Multi-modal fusion material design method and system based on material mechanism constraint

Abstract

The invention discloses a multi-modal fusion material design method and system based on material mechanism constraints, and relates to the technical field of material science and artificial intelligence crossing. According to the system, vertical domain data such as a two-dimensional / three-dimensional structure, a performance index, a cost index, a mass production requirement, an environmental adaptability requirement and a service life requirement of a product are integrated through a multi-modal data input module, and domain rules are converted into quantitative constraints through a material mechanism constraint engine; and then cross-modal feature accurate association is realized through a multi-modal feature fusion module driven by an attention mechanism, and finally, a physical feasible complete design scheme is output by a generative design unit in combination with a multi-objective optimization algorithm. According to the method, the problems of mechanism mismatch and poor process adaptability of a traditional multi-modal model are solved, the development cycle is obviously shortened, the design precision is greatly improved, the development cost and risk of new materials are remarkably reduced in the fields of powder metallurgy and additive manufacturing, and the method is suitable for efficient and innovative design of a complex material system.

Description

Technical Field

[0001] This invention belongs to the field of materials science and engineering technology, and involves the interdisciplinary application of artificial intelligence and materials design. Specifically, it is a multimodal fusion materials design method, system and storage medium based on material mechanism constraints. It is particularly suitable for the innovative material design of complex process systems such as powder metallurgy and additive manufacturing, and can be widely used in the development of key materials in the fields of automobiles, aerospace and high-end equipment. Background Technology

[0002] Current material design relies on a "trial and error" approach, with a single experiment taking 3-6 months, and suffers from the following core problems: 1. Data fusion defects: Multimodal data such as product structure (2D / 3D), performance indicators, and process parameters are isolated from each other and have not been effectively correlated, resulting in a disconnect between the design scheme and actual needs; 2. Lack of mechanism guidance: Existing artificial intelligence models mostly rely on data fitting. Multimodal material design models lack material mechanism constraints and do not incorporate material physical / chemical mechanisms (such as phase equilibrium and diffusion laws). They often generate solutions that "meet the data requirements but are physically infeasible" (such as component combinations that exceed the phase equilibrium range in powder metallurgy). Experimental verification is costly, resulting in low physical feasibility of the design solutions. 3. Modal heterogeneity (such as differences in the structure of image and text data) during multimodal data fusion leads to distortion of feature association; 4. Poor process adaptability: The general multimodal model does not take into account the special design constraints of processes such as powder metallurgy (sintering temperature-composition diffusion coupling) and additive manufacturing (melt pool stability-microstructure correlation), making it difficult to implement the solution for mass production; 4. Difficulty in knowledge accumulation: Design experience relies on the individual abilities of engineers, making it difficult to form a reusable enterprise-level knowledge base, resulting in low efficiency of technology iteration.

[0003] With the advancement of the Materials Genome Initiative and Industry 4.0, there is an urgent need for an integrated design solution that combines "data-driven + mechanism-assured + process-adapted" approaches to address the aforementioned shortcomings and accelerate the development of new materials. Summary of the Invention

[0004] This invention addresses only one of the problems existing in the background art.

[0005] Therefore, this invention provides a multimodal fusion material design method and system based on material mechanism constraints.

[0006] The technical solution adopted by this invention to solve its technical problem is: A multimodal fusion material design system based on material mechanism constraints includes: A multimodal data input module is used to acquire and preprocess vertical domain data related to material design. The vertical domain data includes, but is not limited to, two-dimensional / three-dimensional structural data of the product, material performance indicators, cost indicators, mass production requirements, environmental adaptability requirements, and service life indicators. The material mechanism constraint engine includes a vertical domain material mechanism knowledge base, a constraint transformation unit, and a dynamic verification unit. The constraint transformation unit converts the mechanism rules in the vertical domain material mechanism knowledge base into mathematical constraint formulas, and the dynamic verification unit detects the mechanism compliance of intermediate design results in real time. A multimodal feature fusion module, which includes an attention-based cross-modal fusion network for extracting multimodal features; Generative design unit, which generates material design schemes based on the fusion of features and mechanistic constraints; The design scheme output module outputs the final design scheme in a structured form. The output content includes, but is not limited to, material composition ratio, key process parameters, expected performance parameters, cost accounting report, material microstructure prediction results, process simulation video, mass production prediction report, mechanism feasibility analysis report, and service life assessment report.

