Material data autonomous calling method and system based on multi-source data and multi-modal analysis

Abstract

This invention provides a method and system for autonomously retrieving material data based on multi-source data and multimodal analysis, belonging to the field of material information processing and retrieval. The method first constructs a material operator node library and a multi-source heterogeneous database; it autonomously loads new user-owned data and relevant material data from the user-owned data pool according to the material calculation task; it retrieves relevant material data from the material database pool based on search conditions and performs unified format conversion; it calls at least one material operator node according to the material calculation task, constructs a workflow topology graph, performs initial parameter configuration, and verifies the compliance of the configuration file; after passing the verification, it sorts the workflow topology graph, autonomously schedules and executes material calculation tasks, and monitors and manages logs in real time; it visualizes the execution results; when there are identical material calculation tasks, it directly calls the topology graph to execute the calculation task. This invention improves the autonomy, flexibility, and reusability of the material data analysis process.

Description

Technical Field

[0001] This invention belongs to the field of materials information processing and retrieval, specifically relating to a method and system for autonomous retrieval of materials data based on multi-source data and multimodal analysis. Background Technology

[0002] In recent years, with the rapid accumulation of materials databases and the rapid development of artificial intelligence technology, the paradigm of materials research and development is accelerating its shift towards data-driven approaches. Traditional materials research and development relies on a large number of experiments, resulting in long development cycles and high costs. Therefore, using machine learning methods to model and predict the mapping relationship between material composition, process parameters, and performance has become an important technical path to accelerate materials research and development. However, data-driven materials research and development increases the sensitivity to and frequency of access to materials-related data. How to better acquire, access, and utilize materials data, to a certain extent, determines the progress and quality of computational materials research and development.

[0003] In existing technologies, the modeling management of industrial data generally targets general-purpose data and lacks specific models for material-specific data. Material-specific data has its own unique characteristics, and applying general data management and retrieval methods to materials research and development fails to fully explore material properties and features, hindering the computational processing of materials. For example, Chinese patent CN111310936B discloses a general machine learning training construction method, employing a basic framework of directed acyclic graphs and containerized execution to construct and execute machine learning workflows. However, it does not design dedicated operators for the multimodal characteristics of material data, nor does it support workflow encapsulation, publishing, and model reuse across nodes. Similarly, Chinese patent CN114115857B discloses a method for constructing an automated production line for machine learning models, introducing an operator repository, a model repository, and a complete model publishing mechanism. However, this method is designed for general scenarios and lacks specialized data processing capabilities for the materials field. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method and system for autonomously accessing material data based on multi-source data and multimodal analysis. By constructing a material component operator node mechanism and a machine learning basic model pool, it enables autonomous design of material data analysis tasks, autonomous construction and reuse of models, autonomous loading and preprocessing of multi-source data, and dynamic adjustment and configuration of parameters, thereby enhancing the autonomy, flexibility, and reusability of the material data analysis process.

[0005] To achieve the above objectives, the technical solutions adopted in the embodiments of the present invention are as follows:

[0006] In a first aspect, embodiments of the present invention provide a method for autonomously retrieving material data based on multi-source data and multimodal analysis, the method comprising the following steps:

[0007] Step S1: Construct a material operator node library and simultaneously construct a multi-source heterogeneous database for materials. The multi-source heterogeneous database includes a user-owned data pool and a material database pool.

[0008] Step S2: Based on the material calculation task, new user-owned data and relevant material data from the user-owned data pool are automatically loaded; at the same time, search conditions are set according to the material calculation task, and relevant material data is retrieved from the material database pool based on the search conditions.

[0009] Step S3: Perform a unified format conversion on all loaded multi-source heterogeneous material data, and use the data of each corresponding material calculation task as a data sample;

[0010] Step S4: Determine if there is a duplicate material calculation task; if so, call the workflow topology diagram encapsulated under the corresponding material calculation task, modify the parameters of the corresponding operator nodes according to the actual situation of the material calculation task and data sample, and then proceed to step S7; if not, proceed to step S5.

[0011] Step S5: Based on the material calculation task, call at least one material operator node from the material operator node library to construct the workflow topology graph of the current task. The nodes of the workflow topology graph include all data samples and operator nodes, and the edges of the workflow topology graph include the connection relationship from data samples to operator nodes and the connection relationship between operator nodes. Initial parameter configuration is performed on the workflow topology graph.

[0012] Step S6: Read the configuration file of the workflow topology diagram and perform compliance verification on the configuration file;

[0013] Step S7: Sort the workflow topology diagram, autonomously schedule and execute material calculation tasks, and monitor and manage the running status in real time.

[0014] Step S8: Visualize the task execution results.

[0015] As a preferred embodiment of the present invention, the operator node library covers the entire process of material data analysis, including data fusion node group, data processing node group, feature engineering node group, model node group and visualization node group.

[0016] In a preferred embodiment of the present invention, the user-owned data pool includes user-owned data loaded each time a material calculation task is performed, consisting of local CSV and Excel format numerical data files, MOL, SDF, RXN, and RDF material-specific format structured data files, and material microscopic image files in JPG, PNG, or TIFF format from scanning electron microscopy (SEM) and transmission electron microscopy (TEM); the material database pool includes third-party material databases accessed through a preset application programming interface and a characterization state transition architecture.

[0017] In a preferred embodiment of the present invention, the data flow of the workflow topology is as follows: the proprietary data node and the retrieved data node output multi-source standard data objects to the data fusion node respectively; the data fusion node performs entity recognition and horizontal stitching on the multi-source standard data objects based on the unique material identifier, eliminates attribute conflicts, and generates an initial fusion dataset; then the initial fusion dataset is passed to the data processing node, and then passes through the feature engineering node and the model calling node in sequence, and finally the visualization node outputs the calculation results.

