Trinity platform data research and development closed-loop collaboration method and system

Abstract

The present application relates to the technical field of data collaborative management, in particular to a trinity platform data research and development closed-loop collaboration method and system, a high-throughput computing module for automatically simulating and calculating the high-throughput of various properties of a material system and outputting unified structured data, a material simulation database module pre-provided with a unified data standard for automatically receiving, storing and calculating results and providing a standardized query interface, a material artificial intelligence module for executing data preprocessing, feature engineering, model training and deployment, and a closed-loop collaboration module for capturing artificial intelligence prediction results and automatically converting them into new computing tasks according to task parameter mapping rules and submitting them to the high-throughput computing module. The present application realizes full-process automatic collaboration, self-enhanced intelligent feedback loop and data security and credibility, and significantly improves the material research and development efficiency.

Description

Technical Field

[0001] This invention relates to the field of data collaborative management technology, and in particular to a three-in-one platform data research and development closed-loop collaborative method and system. Background Technology

[0002] With the popularization of the materials genome concept, high-throughput computing, materials databases, and machine learning technologies have become key means to accelerate the discovery and development of new materials. However, in current research and development practices, these technology modules are often isolated from each other and lack effective collaboration.

[0003] First, there is a severe disconnect between high-throughput computation and data management. Researchers generate massive amounts of results after performing high-throughput calculations of thermodynamic and kinetic properties using specialized software such as phase diagrams. These results are scattered across computation nodes in the form of disorganized, inconsistently formatted raw files. Organizing, labeling, and inputting these results into materials databases requires a significant amount of manual work, which is not only inefficient but also highly susceptible to human error, hindering the effective accumulation and reuse of data assets.

[0004] Secondly, data storage and artificial intelligence analysis are disconnected. Even when materials databases are established, their internal data formats, naming standards, and metadata descriptions often lack unified specifications. This makes it difficult for these databases to provide standardized, highly compatible data interfaces required for training AI models. Researchers still face a heavy workload when performing preprocessing tasks such as data cleaning and feature engineering. This process is independent of subsequent model training, evaluation, and deployment, severely hindering the large-scale application of AI technology in materials research and development.

[0005] Finally, there is a lack of feedback between AI predictions and the next round of computational design. Existing AI prediction models are typically used as standalone offline tools, and their predictions (such as the potential superior performance of a component) cannot be directly and automatically fed back to the computational module or experimental design stage. Therefore, the R&D process still follows an open-loop model of "computation → manual analysis → manual decision-making → re-computation," requiring significant manual intervention for each iteration, resulting in a lengthy and inefficient R&D cycle.

[0006] In addition, some existing integrated platforms are complex to operate, require users to have a high level of cross-domain knowledge, and generally lack security verification and access control mechanisms throughout the entire data lifecycle.

[0007] In conclusion, there is an urgent need for a materials research and development platform that can deeply integrate computing, data, and artificial intelligence to achieve fully automated, intelligent, and closed-loop iteration throughout the entire process, in order to solve the current technical problems of isolated technology modules, high dependence on manual labor, and low efficiency of research and development iteration. Summary of the Invention

[0008] The purpose of this invention is to provide a three-in-one platform data research and development closed-loop collaborative method and system, which realizes full-process automated collaboration, self-reinforcing intelligent feedback closed loop, and data security and reliability, significantly improving the efficiency of materials research and development.

[0009] To achieve the above objectives, in a first aspect, the present invention provides a three-in-one platform data R&D closed-loop collaborative system, comprising: The high-throughput computing module is used to perform high-throughput automatic simulation calculations on various properties of material systems and output the calculation results in a unified structured format. The materials simulation database module is pre-set with a unified structured data standard, which is used to automatically receive and store the calculation results output by the high-throughput computing module, automatically label the stored data according to the structured data standard, and provide a standardized data query interface. The materials artificial intelligence module is used to acquire data from the materials simulation database module, perform data preprocessing, feature engineering, model training and deployment, and obtain a materials performance prediction model. The closed-loop collaborative module is used to capture the prediction results of the material artificial intelligence module, automatically convert the prediction results into new calculation task parameters according to the preset task parameter mapping rules, and submit them to the high-throughput computing module for verification calculation.

[0010] The high-throughput computing module includes: The task configuration unit is used to receive the settings of material system, composition space, process parameter space and target calculation properties, and generate a calculation task definition object; The task generation and scheduling unit contains a discretization engine and a computation scheduler. The discretization engine is used to discretize the continuous component space and process parameter space in the computation task definition object into multiple independent computation tasks. The computation scheduler is used to dynamically allocate tasks according to the computing node resources. The computation execution unit is used to receive independent computation tasks, call the phase diagram computation engine to perform simulations, monitor task status, and implement a limited automatic retry strategy. The results collection and standardization unit is used to receive the raw result files returned by each computing node, convert the raw results into a unified structured data format of the platform through the format conversion engine, and push them to the materials simulation database module after attaching computing metadata.

[0011] The material simulation database module includes: The data standard setting unit is used to define structured data standard templates that include material system identifiers, composition lists, process condition sets, calculation result sets, and metadata blocks, and publish them in the form of machine-readable global data contracts. The data receiving and parsing unit is used to monitor data push, receive structured data packets, extract key fields and perform data type validation, and output standardized data objects. An automated tag generation unit maintains a tag generation rule base and is used to extract or derive tag metadata from standardized data objects according to the rule base to form a tag set in the form of key-value pairs. The storage and indexing unit uses a distributed document storage system to merge standardized data objects and tag metadata for storage, and builds an inverted index for tags and a range index for numeric fields. The external data fusion unit includes a data source adapter manager and multiple adapter plugins, which are used to connect to external public material databases. It maps external data fields to internal data standards through a schema mapping table and stores them in the storage and indexing unit after deduplication.

[0012] The material artificial intelligence module includes: The data preprocessing unit is used to obtain the raw dataset from the material simulation database module, and sequentially perform outlier detection and removal, duplicate record removal, numerical normalization or standardization, missing value handling, and save the preprocessing transformation parameters. The feature engineering unit includes a materials science knowledge feature library, a feature dimensionality reducer, and a feature filter. The materials science knowledge feature library is used to construct derived features from composition and process parameters based on atomic property data. The feature dimensionality reducer is used to reduce the feature dimension through principal component analysis or autoencoder. The feature filter is used to filter feature subsets through correlation analysis or model-based feature importance evaluation. The model training and management unit is used to divide the dataset into training, validation and test sets, perform hyperparameter optimization, train the final model, and store the model and its metadata in the model library. The model deployment and inference unit is used to load models from the model library, package them as online services or offline tools, perform the same preprocessing and feature construction on the input data as during training, call the model for inference, and output performance prediction results.

[0013] The closed-loop collaborative module includes: The prediction result capture unit is used to continuously monitor the prediction result messages output by the materials artificial intelligence module; The data parsing and verification unit is used to extract the compositional boundaries, target performance, and confidence level of candidate materials from the message, and to verify their physical rationality and credibility, thereby generating candidate scheme objects. The task parameter mapping rule library is used to define the mapping logic for converting candidate solution objects into specific computation task parameters; The computation task generation unit is used to sequentially call the rule set in the rule base, discretize the candidate scheme object into a list of component points, associate each component point with a combination of process conditions and target computation properties, generate an atomic computation task list, and submit it to the high-throughput computing module in batches through the task submitter. The collaborative fine-tuning and iterative control unit is used to monitor the completion status of computing tasks, trigger new data entry events, initiate incremental model update requests to the materials artificial intelligence module, and determine whether to continue iterating based on preset termination conditions.

[0014] When performing incremental model updates, the collaborative fine-tuning and iterative control unit sends a request containing the new data storage location, the original model identifier, and the iteration round number to the materials artificial intelligence module according to a preset fine-tuning strategy. The materials artificial intelligence module merges the new data with the original training data and performs incremental training or partial parameter fine-tuning on the existing model, and returns the new model identifier.