[0007] Furthermore, the preprocessing steps of the multimodal data input module include: data cleaning, outlier removal, missing value completion, and data standardization, and modal alignment is achieved by establishing a spatial correlation mapping between two-dimensional / three-dimensional structural data and performance indicators.

[0008] Furthermore, the workflow of the dynamic verification unit includes: S41: Real-time acquisition of intermediate solutions output by generative design units; S42: Substitute the intermediate solution into the mathematical constraint formula corresponding to the mechanism rule and determine whether the constraint conditions are met. S43: If the constraints are met, the intermediate solution is passed to the multimodal feature fusion module for feature iteration optimization; if the constraints are not met, the reason for the violation is output, and a correction direction is generated based on the mechanism rules.

[0009] Furthermore, the multimodal feature fusion module includes an encoder, a cross-attention layer, and a feature optimization unit. The encoder is used to perform feature extraction on image files, text, and numerical data respectively. The cross-attention layer learns the association weights of different modal features to strengthen the strong association features of performance-component-process. The feature optimization unit dynamically adjusts the feature weight distribution in combination with the feedback of the mechanism constraint engine.

[0010] Furthermore, the encoder uses a CNN-Transformer hybrid architecture for feature extraction on two-dimensional / three-dimensional structural data. In the CNN-Transformer hybrid architecture, the CNN module extracts local structural features, and the Transformer module captures global structural relationships. The encoder uses a BERT pre-trained model for semantic encoding on textual data. The encoder uses a fully connected layer + batch normalization for feature mapping on numerical data.

[0011] Furthermore, the cross-attention layer calculates the correlation degree of different modal features through an attention weight matrix, wherein the correlation weight coefficient between the performance index modality and the component modality is not less than 0.6, ensuring that performance requirements provide directional guidance for component design.

[0012] Furthermore, the generative design unit includes a target performance analysis module, a scheme generation network, and a multi-objective optimizer. The training process of the scheme generation network includes: first, pre-training using labeled data from a public materials database, and then fine-tuning by combining private manufacturing process data from vertical enterprises. During the training process, a mechanism constraint loss function is introduced to reduce the mechanism violation rate of the generated schemes.

[0013] Furthermore, the design output module supports multiple output formats, including: structured documents, visual charts, 3D model files, and process video files.

[0014] A multimodal fusion material design method based on material mechanism constraints includes the following steps: S1: Multimodal vertical domain data acquisition and preprocessing: Collect data on product structure, performance, cost, environmental adaptability, etc. through the multimodal data input module, and perform cleaning, standardization, and modal alignment processing; S2: Construction of vertical material mechanism constraints: Based on the knowledge base of the material mechanism constraint engine, the rules such as crystal structure stability and phase equilibrium are transformed into mathematical constraints, and a mechanism verification rule base is constructed. S3: Multimodal Feature Fusion and Optimization: Modality-specific encoders are used to extract features of each modality. Feature association weights are learned through cross-attention layers. The feature weights are dynamically adjusted in combination with feedback from the mechanism constraint engine, and the optimized fused features are output. S4: Initial Design Scheme Generation: The target performance analysis module of the generative design unit transforms user requirements into optimization objectives, and the scheme generation network generates multiple sets of initial design schemes based on fusion features and mechanism constraints. S5: Multi-objective optimization and solution output: The multi-objective optimizer performs multi-objective optimization of the initial solution in terms of performance, cost and mass production, selects the non-dominated optimal solution, and outputs the final solution in a structured form by the design solution output module, along with a mechanism feasibility report.

[0015] Furthermore, the multi-objective optimization process in step S5 includes: S51: Set the weight coefficients for each objective, using "performance compliance rate", "cost control rate" and "mass production qualification rate" as optimization targets; S52: Iterative optimization is performed using a non-dominated sorting genetic algorithm, with no less than 100 iterations and a population size of 50-100. S53: Select 3-5 optimal solutions based on congestion distance calculations, allowing users to choose according to their actual needs.

