An artificial intelligence driven additive manufacturing dual-phase high-entropy alloy preparation method and system

Abstract

The application provides an artificial intelligence driven additive manufacturing two-fold heterogeneous high-entropy alloy preparation method and system, and belongs to the field of high-temperature alloy materials and additive manufacturing technology. The method comprises the following steps: constructing an additive manufacturing experiment database; extracting microstructure image features and identifying two-fold heterogeneous structures by using a convolutional neural network; establishing and training a printing parameter-structure mapping model; reversely searching for an optimal printing parameter combination; performing additive manufacturing experiment verification; feeding back experimental data and iteratively updating the model until convergence. The application realizes accurate design of two-level heterogeneous structures through an AI algorithm, accurately regulates microstructures by combining additive manufacturing technology, effectively suppresses the medium-temperature brittleness of high-entropy alloys, realizes synergistic improvement of high strength and excellent plasticity in a wide temperature range of-196 DEG C to 900 DEG C, and simultaneously replaces a traditional trial-and-error method with an AI model, greatly shortens a research and development cycle, and reduces preparation cost.

Description

Technical Field

[0001] This invention belongs to the field of high-temperature alloy materials and additive manufacturing technology, and relates to an artificial intelligence-driven additive manufacturing method and system for preparing dual heterogeneous high-entropy alloys. Background Technology

[0002] Nickel-based high-entropy alloys, with their superior high-temperature stability, synergistic strengthening and toughening properties, and resistance to corrosion in extreme environments due to their high mixing entropy, have become core structural materials for high-end equipment in aerospace, outer space exploration, and deep-sea mining. The structural components of such equipment need to operate under extreme conditions across a wide temperature range of -196℃ to 900℃ for extended periods, placing stringent requirements on materials that combine high strength with excellent ductility. The multi-principal element design characteristics of high-entropy alloys give them a natural advantage in controlling mechanical properties over a wide temperature range, making them a mainstream candidate material for meeting these extreme service requirements and a key material support for upgrading high-end equipment to operate under extreme conditions.

[0003] Currently, precipitation-strengthened nickel-based alloys have achieved a high-temperature tensile strength of over 800 MPa above 750℃ by increasing the volume fraction of the γ′ phase to 55%. After optimization by grain boundary engineering, their elongation at break at 800℃ has increased from 7.1% to over 20%, and they have been successfully used in advanced aero-engine turbine disks and combustion chamber liners. However, while existing nickel-based high-entropy alloy systems each possess their own performance advantages, they still struggle to directly meet the comprehensive performance requirements of high strength, excellent plasticity, and absence of mid-temperature brittleness under extreme working conditions across a wide temperature range of -196℃ to 900℃. Alloys optimized for low / room temperatures (such as Inconel 625) exhibit excellent plasticity (elongation exceeding 25%), but their strength decreases rapidly above 600℃. Alloys optimized for high temperatures (such as RR1000) maintain stable strength above 750℃, but exhibit low plasticity (elongation <10%) from room temperature to 600℃. Even the GH4151 alloy optimized through grain boundary engineering shows a loss of about 15% in strength while plasticity improves in the ultra-high temperature range above 1000℃, indicating that a balance between strength and plasticity across a wide temperature range and all working conditions has not yet been achieved.

[0004] The regulation of heterogeneous structures is the core technical means to solve the above problems. Existing high-entropy alloy single homogeneous structure designs can only improve material properties in specific temperature ranges. At high temperatures, they are prone to grain coarsening and failure of strengthening mechanisms, and cannot fundamentally solve the problems of strain localization and crack propagation at medium temperatures. Heterogeneous structures, on the other hand, can induce heterogeneous deformation-induced strengthening and strain delocalization toughening effects through the differences in deformation behavior of different structural units, strengthening grain boundaries and suppressing or alleviating stress concentration. Back stress promotes the continuous storage of dislocations, uniform strain distribution, delays necking, and improves uniform elongation, achieving a synergistic improvement in strength and plasticity. It can both improve alloy strength by hindering dislocation movement through structural interfaces and guide the dispersion of dislocations to improve plasticity and suppress crack initiation and propagation, achieving a synergistic improvement in strength and plasticity over a wide temperature range.