[0018] In a preferred embodiment of the present invention, step S5, which involves calling the material operator node according to the material calculation task, includes:

[0019] The graphical workflow design interface provides human-computer interaction support; users can instantiate target operator nodes into the workflow canvas by dragging and dropping from the visualized operator node library directory according to the logical requirements of the current computing task, and add edges to the workflow topology graph by drawing lines between nodes.

[0020] In a preferred embodiment of the present invention, the compliance verification includes:

[0021] Topology compliance verification checks for loops in the directed graph to ensure the workflow is a compliant directed acyclic graph.

[0022] Parameter configuration validity verification verifies whether the required parameters of each node have been fully configured.

[0023] Data type compatibility check: Verify whether the output data type and input data type match between adjacent nodes;

[0024] Data source reachability verification verifies whether the data source specified by each data access node can be accessed normally.

[0025] In a preferred embodiment of the present invention, the topology sorting includes: sorting the DAG workflow that has passed the compliance verification according to data dependencies to determine the compliance execution order of each material operator node.

[0026] In a preferred embodiment of the present invention, the autonomous invocation includes a hybrid serial-parallel autonomous scheduling mechanism. Specifically: the workflow task scheduling engine sorts the results by topology and then autonomously invokes the encapsulated processing logic corresponding to each node from the material operator node repository for execution; a workflow dependency graph is constructed based on adjacency lists and inverse adjacency lists, and the node topology hierarchy is calculated through breadth-first search. Depth-first search or transitive closure matrix is ​​used to determine whether there is a direct or indirect data dependency between any two nodes in the ready node set. Nodes with data dependencies are scheduled sequentially, while nodes without direct or indirect data dependencies are automatically identified and concurrently scheduled; each node automatically extracts output data from upstream nodes, performs conversion operations, and loads the processing results to the input port of downstream nodes, thereby realizing automatic pipelined processing of material data.

[0027] In a preferred embodiment of the present invention, the method of visualizing the task execution results includes:

[0028] (1) Numerical results: such as prediction results of continuous properties of materials, such as band gap, hardness, and conductivity, presented in the form of data tables or scatter plots.

[0029] (2) Classification and clustering results: such as discrete classification results such as material phase structure classification and stability discrimination, as well as clustering results of new material category discovery, combined with receiver operation characteristic curves and confusion matrix to present model confidence;

[0030] (3) Image-type results: such as the segmentation or target recognition results of material microscopic images, which are presented in a visual comparison by superimposing color masks or bounding boxes on the original microscopic images;

[0031] (4) Explanatory analysis results: such as the ranking of feature importance output by feature engineering nodes, presented in the form of heat map or model performance improvement map, to help explain the cause mechanism of material properties.

[0032] Secondly, embodiments of the present invention also provide a material data autonomous retrieval system based on multi-source data and multimodal analysis. The system includes: a material database construction module, a material operator node library, a multi-source heterogeneous database, a material calculation task loading module, a data loading module, a data format conversion module, a topology graph retrieval judgment module, a workflow topology graph construction module, an initial parameter configuration module, a compliance verification module, a calculation task execution module, a real-time monitoring module, a log management module, and a visualization module.

[0033] The material database construction module is used to construct a material operator node library and a multi-source heterogeneous database oriented towards materials.

[0034] The material operator node library is used to store material operator nodes;

[0035] The multi-source heterogeneous database is used to store multi-source heterogeneous databases, which include user-owned data pools and material database pools.

[0036] The material calculation task loading module is used to load material calculation tasks;

[0037] The data loading module is used to autonomously load new user-owned data and relevant material data from the user-owned data pool according to the material calculation task; at the same time, it sets search conditions according to the material calculation task and retrieves relevant material data from the material database pool based on the search conditions.

[0038] The data format conversion module is used to perform unified format conversion on all loaded multi-source heterogeneous material data, with each data item corresponding to a material calculation task serving as a data sample.

[0039] The topology graph call judgment module is used to determine whether there is a duplicate material calculation task. If it exists, the workflow topology graph encapsulated under the corresponding material calculation task is called, the parameters of the corresponding operator nodes are modified according to the actual situation of the material calculation task and data sample, and the calculation task execution module is started. If it does not exist, the workflow topology graph construction module is started.

[0040] The workflow topology graph construction module is used to call at least one material operator node from the material operator node library according to the material calculation task to construct the workflow topology graph of the current task. The nodes of the workflow topology graph include all data samples and operator nodes, and the edges of the workflow topology graph include the connection relationship from data samples to operator nodes and the connection relationship between operator nodes.

[0041] The initial parameter configuration module is used to configure the initial parameters of the workflow topology graph;

[0042] The compliance verification module is used to read the configuration file of the workflow topology diagram and perform compliance verification on the configuration file;

[0043] The computation task execution module is used to perform topological sorting on the workflow topology graph and autonomously schedule and execute material calculation tasks.

[0044] The real-time monitoring module is used to monitor the operating status in real time;

[0045] The log management module is used to manage the running status logs;

[0046] The visualization module is used to visualize the results of task execution.