[0016] The beneficial effects of this invention are that the design system in this application integrates vertical domain data such as the two-dimensional / three-dimensional structure, performance indicators, cost indicators, mass production requirements, environmental adaptability requirements, and service life requirements of the product through a multimodal data input module. The domain rules are transformed into quantitative constraints by a material mechanism constraint engine (which has a built-in vertical domain knowledge base covering crystallography, thermodynamics, and kinetics). Then, a multimodal feature fusion module driven by an attention mechanism is used to achieve accurate correlation of cross-modal features. Finally, a generative design unit combined with a multi-objective optimization algorithm outputs a physically feasible complete design scheme.

[0017] This invention solves the problems of "data fitting but mechanism mismatch" and "poor process adaptability" in traditional multimodal models. It significantly shortens the development cycle and greatly improves the design accuracy in the fields of powder metallurgy and additive manufacturing, and significantly reduces the development cost and risk of new materials. It is suitable for efficient and innovative design of complex material systems. Attached Figure Description

[0018] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0019] Figure 1 This is a schematic diagram of the structure of the multimodal fusion material design system based on material mechanism constraints in this invention.

[0020] Figure 2 This is a schematic diagram illustrating the implementation process of the multimodal fusion material design method based on material mechanism constraints in this invention. Detailed Implementation

[0021] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, features defined with "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] A multimodal fusion material design system based on material mechanism constraints is proposed, comprising a five-layer architecture: data layer, mechanism layer, fusion layer, design layer, and output layer. The data layer supports hybrid storage using MySQL (numerical data), MongoDB (unstructured data such as 3D models), and Redis (caching frequently accessed data). The mechanism layer constructs a knowledge graph using Neo4j to achieve efficient rule retrieval and reasoning. The fusion and design layers are deployed on a GPU server (NVIDIA A100) to support parallel computing. The output layer provides web and desktop interfaces, allowing users to adjust design parameters in real time.

[0025] Specifically, the data layer is set up as a multimodal data input module, whose data acquisition scope includes covering two-dimensional drawings (such as CAD files), three-dimensional models (such as STL files), performance indicators (such as hardness, tensile strength, and corrosion resistance), cost limits, mass production capacity requirements (such as daily output ≥ 1000 pieces), environmental adaptability (such as operating temperature of -40~120℃), and service life (such as ≥ 5 years). The multimodal data input module has a data preprocessing function to preprocess the acquired data.

[0026] The mechanism layer is set as a material mechanism constraint engine, which includes a vertical domain knowledge base, constraint transformation unit, and dynamic verification unit. The construction of the vertical domain knowledge base employs ontology modeling technology to build a knowledge graph, encompassing three major categories of core rules: crystallographic rules, thermodynamic rules, and kinetic rules. Crystallographic rules define, for example, the stability criterion for face-centered cubic (FCC) crystal structures (coordination number ≥ 12) and the allowable range of lattice constant deviation (≤ 0.5%). Thermodynamic rules define, for example, the phase equilibrium diagram constraints for alloys such as Fe-Cu and Ti-Al-V, and the condition for minimizing Gibbs free energy. Kinetic rules include, for example, the diffusion coefficient formula (D = D0·exp(-Q / RT), where D0 is the diffusion constant, Q is the diffusion activation energy, R is the gas constant, and T is the absolute temperature) and phase transition rate equations (such as the Avrami equation). The constraint transformation unit converts the above rules into computable mathematical constraint formulas. For example, the phase equilibrium constraint of Fe-Cu alloy is transformed into "Cu content ∈ [8,18]wt% (at T=1050℃)". The dynamic verification unit adopts a three-step process of "real-time interception - mechanism tracing - scheme correction". For example, when the Cu content in the generated powder metallurgy Fe-Cu alloy is 20wt%, the unit automatically identifies that it exceeds the phase equilibrium range, outputs the violation reason "Cu content is too high, which is easy to form brittle Cu-Fe compounds", and suggests adjusting it to 10-15wt%.