[0005] Traditional processes such as smelting, forging, and cold rolling struggle to achieve the directional and gradient construction of heterogeneous structures, often resulting in issues like structural homogenization and interfacial stress concentration. In contrast, laser powder bed fusion (L-PBF) additive manufacturing offers advantages such as layer-by-layer forming and controllable thermal fields, enabling precise matching of design requirements for two heterogeneous structures. By adjusting printing parameters and strategies, precise control over the microstructure can be achieved. However, L-PBF involves numerous adjustable parameters, including laser power, scanning speed, and scanning strategy, making coordinated control difficult. Therefore, machine learning methods are needed for this control, enabling multi-objective collaborative optimization.

[0006] Traditional high-entropy alloy structure design methods mainly rely on the "trial and error" approach, which has a long development cycle, high experimental costs, and low efficiency. It is known in the industry for the pain points of "long cycle, high cost, and low efficiency" in materials research and development. This inefficient research and development model is difficult to adapt to the current high-end equipment's rapid research and development needs for high-entropy alloys for extreme environments. There is an urgent need to introduce a new design paradigm to accelerate the research and development process of high-entropy alloy materials.

[0007] In recent years, with the rapid development of artificial intelligence (AI) technology, data-driven heterogeneous structure design technology has gradually attracted widespread attention from academia and industry. Researchers can leverage the powerful data processing and pattern recognition capabilities of AI algorithms to uncover the complex structure-property relationships in high-entropy alloys, overcoming the limitations of traditional empirical models in complex structure design and enabling a shift from trial-and-error to on-demand customized heterogeneous structures. Among these, convolutional neural networks (CNNs), as one of the core models of deep learning, have demonstrated significant advantages in process optimization and performance prediction in additive manufacturing due to their powerful image feature extraction and nonlinear mapping capabilities. In additive manufacturing processes such as laser powder bed melting, CNNs can accurately analyze microstructure images, automatically extracting key features such as grain size distribution, phase distribution, and microstructure inhomogeneity, thereby establishing a quantitative mapping relationship between process parameters and microstructure, providing data support for precise control of the processing. For example, in related research, CNNs have been successfully applied to scenarios such as molten pool state classification and forming quality prediction. By performing deep learning on image data collected during processing, effective prediction of process stability and component performance has been achieved.

[0008] It is worth noting that although convolutional neural networks (CNNs) have made some progress in the application of additive manufacturing, their deep integration into the entire process of high-entropy alloy heterogeneous structure design and fabrication still has significant shortcomings. Existing research mostly focuses on optimizing single process parameters or predicting local performance, failing to fully leverage the advantages of CNNs in multi-objective collaborative optimization and quantitative design of complex structures. A complete intelligent solution encompassing the entire chain from heterogeneous structure design and process parameter matching to performance verification has not yet been formed. Furthermore, given the unique characteristics of wide-temperature-range performance control in high-entropy alloys, there is a lack of dedicated CNN model training systems, resulting in insufficient depth in the models' understanding of the complex relationships between process, structure, and performance, and making it difficult to meet engineering requirements in terms of prediction accuracy and practicality. Summary of the Invention

[0009] In view of this, the purpose of this invention is to provide an artificial intelligence-driven additive manufacturing method and system for preparing dual heterogeneous high-entropy alloys. The aim is to achieve precise design of the dual heterogeneous structure through AI algorithms, and to achieve precise control of the microstructure of high-entropy alloys by combining additive manufacturing technology. This effectively suppresses the mid-temperature brittleness of high-entropy alloys and achieves a synergistic improvement in high strength and excellent uniform plasticity of the alloy in a wide temperature range of -196℃ to 900℃. At the same time, by replacing the traditional experience-based "trial and error" method with an AI model, the development cycle of high-entropy alloys is significantly shortened and the preparation cost is reduced.