[0047] The solutions of the embodiments of the present invention have the following beneficial effects:

[0048] The material data autonomous retrieval method and system based on multi-source data and multi-modal analysis provided in this invention unifies the processing of multi-source and multi-modal material data using the ETL data flow paradigm. The output of each node is transmitted to the input port of the downstream node in a unified format via data flow connections, shielding the format differences between different types of material data. Through a built-in pool of basic machine learning models for material scenarios and a drag-and-drop graphical task configuration interface, material researchers can autonomously design and configure complex analysis workflows without needing machine learning programming skills. The operating environments of each node are isolated, completely eliminating algorithm library version conflicts, lowering the threshold for building material data analysis tasks, and significantly improving the efficiency and scientific rigor of model construction. It also supports dynamic parameter adjustment and injection during workflow execution. This enhances the flexibility and adaptability of the analysis process; the workflow scheduling engine, based on topology sorting, achieves hybrid serial and parallel scheduling, making full use of computing resources and improving the overall execution efficiency of multi-node workflows; the versioned task flow encapsulation and release mechanism and the cross-node model automatic reuse mechanism enable the inheritance and iteration of high-quality material data analysis workflows and material analysis model assets between teams and tasks, significantly reducing the cost of repetitive modeling; it provides real-time running status monitoring, historical log query and anomaly cause analysis, making the entire process monitorable and traceable, ensuring the transparency and traceability of the analysis process; the command-line studio node provides users with the ability to autonomously extend material operators without mastering container technology, supports custom component operator extensions, and allows third-party material analysis tools to be connected to the ETL data stream via scripts.

[0049] Of course, implementing any product or method of the present invention does not necessarily require achieving all of the advantages described above at the same time. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a flowchart of the material data autonomous retrieval method based on multi-source data and multi-modal analysis as described in an embodiment of the present invention;

[0052] Figure 2 This is a schematic diagram of the workflow topology structure in an embodiment of the present invention. Detailed Implementation

[0053] After discovering the aforementioned problems, the inventors of this application conducted a detailed study of existing methods for acquiring, accessing, and using material data. The study revealed the following issues with existing technologies in supporting online analysis of material data:

[0054] (1) Lack of dedicated support for multimodal data processing of materials. Data generated from materials experiments includes chemical composition data, performance test values, electron microscopy images, and proprietary format files such as MOL, SDF, and CIF from mainstream materials databases. These data types have different format specifications and preprocessing requirements. Existing machine learning analysis platforms lack dedicated processing operators for materials data types, requiring researchers to write a large amount of data processing code, which is both difficult and inefficient.

[0055] (2) The construction of materials data analysis tasks has a high threshold. Materials researchers generally lack machine learning programming skills, and there are version conflicts in the operating environments on which different analysis algorithms depend. As a result, most researchers with a materials background cannot independently complete the construction of machine learning environments and the writing of training code, which severely restricts the analysis and application of materials data.

[0056] (3) Insufficient flexibility in task design and model building. Existing tools lack a standardized component operator system and basic model pool for materials analysis scenarios. Materials researchers cannot easily design analysis processes, select and optimize machine learning models in a visual manner. Parameter configurations can usually only be determined once before the process starts, lacking dynamic adjustment capabilities.

[0057] (4) Difficulty in reusing analysis models across nodes. Analysis models trained for specific material systems need to be retrained from scratch when applied to similar material systems; moreover, existing analysis processes lack a systematic encapsulation and release mechanism, which is not conducive to the reuse of analysis models.

[0058] Therefore, there is a need for an autonomous adjustment technology and device for online analysis of material data to solve the above problems.

[0059] It should be noted that the defects in the above-mentioned prior art solutions are all the result of the inventors' practice and careful research. Therefore, the discovery process of the above problems and the solutions proposed by the embodiments of the present invention in the following text should be the inventors' contributions to the present invention.

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. It should be noted that, without conflict, the embodiments and features in the embodiments of the present invention can also be combined with each other.

[0061] It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of the embodiments of the present invention, the terms "first," "second," "third," "fourth," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance. In addition, sometimes a subscript such as W1 may be written in a non-subscript form such as W1, and their meanings are consistent unless the distinction is emphasized.

[0062] It should be understood that the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. A and B can be singular or plural. Additionally, the character " / " in this article generally indicates an "or" relationship between the preceding and following related objects, but it can also represent an "and / or" relationship. Please refer to the context for a more accurate understanding.

[0063] Following the above in-depth analysis, this invention provides a method and system for autonomously retrieving material data based on multi-source data and multimodal analysis. Addressing different data sources and diverse online analysis application scenarios, it constructs a material component operator node mechanism and a machine learning basic model pool to achieve interactive autonomous design and flexible construction of material data analysis tasks, autonomous model construction and reuse, autonomous loading and preprocessing of multi-source data, dynamic adjustment and configuration of parameters, and encapsulation, release, and cross-node sharing of task flows. Furthermore, through an interactive interface, it fundamentally lowers the technical threshold for material data analysis and enhances the autonomy, flexibility, and reusability of the analysis process.

[0064] like Figure 1 As shown, the method for autonomously retrieving material data based on multi-source data and multi-modal analysis includes the following steps:

[0065] Step S1: Construct a material operator node library and simultaneously construct a multi-source heterogeneous database for materials. The multi-source heterogeneous database includes a user-owned data pool and a material database pool.

[0066] In this step, the operator node library covers the entire material data analysis process, including data fusion node groups, data processing node groups (missing value handling, outlier removal, etc.), feature engineering node groups (feature extraction, feature selection, normalization, etc.), model node groups, and visualization node groups. The data fusion node group is used to fuse multi-source material data after format unification. Specifically, this includes entity alignment: based on the unique identifier features of the material (such as standardized chemical formulas, CAS registry numbers, or fingerprint similarity of the material's crystal structure), identifying and matching sample records from different data sources pointing to the same actual material; lateral stitching: merging and expanding the attribute fields of the same sample from different data sources after successful entity alignment, converting the originally scattered vertical multi-source data into a data matrix with single material samples as rows and multi-dimensional material attributes as columns; conflict resolution: for cases where different values ​​exist for the same attribute field of the same material during lateral stitching, selection or correction is performed according to preset data credibility rules; wide table construction: based on the above processing, an initial fused dataset (i.e., a wide table) containing complete feature dimensions is output for downstream nodes to perform unified statistical analysis or model training. The model node group includes regression analysis models such as linear regression, random forest regression, and XGBoost regression; classification analysis models such as decision tree, support vector machine classification (SVM), and multi-objective classification; clustering analysis models such as K-means clustering and hierarchical clustering; and deep learning models that support material microscopic image segmentation and classification. Further expansion is possible to cater to the specific data characteristics of the materials field. The visualization node group supports the visualization of charts such as heatmaps, receiver operating characteristic curves (ROC curves), and model performance improvement graphs (LIFT graphs).