[0027] The fusion layer includes a multimodal feature fusion module, which contains an encoder, a cross-attention layer, and a feature optimization unit. The encoder is used for modality-specific encoding. In this layer, for two-dimensional / three-dimensional structural data, a CNN-Transformer hybrid architecture is adopted. The CNN (ResNet50) extracts local features (such as the size of the friction surface groove), and the Transformer (ViT-Base) captures global structural correlations (such as the overall stress distribution of the blade). For text data (such as "resistant to 600℃ high temperature"), a BERT pre-trained model is used for semantic encoding, outputting a 768-dimensional semantic feature vector. For numerical data (such as cost of 4.5 yuan / piece and sintering temperature of 1100℃), two fully connected layers + batch normalization are used to output a 64-dimensional numerical feature vector. The cross-attention layer constructs a four-modal attention matrix of "structure-performance-composition-process" and calculates the correlation weight of each modal feature. For example, the correlation weight between "tensile strength of blade tenon" and "Al content in Ti-6Al-4V alloy composition" is set to 0.7 (higher than other correlations) to ensure the directional guidance of performance requirements. The feature optimization unit adjusts the weights based on feedback from the mechanism constraint engine. For example, when a component combination does not satisfy phase equilibrium, the weight of that component's feature is reduced, while the weight of the component's feature that conforms to the mechanism is increased.

[0028] The design layer is a generative design unit, featuring a target performance analysis module, a solution generation network, and a multi-objective optimizer. The target performance analysis module transforms users' qualitative needs into quantitative optimization goals. For example, "high hardness" is transformed into "Vickers hardness ≥ 800HV", and "long life" is transformed into "fatigue life ≥ 10^5 cycles". The scheme generation network adopts an improved generative adversarial network (GAN). The generator outputs continuous variables such as component proportions and process parameters, while the discriminator simultaneously verifies the "data fit" and "mechanism conformity". During the training process, a mechanism constraint loss function (L=αL_data+βL_mechanism, where α=0.4 and β=0.6 to ensure mechanism priority) is introduced. The multi-objective optimizer uses the NSGA-II algorithm, with "performance compliance rate (weight 0.4)", "cost control rate (weight 0.3)" and "mass production qualification rate (weight 0.3)" as objectives. After 100 iterations, it selects 3 optimal solutions, such as the "high hardness and low cost", "balanced performance" and "high mass production rate" solutions for powder metallurgy Fe-Cu alloys.

[0029] The output layer is the design scheme output module. The output content of the output layer includes: composition ratio (accurate to 0.1wt%), process parameters (such as sintering temperature ±5℃, laser power ±10W), performance prediction curves (such as hardness-temperature relationship), microstructure prediction diagrams (such as grain size distribution), cost accounting table (raw materials + processing fees), mass production report (yield prediction, equipment requirements), and mechanism feasibility report (phase equilibrium verification, diffusion process simulation). The output format supports PDF documents, Excel data tables, STL 3D models, and MP4 process videos to meet the needs of different application scenarios (such as process simulation and production guidance).

[0030] A multimodal fusion material design method based on material mechanism constraints, which is implemented based on the above system, specifically includes the following steps: S1, Multimodal Vertical Data Acquisition and Preprocessing The multimodal data input module collects data on product structure, performance, cost, and environmental adaptability, and performs cleaning, standardization, and modal alignment processing. Data cleaning: IQR (interquartile range) method was used to remove outliers in performance data, and KNN algorithm was used to complete missing process parameters; Data standardization: Min-Max normalization is used to map data of different dimensions such as cost (yuan / piece) and temperature (°C) to the [0,1] interval to avoid the impact of differences in magnitude on feature fusion; Modal alignment: Establishing spatial relationships between key areas of the 3D model (such as blade tenons) and performance indicators (such as tenon tensile strength), and achieving a one-to-one correspondence between "structure and performance" through coordinate mapping.

[0031] S2: Construction of Vertical Material Mechanism Constraints Based on the knowledge base of the material mechanism constraint engine, the rules such as crystal structure stability and phase equilibrium are transformed into mathematical constraints (such as diffusion coefficient constraint D≥10^-12m² / s), and a mechanism verification rule base is constructed. S3: Multimodal Feature Fusion and Optimization A modality-specific encoder is used to extract features of each modality. Feature association weights are learned through a cross-attention layer. The feature weights are dynamically adjusted in combination with feedback from the mechanism constraint engine (such as "a certain component combination does not satisfy phase equilibrium"), and the optimized fusion features are output. S4: Initial Design Scheme Generation The target performance analysis module of the generative design unit transforms user requirements (such as "fatigue life of aerospace blades ≥ 10^5 cycles") into optimization objectives, and the scheme generation network generates multiple initial design schemes based on fusion features and mechanism constraints. S5: Multi-objective optimization and solution output The multi-objective optimizer performs performance-cost-mass production optimization on the initial solution, filters out the non-dominated optimal solution, and outputs the final solution in a structured form by the design solution output module, along with a mechanism feasibility report. Specifically, S51: "Performance compliance rate," "cost control rate," and "mass production pass rate" are the optimization objectives, and the weight coefficients for each objective are set (user-customizable). S52: Iterative optimization is performed using the non-dominated sorting genetic algorithm (NSGA-II), with no less than 100 iterations and a population size of 50-100. S53: Select 3-5 optimal solutions based on congestion distance calculations, allowing users to choose according to their actual needs.