[0010] To achieve the above objectives, the present invention provides the following technical solution: An artificial intelligence-driven additive manufacturing method for preparing dual heterogeneous high-entropy alloys, the method specifically includes the following steps: S1. Construct an additive manufacturing experimental database, the database containing multiple samples, each sample including a set of printing parameters, a microstructure image of a high-entropy alloy sample prepared under the set of printing parameters, and microstructure information extracted from the microstructure image; S2. Construct a convolutional neural network model, use the convolutional neural network model to extract features from the microscopic tissue image to obtain tissue feature vectors, and determine whether the microscopic tissue image meets the preset requirements for two heterogeneous tissues based on the tissue feature vectors, and generate two heterogeneous tissue labels. S3. Establish and train a printing parameter-organizational structure mapping model. The model is trained with printing parameters as input and the organization feature vector and the two heterogeneous organization labels as supervision information. It is used to predict the organization features that can be obtained under given printing parameters and the probability of forming two heterogeneous organizations. S4. Set the target organization feature vector, and use the trained printing parameter-organization structure mapping model to perform a reverse search within the feasible domain of printing parameters to obtain the optimal combination of printing parameters that makes the predicted organization feature vector approach the target organization feature vector and maximizes the probability of forming two heterogeneous organizations. S5. Using the optimal combination of printing parameters, a high-entropy alloy sample is prepared by additive manufacturing process, and the prepared sample is characterized by microstructure to obtain experimental microstructure images and experimental structure feature vectors. S6. The optimal combination of printing parameters, experimental microscopic tissue images, experimental tissue feature vectors and corresponding dual heterogeneous tissue labels obtained in step 5 are added as new samples to the database of step 1, and the printing parameter-tissue structure mapping model in step 3 is iteratively updated and trained. S7. Repeat steps 4 to 6 until the error between the model prediction result and the experimental result converges to within the preset threshold, thereby achieving the stable preparation of the dual heterogeneous high-entropy alloy.

[0011] Furthermore, in step S1, each set of printing parameters includes at least laser power, scanning speed, scanning spacing, powder layer thickness, and scanning path mode; the microstructure image undergoes grayscale normalization preprocessing; the structure information includes at least grain size, structure gradient, multi-scale structure, phase distribution, and structure inhomogeneity.

[0012] Furthermore, in step S1, establishing the additive manufacturing experimental database of "printing parameters - microstructure images - structural features" specifically includes: The print parameter vector of the i-th sample is represented as: The scan path is represented using one-hot encoding as follows: Based on this, construct the original database: To eliminate the influence of dimensions, the printing parameters are normalized: Gray-level normalization processing was performed on the microscopic tissue images: in: For the first Print parameter vectors for each sample. For laser power, For scanning speed, For scanning spacing, To achieve a thicker powder layer, Encode vectors for scanning path methods. This represents the total number of categories for scanning path methods. For the first Does the _ sample belong to the _ ... Class scan path marker variable, For the original training database, For the first Microscopic tissue images corresponding to each sample The total number of samples, For the first In the nth sample One printing parameter, These are the normalized printing parameters. For the first Each sample image at pixel coordinates grayscale value at that location These are the minimum and maximum grayscale values ​​of the image. The normalized image grayscale values. These are the pixel coordinates of the image.

[0013] Further, in step S2, the method for determining the requirement of dual heterogeneous structure includes: dividing the microstructure image into multiple sub-regions, calculating the average grain size of each sub-region and the overall average grain size, and constructing a spatial gradient heterogeneity index accordingly; simultaneously, calculating the ratio of the 90th percentile to the 10th percentile of the grain size distribution as a multi-scale heterogeneity index; when both the spatial gradient heterogeneity index and the multi-scale heterogeneity index exceed a preset threshold, it is determined that the requirement of dual heterogeneous structure is met.

