A Material Mixed Variable Optimization Selection Method and Apparatus Based on Heterogeneous Graphical Neural Networks

Abstract

This invention provides a method and apparatus for material selection optimization based on heterogeneous graph neural networks, relating to the field of computer technology. It enables low-cost and high-reliability material selection for entities through hybrid optimization and global optimization of discrete material selection variables and continuous proportioning variables. The method includes: acquiring material configuration schemes for an entity and test results of performance indicators for each scheme; each material configuration scheme includes: an identifier for at least one selected material and proportioning data for each selected material; training a material scheme performance evaluation model using the material configuration schemes and test results; the material scheme performance evaluation model includes a graph neural network and a performance prediction model, the graph neural network including multiple heterogeneous graph models; and using the trained material scheme performance evaluation model as the fitness function of a genetic algorithm model to determine a target material configuration scheme based on performance indicators according to the genetic algorithm model.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and in particular to a method and apparatus for material hybrid variable optimization selection based on heterogeneous graph neural networks. Background Technology

[0002] The selection of physical materials is a common problem in engineering practice. Taking optical connectors as an example, I / O optical connectors are key interconnect devices in high-speed data centers, high-performance computing clusters, and optical communication systems. Their terminals undertake multiple core functions, including photoelectric signal transmission, mechanical contact, and impedance matching. The chemical composition and physical properties of the terminal materials directly determine the signal integrity, mechanical reliability, and environmental adaptability of the connector under long-term high-frequency operating conditions. With the continuous evolution of optical module transmission rates and SerDes (serializer / deserializer) technology, the electrical performance stability, contact consistency, and long-term environmental tolerance of terminal materials under high-frequency conditions have become key technical bottlenecks restricting the overall performance and reliability of optical interconnect systems.

[0003] In practical engineering applications, I / O optical connector terminals often employ various copper-based alloy systems such as beryllium copper alloy and copper-nickel-silicon alloy. By controlling the types and proportions of alloying elements such as copper, nickel, cobalt, silicon, and beryllium, a comprehensive balance is achieved between conductivity, mechanical strength, hardness, wear resistance, and fatigue resistance. However, there is a significant synergistic effect between the selection of the material system and the specific element ratios. The mechanisms and degrees of influence of each element vary under different material systems, resulting in a highly nonlinear and complex mapping relationship of multi-parameter coupling between "material selection—element ratio—comprehensive performance." This multi-level and multi-factor interaction makes it difficult for traditional trial-and-error methods relying on experience to systematically grasp its inherent laws, easily leading to extended design cycles, increased costs, and limited performance improvements.

[0004] On the other hand, there is a clear hierarchy and coupling between material selection and element ratios, and their impact on the final performance of terminals often exhibits nonlinear and multi-objective conflict characteristics. The discrete selection of the material system and the continuous adjustment of the element ratios together constitute a hybrid design space, in which even a small change in any parameter can cause a significant change in performance, making it difficult to accurately predict and optimize terminal performance using traditional linear models or univariate analysis methods. This complex multi-layered relationship between "material type—component composition—process conditions—final performance" further increases the difficulty of intelligent design of terminal materials.

[0005] Currently, the selection and proportioning design of terminal materials in engineering mainly relies on expert experience, limited comparative experiments, or stepwise optimization methods based on a single performance index. These methods have significant shortcomings when facing high-dimensional hybrid design spaces: First, it is difficult to systematically establish a collaborative mapping model between material selection and multi-element proportioning, making it impossible to achieve performance comparison and global optimization across material systems; second, as design variables increase, the number of experimental combinations grows exponentially, leading to lengthy R&D cycles and high costs; third, the utilization rate of historical experimental data is low, lacking the ability to extract potential patterns from the data and making it difficult to form a closed-loop design mechanism for continuous self-optimization. In particular, there is a lack of hybrid optimization techniques capable of simultaneously handling discrete material selection variables and continuous element proportioning variables, which is precisely the core challenge in achieving intelligent design of terminal materials.

[0006] Therefore, there is an urgent need to develop an intelligent design method that can systematically construct the complex mapping relationship between material selection, material ratio and physical comprehensive performance under engineering preparation constraints, and can achieve hybrid variable collaborative optimization under multiple performance index requirements, so as to improve the comprehensive performance, design efficiency and engineering reliability of physical materials (such as I / O optical port connector terminals) and meet the urgent needs for physical material performance. Summary of the Invention

[0007] In view of this, embodiments of the present invention provide a material selection method and apparatus based on heterogeneous graph neural networks, which achieves low-cost and high-reliability material selection by hybrid optimization and global optimization of discrete material selection variables and continuous proportioning variables.

[0008] To achieve the above objectives, according to one aspect of the present invention, a material hybrid variable optimization selection method based on heterogeneous graph neural networks is provided.

[0009] The material hybrid variable optimization selection method based on heterogeneous graph neural networks in this invention includes: acquiring multiple material configuration schemes for an entity and test results of each material configuration scheme on at least one preset performance index; each material configuration scheme includes: the identifier of at least one selected material and the proportion data of each selected material; training a material scheme performance evaluation model using the material configuration scheme and the test results; the material scheme performance evaluation model includes a graph neural network and a performance prediction model, the graph neural network including multiple heterogeneous graph models; during the training process: each material configuration scheme is input into a heterogeneous graph model, the material identifier and proportion data in the material configuration scheme are used as nodes of the heterogeneous graph model, the correspondence between the material identifier and each proportion data, and the collaborative relationship of each proportion data are used as edges of the heterogeneous graph model; each heterogeneous graph model performs independent feature updates to obtain corresponding graph embedding vectors, the graph embedding vectors are input into the performance prediction model to obtain performance prediction results, the performance prediction results and the test results are used to update model parameters based on a preset loss function; the trained material scheme performance evaluation model is used as the fitness function of a genetic algorithm model, and a target material configuration scheme based on the performance index is determined according to the genetic algorithm model.

[0010] Optionally, a heterogeneous graph model includes a material system node and multiple material component nodes; and, after any material configuration scheme is input into a heterogeneous graph model, the material identifier in the material configuration scheme is used as the material system node of the heterogeneous graph model, the proportion data of each material in the material configuration scheme is used as the material component node of the heterogeneous graph model, the material system node and each material component node are connected to form the first type of edge of the heterogeneous graph model, and the different material component nodes are connected to form the second type of edge of the heterogeneous graph model.

[0011] Optionally, each heterogeneous graph model performs independent feature updates, including: any node in any heterogeneous graph model performs multi-level iterative feature updates within the heterogeneous graph model to obtain the final feature representation of the node output by the last layer in the multi-level; wherein, in two adjacent layers in the multi-level, the updated feature representation of the previous layer is used as the initial feature representation of the next layer; during the feature update process of any layer, the same-relationship contribution features of each neighboring node of any heterogeneous graph model based on the same type of edge are obtained, the same-relationship contribution features of each type of edge are aggregated into the cross-relationship contribution features of the node, and the cross-relationship contribution features are fused with the initial feature representation of the node in the layer to form the updated feature representation of the node in the layer.

[0012] Optionally, obtaining the same-relation contribution features of any node in any heterogeneous graph model based on the same type of edge among its neighboring nodes includes: for any node in any heterogeneous graph model based on the same type of edge among its neighboring nodes, determining the linear projection feature of the neighboring node on that type of edge; determining the attention weight of the neighboring node on the node based on the node, the neighboring node, and the type of edge; obtaining the contribution features of the neighboring node on the node based on the linear projection feature and the attention weight; and aggregating the contribution features of the node based on the same-relation contribution features of the neighboring nodes among its neighboring nodes based on that type of edge into the same-relation contribution features of the node based on the same-relation edge among its neighboring nodes.