[0067] Furthermore, the operator node library also has open extensibility, including pre-built command-line studio call extension nodes. Users can configure script commands and input / output port mappings for third-party tools within these command-line call extension nodes to encapsulate external heterogeneous computing tools into custom operator nodes that conform to system data flow specifications, and dynamically register them to the operator node library for workflow invocation.

[0068] The user-owned data pool includes user-owned data loaded each time a materials calculation task is executed. This data consists of local CSV and Excel format numerical data files, material-specific format structure data files such as MOL, SDF, RXN, and RDF, and material microscopic image files in JPG, PNG, or TIFF format from scanning electron microscopy (SEM) and transmission electron microscopy (TEM). The materials database pool includes third-party materials databases accessed through a pre-defined application programming interface (API) and a representational state transfer (RESTful) architecture. These third-party materials databases include at least one of publicly available scientific databases or commercial materials databases, and internally store standardized materials data such as first-principles calculation results, known material crystal structures, molecular topological information, and macroscopic physical properties.

[0069] Step S2: Based on the material calculation task, new user-owned data and relevant material data from the user-owned data pool are automatically loaded; at the same time, search conditions are set according to the material calculation task, and relevant material data is retrieved from the material database pool based on the search conditions.

[0070] Step S3: Perform a unified format conversion on all loaded multi-source heterogeneous material data, and use the data from each corresponding material calculation task as a data sample.

[0071] In this step, the unified format conversion performs format parsing and unified conversion on the loaded multi-source heterogeneous material data to form a unified standard data object. Specifically, for microscopic image files, the system parses and encapsulates them into standard image data objects that conform to the input specifications of the built-in deep learning nodes, shielding the differences between material data from different sources and in different formats, and supporting subsequent unified calls.

[0072] Step S4: Determine if there is a duplicate material calculation task. If so, call the workflow topology diagram encapsulated under the corresponding material calculation task, modify the parameters of the corresponding operator nodes according to the actual situation of the material calculation task and data sample, and then proceed to step S7. If not, proceed to step S5.

[0073] Step S5: Based on the material calculation task, call at least one material operator node from the material operator node library to construct the workflow topology graph of the current task. The nodes of the workflow topology graph include all data samples and operator nodes, and the edges of the workflow topology graph include the connection relationship from data samples to operator nodes and the connection relationship between operator nodes; and perform initial parameter configuration on the workflow topology graph.

[0074] In this step, such as Figure 2As shown, the data flow of the workflow topology is as follows: the proprietary data node and the retrieved data node output multi-source standard data objects to the data fusion node respectively; the data fusion node performs entity recognition and horizontal stitching on the multi-source standard data objects based on the unique material identifier, eliminates attribute conflicts, and generates an initial fusion dataset; then the initial fusion dataset is passed to the data processing node, and then passes through the feature engineering node and the model calling node in sequence, and finally the visualization node outputs the calculation results.

[0075] When material operator nodes are invoked based on material calculation tasks, a graphical workflow design interface provides human-computer interaction support. Users can instantiate target operator nodes into the workflow canvas by dragging and dropping from a visual operator node library directory, based on the logical requirements of the current calculation task. Edges in the workflow topology graph are added through interactive operations such as drawing connections between nodes. The input / output ports of each node in the operator node library are dynamically activated or disabled based on the characteristics of the currently loaded data samples, allowing only valid connection edges that satisfy data dimension and type constraints.

[0076] In one executable embodiment, each material operator node is equipped with an independent parameter configuration panel, which allows users to dynamically adjust node parameters during the workflow design phase, including data column selection parameters, algorithm hyperparameters, output configuration parameters, etc. The system supports a parameter linkage mechanism, where the output feature information of upstream nodes can be automatically pushed and populated with candidate parameter options for downstream nodes.

[0077] After the parameters are configured, the system serializes the node identifiers, directed edge connections between nodes, and configuration parameters of each node in the workflow topology graph to generate a structured workflow configuration file and persists it.

[0078] Step S6: Read the configuration file of the workflow topology diagram and perform compliance verification on the configuration file.

[0079] In this step, the compliance verification includes:

[0080] Topology compliance verification checks for loops in the directed graph to ensure the workflow is a compliant directed acyclic graph.

[0081] Parameter configuration validity verification verifies whether the required parameters of each node have been fully configured.

[0082] Data type compatibility check: Verify whether the output data type and input data type match between adjacent nodes;

[0083] Data source reachability verification verifies whether the data source specified by each data access node can be accessed normally.

[0084] When a problem is detected during verification, the system will display a location notification to the user in the error message area on the interface, and pause execution while waiting for correction.

[0085] Step S7: Sort the workflow topology diagram, autonomously schedule and execute material calculation tasks, and monitor and manage the running status in real time.

