic interactions as edges to construct a material structure topology graph. During message passing within the model, physical consistency constraints based on prior knowledge from the material gene pool are introduced, including: Step S3-1: Based on the structural data of the target new material, analyze the types of atoms, spatial coordinates and bonding relationships between atoms in the material, and construct a material structure topology diagram based on this. In the material structure topology diagram, each atom or group of atoms is defined as a node in the diagram, and the chemical bonds or physical interactions between atoms are defined as edges connecting the nodes. Step S3-2: The material fusion feature representation generated in step S2 is used as the initial node feature and loaded onto the corresponding node of the material structure topology graph. Each edge is assigned an edge feature initialized based on the interatomic distance and bonding type, thereby forming graph structure data with physical entity meaning. Step S3-3: Input the graph structure data into the graph neural network model for message passing and aggregation; during the node update process of each layer of the graph neural network, introduce physical consistency constraints based on prior knowledge of the material gene library; specifically, query the physical interaction potential energy curve and charge distribution law between atoms through the pre-constructed material gene library, convert the queried prior physical quantities into differentiable constraint functions, dynamically adjust the weight coefficients in the message passing process, and force the node features obtained by the model aggregation to not only conform to the data-driven law, but also to be consistent with the known physical and chemical principles; Step S3-4: After completing multi-layer message passing, global pooling is performed on the node features of the entire graph to aggregate local atomic environment information and global topological structure information, generating material depth feature encoding containing physical semantic information for use by the subsequent multi-task performance predictor.

5. The method for predicting the properties of new materials based on artificial intelligence according to claim 1, characterized in that: The material structure topology graph is iteratively updated using a graph neural network model to aggregate local atomic environment information and global topological structure information, generating material deep feature encoding rich in physical semantics, including: Step S4-1: Input the initial node features loaded with material fusion feature representation and the initialized edge features into the first graph convolutional layer of the graph neural network model; In the first graph convolutional layer, based on the message passing paradigm, using the chemical bond type and inter-atomic distance between atoms as control coefficients, aggregate the feature information of the first-order neighbor nodes of the target node to generate preliminary aggregated features containing local coordination environment information, and correct the preliminary aggregated features according to the physical consistency constraint to obtain the node hidden state after the first layer is updated; Step S4-2: Input the updated node hidden state and its corresponding edge features from the first layer into the second graph convolutional layer of the graph neural network model; In the second graph convolutional layer, by introducing a hierarchical attention mechanism based on atomic clusters, dynamically calculate the influence weights of different atomic clusters on the target node, and perform weighted aggregation on the updated node hidden state from the first layer according to the influence weights to capture the mid-range ordered structure information, generate cluster-level aggregated features after attention weighting, and then correct the cluster-level aggregated features again according to the physical consistency constraint to obtain the updated node hidden state from the second layer; Step S4-3: Input the updated node hidden states of the second layer into the pooling layer of the graph neural network model; in the pooling layer, through global summation pooling or global average pooling, aggregate the hidden states of all nodes into a graph-level feature vector of fixed dimension, which is the preliminary material depth feature encoding. Step S4-4: Input the preliminary material deep feature encoding into the feature refinement layer of the graph neural network model; in the feature refinement layer, by introducing a physical property prediction branch based on prior knowledge of the material gene library, several basic physical quantities of the material are predicted in parallel based on the preliminary material deep feature encoding, and the predicted basic physical quantities are compared with the real physical quantities stored in the material gene library. Based on the comparison results, residual connections are constructed, and the preliminary material deep feature encoding is fine-tuned and enhanced using residual connections to generate material deep feature encoding rich in physical semantics.

6. The method for predicting the properties of new materials based on artificial intelligence according to claim 1, characterized in that: Material deep feature encoding is input into a multi-task performance predictor to compute initial predictions of multiple target properties of the new material in parallel, while simultaneously capturing the model cognitive uncertainty of the multi-task performance predictor in generating initial predictions, including: Step S5-1: Input the material depth feature encoding into the shared feature encoding module in the multi-task performance predictor, perform feature dimension compression and nonlinear activation through a fully connected layer, extract common semantic features suitable for multiple performance prediction tasks, and generate an intermediate feature representation shared by multiple tasks. Step S5-2: Input the intermediate feature representation shared by multiple tasks into multiple parallel task branches that correspond one-to-one with the target performance; each task branch is composed of multiple fully connected layers stacked together, and a variable layer based on Bayesian inference is introduced after the last fully connected layer. By applying random sampling to the network weights during the forward propagation process, the output of each task branch presents a probability distribution form. Step S5-3: In each task branch, perform statistical calculations on the probability distribution of the variable hierarchical output, use the expected value of the probability distribution as the initial predicted value of the target performance, and use the variance of the probability distribution as a measure of the model's cognitive uncertainty of the prediction result; where cognitive uncertainty reflects the prediction confidence caused by insufficient training data or out-of-sample distribution. Step S5-4: Collect and align the initial prediction values ​​output by all task branches and their corresponding cognitive uncertainties to form an initial prediction set containing multiple target performance initial prediction results and corresponding uncertainty indicators. Output the set to the credibility assessment module for subsequent credibility judgment and active learning triggering based on uncertainty of the initial prediction values.