[0015] It also includes a security and access control module, which uses the national cryptographic hash algorithm to calculate the hash value of the data and performs the following integrity verification process: Before data is written to storage, the hash value of the data is calculated and attached to the storage as an integrity benchmark value. When data is transmitted between modules via the application programming interface, the sender attaches a hash value to the data, the receiver recalculates the hash value and compares it with the value provided by the sender, and rejects the data if they do not match. When reading data from storage, the hash value of the data is recalculated and compared with the baseline value in storage. If they are inconsistent, the data is prevented from being returned and an alarm is triggered. The operation logs are protected by chained cumulative hashing, where each log entry stores its independent hash value and the cumulative hash value obtained by concatenating the hash value of the previous log entry. Periodically schedule global integrity audit tasks, traverse the stored data, recalculate the hash value and compare it with the baseline value, and generate an audit report.

[0016] The high-throughput calculation module supports high-throughput calculations of at least one of the following properties of materials: thermodynamic phase equilibrium, phase diagram, Gibbs free energy, kinetic diffusion coefficient, thermal conductivity, thermal expansion coefficient, dendrite growth rate, and temperature range of the mushy region.

[0017] In the closed-loop collaborative module, the component discretization rule set in the task parameter mapping rule base includes a uniform grid discretization strategy, a random sampling discretization strategy, and a key point sampling strategy based on materials science knowledge. The process parameter determination rule set automatically associates temperature scanning range, cooling rate, and casting temperature parameters according to the target performance type. The calculation property selection rule set pre-sets multiple sets of property mapping relationships and calculation priorities according to the prediction target.

[0018] Secondly, the present invention provides a three-in-one platform data R&D closed-loop collaborative method, applied to a three-in-one platform data R&D closed-loop collaborative system as provided in the first aspect, comprising the following steps: The first round of simulation calculations is performed using a high-throughput computing module based on an initial material design space, which includes a material composition space and a process parameter space. The calculation results are automatically pushed to the material simulation database module, where they are parsed, tagged, and stored according to the preset unified structured data standard. The material artificial intelligence module extracts data from the material simulation database module, and performs data preprocessing, feature engineering, and model training in sequence to obtain a material performance prediction model. The performance prediction model is used to predict the extended design space and select candidate material schemes. The candidate material schemes are captured by the closed-loop collaborative module. According to the component discretization rules, process parameter determination rules and calculation property selection rules in the task parameter mapping rule library, the candidate schemes are automatically converted into new high-throughput computing task parameters, generating an atomic computing task list and submitting it to the high-throughput computing module for verification calculation. Once the new calculation is completed, the results are automatically stored and a data reflux event is triggered. The materials artificial intelligence module incorporates the new data into the training set and performs incremental fine-tuning on the current model until the preset iteration termination condition is met, and outputs the final materials design scheme.

[0019] This invention discloses a three-in-one platform data research and development closed-loop collaborative method and system. The system includes: a high-throughput computing module for performing high-throughput automatic simulation calculations of various properties of material systems and outputting unified structured data; a material simulation database module with pre-set unified data standards for automatically receiving, tagging, and storing calculation results and providing standardized query interfaces; a material artificial intelligence module for performing data preprocessing, feature engineering, model training, and deployment; and a closed-loop collaborative module for capturing artificial intelligence prediction results and automatically converting them into new computational tasks according to task parameter mapping rules, which are then submitted to the high-throughput computing module. This invention also discloses an intelligent material research and development closed-loop collaborative method based on this system. This invention solves the technical problems of low research and development efficiency and long iteration cycles caused by the isolation of computation, data, and artificial intelligence modules and reliance on manual processes in existing material research and development. It achieves fully automated collaboration, a self-reinforcing intelligent feedback closed loop, and secure and reliable data, significantly improving the efficiency of material research and development. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0021] Figure 1 This is a structural principle diagram of a three-in-one platform data research and development closed-loop collaborative system according to the first embodiment of the present invention.

[0022] Figure 2 This is a schematic diagram of the steps of a three-in-one platform data research and development closed-loop collaborative method according to the second embodiment of the present invention.

[0023] In the diagram: 101-High-throughput computing module, 102-Materials simulation database module, 103-Materials artificial intelligence module, 104-Closed-loop collaboration module, 105-Security and access control module, 1011-Task configuration unit, 1012-Task generation and scheduling unit, 1013-Computation execution unit, 1014-Result collection and standardization unit, 1021-Data standard setting unit, 1022-Data receiving and parsing unit, 1023-Automatic tag generation unit, 1024-Storage and indexing unit, 1025-External data fusion unit, 1031-Data preprocessing unit, 1032-Feature engineering unit, 1033-Model training and management unit, 1034-Model deployment and inference unit, 1041-Prediction result capture unit, 1042-Data parsing and verification unit, 1043-Task parameter mapping rule base, 1044-Computation task generation unit, 1045-Collaborative fine-tuning and iterative control unit. Detailed Implementation

[0024] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0025] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0026] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0027] The first embodiment of this application is as follows: Please see Figures 1-2 This invention provides a three-in-one platform data research and development closed-loop collaborative system, a high-throughput computing module, which is used to perform high-throughput automatic simulation calculations of various properties of material systems and output the calculation results in a unified structured format; The materials simulation database module is pre-set with a unified structured data standard, which is used to automatically receive and store the calculation results output by the high-throughput computing module, automatically label the stored data according to the structured data standard, and provide a standardized data query interface. The materials artificial intelligence module is used to acquire data from the materials simulation database module, perform data preprocessing, feature engineering, model training and deployment, and obtain a materials performance prediction model. The closed-loop collaborative module is used to capture the prediction results of the material artificial intelligence module, automatically convert the prediction results into new calculation task parameters according to the preset task parameter mapping rules, and submit them to the high-throughput computing module for verification calculation.

[0028] In this embodiment, the high-throughput computing module includes: a task configuration unit, a task generation and scheduling unit, a computing execution unit, and a result collection and standardization unit. The technical solutions for each unit are described in detail below.

[0029] The task configuration unit offers both a graphical user interface and script templates to accommodate users of varying skill levels. The unit first guides users to define the target material system and select the list of components it contains. Building on this, the unit provides a composition space definer, allowing users to set the content range of each component in continuous intervals, while also supporting the setting of compositional constraints. For process parameters, the unit supports defining multiple continuous or discrete process dimensions. Users also need to select the physicochemical properties to be simulated in this task from a pre-set computational property library, including but not limited to thermodynamic phase equilibrium, phase diagrams, Gibbs free energy, kinetic diffusion coefficient, thermal conductivity, coefficient of thermal expansion, dendrite growth rate, and the temperature range of the mushy region.

[0030] The task configuration unit incorporates a parameter validity validator. This validator automatically checks whether the boundary values ​​of the composition space and process parameter space are within the effective range of the material system's thermodynamic model, whether the discretization step size will exceed the system's preset maximum task limit, and provides real-time prompts for user-input conflicts. After all configurations are complete, the unit is encapsulated into a complete computational task definition object, which includes the material system identifier, composition space boundaries, process parameter space boundaries, a list of target properties, and user-added metadata. This definition object is then passed to the task generation and scheduling unit.

[0031] The task generation and scheduling unit is responsible for breaking down the aforementioned single computational task definition object into a massive number of independently executable atomic computational tasks, and efficiently distributing them based on the current cluster resource status. Internally, the unit includes a discretization engine. This engine receives the component space boundary and process parameter space boundary from the computational task definition object, and generates a series of discrete parameter combinations in a continuous multidimensional parameter space according to the user-specified discretization strategy (uniform grid, random sampling, or custom point set). Each parameter combination, combined with a user-specified list of target properties, constitutes an independent computational task. Each independent computational task has a globally unique task identifier and records the coordinates of its source parameter combination for easy traceability of subsequent results.

[0032] The massive number of independent computing tasks generated are not submitted directly to the computing nodes, but first enter a persistent task queue manager. This manager persistently stores the tasks in a reliable message broker to prevent task loss due to system failures. The task queue supports priority sorting; tasks submitted by users or closed-loop collaboration modules can be assigned different priorities, with higher-priority tasks being pulled by the scheduler first.