[0032] Example 1 The design of the Fe-Cu alloy friction material matrix for powder metallurgy is as follows. The friction material matrix in the automotive braking system has the following requirements: Vickers hardness ≥800HV, coefficient of friction 0.35-0.45, cost ≤5 yuan / piece, mass production rate ≥90%, and service temperature -30~200℃.

[0033] Specific design methods include: S1, Multimodal data input and preprocessing: Data collected: 2D friction surface groove structure diagram (CAD format, groove depth 2mm, spacing 5mm), 3D matrix model (STL format, diameter 100mm, thickness 15mm), performance indicators (hardness ≥800HV, friction coefficient 0.35-0.45), cost limit 5 yuan / piece, mass production capacity 5000 pieces / day; Preprocessing: Remove outliers >1000HV from the hardness data, normalize the cost (3-5 yuan / piece) to [0,1], and establish the spatial relationship between the "friction surface area" and "friction coefficient" in the three-dimensional model.

[0034] S2, Mechanistic Constraint Construction: The Fe-Cu alloy phase equilibrium rules (when T=1050-1100℃, the Cu content must be ∈[8,18]wt%, otherwise a brittle ε phase will be generated) and diffusion coefficient constraints (the diffusion coefficient of Fe in Cu is D≥1.2×10^-12m² / s, to ensure sintering densification) are called from the knowledge base. Transformed into mathematical constraints: Cu∈[10,15]wt% (with a safety margin), sintering temperature T∈[1050,1100]℃, holding time t∈[2,3]h (derived from D=D0exp(-Q / RT)).

[0035] S3, Multimodal Feature Fusion: Encoding: CNN extracts the groove size features of the friction surface (outputs a 256-dimensional vector), BERT encodes "friction coefficient 0.35-0.45" (outputs a 768-dimensional vector), and a fully connected layer processes the cost / temperature data (outputs a 64-dimensional vector). Cross-attention: Strengthen the correlation between "hardness-Cu content-sintering temperature" (weight 0.65), and weaken irrelevant features (such as matrix edge size, weight 0.1). Optimization: Since the initial Cu content of 18wt% exceeds the phase equilibrium range, the weight of this feature is dynamically reduced, while the weight of the 12wt% Cu feature is increased.

[0036] S4, Initial scheme generation: Target Analysis: The optimization target is to transform "friction coefficient 0.35-0.45" into "friction coefficient deviation ≤ 0.02". GAN generates 30 initial schemes, such as Scheme A: Fe=88wt%, Cu=12wt%, sintering temperature 1080℃, holding time 2.5h, predicted hardness 830HV, cost 4.8 yuan / piece.

[0037] S5, Multi-objective Optimization and Output: After 100 iterations of NSGA-II, the optimal solution was selected: Fe=87wt%, Cu=13wt%, sintering temperature 1090℃, and holding time 2.3h. Output results: Hardness 845HV (meets standards), friction coefficient 0.38 (meets standards), cost 4.7 yuan / piece (meets standards), mass production rate 93%, microstructure predicted as "Fe matrix + uniformly distributed Cu particles (particle size 5-8μm)", mechanism feasibility report shows "consistent with Fe-Cu phase equilibrium and diffusion law".

[0038] Example 2 Additive manufacturing (SLM) Ti-6Al-4V alloy aero-engine blade design. Aero-engine high-pressure turbine blades have the following requirements: tensile strength ≥900MPa, fatigue life ≥10^5 cycles, high temperature resistance of 600℃, mass production cycle ≤7 days, and blade dimensional accuracy ±0.1mm.