[0086] In this step, the topology sorting includes: sorting the DAG workflow that has passed compliance verification according to data dependencies to determine the compliant execution order of each material operator node. The autonomous invocation includes a hybrid serial-parallel autonomous scheduling mechanism. Specifically: the workflow task scheduling engine, based on the topology sorting results, autonomously invokes the encapsulated processing logic corresponding to each node from the material operator node repository for execution; a workflow dependency graph is constructed based on adjacency lists and inverse adjacency lists; the node topology hierarchy is calculated through breadth-first search; depth-first search or transitive closure matrix is ​​used to determine whether there is a direct or indirect data dependency between any two nodes in the ready node set; nodes with data dependencies are scheduled sequentially; nodes without direct or indirect data dependencies are automatically identified and concurrently scheduled; each node automatically extracts output data from upstream nodes, performs transformation operations, and loads the processing results to the input port of downstream nodes, realizing automatic pipelined processing of material data.

[0087] This step may also include: allowing users to modify and dynamically inject parameters of downstream nodes that have not yet been started in real time during workflow execution, without interrupting and restarting the overall process.

[0088] When monitoring the running status in real time, the execution status information of each material operator node is obtained in real time, and the execution progress is displayed in real time in a visual manner in the material data analysis interactive interface, including the status identifier of each node, execution time, resource consumption and analysis of abnormal causes.

[0089] The log management includes the unified collection, organization, and archiving of operation logs, recording the execution status, execution time, and reasons for exceptions of each node; it supports querying historical logs by time range, workflow identifier, node identifier, and other conditions; it can also identify abnormal patterns in the execution process of operator nodes, output the reason analysis of node execution exceptions, display the exception analysis results to the user in the error information area of ​​the interactive interface, and support single-node retry for failed nodes.

[0090] Step S8: Visualize the task execution results.

[0091] The visualization presentation includes displaying the data in the interactive interface using various visualization graphic components, such as line graphs, scatter plots, heatmaps, and bar charts. Specifically, the numerical material performance prediction results can be displayed as a comparison curve between predicted and measured values ​​in the form of a scatter plot or line graph, and a bar chart can be used to show the contribution of each material characteristic to the prediction target.

[0092] In this step, the visualization of the task execution results includes the following methods:

[0093] (1) Numerical results: such as prediction results of continuous properties of materials, such as band gap, hardness, and conductivity, presented in the form of data tables or scatter plots.

[0094] (2) Classification and clustering results: such as discrete classification results such as material phase structure classification and stability discrimination, as well as clustering results for new material category discovery, combined with receiver operating characteristic curve (ROC curve) and confusion matrix to present model confidence;

[0095] (3) Image-type results: such as the segmentation or target recognition results of material microscopic images, which are presented in a visual comparison by superimposing color masks or bounding boxes on the original microscopic images;

[0096] (4) Explanatory analysis results: such as the ranking of feature importance output by feature engineering nodes, presented in the form of heat map or model performance enhancement map (LIFT map), to help explain the cause mechanism of material properties.

[0097] Step S8: Encapsulate the workflow topology diagram for automatic reuse of calculation tasks with the same material.

[0098] In this step, after the workflow topology is executed, the current workflow is encapsulated into a reusable, standardized workflow template. The encapsulated content includes the workflow data flow structure configuration, snapshots of configuration parameters for each operator node, the trained material analysis model file, and the model's dependent environment specifications. This entire template is packaged and stored in the material workflow template repository, supporting public sharing, internal team sharing, and private encrypted use. When executing the same material calculation task, the published workflow template is retrieved and loaded from the workflow template repository. Only the data access configuration and key parameters need to be modified to quickly start the modeling and analysis task for the new material data.

[0099] Based on the same approach, this invention also provides a material data autonomous retrieval system based on multi-source data and multimodal analysis. The system includes: a material database construction module, a material operator node library, a multi-source heterogeneous database, a material calculation task loading module, a data loading module, a data format conversion module, a topology graph retrieval judgment module, a workflow topology graph construction module, an initial parameter configuration module, a compliance verification module, a calculation task execution module, a real-time monitoring module, a log management module, and a visualization module.

[0100] The material database construction module is used to construct a material operator node library and a multi-source heterogeneous database oriented towards materials.

[0101] The material operator node library is used to store material operator nodes;

[0102] The multi-source heterogeneous database is used to store multi-source heterogeneous databases, which include user-owned data pools and material database pools.

[0103] The material calculation task loading module is used to load material calculation tasks;

[0104] The data loading module is used to autonomously load new user-owned data and relevant material data from the user-owned data pool according to the material calculation task; at the same time, it sets search conditions according to the material calculation task and retrieves relevant material data from the material database pool based on the search conditions.

[0105] The data format conversion module is used to perform unified format conversion on all loaded multi-source heterogeneous material data, with each data item corresponding to a material calculation task serving as a data sample.

[0106] The topology graph call judgment module is used to determine whether there is a duplicate material calculation task. If it exists, the workflow topology graph encapsulated under the corresponding material calculation task is called, the parameters of the corresponding operator nodes are modified according to the actual situation of the material calculation task and data sample, and the calculation task execution module is started. If it does not exist, the workflow topology graph construction module is started.

[0107] The workflow topology graph construction module is used to call at least one material operator node from the material operator node library according to the material calculation task to construct the workflow topology graph of the current task. The nodes of the workflow topology graph include all data samples and operator nodes, and the edges of the workflow topology graph include the connection relationship from data samples to operator nodes and the connection relationship between operator nodes.

[0108] The initial parameter configuration module is used to configure the initial parameters of the workflow topology graph;

[0109] The compliance verification module is used to read the configuration file of the workflow topology diagram and perform compliance verification on the configuration file;

[0110] The computation task execution module is used to perform topological sorting on the workflow topology graph and autonomously schedule and execute material calculation tasks.