[0033] The module's built-in compute scheduler continuously monitors the load status of each compute node in the compute cluster, including CPU core utilization, memory availability, and storage I / O status. The scheduler uses a dynamic pull mode: idle compute nodes actively request new tasks from the scheduler, which then retrieves tasks from the head of the queue based on task priority and node affinity and assigns them to that node.

[0034] The computation execution unit runs on each computing node, responsible for local task execution, status monitoring, and initial result processing. An execution agent is deployed on each computing node. This agent continuously listens for task allocation instructions from the scheduler. Upon receiving an independent computation task, the execution agent first creates a working directory locally and downloads the required input files. Then, the agent calls the application programming interface of the encapsulated phase diagram computation engine to convert the composition parameters, process parameters, and target property list in the task into an input format recognizable by the engine, thus initiating the computation process.

[0035] During the computation process, the execution agent implements multi-level real-time monitoring: it monitors the running status of the computation process itself, including whether the process is alive and whether the CPU and memory usage are abnormal. It parses the computation output stream and captures warnings or error messages output by the engine. A watchdog timer is set; if the execution time of a single computation task exceeds a preset reasonable threshold, it is considered a timeout.

[0036] When a computation failure is detected, a limited automatic retry strategy is followed: the task status is marked as "failed and pending retry," and the task and its current retry count are pushed back into the task queue. The maximum number of retries is configured by the system; once this limit is exceeded, the task is marked as "permanently failed," its error log is collected and stored in a dedicated failed task table, and an alarm is triggered. If the computation completes successfully, the execution agent collects the raw result files output by the engine; these files are typically in proprietary binary format or unstructured text format.

[0037] In addition, the execution agent records comprehensive metadata for this computation, including the task identifier, start and end timestamps, the name of the node executing the task, the version number of the phase graph computation engine used, and the hash values ​​of the input parameters. This metadata will be packaged together with the original result file and passed to the next unit.

[0038] The Results Collection and Standardization unit is responsible for converting the raw, heterogeneous results files scattered across various computing nodes into a standardized, structured data format within the platform. This data is then integrated with metadata and finally pushed to the materials simulation database module. This unit runs a Results Collection service, which exposes a receiving endpoint for the execution agents on each computing node to call back after computation. The execution agent pushes the raw results files, metadata data, and task identifier to the Results Collection service. To improve reliability, the Results Collection service employs a message acknowledgment mechanism: an acknowledgment signal is only returned to the execution agent after the data has been successfully received and preliminarily verified, allowing the execution agent to clean up local temporary files. If no acknowledgment is received, the execution agent caches the results locally and waits to retry.

[0039] The results collection service includes a format conversion engine. This engine has built-in parsers for the raw result file formats output by different types of phase diagram calculation engines. The parser reads the raw file and extracts all calculation result values ​​and their corresponding physical conditions. During the conversion process, the engine standardizes units. The converted data is organized into a standardized data record. This record strictly follows a unified data model predefined by the platform, which uses an extensible tag structure to accommodate the following fixed fields: material system identifier, precise composition list, process condition list, calculated property name and corresponding numerical result, calculation metadata, and data integrity check placeholders.

[0040] The materials simulation database module is internally divided into five unit structures: data standard setting unit, data receiving and parsing unit, automated label generation unit, storage and indexing unit, and external data fusion unit.

[0041] The data standards development unit first defines a set of structured data standard templates. This template describes, in an extensible hierarchical structure, the fixed set of fields that a material data record should include, along with their data types, value ranges, and constraints. Fixed fields include at least: a material system identifier, a composition list (each element and its content), a set of process conditions (temperature, pressure, cooling rate, etc.), a set of calculation results (property names and values), and metadata blocks (timestamp, calculation software version, data source, data integrity check placeholders, etc.). To accommodate different material systems that may have different numbers of components and property types, the standard template adopts a flexible list structure, allowing the lengths of the composition list, condition list, and result list to change dynamically, but the sub-field structure within each list remains strictly fixed.

[0042] In addition to the fixed fields, the standard template also defines an optional extended field area for recording information specific to certain calculation types. This extended area follows the "name-type-unit" triple specification, allowing newly added calculation properties to be included without modifying the core template structure.

[0043] The data standards development unit serializes the template in a machine-readable format and publishes it as the platform's global data contract. All other units must reference this contract when processing data. When the platform needs to upgrade the data standards, this unit provides a version management mechanism, allowing old and new versions of the standards to coexist and be distinguished by version numbers, ensuring that historical data is not destroyed, and supporting data migration by version.

[0044] The core component of the data receiving and parsing unit is a data listener. The data listener runs as a persistent service, exposing a data receiving endpoint that conforms to the platform's security specifications. This endpoint employs a high-performance asynchronous input / output model, capable of handling a large number of concurrent data push requests simultaneously. The listener implements a message acknowledgment mechanism: upon receiving a data packet, it first performs an integrity check; if the check passes, it returns an acknowledgment to the sender; otherwise, it requests a retransmission.

[0045] The received raw data is encapsulated in a platform-defined portable data exchange format. The data receiving and parsing unit contains a format parsing engine. This engine calls the corresponding parser based on the format version declared in the data packet header. The parser is responsible for extracting various fields from the data packet, including material system identifiers, composition lists, process condition lists, calculation result lists, and metadata blocks. During parsing, the engine performs preliminary data type validation: for example, ensuring that the content values ​​in the composition list are valid numeric types, that the timestamp conforms to date and time specifications, and that the property name is not empty. For missing or incorrectly formatted fields, the parsing engine handles them according to a preset error tolerance strategy: if the missing field is a non-critical field, it is filled with a null value; if the missing field is a critical field, the data is rejected and an error log is recorded.

[0046] The automated tag generation unit is responsible for automatically creating a set of descriptive tag metadata for each standardized data object. Internally, this unit maintains a tag generation rule base. The rule base defines the mapping relationship from fixed fields of the standardized data object to tag fields. The specific generation process is as follows: After receiving the standardized data object from the parsing unit, the automated tag generation unit iterates through each rule in the tag generation rule base. For each rule, the unit extracts the original value from the corresponding field of the data object and processes it according to the optional transformation function defined in the rule. In addition to tags directly mapped from data fields, the rule base also supports generating derived tags. Derived tags are calculated based on a combination of multiple fields.

[0047] All generated tags are organized into a tag metadata object, which is a flat collection of key-value pairs, where the key is the tag name and the value is the tag content. The tag metadata object is associated with the original data object through a unique data record identifier. The automated tag generation unit also adds a data source identifier and a data precision level identifier to the tags.

[0048] The storage and indexing unit is responsible for persistently storing standardized data and its associated tag metadata, building efficient retrieval indexes, and simultaneously invoking the security and access control module to ensure data integrity. This unit uses a distributed document storage system as its underlying storage engine. Each standardized data object and its corresponding tag metadata object are merged into a complete "data document," stored in a format compatible with the global data contract. During storage, the unit first invokes the integrity verification service provided by the security and access control module to calculate the hash value of the entire data document and writes this hash value as a metadata field of the document. This step ensures the baseline integrity of the data at the time of writing.

[0049] After a document is written to the storage system, the storage and indexing unit triggers the index builder. The index builder parses the tag metadata objects in the document and builds an inverted index for each tag key-value pair. The inverted index data structure is as follows: for each unique tag value, it records a list of identifiers for all data documents containing that tag value. This indexing method allows subsequent query operations to locate data documents that meet specific tag combinations very quickly.

[0050] In addition to tag-based inverted indexes, the storage and indexing unit also provides users with full-text search capabilities. The unit builds full-text indexes for text fields such as component lists and calculated property names in standardized data objects, supporting fuzzy matching and range queries. Simultaneously, for calculation results of numeric types, the unit builds range indexes, allowing users to filter by numerical ranges. All indexed data is stored separately from the original data documents but is kept in real-time synchronization: whenever a new document is written or an existing document is updated, the index builder automatically updates the corresponding index structure. The storage and indexing unit also exposes a unified query application interface that accepts query expressions based on tag filtering, full-text search, and combinations of numerical ranges, and returns a list of data documents that meet the conditions, along with their tag information.