[0014] Further, in step S2, the obtained microscopic tissue image is input into a convolutional neural network, and layer-by-layer feature extraction is performed on the image using convolutional layers, pooling layers, and nonlinear activation functions to obtain quantitative expressions of tissue features such as grain size, tissue gradient, multi-scale structure, phase distribution, and tissue inhomogeneity, including: The feature extraction process of a convolutional layer can be represented as follows: The output of the pooling layer is: After multiple convolutions and pooling, a high-dimensional feature vector of the microscopic tissue image is obtained: To achieve the discrimination of two heterogeneous tissues, spatial gradient heterogeneity index and multi-scale heterogeneity index are further constructed to divide the microscopic tissue image into... Each sub-region has an average grain size of [number] grains. The overall average grain size is: The spatial gradient heterogeneity index is defined as: Further define the two heterogeneous organization tags: in: : No. The first in the layer Feature maps output by each convolutional kernel; : No. The first in the layer One input feature map; : No. Layer connection The input channel and the first The convolutional kernel weights for each output channel; : No. Layer The bias term corresponding to each output channel; : Non-linear activation function; : No. Number of input channels in the layer; : No. Layer Each pooled output feature map; Pooling function; Convolutional neural network encoder; Model parameters of a convolutional neural network encoder; : No. The tissue feature vector of each sample; : The normalized first Microscopic tissue images of individual samples; The number of sub-regions in the image; : No. The sample at the th Average grain size in each sub-region; : No. Overall average grain size of the sample; : No. Spatial gradient heterogeneity index for each sample; : No. The 90th percentile value in the grain size distribution of each sample; : No. The 10th percentile value in the grain size distribution of each sample; : No. Multiscale heterogeneous indices for a sample.

[0015] Furthermore, steps S3 and S4 specifically include: Extracted tissue feature vectors and two heterogeneous tissue labels As supervisory information, a prediction model is established with printing parameters as input and tissue features and two heterogeneous discrimination results as output. The model preferably has a structure of "parameter input layer + fully connected mapping layer + classification output layer," where a convolutional neural network is used for image feature encoding, and a parameter mapping network is used to predict the tissue features from printing parameters. Specifically, this includes: The mapping relationship between printing parameters and tissue characteristics is as follows: The probability of the formation of two heterogeneous tissues is: The organizational feature regression loss is: The classification loss for two heterogeneous tissues is: The total loss function is: in: After training is completed, the target dual heterogeneous organization feature vector is defined. The optimal combination of printing parameters is obtained through parameter space search: And satisfy: in: : A mapping model from printing parameters to organizational features; : Mapping model parameters; The model predicts the first Each sample organization feature vector; Two-dimensional heterogeneous tissue classifier; Classifier parameters; The model predicts the first The probability that a sample forms two heterogeneous organizations; Organizational feature regression loss; : L2 norm; Organizational classification loss; The model's total loss function; The weighting coefficients for regression loss and classification loss; Regularization coefficient; The complete set of parameters to be optimized in the model; : Feature vector of the target's two heterogeneous organizations; The optimal combination of printing parameters obtained by reverse calculation; : Print the feasible field of parameters; : Weighting coefficients of the probability constraint term for the formation of two heterogeneous organizations; Weighting coefficients for process stability constraints; : Refer to the printing parameter vector to limit the search results from deviating too much from the existing stable process window; : The lower and upper bounds of the values ​​for each printing parameter.

[0016] Furthermore, in step S5, the obtained optimal combination of printing parameters is used. Additive manufacturing experiments were conducted to prepare new high-entropy alloy samples, and experimental microstructure images were obtained. This includes: inputting the image into a convolutional neural network encoder to extract the feature vector of the experimental tissue. Simultaneously, the corresponding predicted tissue feature vector is obtained from the prediction model: The relative error between the experimental results and the model predictions is defined as: The dual isomeric tissue labels of the experimental samples were calculated according to the aforementioned criteria. ;when and If the selected combination of printing parameters is valid, then the parameter search range is corrected, and the parameter search and experimental verification are re-executed. The correction for the parameter search range is expressed as follows: in: Microscopic images of the experimentally prepared samples; Normalized experimental microscopic tissue images; : Tissue feature vectors extracted from experimental samples; The model prints the optimal combination of parameters. The predicted tissue feature vector; The relative error between experimental results and predicted results; : Allowable error threshold; : The dual isomeric tissue label of the experimental sample; : No. In the first round of search The upper and lower boundaries of each printing parameter; The first in the optimal combination of printing parameters Each parameter value; : No. Wheel of Life The search half-width for each parameter; Iteration rounds; Print parameter dimension number.