[0013] Optionally, obtaining the corresponding graph embedding vector includes: performing pooling on the final feature representation of any material component node in any heterogeneous graph model; and fusing the pooled final feature representation with the final feature representation of the material system node in the heterogeneous graph model to obtain the graph embedding vector of the heterogeneous graph model.

[0014] Optionally, determining the target material configuration scheme based on the performance index according to the genetic algorithm model includes: selecting multiple material configuration schemes as individuals of the genetic algorithm model based on preset constraints; performing a selection operation on the individuals according to the fitness function; performing a crossover operation on the selected individuals to obtain new individuals; performing a mutation operation on the new individuals to obtain the next generation of individuals; iteratively performing the selection operation, the crossover operation, and the mutation operation until a preset termination condition is met; and determining the target material configuration scheme from the material configuration schemes that meet the termination condition.

[0015] Optionally, the material includes: elemental metals and / or elemental non-metals; and / or, the entity includes: optical connector terminals; and / or, the performance indicators include at least one of the following: mechanical strength, insulation resistance, withstand voltage, and environmental resistance; and / or, the performance prediction model includes: a multilayer perceptron model.

[0016] To achieve the above objectives, according to another aspect of the present invention, a material hybrid variable optimization and selection device based on heterogeneous graph neural networks is provided.

[0017] The material hybrid variable optimization and selection device based on heterogeneous graph neural networks according to this invention includes: a data preparation unit, used to acquire multiple material configuration schemes of an entity and test results of each material configuration scheme on at least one preset performance index, wherein each material configuration scheme includes: an identifier of at least one selected material and the proportion data of each selected material; a model training unit, used to train a material scheme performance evaluation model using the material configuration schemes and the test results, wherein the material scheme performance evaluation model includes a graph neural network and a performance prediction model, and the graph neural network includes multiple heterogeneous graph models; during the training process: each material configuration scheme inputs a heterogeneous graph. The model uses material identifiers and proportion data in the material configuration scheme as nodes of the heterogeneous graph model, and the correspondence between material identifiers and proportion data, as well as the synergistic relationship between proportion data, as edges of the heterogeneous graph model. Each heterogeneous graph model performs independent feature updates to obtain corresponding graph embedding vectors. The graph embedding vectors are input into the performance prediction model to obtain performance prediction results. The performance prediction results and the test results are used to update model parameters based on a preset loss function. The computation unit is used to use the trained material scheme performance evaluation model as the fitness function of the genetic algorithm model, and to determine the target material configuration scheme based on the performance index according to the genetic algorithm model.

[0018] To achieve the above objectives, according to another aspect of the present invention, an electronic device is provided.

[0019] An electronic device according to the present invention includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to execute the material hybrid variable optimization selection method based on heterogeneous graph neural networks provided by the present invention.

[0020] To achieve the above objectives, according to another aspect of the present invention, a non-transitory computer-readable storage medium is provided.

[0021] The present invention provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the material hybrid variable optimization selection method based on heterogeneous graph neural networks provided by the present invention.

[0022] According to the technical solution of the present invention, one embodiment of the above invention has the following advantages or beneficial effects: After obtaining multiple material configuration schemes for an entity and the test results of each material configuration scheme on at least one performance index, a material scheme performance evaluation model is trained using the material configuration schemes and test results. The material scheme performance evaluation model includes a graph neural network and a performance prediction model. The graph neural network includes multiple heterogeneous graph models. During training: any material configuration scheme is input into a heterogeneous graph model. The material identifiers and proportion data in the material configuration scheme are used as nodes of the heterogeneous graph model. The correspondence between the material identifiers and each proportion data, as well as the collaborative relationship between each proportion data, are used as edges of the heterogeneous graph model. Each heterogeneous graph model performs independent feature updates to obtain corresponding graph embedding vectors. The graph embedding vectors are input into the performance prediction model to obtain performance prediction results. The performance prediction results and test results are used to update the model parameters based on a preset loss function. Subsequently, the trained material scheme performance evaluation model is used as the fitness function of a genetic algorithm model. The target material configuration scheme based on the performance index is determined according to the genetic algorithm model. Thus, by establishing a heterogeneous graph model and graph neural network adapted to material configuration schemes, along with a node feature update mechanism for the graph neural network, and simultaneously handling the hybrid optimization of discrete material selection variables (selected material identifiers) and continuous proportion variables (the proportions of each material), the problem of difficulty in performance mapping modeling in existing technologies is solved. This achieves independent feature extraction schemes based on various heterogeneous graph models, generating strong robustness and generalization features that reflect the relationships between nodes and the global structure of the graph, contributing to the formation of high-performance, high-accuracy material scheme performance evaluation models. Furthermore, this embodiment of the invention combines graph neural networks with genetic algorithm models, using the above material scheme performance evaluation model as the fitness function in the genetic algorithm model to perform global optimization of material configuration schemes, achieving a balance between the comprehensive performance of physical materials, design efficiency, and engineering reliability.

[0023] The further effects of the aforementioned unconventional alternative methods will be explained below in conjunction with specific implementation methods. Attached Figure Description

[0024] The accompanying drawings are provided to better understand the invention and are not intended to unduly limit the scope of the invention. Wherein: Figure 1 This is a schematic diagram of the main steps of the material hybrid variable optimization and selection method based on heterogeneous graph neural network in an embodiment of the present invention; Figure 2 This is a schematic diagram of the heterogeneous graph neural network-enhanced genetic optimization method for selecting I / O optical port connector terminal materials according to an embodiment of the present invention; Figure 3a This is a schematic diagram of the first physical structure of the high-speed I / O connector in an embodiment of the present invention; Figure 3b This is a schematic diagram of the second physical structure of the high-speed I / O connector in an embodiment of the present invention; Figure 3c This is a schematic diagram of the third physical structure of the high-speed I / O connector in an embodiment of the present invention; Figure 4 This is a schematic diagram of the convergence curve of the genetic algorithm in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the overall performance of terminals before and after material ratio optimization in an embodiment of the present invention; Figure 6a This is a schematic diagram comparing the consistency between the predicted and measured values ​​of the mechanical strength index in an embodiment of the present invention. Figure 6b This is a schematic diagram comparing the consistency between the predicted and measured values ​​of the insulation resistance index in an embodiment of the present invention. Figure 6c This is a schematic diagram comparing the consistency between the predicted and measured values ​​of the pressure resistance performance index in an embodiment of the present invention. Figure 6d This is a schematic diagram comparing the consistency between the predicted and measured values ​​of the environmental resistance performance index in an embodiment of the present invention. Figure 7 This is a schematic diagram of the main parts of the material hybrid variable optimization and selection device based on heterogeneous graph neural network in an embodiment of the present invention; Figure 8 This is a schematic diagram of the hardware structure of an electronic device used to implement the material hybrid variable optimization selection method based on heterogeneous graph neural networks in the embodiments of the present invention. Detailed Implementation

[0025] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0026] Figure 1 This is a schematic diagram illustrating the main steps of the material hybrid variable optimization and selection method based on heterogeneous graph neural networks in an embodiment of the present invention. Please refer to [link / reference]. Figure 1 The material hybrid variable optimization selection method based on heterogeneous graph neural network in this embodiment of the invention performs the following steps.

[0027] S101: Obtain multiple material configuration schemes for the entity and the test results of each material configuration scheme at at least one preset performance index. The entity can be any physical target; this specification uses an I / O optical port connector terminal as an example. Any material configuration scheme may include: the identifier of at least one selected material and the proportion data of each selected material. For example, the material can be a metallic element, a non-metallic element, a compound, or a mixture; the proportion data can be the mass percentage of a certain material in the entity. It is understood that the sum of the proportion data of each material in the same material configuration scheme is 1. The performance index can be determined according to actual needs. For example, for an I / O optical port connector terminal, the performance index can be at least one of the following: mechanical strength, insulation resistance, withstand voltage performance, and environmental resistance performance.