[0051] The external data fusion unit is responsible for interfacing with publicly available external material databases, automatically retrieving external data, and mapping, transforming, and fusion it according to the platform's internal data standards to enrich the platform's data assets. This unit includes a data source adapter manager for managing the connection configurations of multiple external data sources. Each external data source typically has its own data model, access protocol, and authentication method. The adapter manager provides a corresponding adapter plugin for each data source. Each adapter plugin implements the following standard functions: connection establishment and authentication, data query interface adaptation, result set pagination retrieval, error retrieval, and rate limiting control.

[0052] During runtime, the external data fusion unit initiates an external data retrieval task according to the system's configured synchronization strategy (e.g., scheduled execution or event triggering). For each configured external data source, the unit calls its adapter plugin to execute a preset data query statement. This query statement is typically used to retrieve data records that have been added or updated since the last synchronization.

[0053] After the adapter plugin returns external data records, the external data fusion unit enters the data mapping and standardization phase. Internally, the unit maintains a schema mapping table that records the mapping relationships between the original field names and data types of the external data source and the corresponding fields in the platform's internal structured data standard template. The mapping process includes: field name mapping, data type conversion, unified unit conversion, and filling in default values ​​for missing fields.

[0054] After completing the field mapping, the unit organizes the transformed data into standardized data objects that conform to the platform's global data contract. However, external data often lacks platform-specific metadata. For these missing metadata fields, the external data fusion unit adopts a reasonable default value filling strategy and explicitly marks the "data source" label of the data as the corresponding external database name, while also adding a "data credibility" label.

[0055] Subsequently, the standardized external data objects are sent to the automated tag generation unit, which generates tag metadata according to the same process as the internally calculated data. Finally, the storage and indexing unit writes this metadata to the database. Throughout the process, the external data fusion unit also performs deduplication: before writing, it compares the new data with existing data using the "material system + composition characteristics + process conditions + property type" key to determine if there is duplicate data. If duplicates are detected, they are handled according to a preset conflict resolution strategy.

[0056] The materials artificial intelligence module consists of: a data preprocessing unit, a feature engineering unit, a model training and management unit, and a model deployment and inference unit.

[0057] The data preprocessing unit queries and obtains the original dataset from the material simulation database module, and then performs operations such as outlier detection and removal, duplicate record removal, numerical normalization or standardization, and missing value handling in sequence, finally outputting a clean, standardized, and complete data table.

[0058] The unit contains a dataset query builder. This builder generates requests conforming to the database query interface specification based on filtering criteria specified by the user or higher-level processes, and retrieves raw data in batches from the materials simulation database module. During the retrieval process, the system supports pagination to handle large datasets and performs integrity checks on network transmission. The retrieved data is loaded into an in-memory data frame structure for subsequent row-by-row processing.

[0059] The unit incorporates an anomaly detection processor. This processor supports two optional statistical detection methods: box plot method and standard score method. In the box plot method, the processor calculates the quartiles and interquartile ranges for each numerical feature, classifying values ​​below the first quartile minus 1.5 times the interquartile range, or above the third quartile plus 1.5 times the interquartile range, as anomalies. In the standard score method, the processor calculates the mean and standard deviation for each feature, classifying data points whose absolute standard score exceeds a preset threshold as anomalies. This threshold is set to three by default and can be adjusted by the user based on their knowledge of the specific material system.

[0060] For data records identified as anomalous, the processor does not delete them directly. Instead, it executes one of the following strategies based on the system configuration: directly remove the entire record, mark the outlier as a missing value for later processing, or only record a warning but retain the original value. The default strategy is direct removal to ensure the purity of the training data.

[0061] The materials database may contain duplicate records due to duplicate calculations or external data fusion. The deduplication processor in the cell identifies duplicate records based on a combination of key fields. These key fields include the material system identifier, a standardized string representation of the composition list, a numerical sequence of process conditions, and the name of the calculated property. For multiple records with the same combination of key fields, the deduplication processor retains the record with the most recent timestamp by default, or selects based on user-specified priorities. After deduplication, the discarded duplicate records are moved to an isolation table to prevent permanent data loss.

[0062] To eliminate the impact of dimensional differences between different features on model training, the unit integrates a numerical transformation processor. This processor provides two main transformation strategies for different types of features.

[0063] For features with well-defined boundaries, such as component content, the processor employs a minimum-maximum normalization method to linearly map the original values ​​to a closed interval between zero and one. During the transformation, the global minimum and maximum values ​​from the training set are used. The processor calculates and caches these boundary values ​​during the first data iteration to ensure that subsequent transformations of new data use the same parameters.

[0064] For continuous features such as temperature and pressure with unbounded theoretical ranges, the processor uses a standard score normalization method, subtracting the mean from the original values ​​and then dividing by the standard deviation, so that the transformed data has zero mean and unit variance. Similarly, the mean and standard deviation are calculated from the training set and persistently stored.

[0065] Missing values ​​may exist in the dataset due to computational failures or incomplete external data. The missing value handler within a cell employs different strategies based on the field type and missing value rate. For critical attributes such as component content, key process parameters, or main calculation results, if the missing value rate is below a system preset threshold, the handler uses interpolation methods to fill in the missing values. For continuous numerical features, the handler uses linear interpolation or a weighted average method based on neighboring samples; for discrete features, the handler uses mode imputation. If the missing value rate exceeds the threshold, or if the missing value of a critical field prevents a reasonable value from being obtained through interpolation, the handler directly deletes the record.

[0066] After processing, the data preprocessing unit outputs a complete dataset with uniform numerical range, no duplicates, and all anomalies handled, along with a record of all parameters used in this preprocessing, for subsequent units to reproduce.

[0067] The feature engineering unit incorporates a materials science knowledge feature library, which pre-configures various algorithms for calculating derived features based on material composition and process conditions in the form of rule sets. When the user specifies a target material system, the feature constructor automatically matches applicable derived feature definitions from the feature library and calculates for each component and process data. Specifically, the feature constructor can start from the atomic scale, utilizing intrinsic properties such as atomic radius, electronegativity, and valence electron concentration of each element, combined with the content proportions in the composition list, to calculate statistical quantities describing the overall characteristics of the alloy. The feature constructor can also estimate the trend of enthalpy change in composition based on thermodynamic theory, using the enthalpy of mixing parameters of each element. For process parameters, the feature constructor supports generating nonlinear derived features such as the reciprocal of temperature and the logarithmic transformation of cooling rate, to better capture physical laws related to diffusion or phase transformation kinetics.

[0068] The calculation of all derived features relies on an embedded atomic property database, which stores basic data such as atomic radius, electronegativity, valence electron concentration, and mixing enthalpy coefficient for common metallic elements. When encountering elements not included in the database, the system allows users to manually enter them or automatically retrieve them from an external materials database.

[0069] When the number of original or derived features is excessive and highly correlated, the feature dimensionality reducer in the unit is responsible for reducing feature dimensionality while preserving the main information. The dimensionality reducer supports two modes: linear and nonlinear dimensionality reduction. In linear dimensionality reduction mode, the reducer uses principal component analysis (PCA) to linearly transform the feature matrix, projecting the original features onto several orthogonal directions with the largest variance. The system automatically calculates the variance contribution rate of each principal component and selects the top few principal components whose cumulative contribution rate reaches a preset threshold as new features; this threshold defaults to 95%. In nonlinear dimensionality reduction mode, the reducer uses an autoencoder neural network method, training a neural network with a narrow bottleneck layer to learn a nonlinear compressed representation of the features. The autoencoder consists of an encoding layer and a decoding layer, and the training objective is to minimize the reconstruction error. After training, the output of the bottleneck layer is extracted as the dimensionality-reduced feature vector. This method is suitable for scenarios where there are complex interactions between features.