[0111] The real-time monitoring module is used to monitor the operating status in real time;

[0112] The log management module is used to manage the running status logs;

[0113] The visualization module is used to visualize the results of task execution.

[0114] The system or device for executing the method in this embodiment of the invention can be a terminal or a server. The system includes a processor, a memory, and / or a transceiver, etc., and is connected via a communication bus. Each module can be implemented by a processor, a memory, and / or a transceiver, etc. The processor can be, but is not limited to, one or more microprocessors (MPUs), central processing units (CPUs), network processors (NPs), digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), and other programmable logic devices, discrete gates, transistor logic devices, discrete hardware components, etc., or can be configured to implement one or more integrated circuits of this invention. The processor can perform various functions by running or executing software programs in the memory and calling data in the memory. The memory includes Random Access Memory (RAM), Read-Only Memory (ROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Compact Disc Read-Only Memory (CD-ROM), and / or Non-Volatile Memory (NVM), etc. The transceiver is used to communicate with network devices or terminal devices, and includes a receiver and a transmitter. The memory and transceiver can be integrated with the processor or exist independently.

[0115] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0116] It should also be noted that the material data autonomous retrieval system based on multi-source data and multi-modal analysis described in this embodiment corresponds to the material data autonomous retrieval method based on multi-source data and multi-modal analysis. The description and limitations of the method also apply to the system, and will not be repeated here.

[0117] The method and system for autonomous material data retrieval based on multi-source data and multi-modal analysis described in the embodiments of the present invention are applied to application scenarios such as predicting the yield strength of a high-entropy alloy, detecting microscopic defects and predicting the mechanical properties of composite materials, and calculating the adsorption energy of catalyst surfaces and predicting catalytic activity, respectively, to further illustrate the method of the present invention.

[0118] First Embodiment

[0119] When predicting the yield strength of a high-entropy alloy, the material calculation task and retrieval conditions are first determined. Based on this, a CSV data file containing local alloy experimental information, including composition ratios and measured yield strength, is imported into the resource panel of the interactive interface, and / or retrieval parameters are configured. First-principles calculation data and thermodynamic parameters of the corresponding composition system are retrieved and loaded from an external material database pool via a pre-defined RESTful interface. The loaded local tabular data and the structured data returned by the external interface are parsed and ETL-converted to uniformly map heterogeneous data to standard data objects within the system, masking differences in data formats from different sources, and uniformly outputting to the input port of downstream nodes.

[0120] In the data fusion node, the system executes multi-source fusion logic: entity alignment is performed using the "standardized alloy chemical formula" as the unique identifier, matching local experimental data with records in external computational data that point to the same alloy composition; then, horizontal stitching is performed, merging the experimentally measured mechanical property fields with the externally retrieved thermodynamic parameter fields to generate an initial fusion dataset (wide table) with single alloy samples as rows and multidimensional attributes as columns; for individual conflicting melting point values, conflict resolution is performed according to the credibility rule of "experimental data takes precedence over computational data".

[0121] The system determines that no identical material calculation task exists. Through a graphical interface, users sequentially select data preprocessing nodes, feature engineering nodes, model training nodes, and model evaluation nodes from the material operator node library, dragging them to the engineering panel. Directed lines define the data flow between nodes, abstracting the alloy yield strength performance prediction and analysis task into a complete directed acyclic graph. In the parameter configuration panels for each node, users configure missing value handling strategies, feature extraction and normalization methods, dataset partitioning ratios, and model hyperparameters. After the upstream nodes are configured, the alloy physical feature columns output by the upstream feature extraction node are automatically pushed and populated into the feature selection options of the downstream model nodes, achieving linked parameter configuration.

[0122] During workflow construction, the system continuously verifies the data type compatibility between adjacent nodes in real time. After parameter configuration, the system serializes the aforementioned topology and parameters into a workflow configuration file. The engine reads the workflow configuration file for multi-dimensional verification. The system detects that there are no cycles in the current directed acyclic graph; verifies that the required hyperparameters such as the number of trees and maximum depth of model nodes are fully configured; and verifies that the external database interface is connected normally. Any anomalies detected during verification are displayed in real time in the error message area of ​​the interface, helping users to promptly identify and correct configuration problems.

[0123] After successful verification, the workflow scheduling and execution engine sorts the directed acyclic graph according to node data dependencies. The engine calculates the node topology hierarchy using breadth-first search, automatically identifying and concurrently scheduling nodes without data dependencies, while scheduling dependent backbone pipelines serially. Each node automatically extracts output data from upstream, processes it, and loads the results into the input ports of downstream nodes, forming an automated pipelined processing system. During execution, the log management module updates the execution status, execution time, and resource consumption of each node in real time.

[0124] After the task is completed, users can choose to view the multidimensional visualization results: (1) Numerical results: The comparison curve between the model-predicted yield strength and the measured yield strength is shown in the form of a scatter plot; (2) Interpretive analysis results: The content of Fe, Co, Ni and other alloying elements and thermodynamic parameters are displayed in the form of a thermogram to show the characteristic importance ranking of the yield strength prediction target, which helps material experts understand the performance cause mechanism.

[0125] After the workflow is completed, the system packages the current directed acyclic graph structure containing data fusion logic, snapshots of parameters for each node, the trained optimal machine learning model file, and Python dependency environment specifications into a "High-Entropy Alloy Strength Prediction Standard Template" and stores it in the workflow template repository for internal team sharing. Subsequently, when faced with similar tasks such as "predicting the plasticity of another high-entropy alloy," users can directly load this template from the repository, only needing to modify the data input source and prediction target column. This allows for rapid startup of new tasks without redesigning the workflow, enabling autonomous reuse of the analysis process.