[0028] S102: Train a material scheme performance evaluation model using the material configuration scheme and test results. In one embodiment, the material scheme performance evaluation model may include a graph neural network (GNN) and a performance prediction model. The graph neural network may include multiple heterogeneous graph models. For example, the performance prediction model may employ a multilayer perceptron (MLP) model, such as a multilayer perceptron regression model.

[0029] Specifically, in the training process of the material scheme performance evaluation model: any material configuration scheme is input into a heterogeneous graph model. The material identifiers and proportion data in the material configuration scheme serve as nodes of the heterogeneous graph model, and the correspondence between the material identifiers and each proportion data, as well as the synergistic relationships between each proportion data, serve as edges of the heterogeneous graph model. These synergistic relationships can be the relationship between the proportion data of any two different materials in the material configuration scheme. Each heterogeneous graph model performs independent feature updates to obtain corresponding graph embedding vectors. The graph embedding vectors are input into the performance prediction model to obtain performance prediction results. The performance prediction results and test results are used to update the model parameters based on a preset loss function. In the above graph neural network, the feature update process of each heterogeneous graph model is independent of each other, and the nodes of each heterogeneous graph model perform feature updates based on a message passing mechanism within the model to achieve feature extraction. Through the above steps, it is possible to simultaneously handle the mixed optimization of discrete material selection variables (selected material identifiers) and continuous proportion variables (proportions of each material), solving the problem of difficulty in performance mapping modeling in existing technologies.

[0030] In one embodiment, a heterogeneous graph model includes a material system node and multiple material component nodes. After any material configuration scheme is input into a heterogeneous graph model, the material identifiers in the material configuration scheme are used together as the material system node of the heterogeneous graph model, and the proportion data of each material in the material configuration scheme are used as the material component nodes of the heterogeneous graph model. The material system node and each material component node are connected to form the first type of edge of the heterogeneous graph model, and the different material component nodes are connected to form the second type of edge of the heterogeneous graph model.

[0031] In one optional technical solution, each heterogeneous graph model can perform feature updates based on the Heterogeneous Graph Transformer (HGT). For example, any node in any heterogeneous graph model can perform multi-level iterative feature updates within the model to obtain the final feature representation of that node as output from the last layer. Exemplarily, in adjacent layers within a multi-level model, data is transferred from the previous layer to the next, and the updated feature representation of the previous layer serves as the initial feature representation of the next layer. During the feature update process of any layer, the same-relationship contribution features of each neighboring node of the same type of edge in any heterogeneous graph model are obtained. These same-relationship contribution features are aggregated into the cross-relationship contribution features of that node, and the cross-relationship contribution features are fused with the initial feature representation of that node in that layer to form the updated feature representation of that node in that layer.

[0032] As an optional approach, the above-mentioned same-relationship contribution features can be determined according to the following steps: For any node in any heterogeneous graph model, determine the linear projection feature of the neighboring node on the edge of the same type; determine the attention weight of the neighboring node on the node based on the node, the neighboring node, and the edge of the same type; obtain the contribution feature of the neighboring node on the node based on the linear projection feature and the attention weight; aggregate the contribution features of each neighboring node of the node based on the edge of the same type into the same-relationship contribution features of each neighboring node of the node based on the edge of the same type.

[0033] In one embodiment, after determining the final feature representation of each node in each heterogeneous graph model, pooling can be performed on the final feature representation of any material component node in any heterogeneous graph model, and the pooled final feature representation can be fused with the final feature representation of the material system node in the heterogeneous graph model to obtain the graph embedding vector of the heterogeneous graph model.

[0034] S103: Use the trained material scheme performance evaluation model as the fitness function of the genetic algorithm model, and determine the target material configuration scheme based on performance indicators according to the genetic algorithm model.

[0035] In an optional embodiment, multiple material configuration schemes can be selected as individuals in the genetic algorithm model based on preset constraints. A selection operation is performed on the individuals according to the fitness function of the material scheme performance evaluation model. A crossover operation is performed on the selected individuals to obtain new individuals. A mutation operation is performed on the new individuals to obtain the next generation of individuals. The above selection operation, crossover operation and mutation operation are iteratively performed until a preset termination condition is met. After that, the target material configuration scheme is determined from the material configuration schemes that meet the termination condition.

[0036] The following uses the material selection of I / O optical port connector terminals as an example to illustrate a specific embodiment of the present invention. The execution flow is described in [link to implementation details]. Figure 2 .

[0037] Step 1: Determining the feasible range of candidate material systems and their material composition ratios. This invention focuses on I / O optical connector terminals and identifies several promising candidate material systems to meet practical application requirements. These include, but are not limited to, multi-element copper-based alloys such as beryllium copper alloy, copper-nickel-silicon alloy, and copper-cobalt-beryllium alloy. Each material system utilizes controlled alloying elements and their proportions to meet comprehensive requirements for conductivity, mechanical strength, wear resistance, and environmental resistance.

[0038] For each candidate material system, based on materials science theory, phase diagram analysis, and engineering preparation experience, feasible ranges for the proportions of key alloying elements are defined. For example, for a typical copper-nickel-silicon alloy system, the feasible ranges for the proportions of key elements copper (Cu), nickel (Ni), and silicon (Si) can be set as Cu: 94.0% ~ 98.0%, Ni: 1.5% ~ 4.0%, and Si: 0.5% ~ 2.0%, respectively. For a beryllium-copper alloy system, the feasible ranges for the proportions of key elements copper (Cu), beryllium (Be), cobalt (Co), or nickel (Ni) need to be set separately according to specific performance targets, such as Be: 0.5% ~ 2.0%, Co: 0.1% ~ 1.5%, etc. All element proportion ranges are expressed as mass percentages and must meet the normalization constraint of a total of 100%, ensuring good formability and stability of the material in subsequent engineering preparation processes such as powder preparation, melting, casting, and rolling.

[0039] By setting scientifically reasonable feasible ranges for the proportions of key alloying elements within each material system, terminal material design can achieve sufficient design freedom while meeting the requirements of engineering feasibility, cost control, and service reliability. The aforementioned material systems and their corresponding feasible ranges for element proportions collectively constitute a mixed-variable design space for subsequent experimental design, model training, and optimization search, providing a fundamental parameter framework for the implementation of the entire optimization method.

[0040] Step 2, Preparation process of terminal alloy material After determining the candidate system of the terminal multi-element alloy material and the feasible range of the mass fraction ratio of each component, the material is synthesized and formed using a unified and determined engineering preparation process based on the material ratio parameters of each group, so as to ensure the comparability and consistency of the preparation process between samples with different ratios.

[0041] Specifically, the standardized preparation process of terminal multi-element alloy materials includes the following steps: (1) Powder preparation and mixing: Select metal raw materials of appropriate purity (such as copper, nickel, cobalt, silicon, beryllium, etc.) according to the material ratio parameters. Use mechanical alloying or gas atomization technology to make the raw materials into metal powder with controllable particle size, and after accurate weighing according to the ratio, perform thorough mechanical mixing to ensure that each component achieves uniform distribution at both the macroscopic and microscopic scales in the mixed powder, so as to provide a stable precursor for the subsequent smelting process.

[0042] (2) Melting: Place the uniformly mixed metal powder in a melting equipment under vacuum or inert gas protection, heat it to the set melting temperature (usually 1100℃~1300℃, the specific temperature is adjusted according to the alloy system), and keep it at the temperature for a certain time to allow the metal elements to fully dissolve and diffuse to form a uniform alloy melt.

[0043] (3) Casting: The alloy melt with uniform composition is quickly poured into a preheated metal mold or water-cooled copper mold, so that it solidifies and forms under controlled cooling conditions to obtain an alloy ingot or billet with accurate chemical composition and dense structure.