[0017] Furthermore, in step S6, the newly obtained experimental data, including the optimal combination of printing parameters, corresponding microscopic tissue images, tissue feature vectors, and two heterogeneous tissue labels, are added to the original training database to construct the expanded database: The model parameters are retrained based on the expanded database, and the parameter update process is represented as follows: To determine whether the closed-loop optimization process has converged, the rate of change of loss between two adjacent training rounds is defined as: When the following conditions are met: When the closed-loop iteration between the model and the experiment reaches a convergence state, the stable preparation of the dual heterogeneous high-entropy alloy microstructure is achieved. in: : No. The training database corresponding to each iteration; : No. Print the parameter vector of each newly added experimental sample; : No. Microscopic tissue images of newly added experimental samples; : No. The tissue feature vector of the newly added experimental samples in the round; : No. Two heterogeneous tissue labels for newly added experimental samples; , : No. Wheel and First Wheel model parameters; Learning rate; Regarding model parameters The gradient operator; , : No. Wheel and First The loss function value obtained from the first round of training; : Rate of change of loss between two adjacent rounds; Loss convergence threshold; The relative error between the experiment and the prediction; Error tolerance threshold.

[0018] This invention also provides an artificial intelligence-driven additive manufacturing system for preparing dual heterogeneous high-entropy alloys, used to perform the above-described method. The system includes: The database construction module is used to collect and store printing parameters, microscopic images and tissue structure information to build an additive manufacturing experimental database. The image feature extraction module has a built-in convolutional neural network model for extracting features from microscopic tissue images and outputting tissue feature vectors and the results of two heterogeneous tissue discrimination. The mapping model module is used to establish and store the mapping relationship between printing parameters, tissue feature vectors, and labels of two heterogeneous tissues, and can predict the output tissue features and the probability of forming two heterogeneous tissues based on the input printing parameters. The parameter optimization module is used to organize feature vectors according to the set target and use the mapping model module to search for the optimal combination of printing parameters in reverse. An additive manufacturing execution module is used to prepare high-entropy alloy samples using laser powder bed melting technology based on the optimal combination of printing parameters output by the parameter optimization module. The feedback update module is used to supplement the database construction module with newly added experimental data and trigger the iterative retraining of the mapping model module until the prediction error converges.

[0019] Furthermore, the convolutional neural network model in the image feature extraction module includes multiple alternating stacked convolutional and pooling layers, as well as a fully connected encoder; the convolutional layers are used to extract local features of microscopic tissue images, the pooling layers are used to reduce the feature dimension, and the fully connected encoder is used to map the extracted multidimensional features into a fixed-length tissue feature vector.

[0020] The beneficial effects of this invention are as follows: 1) This invention effectively suppresses the nucleation and propagation of intergranular cracks in the 900℃ temperature range by precisely constructing a dual heterogeneous structure driven by AI, and significantly improves the uniform elongation of the alloy in this temperature range, achieving several times the improvement compared to traditional homogeneous high-entropy alloys, and successfully achieving the transformation from brittle to tough. The alloy has both high strength and excellent uniform plasticity in a wide temperature range from ultra-low temperature to 900℃. The strength-plasticity product in the 900℃ temperature range is significantly improved compared to traditional homogeneous high-entropy alloys, realizing the synergistic optimization of strength and plasticity in a wide temperature range.

[0021] 2) This invention replaces the traditional trial-and-error method with an AI model, realizing the efficient design of the dual heterogeneous structure of high-entropy alloys and the precise optimization of process parameters, which greatly shortens the R&D cycle, reduces experimental costs, and ensures the stability and consistency of the performance of high-entropy alloy products. The dual heterogeneous structure can generate a synergistic heterogeneous deformation-induced strengthening effect and a two-level strain delocalization toughening effect, which effectively solves the problem of the effectiveness decay of a single strengthening mechanism in a wide temperature range, and ensures that the alloy maintains stable mechanical properties under extreme temperature conditions.