[0070] The feature filter offers two filtering algorithms. The first is correlation-based filtering, which calculates the Pearson correlation coefficient between each feature and the target performance variable, and selects features whose absolute values ​​are higher than a preset threshold. The second is model-based feature importance evaluation, which uses the random forest algorithm to train an initial model on the training data, then outputs the splitting contribution of each feature in the model, and selects the features with the highest importance ranking.

[0071] Users can set an upper limit on the number of features to be retained. The filter will then select features based on relevance or importance. After filtering, the feature engineering unit outputs a concise, highly discriminative feature matrix, which, together with the target performance variables, constitutes the model training dataset.

[0072] The model training and management unit first invokes the dataset partitioner to randomly divide the input model training dataset into three subsets according to a preset ratio: training set, validation set, and test set. The system's default partition ratio is 70% training set, 15% validation set, and 15% test set. The partitioner employs a stratified sampling strategy to ensure that the distribution of the target performance variables in each subset remains consistent with the original dataset, avoiding data distribution shifts caused by random sampling. After partitioning, the three subsets are persistently stored, and their partition index records are saved for subsequent reproduction.

[0073] The unit provides an algorithm library that encapsulates various machine learning and deep learning algorithms, including Support Vector Machines, Random Forests, Extreme Gradient Boosting, and Deep Neural Networks based on self-attention mechanisms. Users can specify the algorithm type through a visual interface or scripts, or enable an algorithm recommender—this recommender automatically suggests the most suitable algorithm and initial hyperparameter combination from the library based on the dataset's sample size, feature dimensions, target variable type (regression or classification), and performance requirements.

[0074] After selecting an algorithm, the unit initializes the model instance, using the algorithm's preset default hyperparameter values ​​as the starting point. Hyperparameters include learning rate, tree depth, regularization coefficient, number of neural network layers, and number of neurons.

[0075] To achieve optimal model performance, the hyperparameter optimizer searches the hyperparameter space on the validation set. The optimizer supports both grid search and Bayesian optimization strategies.

[0076] In grid search mode, the user specifies a set of candidate values ​​for each hyperparameter to be optimized. The optimizer exhaustively searches all candidate combinations, trains the model separately, calculates performance metrics on the validation set, and finally selects the hyperparameter combination with the best performance. This mode is suitable for situations with a small number of hyperparameters.

[0077] In Bayesian optimization, the optimizer uses probabilistic surrogate models (such as Gaussian processes) to model the relationship between hyperparameters and validation set performance. By balancing exploration and utilization through a sampling function, iteratively selecting the most promising hyperparameter combinations for evaluation. This approach can find near-optimal hyperparameter combinations with fewer evaluation iterations and is suitable for high-dimensional hyperparameter spaces.

[0078] Performance evaluation metrics are automatically selected based on the task type: for regression tasks, root mean square error or mean absolute error is used by default; for classification tasks, accuracy or area under the receiver operating characteristic (AUC) curve is used by default. After optimization, the optimal hyperparameter combination is recorded.

[0079] After determining the optimal hyperparameters, the model trainer uses the complete training set (excluding the validation set) to perform final training on the model. During training, the trainer continuously monitors the change in the loss function in each iteration. For deep learning models, the trainer employs early stopping: if the performance on the validation set fails to improve for several consecutive iterations, training is automatically terminated to prevent overfitting, and the model parameters are rolled back to the optimal parameters from the previous iteration. After training is complete, the model trainer outputs a serialized model object, which contains the model's network structure, weight parameters, hyperparameters used in training, and training process logs.

[0080] The model evaluator independently evaluates the trained model objects using the previously partitioned test set. The evaluator calculates multiple performance metrics, including but not limited to the coefficient of determination, root mean square error, mean absolute percentage error (for regression tasks), or accuracy, precision, recall, and F1 score (for classification tasks). The evaluation results, along with information such as the model object, hyperparameter combinations, training timestamps, and training and test set version identifiers, are packaged together to form a model metadata record.

[0081] Subsequently, the unit invokes the model repository manager to store the model's metadata in a centralized model library. The model library supports multi-version management, each model has a unique identifier, and can be retrieved by tags such as material system, target performance type, and algorithm type. The model repository manager is also responsible for invoking the security and access control module to perform integrity verification and encrypted storage of model files.

[0082] The model deployment and inference unit is responsible for transforming the model objects produced by the model training and management unit into practical applications that can provide prediction services. It receives inputs of new material composition and process parameters and returns performance prediction results.

[0083] The model loader within the unit retrieves the corresponding model object and its complete metadata record from the model repository based on the user-specified model identifier. After verifying the model's integrity, the loader loads it into memory. Subsequently, the service wrapper packages the model and its dependent preprocessing and feature construction logic into an independent prediction unit. This wrapper supports two output modes: online service mode and offline batch processing mode. In online service mode, the wrapper generates a lightweight network service container containing application programming interface endpoint definitions; in offline batch processing mode, the wrapper generates an independently executable command-line tool or script that supports reading input from a file and outputting results.

[0084] In online service mode, the unit exposes a standardized prediction interface that accepts Hypertext Transfer Protocol (HTTP) requests. The request body uses the platform's unified portable data exchange format. The interface requires input fields to include a list of material composition (elements and their amounts) and a list of process conditions (temperature, pressure, cooling rate, etc.). Internally, the interface includes an input validator that verifies the completeness of the input data and whether the numerical range is within a reasonable range seen during model training. If the values ​​exceed the range, a warning is issued, but prediction is still allowed.

[0085] Upon receiving a prediction request, the unit invokes the inference preprocessing engine. This engine first loads the preprocessing transformation parameters associated with the model, including the minimum, maximum, mean, and standard deviation used for normalization / standardization. Then, following the exact same order and algorithm as the data preprocessing unit, the engine performs anomaly labeling (but not removal), missing value imputation (using the imputation strategy used during training), and numerical normalization or standardization on the input data. This step ensures that the prediction input and training data are in the same numerical distribution space.

[0086] After basic preprocessing, the feature construction engine loads the feature engineering parameters associated with the model, including the formula definitions of derived features, dimensionality reduction matrices or autoencoder parameters, and a filtered feature list. Based on these definitions, the engine automatically constructs domain features from the preprocessed input data and applies the same dimensionality reduction transformation and feature filtering to generate feature vectors with dimensions completely consistent with the model input.

[0087] The model inference engine inputs the constructed feature vectors into the loaded model object, performs forward propagation calculations, and outputs the prediction results. For models that support uncertainty estimation, the inference engine also calculates the confidence interval or prediction variance of the predicted values. The inference engine simultaneously records the timestamp and input summary of this inference operation for subsequent auditing.

[0088] The result formatter encapsulates the numerical results output by the model, along with optional confidence intervals, into a standardized response structure. This structure includes the predicted value, upper and lower bounds of the confidence interval, the model identifier used for the prediction, and the model version. The response is returned to the caller via an application programming interface (API). For batch prediction requests, the inference unit uses an asynchronous task queue: the caller receives a task identifier immediately after submitting the batch input file, the system processes each sample sequentially in the backend, stores the results in a downloadable result file, and notifies the caller.

[0089] The unit includes a built-in model version controller for managing the switching and routing between different versions of the same model. The controller supports blue-green deployment and canary release strategies: a small portion of forecast traffic can be redirected to the new model version, and after verification, all traffic can be gradually switched over. Simultaneously, the controller periodically checks the health status of deployed models (response time, error rate) and automatically rolls back unhealthy model instances.

[0090] The closed-loop collaborative module includes: a prediction result capture unit, a data parsing and verification unit, a task parameter mapping rule base, a computation task generation unit, and a collaborative fine-tuning and iterative control unit.