[0044] (4) Rolling: The alloy material after forming is plastically processed by a combination of hot rolling and cold rolling to gradually roll it into alloy sheets or strips of a set thickness in order to improve the material’s density, mechanical properties and meet dimensional accuracy requirements.

[0045] (5) Cutting: Finally, the rolled alloy sheet or strip is precisely cut to obtain standardized terminal material samples that meet the structural size and shape requirements of I / O optical connector terminals.

[0046] Through the above-mentioned standardized preparation process, terminal material samples with different material ratio schemes can be prepared under the same process conditions, providing experimental samples with consistent process basis and comparability for subsequent comprehensive performance testing.

[0047] Step 3: Terminal sample acquisition and comprehensive performance testing After completing the standard preparation of the terminal alloy materials, experimental samples corresponding to each candidate material system and its specific element ratio scheme were obtained. To systematically quantify the impact of material selection and element ratio on the overall performance of the terminals, and to ensure the representativeness and validity of subsequent modeling data, this invention first uses the Latin hypercube sampling method to efficiently generate experimental sample points in a defined hybrid design space, and then performs a unified standard comprehensive performance test on all samples.

[0048] (1) Latin hypercube sampling method Latin hypercube sampling is a highly efficient multidimensional spatial uniform sampling technique. Its core lies in uniformly stratifying the value range of each input variable and ensuring that each stratified interval has exactly one sample point. This avoids the sample point clustering phenomenon that may occur in traditional random sampling, significantly improving the uniform coverage of sampling points in the design space. For the hybrid design space of this invention (containing discrete material selection variables and continuous element ratio variables), the sampling implementation steps are as follows: Implementation steps: 1) Variable definition: Determine the sampling variables, including discrete material selection variables and continuous variables of the proportion (mass fraction) of key alloying elements (such as Cu, Ni, Si, Be, Co, etc.) in each material system.

[0049] 2) Sampling quantity setting: Determine the total sampling quantity M based on the dimensions of the design space and empirical criteria. For example, for a design space containing several material systems and multiple elements, the value of M can be set between 50 and 100 to balance experimental costs and space coverage.

[0050] 3) Interval partitioning and sampling: For each continuous variable (i.e., the range of proportions of a certain element), independently and uniformly divide it into M non-overlapping equally probable intervals. Randomly select a value from each interval as a sample of that variable in that interval.

[0051] 4) Sample combination and design point generation: Randomly pair the sampled values ​​of each continuous variable and combine them with the corresponding material selection variables to form M complete "material selection-element ratio" experimental design points.

[0052] 5) Mathematical expression: For the j-th continuous variable (element ratio), its feasible interval is: Then the M sample values ​​generated within this interval can be represented as: (1) in These are random numbers uniformly distributed within the interval [0,1]. This sampling strategy ensures that a limited number of experimental points can effectively cover the entire high-dimensional, hybrid design space.

[0053] (2) Standardized performance testing process For each set of terminal samples prepared according to the Latin hypercube sampling scheme, systematic testing of its four core performance indicators was conducted under strictly controlled uniform environmental conditions (temperature: 23±2℃, relative humidity: 50±10%), in accordance with industry standards and using calibrated testing equipment. 1) Mechanical strength test: Performed according to ISO 6892-1 standard for tensile testing of metallic materials using a universal testing machine. Apply a load to the standard tensile specimen at a constant loading rate (e.g., 2 mm / min) until fracture, record the maximum load-bearing capacity, and calculate the tensile strength.

[0054] 2) Insulation resistance test: Measure the insulation resistance using a high-resistance meter at the terminal's rated operating voltage (e.g., 500V DC). After applying the voltage, wait for the reading to stabilize (usually 60 seconds) and then record the insulation resistance value to evaluate the material's insulation performance.

[0055] 3) Withstand voltage test: The step-up voltage method is used to apply AC or DC voltage to the sample at a constant rate (e.g., 100 V / s) until dielectric breakdown occurs or the preset upper limit is reached. The breakdown voltage value is recorded to evaluate the high voltage resistance of the material.

[0056] 4) Environmental Resistance Testing: To assess the material's stability under harsh environments, high-temperature storage tests (e.g., storage at 125°C for 168 hours), damp heat cycling tests (e.g., 10 cycles at 40°C / 93%RH), and salt spray corrosion tests (e.g., spraying with 5% NaCl solution at 35°C for 48 hours) were conducted sequentially. Before and after each environmental test, key properties of the samples (e.g., mechanical strength, insulation resistance) were tested, and the performance retention rate was calculated to comprehensively evaluate its environmental resistance.

[0057] To ensure the statistical reliability of the test results, at least three parallel samples are usually prepared for each "material selection-proportioning" scheme. Each performance index is measured three times under the same test conditions, and the average value is used as the performance characterization value of the scheme. This lays a solid data foundation for building a high-quality "material selection-material proportioning-comprehensive performance" database.

[0058] Step 4: Construction of the "Material Selection - Material Proportioning - Comprehensive Performance" database After completing the comprehensive performance test of the terminal materials, the material selection information defined in step 1, the specific element ratio parameters on which the sample was prepared in step 2, and the various performance test results obtained in step 3 are systematically linked and unified to build a structured database of "material selection-material ratio-comprehensive performance".

[0059] (1) Database structure definition Each record in the database corresponds to a complete set of experimental samples and contains the following three types of core information: 1) Material selection identifier: Used to uniquely identify the material system to which the sample belongs (such as beryllium copper alloy, copper nickel silicon alloy, etc.).

[0060] 2) Material proportion vector: Represents the specific elemental composition of the sample. For the i-th sample, its proportion vector is expressed as: (2) Where m is the number of key alloying elements defined in this material system (such as Cu, Ni, Si, Be, Co, etc.). Let j represent the mass fraction (proportion) of the j-th element in the sample, and let the sum of the mass fractions of all elements satisfy the following condition: .

[0061] 3) Comprehensive performance index vector: For the i-th sample, its performance vector is represented as: (3) In the formula, The mechanical strength index is expressed as: (4) in, Let be the maximum load-bearing capacity of the i-th group of specimens in the tensile or bending test. This represents the effective cross-sectional area of ​​the sample under stress.

[0062] Insulation resistance is an indicator, expressed as: (5) in, The volume resistivity of the material. This is the insulation path length. The effective insulation cross-sectional area.

[0063] The pressure resistance performance index is expressed as: (6) in, The dielectric breakdown field strength of the material, This is the equivalent thickness of the insulation layer.

[0064] Environmental resistance performance indicators are expressed as: (7) in, and These represent the specified performance parameter values ​​(usually mechanical strength or insulation resistance) of the i-th sample before and after the k-th environmental test (such as high temperature storage, damp heat cycling, salt spray corrosion). For the corresponding environmental test weighting coefficients, satisfying .

[0065] (2) Database generation By aggregating complete information from all N sets of experimental samples (generated by Latin hypercube sampling design), a final structured database of "material selection - material ratio - comprehensive performance" is constructed, the mathematical expression of which is: (8) in, This indicates the material selection for the i-th sample. This is the corresponding material proportioning parameter vector. This is a comprehensive performance index vector. The database comprehensively records the related data from material selection and elemental ratios to various performance characteristics, providing a structured, high-quality data foundation for the subsequent training of heterogeneous graph neural network models.

[0066] Step 5, Heterogeneous Graph Neural Network Modeling and Training To accurately model the complex nonlinear system in which "material selection" (discrete categorical variable) and "element ratio" (continuous numerical variable) jointly affect "comprehensive performance," this invention constructs and trains a heterogeneous graph neural network. Its core employs a heterogeneous message passing mechanism based on relation-based and hierarchical attention to map each experimental sample to a heterogeneous graph and perform high-order information fusion.