[0126] Second Embodiment

[0127] This embodiment employs the aforementioned method and system for autonomous material data retrieval based on multi-source data and multimodal analysis to perform microscopic defect detection and mechanical property prediction tasks for composite materials. This embodiment focuses on illustrating the system's multimodal data processing capabilities and its efficient workflow template reuse mechanism.

[0128] The high-entropy alloy strength prediction standard template published in the first embodiment is retrieved from the workflow template library and loaded. The engineering panel automatically restores the directed acyclic graph structure, which includes data fusion, preprocessing, model training, and visualization, as well as a snapshot of the parameter configurations of all nodes. Users only need to replace the data extraction node with experimental data files related to composite materials, including material composition tables and material microscopic images; no changes are required to the configurations of other nodes. Since the performance of composite materials depends not only on the composition table data but also on the porosity and defects in the microscopic images, users do not need to modify the original table processing logic. They only need to drag and drop a built-in deep learning image segmentation node, such as the U-Net model, from the material operator library and connect it to the original "data fusion node".

[0129] During the workflow execution phase, a dual-modal parallel processing flow is formed: one path processes the material composition table data; the other path, with its built-in U-Net node, receives microscopic images and automatically outputs standard data objects containing morphological features such as "pore area" and "aspect ratio". By reusing the template backbone and plugging and unplugging the built-in multimodal nodes, the workflow construction time is significantly shortened.

[0130] Third Embodiment

[0131] This embodiment employs the aforementioned method and system for autonomously retrieving material data based on multi-source data and multimodal analysis to perform tasks such as calculating catalyst surface adsorption energy and predicting catalytic activity. This embodiment focuses on illustrating the system's ability to extend custom operators for third-party professional computing tools and its capacity for scheduling complex tasks.

[0132] In catalysis research, it is necessary to obtain electronic structure data of materials, which relies on first-principles calculation software (such as VASP). Such heavy scientific computing software cannot be built into the analysis platform as a general node.

[0133] To this end, users can drag and instantiate a "Command Line Studio" custom node from the operator node library. In its parameter configuration panel, users can fill in the script commands for calling the VASP software, the SSH connection parameters for the target high-performance computing cluster (HPC), and the computing power queue requirements, and define standard numerical data types for physical quantities such as "adsorption energy" and "d-band center" in the output port.

[0134] Users integrate this custom node into the workflow. During the scheduling and execution phase, since VASP calculations typically take several hours, the workflow scheduling engine schedules it concurrently with regular table preprocessing nodes to avoid resource idleness.

[0135] During execution, the system's log management module provides real-time feedback on the cluster queue status and the execution time of custom nodes on the interface. If VASP script execution is interrupted due to external cluster network fluctuations, the log module automatically identifies the abnormal pattern and outputs an analysis of the cause of "remote connection timeout" in the error message area.

[0136] At this point, users do not need to interrupt or restart the entire workflow. Instead, they can directly utilize the system's single-node retry mechanism to resubmit computations only for the failed "command-line studio node." Upon successful retry, the node automatically parses the file, converts it into a standard data object, and passes it to downstream nodes, ensuring the robustness and data flow integrity of long-cycle, multi-source computation tasks.

[0137] In summary, this invention takes ETL data flow as its core paradigm and integrates multi-source material data access, autonomous design of analysis tasks, construction and reuse of machine learning models, expansion of custom operators, and visualization of results into a unified framework through a visual workflow engine. Materials researchers do not need to have programming skills to independently complete the design and execution of complex material data analysis tasks with the help of this invention, and achieve continuous accumulation and sharing of analysis assets through versioned encapsulation of analysis processes and cross-task model reuse.

[0138] It should be understood that, in various embodiments of the present invention, the order of the above-mentioned process numbers does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0139] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed, and is not intended to limit the scope of the claimed invention, but only to illustrate preferred embodiments of the invention. Those skilled in the art should understand that the scope of the invention is not limited to the specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the inventive concept. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

Claims

1. A method for autonomously retrieving material data based on multi-source data and multimodal analysis, characterized in that, The method includes the following steps: Step S1: Construct a material operator node library and simultaneously construct a multi-source heterogeneous database for materials. The multi-source heterogeneous database includes a user-owned data pool and a material database pool. Step S2: Based on the material calculation task, new user-owned data and relevant material data from the user-owned data pool are automatically loaded; at the same time, search conditions are set according to the material calculation task, and relevant material data is retrieved from the material database pool based on the search conditions. Step S3: Perform a unified format conversion on all loaded multi-source heterogeneous material data, and use the data of each corresponding material calculation task as a data sample; Step S4: Determine if there is a duplicate material calculation task; If it exists, call the workflow topology diagram encapsulated under the corresponding material calculation task, modify the parameters of the corresponding operator node according to the actual situation of the material calculation task and data sample, and then proceed to step S7; if it does not exist, proceed to step S5. Step S5: Based on the material calculation task, call at least one material operator node from the material operator node library to construct the workflow topology graph of the current task. The nodes of the workflow topology graph include all data samples and operator nodes, and the edges of the workflow topology graph include the connection relationship from data samples to operator nodes and the connection relationship between operator nodes. Initial parameter configuration is performed on the workflow topology graph. Step S6: Read the configuration file of the workflow topology diagram and perform compliance verification on the configuration file; Step S7: Sort the workflow topology diagram, autonomously schedule and execute material calculation tasks, and monitor and manage the running status in real time. Step S8: Visualize the task execution results.

2. The method according to claim 1, characterized in that, The operator node library covers the entire material data analysis process, including data fusion node groups, data processing node groups, feature engineering node groups, model node groups, and visualization node groups.