7. The method for predicting the properties of new materials based on artificial intelligence according to claim 1, characterized in that: The credibility of the initial prediction is evaluated based on the model's cognitive uncertainty: if the credibility meets the preset conditions, the initial prediction is output as the final prediction result. If the confidence level is lower than a preset threshold, the current sample data and its corresponding multi-source heterogeneous initial data are marked as high-value samples, triggering an active learning feedback mechanism. The high-value samples are added to the training set to iteratively optimize the multimodal feature encoder, graph neural network model, and multi-task performance predictor. Steps S2 to S6 are then re-executed using the optimized model until a final prediction result meeting the confidence level requirement is obtained, including: Step S6-1: Receive the initial prediction set output by the multi-task performance predictor. The initial prediction set contains the initial prediction values ​​of multiple target performances of the target new material and the model cognitive uncertainty measure corresponding to each initial prediction value. Normalize the model cognitive uncertainty measure of each target performance and perform weighted fusion according to the preset weight coefficient to calculate the comprehensive credibility index used to characterize the overall credibility of the current prediction result. Step S6-2: Compare the comprehensive credibility index with the preset credibility acceptance threshold: If the comprehensive credibility index is greater than or equal to the credibility acceptance threshold, the current prediction result is determined to be credible, and all the initial prediction values ​​in the initial prediction set are directly output as the final prediction result, and the prediction process ends. Step S6-3: If the overall confidence index is lower than the confidence acceptance threshold, then further iterate through the model cognitive uncertainty measure corresponding to each target performance, filter out the specific performance whose uncertainty measure exceeds the single task uncertainty threshold, and mark the specific performance and its corresponding initial prediction value together as low confidence prediction items. Step S6-4: In response to the labeling results of low-confidence predictions, the identification information of the target new material being processed, the initial data of multi-source heterogeneity, the initial prediction value, and the low-confidence predictions are encapsulated to generate a complete high-value sample record, and the high-value sample record is written into the database to be labeled to trigger the active learning feedback mechanism. Step S6-5: In the active learning feedback mechanism, the high-value sample records in the database to be labeled are pushed to the user terminal or expert review platform. The user or material expert labels and confirmations of the actual performance values ​​corresponding to the low confidence prediction items are received through the human-computer interaction interface. The confirmed actual performance values ​​are used as supervision labels and combined with the encapsulated multi-source heterogeneous initial data to form a labeled high-value training sample. Step S6-6: Add the labeled high-value training samples to the original model training set to expand the training data scale, and perform joint fine-tuning and parameter optimization on the multimodal feature encoder, graph neural network model and multi-task performance predictor based on the expanded training set to reduce the model's cognitive uncertainty of the labeled high-value training samples. Steps S6-7: After completing the model optimization, reload the initial multi-source heterogeneous data of the current target new material and jump to step S2. Use the optimized model to execute the entire process from feature encoding to performance prediction again until the newly calculated comprehensive credibility index meets the credibility acceptance threshold and outputs the final prediction result that meets the credibility requirements.

8. A novel material performance prediction system based on artificial intelligence, characterized in that: The system is used to execute the artificial intelligence-based new material performance prediction method according to any one of claims 1-7, the system comprising: The multi-source heterogeneous data acquisition module is used to acquire the initial multi-source heterogeneous data of the target new material; the initial multi-source heterogeneous data includes composition data, structural data and microstructure data; The multimodal feature encoding module has a built-in pre-trained multimodal feature encoder, which is used to input multi-source heterogeneous initial data into the multimodal feature encoder, perform cross-modal feature alignment and fusion, and generate a material fusion feature representation for characterizing the target new material; The physics-enhanced graph neural network module has a pre-built physics-enhanced graph neural network model built in, which is used to load the material fusion feature representation as the initial feature of the node into the graph neural network model. The graph neural network model uses atoms or atomic groups in the target new material as nodes and inter-atomic interactions as edges to construct a material structure topology graph. In the process of message passing in the model, physical consistency constraints based on prior knowledge of the material gene library are introduced. The deep feature encoding generation module is used to iteratively update the material structure topology graph through a graph neural network model, so as to aggregate local atomic environment information and global topological structure information to generate material deep feature encoding rich in physical semantics. The multi-task performance prediction module has a built-in multi-task performance predictor. It is used to input the material deep feature encoding into the multi-task performance predictor and calculate the initial predicted values ​​of multiple target properties of the target new material in parallel. At the same time, it captures the model cognitive uncertainty of the multi-task performance predictor when generating the initial predicted values. The credibility assessment and active learning feedback module is used to assess the credibility of the initial prediction value based on the model's cognitive uncertainty. If the credibility meets the preset conditions, the initial prediction value is output as the final prediction result. If the credibility is lower than the preset threshold, the current sample data and its corresponding multi-source heterogeneous initial data are marked as high-value samples, triggering the active learning feedback mechanism. The high-value samples are added to the training set to iteratively optimize the multimodal feature encoder, graph neural network model, and multi-task performance predictor. The optimized model is then used to re-trigger the multimodal feature encoding module, the physics-enhanced graph neural network module, the deep feature encoding generation module, the multi-task performance prediction module, and the credibility assessment and active learning feedback module to perform their respective functions until the final prediction result that meets the credibility requirements is obtained. The central control module is used to schedule and control the working processes of the multi-source heterogeneous data acquisition module, the multimodal feature encoding module, the physical knowledge-enhanced graph neural network module, the deep feature encoding generation module, the multi-task performance prediction module, and the credibility assessment and active learning feedback module.

Images