[0067] (1) Formal definition of heterogeneous sample graphs For databases For any sample i, its corresponding attribute heterogeneity graph is defined as follows: . It is a collection of nodes, containing two types: Material System Node (MT): A single node its type Its initial characteristics For the learnable type embedding vector of the corresponding material system .

[0068] Elemental composition nodes (E): m nodes Corresponding to m alloying elements, type .node The initial characteristic is a scalar ratio value. Embedding: (9) in For the embedding matrix, MLP( The one-dimensional ratio is converted into a d′-dimensional feature.

[0069] For a set of edges, define two types of relations. : Composition (Comp): Connects each element node To the material system node ,Right now relation type .

[0070] Co-Composition (Co-Comp): Establishes fully connected edges between all element nodes, i.e. relation type .

[0071] This graph structure will display the data table. Transform it into a set containing N independent heterogeneous graphs. .

[0072] (2) Relationship-aware heterogeneous message passing The model employs the design concept of a Heterogeneous Graph Transformer (HGT), performing type- and relation-aware message passing at each layer l. For node u, its contribution to target node v under relation r is calculated through the following steps: 1) Relationship-specific linear projection: (10) in It is specifically designed for source node types (u) and the learnable projection matrix defined by relation r at layer l.

[0073] 2) Meta-relation triple attention calculation: attention weights It is determined by the query (target node v), the key (source node u), and the relationship r between them: (11) in: Let be the learnable query and key projection matrices corresponding to relation r, respectively.

[0074] It is a learnable scalar importance factor defined for meta-relation triples (target type-relation-source type) to explicitly model the inherent semantic strength of different interaction patterns.

[0075] 3) Aggregation of relation messages: Aggregate all neighbor messages from relation r: (12) Where ⊕ represents an aggregation operation.

[0076] (3) Hierarchical aggregation and node update The final update of node v at level l is accomplished by aggregating information from all relations and combining it with the residual join: 1) Cross-relationship aggregation: Merge all relationship-specific message vectors: (13) Aggregation functions can employ attention mechanisms or be directly concatenated followed by linear transformation.

[0077] Residual connectivity and nonlinear activation: The aggregated message is added to the node's own features after type-specific mapping, and an activation function is applied. (14) in, This is the residual connection weight matrix for the corresponding node type. This is the ReLU activation function.

[0078] After L layers of iteration, each node obtains a final representation that integrates the high-order structural semantics and attribute information of the entire graph. .

[0079] (4) Graph-level representation readout and performance prediction To perform performance regression prediction, the entire graph needs to be... Encode as a fixed-dimensional global representation vector .

[0080] 1) Target node pooling and feature fusion: based on material system nodes The final representation As the core of the graph representation, all element nodes are also included. The final representation is then pooled (e.g., attention pooling): (15) Where q and W are learnable parameters.

[0081] 2) Generate graph embedding: Merge the two to form the final graph embedding: (16) 3) Comprehensive performance prediction: Input the graph embedding vector into a multilayer perceptron regression head to predict four performance metrics: (17) The last layer of the regression head MLP uses a linear activation function.

[0082] (5) Model training, optimization and evaluation 1) Loss function: Mean squared error loss is used, and L2 regularization is added to prevent overfitting.

[0083] (18) in For the set of all trainable parameters of the model, is the regularization coefficient.

[0084] 2) Training and Validation: The dataset... Divide into training sets according to a ratio (e.g., 70:15:15). Validation set The test set and the Adam optimizer are used with early stopping. Training is terminated when the validation set loss does not decrease for P consecutive training epochs, and the model parameters with the lowest validation loss are rolled back.

[0085] 3) Model evaluation: on the test set The final model performance was evaluated using the following metrics: Root Mean Square Error (RMSE): (19) Coefficient of determination (R²): (20) Model acceptance criteria are usually as follows Furthermore, the RMSE is less than 10% of the actual value range of the corresponding performance index, in order to meet the requirements of the high-precision proxy model.

[0086] Step 6: Multi-objective genetic optimization of material selection and proportioning based on heterogeneous graph neural network surrogate model Based on a high-precision heterogeneous graph neural network mapping model of "material selection-material ratio-comprehensive performance", this invention introduces a multi-objective genetic algorithm as an optimization engine. This algorithm performs a global search and adaptive optimization of material selection and element ratio parameters within the engineering feasible range to identify the optimal material selection and ratio scheme that meets specific target performance requirements. The genetic algorithm simulates the evolutionary mechanism of "selection, crossover, and mutation" in nature, enabling efficient handling of high-dimensional, nonlinear, and multi-objective optimization problems while avoiding getting trapped in local optima.

[0087] (1) Definition of multi-objective optimization problem Given that multiple objectives, such as mechanical strength, insulation resistance, withstand voltage, and environmental resistance, need to be comprehensively balanced in terminal materials, this invention constructs the material selection and proportioning optimization problem as a multi-objective optimization problem. Assume that the comprehensive performance of the terminal materials is predicted by a trained heterogeneous graph neural network model: (twenty one) in This represents a complete material scheme, including the material selection identifier t and the element ratio vector. .

[0088] Define the target vector that needs to be optimized simultaneously as: (twenty two) in, Predicted mechanical strength value; Predicted insulation resistance value; Predicted pressure resistance value; : Predicted environmental resistance performance.

[0089] The optimization problem is defined as: (twenty three) (2) Implementation of genetic algorithm 1) Mixed encoding and initialization A hybrid coding scheme is adopted: Discrete part: Material selection t uses integer encoding.

[0090] Continuous part: Element proportion vector Real number encoding is used.

[0091] Initialize the population size to N pop =80. The discrete part of each individual is uniformly and randomly selected from its candidate set; each gene of the continuous part is uniformly and randomly generated within the feasible interval of its corresponding material, and normalization is performed to ensure that the sum of the quality fractions of all elements is 100%.

[0092] 2) Fitness assessment For each individual The fitness assessment process is as follows: Neural network performance prediction: (twenty four) in This is the heterogeneous graph neural network model trained in step 5.

[0093] Overall fitness calculation: (25) in For performance weighting coefficients, satisfying This reflects the relative importance of each performance characteristic.

[0094] 3) Genetic operators Selection operator: The tournament selection method is used, with a tournament size of k=3, while retaining the optimal N. elite =3 elite individuals directly enter the next generation.

[0095] Crossover operator: Discrete component (material selection): based on probability Perform uniform crossover, and the offspring randomly inherit from one of the parent generations.

[0096] Continuous portion (elemental ratio): using arithmetic crossover: Using arithmetic cross: (26) in Crossover probability .

[0097] Mutation operator: Discrete part: in terms of probability The material system is randomly switched to another type.

[0098] Continuous portion: using boundary variation: (27) in Probability of mutation .

[0099] 4) Constraint handling mechanism Boundary constraint repair: (28) Normalization process: Normalization of non-copper elements: (29) Then calculate: (30) (3) Genetic Algorithm Optimization Process and Termination Determination Mechanism In this embodiment of the invention, based on the trained heterogeneous graph neural network model, the genetic algorithm is used to iteratively search and adaptively optimize the material selection and element ratio parameters within the feasible range. The overall optimization process includes steps such as initialization, performance prediction, fitness evaluation, genetic evolution, and termination determination.

[0100] Specifically, the optimization process of a genetic algorithm includes the following steps: First, within a pre-defined set of material selections and feasible ranges for element ratios, an initial material scheme population is randomly generated using a hybrid coding method. Each individual corresponds to a set of candidate material selections and ratio parameters. Through constraint repair and normalization, it is ensured that all individuals in the initial population meet the engineering preparation constraints.

[0101] Secondly, for each set of material scheme parameters in the current population, the trained heterogeneous graph neural network model is called to predict the comprehensive performance of the terminals, and the corresponding predicted values ​​of mechanical strength, insulation resistance, withstand voltage performance and environmental resistance performance are obtained.