[0091] The prediction result capture unit is responsible for continuously listening to and receiving model prediction outputs from the materials artificial intelligence module, using them as trigger signals for closed-loop iteration. This unit runs as a persistent service, internally maintaining a message subscriber. The subscriber establishes a persistent connection with the platform's asynchronous message bus, subscribing to the prediction result channel output by the materials artificial intelligence module. When the artificial intelligence module completes batch predictions for candidate material schemes and outputs results, the prediction results are encapsulated into a standardized message structure and published to this topic. The message subscriber captures these messages in real time and performs preliminary format validation to ensure that the messages contain necessary information fields. For messages that do not conform to the format specifications, the subscriber moves them to an error queue and records an alarm log, preventing them from entering the subsequent processing flow.

[0092] The data parsing and verification unit is responsible for deep parsing, content extraction, and rationality verification of the captured prediction result messages, transforming them into structured objects that can be processed within the module. The unit contains a message parsing engine. This engine extracts key information from the original message field by field according to a predefined message protocol, including but not limited to: the compositional boundary description of the candidate material, the target performance index for the prediction, the predicted performance value and confidence level, and the metadata of the AI ​​model that generated the prediction. The parsing engine also extracts the iterative context information accompanying the message.

[0093] After parsing, the data parsing and verification unit calls the result rationality validator to verify the extracted information. The validator performs multi-level checks: First, it checks whether the component boundaries are within the range of chemical elements supported by the platform, and whether the content of each element is within a reasonable physical range; second, it checks whether the predicted value of the target performance exceeds the preset threshold of the material's physical limit, and marks results that exceed the reasonable range; third, it checks whether the confidence level is lower than the system's set confidence lower limit, and if the confidence level is too low, the prediction result is determined to be unreliable. For results that pass the rationality verification, the validator marks them as "valid candidates"; for those that fail, they are marked as "questionable candidates" and the reason is recorded. At the same time, it decides whether to discard them or submit them for manual review based on the system configuration. Finally, the valid candidate information is encapsulated into a standardized candidate scheme object and passed to the task parameter mapping rule base.

[0094] The task parameter mapping rule base is internally organized into three sub-rule sets according to functional dimensions: component discretization rule set, process parameter determination rule set, and calculation property selection rule set.

[0095] The component discretization rule set defines methods for converting continuous component intervals into discrete component points. The rule set predefines various discretization strategies, such as uniform grid discretization, random sampling discretization, and keypoint sampling based on materials science knowledge. Each rule also includes constraint parameters such as upper limits on step size and the number of discrete points to prevent generating excessive computational tasks that exceed the system's capacity.

[0096] The process parameter determination rule set defines which process conditions need to be simulated for a given candidate component, as well as the range of values ​​for these conditions. Rules can automatically associate process parameter domains based on the type of target performance: for example, for performance prediction related to temperature sensitivity, the rule set automatically generates the temperature scan range and step size. The rule set also supports inheritance and overriding mechanisms, allowing users to modify or add rules based on default rules.

[0097] The calculation property selection rule set defines which specific material properties should be calculated for different candidate scenarios and prediction objectives. The rule set pre-defines multiple sets of mapping relationships: for example, when the prediction objective is mechanical properties, the rule set suggests calculating phase composition, Gibbs free energy, and diffusion coefficient. Each rule also includes calculation priorities to prioritize simulations of critical properties when computational resources are limited.

[0098] All rules in the rule base are stored in a declarative language and provide a graphical editing interface, allowing materials experts to adjust the mapping logic without programming. The rule base also supports version management, recording the change history for each modification, facilitating retrospective analysis.

[0099] The computation task generation unit is responsible for calling the task parameter mapping rule base to convert candidate solution objects into a series of specific computation tasks that can be directly executed by the high-throughput computing module, and submitting them in a standard format. The unit first receives candidate solution objects from the data parsing unit. Then, it sequentially accesses the three sub-rule sets in the rule base and performs the following transformation process: First, the component discretization rule set is invoked. Based on the component boundary descriptions carried in the candidate schemes, the cell matches the applicable discretization strategy and generates a list of discrete component points. Each discrete component point contains the precise content value of each element. During the discretization process, the cell performs deduplication to ensure that adjacent grid points do not overlap due to computational accuracy issues.

[0100] Second, the rule set is determined by calling the process parameters. For each component point in the list of discrete component points, the unit generates a corresponding list of process condition combinations based on the target performance type and the preset process parameter template. This list may contain multiple different sets of process conditions, forming a two-dimensional parameter combination space.

[0101] Third, the calculation property selection rule set is invoked. The unit determines the list of material properties to be simulated based on the prediction target and the priority of candidate schemes, and sorts them according to priority.

[0102] After completing the above three steps, the computational task generation unit obtains a complete three-dimensional task space: component point dimension, process condition dimension, and computational property dimension. The unit performs a Cartesian product expansion on this space to generate each atomic computational task. Each task explicitly records the component points, process conditions, and the properties to be computed. Each atomic task is assigned a globally unique task identifier and includes metadata inherited from the candidate solutions, such as the source model identifier, iteration round number, and prediction confidence, for subsequent traceability.

[0103] The generated list of atomic tasks is then sent to the task submitter. The task submitter encapsulates the task list in batches into request bodies conforming to the high-throughput computing module's application programming interface (API) specification, and submits them to the high-throughput computing module's task queue via a secure channel. The submitter employs an exponential backoff retry strategy, automatically retrying in the event of network jitter or temporary module unavailability to ensure no tasks are lost. Upon successful submission, the task submitter associates the returned task identifier with its local task record and updates the task status to "submitted."

[0104] The Cooperative Fine-tuning and Iterative Control Unit is responsible for managing the entire closed-loop iteration lifecycle, including monitoring the execution progress of computational tasks, triggering the backflow of representation data, coordinating the incremental learning of the artificial intelligence module (cooperative fine-tuning), and determining the iteration termination conditions.

[0105] The unit has a built-in task status monitor that periodically retrieves the current status of the submitted task set through the query interface of the high-throughput computing module. The monitor maintains a local task status table, recording the status transition history of each task. When it detects that a batch of tasks has been completed, the status monitor sends a completion event to the data feedback trigger.

[0106] Upon receiving the completion event, the data backflow trigger subscribes to newly added data events from the materials simulation database module. Since the results from the high-throughput calculation module automatically trigger a "new data arrived" message in the database after being added, the backflow trigger waits for this message and confirms that the task identifier corresponding to the received data matches the task identifier submitted in this closed-loop iteration. After confirmation, the trigger encapsulates the storage location information of this batch of new data into a data ready notification and sends it to the collaborative learning coordinator.

[0107] After receiving the data-ready notification, the collaborative learning coordinator sends an incremental model update request to the materials AI module. This request includes the following information: the storage location of the new data, the iteration number of the current iteration, and the identifier of the original model. Upon receiving the request, the AI ​​module performs the following operations: loads the new data from the database, merges it with the existing training data, and then performs incremental training or fine-tuning on the existing model. During fine-tuning, the collaborative learning coordinator can configure fine-tuning strategies, such as adjusting only the parameters of the last few layers of the model (for deep neural networks) or partially retraining the random forest model using the new data. After fine-tuning is complete, the AI ​​module returns the identifier of the new model and a summary of the performance evaluation.

[0108] The iteration termination judge in the unit is responsible for evaluating whether to continue to the next round of closed-loop iteration. The judge evaluates according to the user-preset termination conditions, which may include: the target performance prediction value has reached a set threshold, the performance improvement of the best candidate solution in two consecutive iterations is less than the set minimum improvement threshold, or the maximum number of iterations executed has been reached. If any termination condition is met, the iteration termination judge reports the iteration completion to the platform and outputs the final optimized material design scheme list. If the condition is not met, the judge sends a continue iteration signal to the prediction result capture unit to start a new round of "prediction → transformation → calculation → fine-tuning" loop.

[0109] Throughout the closed-loop iteration process, the collaborative fine-tuning and iterative control unit incorporates an exception handler. The exception handler will interrupt the current iteration when the following occurs: widespread computational task failures (failure rate exceeds system threshold), data reflux timeout (new data not entered into the database within a specified time), or a significant performance degradation after fine-tuning the AI ​​model. The exception handler records the complete fault context, generates an alarm notification, and switches the closed-loop status to "paused awaiting manual intervention." Users can view fault details through the platform interface and manually resume or terminate the iteration after resolving the issue.