[0022] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of the closed-loop mode technology of the method of the present invention. Detailed Implementation

[0024] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0025] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0026] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present 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. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0027] This invention presents an innovative method that integrates artificial intelligence and additive manufacturing technologies into the entire process of designing, preparing, and controlling the performance of heterogeneous structures. It constructs a closed-loop optimization procedure of "AI heterogeneous structure design - in-situ additive manufacturing - performance characterization - model feedback optimization". Through multi-round iterative optimization using convolutional neural networks, it achieves precise control of the dual heterogeneous structure and preparation process of high-entropy alloys, providing a new data-driven path for the research and development of high-entropy alloys for extreme environments.

[0028] Figure 1 This is a flowchart illustrating the closed-loop mode technology of the method of the present invention. The present invention deeply integrates artificial intelligence and additive manufacturing technology, and applies it to the entire process of designing and preparing multiple heterogeneous structures of high-entropy alloys, providing a novel method for optimizing the performance of high-entropy alloys. As shown in the figure, the specific steps are as follows: Step 1: Alloy ingot smelting and powder sample preparation Weigh out raw materials such as Ni, Co, Cr, Fe, and Al with a purity ≥ 99.95 wt.% according to the alloy composition ratio. Place the pretreated raw materials into the copper crucible of the electric arc melting furnace and evacuate to 10 °C. -3 Below Pa, 99.999% high-purity argon gas is introduced, and Ti blocks are used as oxygen scavengers (Ti as the base) for arc discharge melting. After the raw materials are completely melted, they are held at the temperature for 2 minutes, cooled, and then the ingot is turned over and melted repeatedly 8 times to ensure the uniformity of the alloy composition. The ingot is then sent to a powder-making device filled with an inert gas protective atmosphere to prepare spherical powder with a particle size of 15-53 μm.

[0029] Step 2: Construction of Additive Manufacturing Experimental Database Through literature review and preliminary experiments, parameters such as different laser powers, scanning speeds, scanning intervals, slice thicknesses, and scanning paths were collected. Based on the scanning method, corresponding microscopic tissue images and tissue structure information were established, creating a training database of "printing parameters - microscopic tissue images - tissue structure characteristics".

[0030] The print parameter vector of the i-th sample is represented as: The scan path is represented using one-hot encoding as follows: Based on this, construct the original database: To eliminate the influence of dimensions, the printing parameters are normalized: Gray-level normalization processing was performed on the microscopic tissue images: in: : No. Print parameter vectors for each sample. Laser power. Scanning speed. Scanning interval. Thick powder layer. : Scan path encoding vector. Total number of categories of scanning path methods. : No. Does the _ sample belong to the _ ... A marker variable for the class scan path. : Original training database. : No. Microscopic tissue images corresponding to each sample. Total number of samples. : No. In the nth sample Print parameters. Normalized print parameters. : No. Each sample image at pixel coordinates The grayscale value at that location. Minimum and maximum grayscale values ​​of the image. : The normalized grayscale value of the image. Image pixel coordinates.

[0031] Step 3: Convolutional Neural Network Model Construction and Feature Extraction In this invention, the microscopic tissue image is a single-channel or multi-channel image I acquired under the same field of view and registered. i∈R^(H×W×C), where H and W are the image height and width, respectively, and C is the number of image channels. The image channels include one or more of the following: SEM image, BSE image, EBSD grain orientation map, EBSD phase diagram, or EDS element distribution map. After size unification, field registration, and normalization of each image channel, it is used as the input to the convolutional neural network. The training labels corresponding to the microstructure image include the grain boundary mask B. i (u,v), Phase type mask S i (u,v), average grain size of subregion Phase area fraction f i k and the spatial gradient heterogeneity index G calculated from the above physical quantities. i Multiscale heterogeneous index M i and tissue heterogeneity index U i .

[0032] The convolutional neural network includes a shared encoder, a grain size regression branch, a phase distribution segmentation branch, an organization feature vector regression branch, and a dual heterogeneous classification branch. The grain size regression branch outputs a pixel-level grain size map D. i The average grain size of (u,v) or R subregions The pixel-level probability P of K-type phases output by the phase distribution segmentation branch. i ^phase(u,v,k) is used to obtain the phase category mask S. i (u,v); Output of the regression branch of the organizational feature vector The output of the dual heterogeneous classification branch is the probability that the i-th sample meets the dual heterogeneous organization requirement. .