[0102] Then, based on the performance prediction results, the fitness of each individual in the current population is evaluated according to the preset comprehensive fitness evaluation function to reflect the comprehensive advantages and disadvantages of different material schemes under multiple performance objective constraints.

[0103] After completing the fitness assessment, genetic evolutionary operations are performed on the population, including individual selection, crossover and recombination, and mutation updates. Specifically, selection operations retain high-fitness, superior individuals; crossover and mutation operations generate new candidate solutions for material selection; and constraint repair processing is applied to the newly generated individuals to ensure they meet engineering feasibility requirements.

[0104] The aforementioned performance prediction, fitness assessment, and genetic evolution process are repeated across multiple generations, enabling the material scheme to gradually evolve towards better overall performance under the guidance of the neural network surrogate model.

[0105] The termination condition of a genetic algorithm is set to end the optimization process when any of the following conditions are met: 1) When the number of generations reaches the preset maximum number of generations, the iteration process of the genetic algorithm is terminated; 2) When the improvement in the overall fitness of the optimal material scheme in several consecutive generations is lower than the preset threshold, the algorithm is considered to have converged and the optimization is terminated; 3) When the predicted performance of the current optimal material solution meets the preset target requirements in key indicators such as mechanical strength, insulation resistance, withstand voltage performance and environmental resistance, the optimization process is terminated in advance.

[0106] After the termination conditions are met, the material selection and proportion scheme with the best overall adaptability is output as the global optimization result, which is used for subsequent terminal material preparation and performance verification.

[0107] (4) Optimize the output and verification of results After the algorithm terminates, it outputs the globally optimal material selection and proportioning scheme s. and its neural network prediction performance y .

[0108] Experimental verification of the optimization results: 1) According to the optimal ratio s Preparation of terminal material samples 2) Determine the actual performance yexp according to the standard test method. 3) Calculate the prediction error: 4) If all If the optimization result is valid, then the optimization result is considered valid. Using the above method, this embodiment achieves intelligent collaborative optimization of material selection and element ratio for I / O optical connector terminals under given engineering fabrication process constraints, significantly improving the efficiency, accuracy, and repeatability of material design. The complete process is as follows: Figure 2 As shown.

[0109] The method proposed in this embodiment is used to select the optimal parameters for key materials of high-speed I / O connector terminals, and the resulting high-speed I / O connector structure is shown below. Figure 3a , Figure 3b , Figure 3c As shown.

[0110] The following describes the hybrid design space definition and data sampling in this embodiment. After completing the connector terminal structure design, it is necessary to systematically analyze the influencing factors of terminal performance from a material perspective. Given that the overall performance of the terminal is largely controlled by material selection and composition ratio, this case selects two representative copper-based alloy systems as candidate materials: beryllium copper alloy (System A) and copper-nickel-silicon alloy (System B). For each system, the initial range of its key alloying elements and engineering-feasible proportions (mass fractions) is defined, as shown in Table 1.

[0111] Table 1 Initial range of multi-element alloy material ratios for high-speed I / O connector terminals Based on this design space, the Latin hypercube sampling method was used to generate 35 experimental schemes for each system, and a total of 70 samples were prepared according to standard processes. All samples underwent four comprehensive performance tests under uniform conditions, and a database containing 70 samples of "material selection-material ratio-comprehensive performance" was constructed.

[0112] The following describes the model construction, optimization, and results of this embodiment. Using the aforementioned database, a heterogeneous graph neural network (HGNN) model was constructed and trained. This model exhibits a prediction determination coefficient (R²) greater than 0.96 on the test set, demonstrating extremely high prediction accuracy and serving as a reliable performance proxy model. Subsequently, this model was used as the fitness evaluation function of a genetic algorithm (GA) to perform multi-objective collaborative optimization of material selection and element ratios. The optimization objective was to simultaneously maximize mechanical strength (S), insulation resistance (R), withstand voltage performance (V), and environmental resistance performance (E). After iterative search, the algorithm generated a series of non-dominated Pareto optimal solutions. After engineering trade-offs, a solution with outstanding and balanced comprehensive performance was finally selected from the Pareto front as the optimal solution.

[0113] The optimization results are as follows. The selected optimal solution is: the material selection is copper-nickel-silicon alloy (system B), and the specific element ratio is shown in Table 2.

[0114] Table 2 Optimized Terminal Selection and Ratio The following describes the optimization process and performance evolution of the genetic algorithm in this embodiment. Figure 4 The convergence curve of the comprehensive objective function value (calculated by the weighted fitness function) with the number of iterations is shown in the genetic algorithm for optimizing material proportions. It can be seen that the algorithm tends to stabilize after about 30 generations, indicating that the proposed method has good convergence and optimization efficiency. Figure 5 The overall performance of the terminal materials before and after optimization was further compared (expressed as the normalized performance index H / OHB). The optimized solution outperformed the unoptimized solution in all key performance dimensions, intuitively demonstrating the comprehensive benefits of multi-objective optimization.

[0115] The following describes the experimental verification and comparative analysis of this embodiment. To verify the effectiveness and superiority of the optimized scheme, a comparative experiment was conducted with the initial scheme designed based on engineering experience. The initial scheme also used a copper-nickel-silicon alloy (system B), with a ratio of Cu: 97.0%, Ni: 2.0%, and Si: 1.0%. Terminal material samples were prepared according to the two schemes, and comprehensive performance tests were conducted under a unified standard. The key results are compared in Table 3.

[0116] Table 3. Comparison of measured overall performance between the optimized and initial schemes The data in Table 3 clearly show that the material scheme optimized by the method in this embodiment achieves significant improvements in all four key performance indicators compared to traditional empirical design schemes. This strongly demonstrates that the "heterogeneous graph neural network surrogate model-multi-objective genetic algorithm" framework constructed by this method can effectively overcome the limitations of empirical design and systematically find a global solution with better overall performance in a complex hybrid design space, thus providing a reliable and efficient intelligent new approach for the material design of high-performance I / O optical port connector terminals.

[0117] To further verify the accuracy of the performance prediction of the optimization scheme, Figure 6a , Figure 6b , Figure 6c , Figure 6d The charts show a comparison between predicted and measured values ​​for four key performance indicators (mechanical strength, insulation resistance, withstand voltage, and environmental resistance). Each chart is presented as a scatter plot with a fitted line to visually reflect the prediction accuracy.

[0118] Figure 6a The units of the horizontal and vertical axes are mechanical strength / MPa. Figure 6b The units of the x and y coordinates are log. 10 (Insulation resistance) value, Figure 6c The units for the horizontal and vertical axes are pressure resistance / MPa. Figure 6d The units of the horizontal and vertical axes are environmental resistance performance / %. It can be seen that the predicted values ​​of each performance index are highly consistent with the measured values, and the fitted line is close to the ideal straight line y = x. This indicates that the optimization model established in this embodiment has good prediction accuracy and reliability, and can provide effective guidance for actual material design.

[0119] This invention enables global material optimization for various entities. Taking an optical connector terminal as an example, under the constraints of actual engineering manufacturing processes, this invention defines commonly used connector terminal materials, such as beryllium copper alloys and copper composite alloys, along with their corresponding copper, nickel, cobalt, and silicon elements, as selectable parameters. Then, a typical experimental scheme is generated in the parameter design space using the Latin hypercube sampling method. Samples are prepared and their performance is tested according to the scheme, thereby constructing a "material selection-material ratio-comprehensive performance" database, where performance indicators include mechanical strength, insulation resistance, withstand voltage, and environmental resistance. Next, a heterogeneous graph neural network model is constructed, defining material selection and material ratio and forming a hierarchical heterogeneous graph for model training, to accurately establish the nonlinear mapping relationship from material selection and material ratio to various performance indicators. Finally, the trained network model is used as the fitness function of a multi-objective genetic optimization algorithm to perform global automated search and iterative optimization of the material selection and material ratio parameters within their feasible range, ultimately outputting the optimal material selection and material ratio scheme that satisfies the comprehensive performance objective. The above methods can effectively guide material selection and material ratio design under given process constraints, achieving a significant improvement in the overall performance of terminals. The general, data-driven method for optimizing the selection and material ratio of optical connector terminals provided in this invention has clear engineering application value for reducing R&D costs and shortening development cycles.