3. The method according to claim 1, characterized in that, The user-owned data pool includes user-owned data loaded each time a material calculation task is executed. It consists of local CSV and Excel format numerical data files, MOL, SDF, RXN, and RDF material-specific format structured data files, as well as material microscopic image files in JPG, PNG, or TIFF format from scanning electron microscopy (SEM) and transmission electron microscopy (TEM). The material database pool includes third-party material databases accessed through a preset application programming interface and a characterization state transition architecture.

4. The method according to claim 1, characterized in that, The data flow of the workflow topology is as follows: the proprietary data node and the retrieved data node output multi-source standard data objects to the data fusion node respectively; the data fusion node performs entity recognition and horizontal stitching on the multi-source standard data objects based on the unique material identifier, eliminates attribute conflicts, and generates an initial fusion dataset. The initial fused dataset is then passed to the data processing node, and subsequently passes through the feature engineering node and the model calling node in sequence. Finally, the visualization node outputs the calculation results.

5. The method according to claim 1, characterized in that, Step S5, which involves calling the material operator node according to the material calculation task, includes: The graphical workflow design interface provides human-computer interaction support; users can instantiate target operator nodes into the workflow canvas by dragging and dropping from the visualized operator node library directory according to the logical requirements of the current computing task, and add edges to the workflow topology graph by drawing lines between nodes.

6. The method according to claim 1, characterized in that, The compliance verification includes: Topology compliance verification checks for loops in the directed graph to ensure the workflow is a compliant directed acyclic graph. Parameter configuration validity verification verifies whether the required parameters of each node have been fully configured. Data type compatibility check: Verify whether the output data type and input data type match between adjacent nodes; Data source reachability verification verifies whether the data source specified by each data access node can be accessed normally.

7. The method according to claim 1, characterized in that, The topology sorting includes: sorting the DAG workflows that have passed compliance verification according to data dependencies to determine the compliance execution order of each material operator node.

8. The method according to claim 1, characterized in that, The autonomous invocation includes a hybrid serial-parallel autonomous scheduling mechanism. Specifically: the workflow task scheduling engine sorts the results by topology and autonomously invokes the encapsulated processing logic corresponding to each node from the material operator node repository for execution in sequence; a workflow dependency graph is constructed based on adjacency lists and inverse adjacency lists, and the node topology hierarchy is calculated through breadth-first search. Depth-first search or transitive closure matrix is ​​used to determine whether there is a direct or indirect data dependency between any two nodes in the ready node set. Nodes with data dependencies are scheduled sequentially, while nodes without direct or indirect data dependencies are automatically identified and scheduled concurrently; each node automatically extracts output data from upstream nodes, performs conversion operations, and loads the processing results to the input port of downstream nodes, realizing automatic pipelined processing of material data.

9. The method according to claim 1, characterized in that, The methods for visualizing the task execution results include: (1) Numerical results: such as prediction results of continuous properties of materials, such as band gap, hardness, and conductivity, presented in the form of data tables or scatter plots. (2) Classification and clustering results: such as discrete classification results such as material phase structure classification and stability discrimination, as well as clustering results of new material category discovery, combined with receiver operation characteristic curves and confusion matrix to present model confidence; (3) Image-type results: such as the segmentation or target recognition results of material microscopic images, which are presented in a visual comparison by superimposing color masks or bounding boxes on the original microscopic images; (4) Explanatory analysis results: such as the ranking of feature importance output by feature engineering nodes, presented in the form of heat map or model performance improvement map, to help explain the cause mechanism of material properties.

10. A material data autonomous retrieval system based on multi-source data and multi-modal analysis, characterized in that, The system includes: a materials database construction module, a materials operator node library, a multi-source heterogeneous database, a materials calculation task loading module, a data loading module, a data format conversion module, a topology graph call and judgment module, a workflow topology graph construction module, an initial parameter configuration module, a compliance verification module, a calculation task execution module, a real-time monitoring module, a log management module, and a visualization module; among which... The material database construction module is used to construct a material operator node library and a multi-source heterogeneous database oriented towards materials. The material operator node library is used to store material operator nodes; The multi-source heterogeneous database is used to store multi-source heterogeneous databases, which include user-owned data pools and material database pools. The material calculation task loading module is used to load material calculation tasks; The data loading module is used to autonomously load new user-owned data and relevant material data from the user-owned data pool according to the material calculation task; at the same time, it sets search conditions according to the material calculation task and retrieves relevant material data from the material database pool based on the search conditions. The data format conversion module is used to perform unified format conversion on all loaded multi-source heterogeneous material data, with each data item corresponding to a material calculation task serving as a data sample. The topology graph call judgment module is used to determine whether there is a duplicate material calculation task. If it exists, the workflow topology graph encapsulated under the corresponding material calculation task is called, the parameters of the corresponding operator nodes are modified according to the actual situation of the material calculation task and data sample, and the calculation task execution module is started. If it does not exist, the workflow topology graph construction module is started. The workflow topology graph construction module is used to call at least one material operator node from the material operator node library according to the material calculation task to construct the workflow topology graph of the current task. The nodes of the workflow topology graph include all data samples and operator nodes, and the edges of the workflow topology graph include the connection relationship from data samples to operator nodes and the connection relationship between operator nodes. The initial parameter configuration module is used to configure the initial parameters of the workflow topology graph; The compliance verification module is used to read the configuration file of the workflow topology diagram and perform compliance verification on the configuration file; The computation task execution module is used to perform topological sorting on the workflow topology graph and autonomously schedule and execute material calculation tasks. The real-time monitoring module is used to monitor the operating status in real time; The log management module is used to manage the running status logs; The visualization module is used to visualize the results of task execution.

Images