[0039] Specific design methods include: S1, Multimodal data input and preprocessing: Data collected: 3D blade model (STL format, length 150mm, tip thickness 3mm), performance indicators (tensile strength ≥900MPa, fatigue life ≥10^5 cycles), environmental requirements (600℃ high temperature), mass production cycle ≤7 days, dimensional accuracy ±0.1mm; Preprocessing: The tensile strength (800-1000MPa) is normalized to [0,1]. The correlation between the blade tip region (high-temperature service area) and the high-temperature strength is established through coordinate mapping, and missing SLM process parameters (such as scanning speed) are supplemented.

[0040] S2, Mechanistic Constraint Construction: The knowledge base is used to call the Ti-6Al-4V phase transformation rules (the β→α phase transformation rate must satisfy the Avrami equation: X(t)=1-exp(-ktⁿ), n=2.5, to ensure that the microstructure is a fine-grained α+β phase) and the molten pool solidification constraints (the molten pool temperature must be ∈[1650,1750]℃ to avoid overheating and resulting in coarse grains). Transformed into mathematical constraints: Al∈[5.5,6.5]wt%, V∈[3.5,4.5]wt% (to ensure alloy phase stability), laser power P∈[180,220]W, scanning speed v∈[800,1000]mm / s (derived from the molten pool temperature formula).

[0041] S3, Multimodal Feature Fusion: Encoding: Transformer extracts the 3D structural features of the blade (outputting a 512-dimensional vector, focusing on the tip thickness), wavelet transform processes the temperature-intensity data (outputting a 128-dimensional vector), and a fully connected layer processes the laser power / scanning speed (outputting a 64-dimensional vector). Cross-attention: Strengthen the correlation between "tip high temperature intensity - Al content - laser power" (weight 0.7) to ensure high-temperature service performance; Optimization: Due to the initial laser power of 230W causing excessively high molten pool temperature, the feature weight was dynamically reduced to 200W.

[0042] S4, Initial scheme generation: Target Analysis: The optimization target is to transform "withstand high temperature of 600℃" into "tensile strength ≥ 850MPa at 600℃". The variational autoencoder generates 25 initial schemes, such as scheme B: Al=6.2wt%, V=4.0wt%, laser power 200W, scanning speed 900mm / s, layer thickness 40μm, predicted tensile strength 920MPa, fatigue life 1.1×10^5 cycles.

[0043] S5, Multi-objective Optimization and Output: After 100 iterations of NSGA-II, the optimal solution was selected: Al=6.0wt%, V=4.2wt%, laser power 210W, scanning speed 850mm / s, and layer thickness 35μm; Output results: tensile strength 930MPa (meets standards), fatigue life 1.2×10^5 cycles (meets standards), high temperature strength at 600℃ 860MPa (meets standards), mass production cycle 6 days (meets standards), dimensional accuracy ±0.08mm (meets standards), microstructure predicted as "fine-grained α+β phase (grain size ≤5μm)", process video simulation shows "stable molten pool, no spheroidization defects".

[0044] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined by the scope of the claims.

Claims

1. A multimodal fusion material design system based on material mechanism constraints, characterized in that, include: A multimodal data input module is used to acquire and preprocess vertical domain data related to material design. The vertical domain data includes, but is not limited to, two-dimensional / three-dimensional structural data of the product, material performance indicators, cost indicators, mass production requirements, environmental adaptability requirements, and service life indicators. The material mechanism constraint engine includes a vertical domain material mechanism knowledge base, a constraint transformation unit, and a dynamic verification unit. The constraint transformation unit converts the mechanism rules in the vertical domain material mechanism knowledge base into mathematical constraint formulas, and the dynamic verification unit detects the mechanism compliance of intermediate design results in real time. A multimodal feature fusion module, which includes an attention-based cross-modal fusion network for extracting multimodal features; Generative design unit, which generates material design schemes based on the fusion of features and mechanistic constraints; The design scheme output module outputs the final design scheme in a structured form. The output content includes, but is not limited to, material composition ratio, key process parameters, expected performance parameters, cost accounting report, material microstructure prediction results, process simulation video, mass production prediction report, mechanism feasibility analysis report, and service life assessment report.

2. The multimodal fusion material design system based on material mechanism constraints according to claim 1, characterized in that, The preprocessing steps of the multimodal data input module include: data cleaning, outlier removal, missing value completion, and data standardization. Modal alignment is achieved by establishing a spatial correlation mapping between two-dimensional / three-dimensional structural data and performance indicators.