[0120] Figure 7 This is a schematic diagram of the main parts of a material hybrid variable optimization and selection device based on a heterogeneous graph neural network according to an embodiment of the present invention.

[0121] like Figure 7As shown, the material hybrid variable optimization and selection device 70 based on heterogeneous graph neural networks in this embodiment of the invention includes: a data preparation unit 71, used to acquire multiple material configuration schemes of an entity and test results of each material configuration scheme on at least one preset performance index, wherein any material configuration scheme includes: the identifier of at least one selected material and the proportion data of each selected material; a model training unit 72, used to train a material scheme performance evaluation model using the material configuration schemes and the test results, wherein the material scheme performance evaluation model includes a graph neural network and a performance prediction model, and the graph neural network includes multiple heterogeneous graph models; during the training process: any material configuration scheme inputs a A heterogeneous graph model is used, where material identifiers and proportion data in the material configuration scheme serve as nodes, and the correspondence between material identifiers and proportion data, as well as the synergistic relationships among proportion data, serve as edges. Each heterogeneous graph model performs independent feature updates to obtain corresponding graph embedding vectors. These graph embedding vectors are input into the performance prediction model to obtain performance prediction results. The performance prediction results and the test results are used to update model parameters based on a preset loss function. A computation unit 73 is used to use the trained material scheme performance evaluation model as the fitness function of a genetic algorithm model, and to determine the target material configuration scheme based on the performance indicators according to the genetic algorithm model.

[0122] In one embodiment, a heterogeneous graph model includes a material system node and multiple material component nodes; and, after any material configuration scheme is input into a heterogeneous graph model, the material identifier in the material configuration scheme serves as the material system node of the heterogeneous graph model, the proportion data of each material in the material configuration scheme serves as the material component nodes of the heterogeneous graph model, the material system node is connected to each material component node to form a first type of edge of the heterogeneous graph model, and the different material component nodes are connected to form a second type of edge of the heterogeneous graph model.

[0123] In one embodiment, the model training unit 72 is further configured to: perform multi-level iterative feature updates on any node in any heterogeneous graph model to obtain the final feature representation of the node output by the last layer in the multi-level model; wherein, in two adjacent layers in the multi-level model, the updated feature representation of the previous layer is used as the initial feature representation of the next layer; during the feature update process of any layer, obtain the same-relationship contribution features of each neighboring node of any heterogeneous graph model based on the same type of edge, aggregate the same-relationship contribution features of each type of edge into the cross-relationship contribution features of the node, and fuse the cross-relationship contribution features with the initial feature representation of the node in the layer into the updated feature representation of the node in the layer.

[0124] In one embodiment, the model training unit 72 is further configured to: for any node in any heterogeneous graph model, determine the linear projection feature of the neighbor node on the edge of the same type; determine the attention weight of the neighbor node on the node based on the node, the neighbor node, and the edge of the same type; obtain the contribution feature of the neighbor node on the node based on the linear projection feature and the attention weight; and aggregate the contribution features of each neighbor node of the node based on the edge of the same type into the same relation contribution feature of each neighbor node of the node based on the edge of the same type.

[0125] In one embodiment, the model training unit 72 is further configured to: perform pooling on the final feature representation of any material component node in any heterogeneous graph model; and fuse the pooled final feature representation with the final feature representation of the material system node in the heterogeneous graph model to obtain the graph embedding vector of the heterogeneous graph model.

[0126] In one embodiment, the computing unit 73 is further configured to: select multiple material configuration schemes as individuals of the genetic algorithm model based on preset constraints; perform a selection operation on the individuals according to the fitness function; perform a crossover operation on the selected individuals to obtain new individuals; perform a mutation operation on the new individuals to obtain the next generation of individuals; iteratively execute the selection operation, the crossover operation, and the mutation operation until a preset termination condition is met; and determine a target material configuration scheme from the material configuration schemes that meet the termination condition.

[0127] In one embodiment, the material comprises: a metallic element and / or a non-metallic element; and / or, the entity comprises: an optical connector terminal; and / or, the performance indicators comprise at least one of the following: mechanical strength, insulation resistance, withstand voltage performance, and environmental resistance performance; and / or, the performance prediction model comprises: a multilayer perceptron model.

[0128] It should be noted that the material hybrid variable optimization and selection device based on heterogeneous graph neural network in this embodiment of the invention is software and can be installed in devices such as computers and mobile terminals.

[0129] This embodiment enables global material optimization for various entities. Taking optical connector terminals as an example, this embodiment can systematically construct a complex mapping relationship between material selection, element ratio, and overall terminal performance under engineering fabrication constraints. It can also achieve a smart design method for hybrid variable collaborative optimization under multiple performance requirements, thereby improving the overall performance, design efficiency, and engineering reliability of I / O optical connector terminal materials and meeting the urgent need for high-performance terminal materials in high-speed optical interconnect systems.

[0130] It should be noted that the collection, analysis, use, transmission, and storage of user personal information involved in the technical solution of this invention all comply with the provisions of relevant laws and regulations, are used for legal and reasonable purposes, and are not shared, disclosed, or sold outside of these legal uses, and are subject to supervision and management by regulatory authorities. Necessary measures should be taken to prevent unauthorized access to such personal information data, ensure that personnel authorized to access personal information data comply with the provisions of relevant laws and regulations, and ensure the security of user personal information. Once this user personal information data is no longer needed, the risk should be minimized by restricting or even prohibiting data collection and / or deleting the data. When used, including in certain related applications, user privacy is protected by de-identifying the data, for example, by removing specific identifiers (e.g., date of birth), controlling the amount or specificity of stored data (e.g., collecting location data at the city level rather than the specific address level), controlling how the data is stored, and / or other de-identification methods.

[0131] According to embodiments of the present invention, the present invention also provides an electronic device and a readable storage medium.

[0132] The electronic device of the present invention includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to cause the at least one processor to perform the material hybrid variable optimization selection method based on heterogeneous graph neural networks provided by the present invention.

[0133] The non-transitory computer-readable storage medium of the present invention stores computer instructions, which are used to cause the computer to execute the material hybrid variable optimization selection method based on heterogeneous graph neural network provided by the present invention.

[0134] like Figure 8 The diagram shown is a hardware structure schematic of an electronic device used to implement the method of an embodiment of the present invention. Figure 8 The electronic device includes: one or more processors 81 and memory 82. Figure 8 Taking a processor 81 as an example, the memory 82 is the non-transitory computer-readable storage medium provided by this invention.

[0135] The electronic device of the present invention may further include an input device 83 and an output device 84.

[0136] The processor 81, memory 82, input device 83, and output device 84 can be connected via a bus or other means. Figure 8 Taking the example of a connection between China and Israel via a bus.

[0137] The memory 82, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs, and modules, such as the program instructions / modules corresponding to the method in this embodiment of the invention. The processor 81 executes various server functions and data processing by running the non-transitory software programs, instructions, and modules stored in the memory 82, thereby implementing the material hybrid variable optimization selection method based on heterogeneous graph neural networks described in the above method embodiment.

[0138] The memory 82 may include a program storage area and a data storage area. The program storage area may store the operating system and applications required for at least one function; the data storage area may store data created based on the use of the device. Furthermore, the memory 82 may include high-speed random access memory and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, the memory 82 may optionally include memory remotely located relative to the processor 81, and these remote memories may be connected to the device via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0139] Input device 83 can receive input digital or character information, and generate key signal inputs related to user settings and function control of the device. Output device 84 may include display devices such as a display screen.