[0110] The security and access control module has a built-in hash calculation engine that encapsulates a complete implementation of the national cryptographic hash algorithm. Regardless of the data type (structured data records, model files, log entries, etc.), the hash value calculation follows a unified process.

[0111] When the hash value of a data block needs to be calculated, the hash calculation engine submits the data block to be calculated and its metadata description. The hash calculation engine first performs normalization processing on the data block: for binary data (such as model files), it directly reads its raw byte stream; for structured data (such as database records), the engine serializes the values ​​of each field into a normalized byte sequence according to a predefined field order, ensuring that data with the same content always produces the same serialization result, avoiding inconsistencies in hash values ​​due to minor adjustments in field order or format.

[0112] The engine then inputs the normalized byte sequence into the national cryptographic hash algorithm function. This algorithm compresses input data of arbitrary length into a fixed-length hash value output. After calculation, the engine outputs the hash value and associates it with the identification information of the data block for subsequent verification. The entire calculation process does not rely on any external key, ensuring that the hash value itself is uniquely determined solely by the data content.

[0113] Before data is written to the material simulation database module or file system, the security and access control module performs a pre-storage verification process.

[0114] First, the data integrity preprocessor intercepts all write requests. For the data object to be stored, the preprocessor extracts its complete byte representation or normalized serialization form and calls the hash calculation engine to calculate the hash value of the data. This hash value, serving as the data's "digital fingerprint," is not stored encrypted but is appended to the data object's metadata area as a plaintext field. Specifically, a fixed "integrity checksum" field is added to the database record structure; in the file system, a metadata file or extended attribute with the same name is generated for each file to store the hash value.

[0115] After the hash value is appended, the data and its hash value are written to persistent storage. At this point, a one-to-one binding relationship is established between the data and its integrity baseline. This design ensures that any subsequent modifications to the data content will result in a discrepancy between its current hash value and the stored baseline, thus making it detectable.

[0116] When modules within the platform exchange data via application programming interfaces (APIs), the security and access control module implements end-to-end transmission integrity protection.

[0117] At the sending end, when constructing a data packet, the transport security encapsulator, in addition to including the original data content, calls a hash calculation engine to calculate the hash value of the data and appends this hash value to the header or extended fields of the data packet. The encapsulator then optionally encrypts the data packet using a national standard symmetric encryption algorithm before sending it over the network.

[0118] At the receiving end, after receiving the data packet, the secure decryptor first decrypts it, then extracts the original data content and the hash value provided by the sender. The decryptor independently calls the hash calculation engine to recalculate the hash value of the received original data content. Subsequently, the integrity comparator compares the recalculated hash value bit by bit with the hash value appended by the sender.

[0119] If the two are completely consistent, it proves that the data has not been tampered with or damaged during transmission. The desealer submits the original data content to the upper-layer business processing unit and returns an acknowledgment response to the sender. If the two are inconsistent, the desealer determines that the data integrity has been compromised, refuses to receive the data packet, requests a retransmission from the sender, and records a security event log, including timestamps, sender and receiver identifiers, data packet digests, and other information, for security auditing purposes.

[0120] When a user or module reads data from a database or file system, the security and access control module performs read-time verification to ensure that the returned data is consistent with the original state at the time of storage.

[0121] The read verification interceptor intervenes in every data query or file read operation. After the data is retrieved from the storage medium, the interceptor first extracts the original hash base value (i.e., the value calculated and saved before storage) attached to the data object. Simultaneously, the interceptor recalculates the hash value of the currently read data content. Then, it calls the dynamic comparison service to compare the newly calculated hash value with the stored base value.

[0122] Based on the comparison results, the dynamic comparison service triggers different follow-up actions: If the hash values ​​match, it proves that the data has not been modified since it was stored, and the data is complete and trustworthy. The interceptor returns the original data content normally to the requester. If the hash values ​​do not match, it determines that the data may have been tampered with, corrupted, or the storage medium may have failed. The interceptor immediately blocks the return of the data, returns an error message "Data integrity verification failed" to the requester, marks the event as a serious security alert, records detailed fault information, and triggers a system administrator notification.

[0123] Furthermore, the read verification interceptor also performs the above process for data imported from external data fusion units. Since the external data has already been standardized and a hash base value has been generated during import, any subsequent read operation will verify its integrity, thereby incorporating the external data into a unified security trust system.

[0124] To ensure the immutability and traceability of operation logs, the security and access control module extends the hash verification mechanism to the log system and employs a chained cumulative hash protection strategy. All critical operations of the platform are captured by a security logger and written to the log storage system. Before each log entry is written, the logger performs the following steps: Extract the content of the log entry and serialize it into a normalized byte sequence. Call the hash calculation engine to calculate the hash value of this log entry. Read the hash value of the previous log entry, concatenate the hash value of this log entry with the hash value of the previous log entry, and then calculate the hash value of the concatenated result again to obtain the chained cumulative hash value.

[0125] The chained cumulative hash value is stored along with the individual hash value of this log entry. When log integrity needs to be verified, the verification program starts from the first log entry, recalculates the chained cumulative hash value for each entry, and compares it with the stored cumulative value. Any modification to a historical log will destroy the chained cumulative hash value of all subsequent log entries, thus being detected quickly. This design makes it impossible for attackers to tamper with any historical log entry undetected.

[0126] The security and access control module also includes a periodic audit scheduler for proactive, systematic integrity scans of the data stored across the entire platform.

[0127] The audit scheduler initiates a global integrity audit task according to the system's configured time period. During task execution, the audit scheduler traverses all data records in the database and all critical files in the file system. For each data object, the audit agent performs the following operations: reads the current data content and its accompanying original hash base value from the storage medium; recalculates the hash value of the current data content; and compares the recalculated hash value with the original base value. For the log system, the audit agent performs a record-by-recalculation and comparison of chained cumulative hashes.

[0128] All verification results are aggregated into a comprehensive audit report, which details the number of data points that passed verification, the list of data points that failed verification, and details of any anomalies. If any mismatched data objects are found, the audit scheduler immediately triggers an emergency alarm and highlights the affected data on the platform's management interface, while also recommending that the administrator restore from backup or conduct a manual investigation.

[0129] In addition, the results of regular audits can be used to demonstrate compliance, showcasing the integrity of platform data and the effectiveness of security controls to internal auditors or external regulators.

[0130] The second embodiment of this application is as follows: Please see Figure 2 This invention provides a three-in-one platform data R&D closed-loop collaborative method, applied to a three-in-one platform data R&D closed-loop collaborative system as provided in the first embodiment, including the following steps: S1. The first round of simulation calculation is performed using the high-throughput computing module based on the initial material design space, which includes the material composition space and the process parameter space. S2. The calculation results are automatically pushed to the material simulation database module, where they are parsed, tagged, and stored according to the preset unified structured data standard. S3. Extract data from the material simulation database module through the material artificial intelligence module, and perform data preprocessing, feature engineering and model training in sequence to obtain a material performance prediction model; S4. The performance prediction model is used to predict the extended design space and filter candidate material schemes; wherein, the extended design space is a preset extended search space whose range is larger than or different from the initial material design space; S5. The candidate material scheme is captured by the closed-loop collaborative module. According to the component discretization rules, process parameter determination rules and calculation property selection rules in the task parameter mapping rule library, the candidate scheme is automatically converted into new high-throughput computing task parameters, an atomic computing task list is generated and submitted to the high-throughput computing module for verification calculation. S6. After the new calculation is completed, the result is automatically stored in the database and a data return event is triggered. The materials artificial intelligence module will incorporate the new data into the training set and perform incremental fine-tuning on the current model until the preset iteration termination condition is met, and output the final materials design scheme.

[0131] Regarding the methods in the above embodiments, the specific manner in which each step is performed has been described in detail in the embodiments of the system, and will not be elaborated here.