3. The multimodal fusion material design system based on material mechanism constraints according to claim 1, characterized in that, The workflow of the dynamic verification unit includes: S41: Real-time acquisition of intermediate solutions output by generative design units; S42: Substitute the intermediate solution into the mathematical constraint formula corresponding to the mechanism rule and determine whether the constraint conditions are met. S43: If the constraints are met, the intermediate solution is passed to the multimodal feature fusion module for feature iteration optimization; if the constraints are not met, the reason for the violation is output, and a correction direction is generated based on the mechanism rules.

4. The multimodal fusion material design system based on material mechanism constraints according to claim 1, characterized in that, The multimodal feature fusion module includes an encoder, a cross-attention layer, and a feature optimization unit. The encoder is used to perform feature extraction on image files, text, and numerical data respectively. The cross-attention layer learns the association weights of different modal features to strengthen the strong correlation features of performance-component-process. The feature optimization unit dynamically adjusts the feature weight distribution in combination with the feedback of the mechanism constraint engine.

5. The multimodal fusion material design system based on material mechanism constraints according to claim 4, characterized in that, The encoder uses a CNN-Transformer hybrid architecture for feature extraction on two-dimensional / three-dimensional structural data. In the CNN-Transformer hybrid architecture, the CNN module extracts local structural features, and the Transformer module captures global structural relationships. The encoder uses a BERT pre-trained model for semantic encoding on textual data. The encoder uses a fully connected layer + batch normalization for feature mapping on numerical data.

6. The multimodal fusion material design system based on material mechanism constraints according to claim 4, characterized in that, The cross-attention layer calculates the correlation degree of different modal features through the attention weight matrix, wherein the correlation weight coefficient between the performance index modality and the component modality is not less than 0.6, ensuring that performance requirements guide the component design in a targeted manner.

7. The multimodal fusion material design system based on material mechanism constraints according to claim 1, characterized in that, The generative design unit includes a target performance analysis module, a scheme generation network, and a multi-objective optimizer. The training process of the scheme generation network includes: first, pre-training using labeled data from a public materials database, and then fine-tuning using private manufacturing process data from vertical enterprises. During the training process, a mechanism constraint loss function is introduced to reduce the mechanism violation rate of the generated schemes.

8. The multimodal fusion material design system based on material mechanism constraints according to claim 1, characterized in that, The design output module supports multiple output formats, including: structured documents, visual charts, 3D model files, and process video files.

9. A multimodal fusion material design method based on material mechanism constraints, characterized in that, Includes the following steps: S1: Multimodal vertical domain data acquisition and preprocessing: Collect data on product structure, performance, cost, environmental adaptability, etc. through the multimodal data input module, and perform cleaning, standardization, and modal alignment processing; S2: Construction of vertical material mechanism constraints: Based on the knowledge base of the material mechanism constraint engine, the rules such as crystal structure stability and phase equilibrium are transformed into mathematical constraints, and a mechanism verification rule base is constructed. S3: Multimodal Feature Fusion and Optimization: Modality-specific encoders are used to extract features of each modality. Feature association weights are learned through cross-attention layers. The feature weights are dynamically adjusted in combination with feedback from the mechanism constraint engine, and the optimized fused features are output. S4: Initial Design Scheme Generation: The target performance analysis module of the generative design unit transforms user requirements into optimization objectives, and the scheme generation network generates multiple sets of initial design schemes based on fusion features and mechanism constraints. S5: Multi-objective optimization and solution output: The multi-objective optimizer performs multi-objective optimization of the initial solution in terms of performance, cost and mass production, selects the non-dominated optimal solution, and outputs the final solution in a structured form by the design solution output module, along with a mechanism feasibility report.

10. The multimodal fusion material design method based on material mechanism constraints according to claim 9, characterized in that, The multi-objective optimization process in step S5 includes: S51: Set the weight coefficients for each objective, using "performance compliance rate", "cost control rate" and "mass production qualification rate" as optimization targets; S52: Iterative optimization is performed using a non-dominated sorting genetic algorithm, with no less than 100 iterations and a population size of 50-100. S53: Select 3-5 optimal solutions based on congestion distance calculations, allowing users to choose according to their actual needs.

Images