[0140] One or more of the above modules are stored in the memory 82. When executed by one or more processors 81, the material hybrid variable optimization selection method based on heterogeneous graph neural network of any of the above method embodiments is executed.

[0141] The above-described product can execute the material mixed variable optimization selection method based on heterogeneous graph neural networks provided in the embodiments of the present invention, and has the corresponding functional modules and beneficial effects for executing the material mixed variable optimization selection method based on heterogeneous graph neural networks. Technical details not described in detail in this embodiment can be found in the material mixed variable optimization selection method based on heterogeneous graph neural networks provided in the embodiments of the present invention.

[0142] This invention provides an enhanced genetic optimization model based on heterogeneous graph neural networks, aiming to solve the problems of difficulty in co-optimizing discrete material selection and continuous proportioning parameters, performance mapping modeling difficulties, and low efficiency of multi-objective optimization under complex engineering constraints, thereby achieving systematic optimization and improvement of the comprehensive performance of physical materials. Taking optical connector terminals as an example, this invention first uses various candidate material systems such as beryllium copper alloy and copper composite alloy, and their key alloying elements (such as copper, nickel, cobalt, and silicon), under the constraints of engineering manufacturing processes, to treat material selection and proportioning as optimizable parameters. A typical experimental scheme is generated in the design space using the Latin hypercube sampling method. Through standardized sample preparation and performance testing, a structured database of "material selection-material proportioning-comprehensive performance" is constructed. Secondly, a heterogeneous graph neural network model is constructed to encode material selection and element proportioning as hierarchical heterogeneous nodes to accurately capture the complex nonlinear mapping relationship between "selection-proportioning" and various comprehensive performance indicators. Finally, the trained heterogeneous graph neural network model is used as the fitness evaluation function of the multi-objective genetic algorithm to perform global automated search and iterative optimization on the mixed variable (discrete selection and continuous ratio) design space, and finally outputs the optimal material selection and ratio scheme that meets the comprehensive performance target requirements, thereby achieving a systematic improvement in the performance of terminal materials.

[0143] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A material mixing variable optimization selection method based on a heterogeneous graph neural network, characterized in that, include: Obtain multiple material configuration schemes for an entity and the test results of each material configuration scheme on at least one preset performance index. Each material configuration scheme includes: the identifier of at least one selected material and the proportion data of each selected material. The material configuration scheme and the test results are used to train a material scheme performance evaluation model, which includes a graph neural network and a performance prediction model, and the graph neural network includes multiple heterogeneous graph models. During the training process: any material configuration scheme is input into a heterogeneous graph model. The material identifiers and proportion data in the material configuration scheme are used as nodes of the heterogeneous graph model, and the correspondence between the material identifiers and each proportion data, as well as the synergistic relationship of each proportion data, are used as edges of the heterogeneous graph model. Each heterogeneous graph model performs independent feature updates to obtain corresponding graph embedding vectors. The graph embedding vectors are input into the performance prediction model to obtain performance prediction results. The performance prediction results and the test results are used to update the model parameters based on a preset loss function. The trained material scheme performance evaluation model is used as the fitness function of the genetic algorithm model, and the target material configuration scheme based on the performance index is determined according to the genetic algorithm model.

2. The method of claim 1, wherein, A heterogeneous graph model includes a material system node and multiple material component nodes; and, After inputting a heterogeneous graph model with any material configuration scheme, the material identifiers in the material configuration scheme are used as material system nodes of the heterogeneous graph model, and the proportion data of each material in the material configuration scheme are used as material component nodes of the heterogeneous graph model. The material system nodes and each material component node are connected to form the first type of edge of the heterogeneous graph model, and the different material component nodes are connected to form the second type of edge of the heterogeneous graph model.

3. The method of claim 2, wherein, Each heterogeneous graph model performs independent feature updates, including: In any heterogeneous graph model, any node undergoes multi-level iterative feature updates within the model to obtain the final feature representation of that node, output by the last layer. In the multi-layered system, the updated feature representation of the previous layer is used as the initial feature representation of the next layer. During the feature update process of any layer, the same-relationship contribution features of each neighboring node of any heterogeneous graph model based on the same type of edge are obtained. The same-relationship contribution features of each type of edge are aggregated into the cross-relationship contribution features of the node. The cross-relationship contribution features are fused with the initial feature representation of the node in the layer to form the updated feature representation of the node in the layer.

4. The method of claim 3, wherein, The step of obtaining the same relation contribution features of each node in any heterogeneous graph model based on the same type of edge among its neighboring nodes includes: For any node in any heterogeneous graph model, determine the linear projection feature of that neighbor node on the edge of the same type. The attention weight of any neighbor node to any given node is determined based on the given node, the given neighbor node, and the edge of that type. The contribution characteristics of any neighbor node to any given node are obtained based on the linear projection features and the attention weights; and, The contribution features of any node based on the edge of that type are aggregated into the same relation contribution features of any node based on the edge of that type.

5. The method according to claim 3, characterized in that, The process of obtaining the corresponding graph embedding vector includes: Pooling is performed on the final feature representation of any material component node in any heterogeneous graph model; The final feature representation after pooling is fused with the final feature representation of the material system nodes in the heterogeneous graph model to obtain the graph embedding vector of the heterogeneous graph model.

6. The method of claim 1, wherein, The step of determining the target material configuration scheme based on the performance index according to the genetic algorithm model includes: Multiple material configuration schemes are selected as individuals in the genetic algorithm model based on preset constraints. A selection operation is performed on the individuals according to the fitness function. A crossover operation is performed on the selected individuals to obtain new individuals. A mutation operation is performed on the new individuals to obtain the next generation of individuals. The selection operation, the crossover operation, and the mutation operation are iteratively performed until a preset termination condition is met. The target material configuration scheme is determined from the material configuration schemes that meet the termination conditions.

7. The method according to any one of claims 1 to 6, characterized in that, The material includes: elemental metals and / or elemental nonmetals; and / or, The entity includes: optical port connector terminals; and / or, The performance indicators include at least one of the following: mechanical strength, insulation resistance, withstand voltage performance, environmental resistance performance; and / or, The performance prediction model includes a multilayer perceptron model.

8. A material mixing variable optimization selection device based on a heterogeneous graph neural network, characterized in that, include: The data preparation unit is used to acquire multiple material configuration schemes of the entity and the test results of each material configuration scheme on at least one preset performance index. Each material configuration scheme includes: the identifier of at least one selected material and the proportion data of each selected material. A model training unit is used to train a material scheme performance evaluation model using the material configuration scheme and the test results. The material scheme performance evaluation model includes a graph neural network and a performance prediction model. The graph neural network includes multiple heterogeneous graph models. During the training process: any material configuration scheme is input into a heterogeneous graph model. The material identifiers and proportion data in the material configuration scheme are used as nodes of the heterogeneous graph model. The correspondence between the material identifiers and each proportion data, as well as the synergistic relationship of each proportion data, are used as edges of the heterogeneous graph model. Each heterogeneous graph model performs independent feature updates to obtain corresponding graph embedding vectors. The graph embedding vectors are input into the performance prediction model to obtain performance prediction results. The performance prediction results and the test results are used to update the model parameters based on a preset loss function. The computational unit is used to use the trained material scheme performance evaluation model as the fitness function of the genetic algorithm model, and to determine the target material configuration scheme based on the performance index according to the genetic algorithm model.

9. An electronic device, comprising: include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform the method of any one of claims 1-7.

10. A non-transitory computer-readable storage medium, comprising: The non-transitory computer-readable storage medium stores computer instructions for causing the computer to perform the method of any one of claims 1-7.

Images