[0132] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0133] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A three-in-one platform data R&D closed-loop collaborative system, characterized in that, include: The high-throughput computing module is used to perform high-throughput automatic simulation calculations on various properties of material systems and output the calculation results in a unified structured format. The materials simulation database module is pre-set with a unified structured data standard, which is used to automatically receive and store the calculation results output by the high-throughput computing module, automatically label the stored data according to the structured data standard, and provide a standardized data query interface. The materials artificial intelligence module is used to acquire data from the materials simulation database module, perform data preprocessing, feature engineering, model training and deployment, and obtain a materials performance prediction model. The closed-loop collaborative module is used to capture the prediction results of the material artificial intelligence module, automatically convert the prediction results into new calculation task parameters according to the preset task parameter mapping rules, and submit them to the high-throughput computing module for verification calculation.

2. The three-in-one platform data R&D closed-loop collaborative system as described in claim 1, characterized in that, The high-throughput computing module includes: The task configuration unit is used to receive the settings of material system, composition space, process parameter space and target calculation properties, and generate a calculation task definition object; The task generation and scheduling unit contains a discretization engine and a computation scheduler. The discretization engine is used to discretize the continuous component space and process parameter space in the computation task definition object into multiple independent computation tasks. The computation scheduler is used to dynamically allocate tasks according to the computing node resources. The computation execution unit is used to receive independent computation tasks, call the phase diagram computation engine to perform simulations, monitor task status, and implement a limited automatic retry strategy. The results collection and standardization unit is used to receive the raw result files returned by each computing node, convert the raw results into a unified structured data format of the platform through the format conversion engine, and push them to the materials simulation database module after attaching computing metadata.

3. The three-in-one platform data R&D closed-loop collaborative system as described in claim 1, characterized in that, The material simulation database module includes: The data standard setting unit is used to define structured data standard templates that include material system identifiers, composition lists, process condition sets, calculation result sets, and metadata blocks, and publish them in the form of machine-readable global data contracts. The data receiving and parsing unit is used to monitor data push, receive structured data packets, extract key fields and perform data type validation, and output standardized data objects. An automated tag generation unit maintains a tag generation rule base and is used to extract or derive tag metadata from standardized data objects according to the rule base to form a tag set in the form of key-value pairs. The storage and indexing unit uses a distributed document storage system to merge standardized data objects and tag metadata for storage, and builds an inverted index for tags and a range index for numeric fields. The external data fusion unit includes a data source adapter manager and multiple adapter plugins, which are used to connect to external public material databases. It maps external data fields to internal data standards through a schema mapping table and stores them in the storage and indexing unit after deduplication.

4. The three-in-one platform data R&D closed-loop collaborative system as described in claim 1, characterized in that, The material artificial intelligence module includes: The data preprocessing unit is used to obtain the raw dataset from the material simulation database module, and sequentially perform outlier detection and removal, duplicate record removal, numerical normalization or standardization, missing value handling, and save the preprocessing transformation parameters. The feature engineering unit includes a materials science knowledge feature library, a feature dimensionality reducer, and a feature filter. The materials science knowledge feature library is used to construct derived features from composition and process parameters based on atomic property data. The feature dimensionality reducer is used to reduce the feature dimension through principal component analysis or autoencoder. The feature filter is used to filter feature subsets through correlation analysis or model-based feature importance evaluation. The model training and management unit is used to divide the dataset into training, validation and test sets, perform hyperparameter optimization, train the final model, and store the model and its metadata in the model library. The model deployment and inference unit is used to load models from the model library, package them as online services or offline tools, perform the same preprocessing and feature construction on the input data as during training, call the model for inference, and output performance prediction results.

5. The three-in-one platform data R&D closed-loop collaborative system as described in claim 1, characterized in that, The closed-loop collaborative module includes: The prediction result capture unit is used to continuously monitor the prediction result messages output by the materials artificial intelligence module; The data parsing and verification unit is used to extract the compositional boundaries, target performance, and confidence level of candidate materials from the message, and to verify their physical rationality and credibility, thereby generating candidate scheme objects. The task parameter mapping rule library is used to define the mapping logic for converting candidate solution objects into specific computation task parameters; The computation task generation unit is used to sequentially call the rule set in the rule base, discretize the candidate scheme object into a list of component points, associate each component point with a combination of process conditions and target computation properties, generate an atomic computation task list, and submit it to the high-throughput computing module in batches through the task submitter. The collaborative fine-tuning and iterative control unit is used to monitor the completion status of computing tasks, trigger new data entry events, initiate incremental model update requests to the materials artificial intelligence module, and determine whether to continue iterating based on preset termination conditions.

6. The three-in-one platform data R&D closed-loop collaborative system as described in claim 5, characterized in that, When performing incremental model updates, the collaborative fine-tuning and iterative control unit sends a request to the materials artificial intelligence module containing the new data storage location, the original model identifier, and the iteration round number according to a preset fine-tuning strategy. The materials artificial intelligence module merges the new data with the original training data and performs incremental training or partial parameter fine-tuning on the existing model, and returns the new model identifier.

7. The three-in-one platform data R&D closed-loop collaborative system as described in claim 1, characterized in that, It also includes a security and access control module, which uses the national cryptographic hash algorithm to calculate the hash value of the data and performs the following integrity verification process: Before data is written to storage, the hash value of the data is calculated and attached to the storage as an integrity benchmark value. When data is transmitted between modules via the application programming interface, the sender attaches a hash value to the data, the receiver recalculates the hash value and compares it with the value provided by the sender, and rejects the data if they do not match. When reading data from storage, the hash value of the data is recalculated and compared with the baseline value in storage. If they are inconsistent, the data is prevented from being returned and an alarm is triggered. The operation logs are protected by chained cumulative hashing, where each log entry stores its independent hash value and the cumulative hash value obtained by concatenating the hash value of the previous log entry. Periodically schedule global integrity audit tasks, traverse the stored data, recalculate the hash value and compare it with the baseline value, and generate an audit report.

8. The three-in-one platform data R&D closed-loop collaborative system as described in claim 1, characterized in that, The high-throughput calculation module supports high-throughput calculations of at least one of the following properties of materials: thermodynamic phase equilibrium, phase diagram, Gibbs free energy, kinetic diffusion coefficient, thermal conductivity, thermal expansion coefficient, dendrite growth rate, and temperature range of the mushy region.

9. The three-in-one platform data R&D closed-loop collaborative system as described in claim 1, characterized in that, In the closed-loop collaborative module, the component discretization rule set in the task parameter mapping rule base includes a uniform grid discretization strategy, a random sampling discretization strategy, and a key point sampling strategy based on materials science knowledge. The process parameter determination rule set automatically associates temperature scanning range, cooling rate, and casting temperature parameters according to the target performance type. The calculation property selection rule set pre-sets multiple sets of property mapping relationships and calculation priorities according to the prediction target.

10. A three-in-one platform data R&D closed-loop collaborative method, applied to the three-in-one platform data R&D closed-loop collaborative system as described in claim 1, characterized in that, Includes the following steps: The first round of simulation calculations is performed using a high-throughput computing module based on an initial material design space, which includes a material composition space and a process parameter space. The calculation results are automatically pushed to the material simulation database module, where they are parsed, tagged, and stored according to the preset unified structured data standard. The material artificial intelligence module extracts data from the material simulation database module, and performs data preprocessing, feature engineering, and model training in sequence to obtain a material performance prediction model. The performance prediction model is used to predict the extended design space and select candidate material schemes. The candidate material schemes are captured by the closed-loop collaborative module. According to the component discretization rules, process parameter determination rules and calculation property selection rules in the task parameter mapping rule library, the candidate schemes are automatically converted into new high-throughput computing task parameters, generating an atomic computing task list and submitting it to the high-throughput computing module for verification calculation. Once the new calculation is completed, the results are automatically stored and a data reflux event is triggered. The materials artificial intelligence module incorporates the new data into the training set and performs incremental fine-tuning on the current model until the preset iteration termination condition is met, and outputs the final materials design scheme.

Images