[0033] During the training sample annotation process, a grain boundary mask B is first obtained based on the EBSD grain orientation map or the etched metallographic / SEM image. i (u,v) is used to determine the area A of each grain by combining automatic segmentation and manual correction. The grain size is calculated using the equivalent circle diameter d = 2√(A / π). Then, the microstructure image is divided into R sub-regions, and the average grain size in each sub-region is calculated. Phase type mask S i (u,v) are determined by EBSD phase diagrams, EDS element distribution maps, or BSE image grayscale / texture segmentation results, and are used as phase distribution segmentation labels after manual correction. (Based on S...) i (u,v) Calculate the area fraction f of the k-th phase in the overall image and in each sub-region. i k and f i,r,k. Spatial gradient heterogeneity index G i Multiscale heterogeneous index M i and tissue heterogeneity index U i All results were calculated from the above-mentioned annotations using formulas.

[0034] Model training uses a multi-task joint loss function L total =λ_D L_D+λ_S L_S+λ_z L_z+λ_y L_y+λ_R||Θ|| 2 Where L_D is the mean squared error loss for grain size regression, L_S is the cross-entropy loss or Dice loss for phase distribution segmentation, L_z is the loss for microstructure feature vector regression, L_y is the cross-entropy loss for two-fold heterogeneous microstructure classification, λ_D, λ_S, λ_z, λ_y, and λ_R are the weight coefficients of each loss term, and Θ is the set of parameters to be trained on the model. Through the above joint training, a supervised correspondence is established between the model output and measurable microstructure parameters such as grain size, phase distribution, microstructure gradient, multi-scale structure, and microstructure inhomogeneity.

[0035] The microstructure image obtained in step 2 is input into a convolutional neural network. Convolutional layers, pooling layers, and nonlinear activation functions are used to extract features from the image layer by layer to obtain quantitative expressions of structural features such as grain size, structure gradient, multi-scale structure, phase distribution, and structure inhomogeneity.

[0036] The feature extraction process of a convolutional layer can be represented as follows: The output of the pooling layer is: After multiple convolutions and pooling, a high-dimensional feature vector of the microscopic tissue image is obtained: To distinguish between "two heterogeneous organizations", spatial gradient heterogeneity index and multi-scale heterogeneity index can be further constructed.

[0037] This step employs a "block-based sensing" strategy, dividing the microstructure image into R sub-regions. The regression branch acts as a "physical quantity decoder," quantitatively mapping the feature vector Zi in physical space and directly outputting the average grain size of each sub-region. .Should Instead of being measured manually, the physical parameters are automatically identified and quantified by the AI ​​model, establishing a direct quantitative correlation between pixel features and physical indicators.

[0038] The overall average grain size is: The spatial gradient heterogeneity index is defined as: Further define the two heterogeneous organization tags: in: : No. The first in the layer Convolution Feature map of kernel output. : No. The first in the layer Input feature maps. : No. Layer connection The input channel and the first The weights of the convolution kernels for each output channel. : No. Layer The bias term corresponding to each output channel. : Non-linear activation function. : No. Number of input channels for the layer. : No. Layer Each pooled output feature map. Pooling function. : Convolutional Neural Network Encoder. : Model parameters of the convolutional neural network encoder. : No. The tissue feature vector of each sample. : The normalized first Microscopic tissue images of individual samples. The number of sub-regions into which the image is divided. : No. The sample at the th Average grain size in each sub-region. : No. The overall average grain size of the sample. : No. The 90th percentile value of the grain size distribution of each sample. : No. The 10th percentile value of the grain size distribution of each sample.

[0039] In this invention, the multi-scale heterogeneous index M i This is used to characterize the size span between coarse-grained and fine-grained regions within the same sample. The 90th percentile value d90 of the grain size distribution is obtained by statistically analyzing all grain sizes d of the i-th sample. i And the 10th percentile value d10, iAnd according to Mᵢ=(d90, i +ε) / (d10, i +ε) Calculates the multiscale heterogeneity index, where ε is a very small positive number to prevent the denominator from being zero. M i The larger the value, the more significant the size difference between coarse and fine grains in the microstructure. Spatial gradient isomerism index G i M is used to characterize grain size differences in different spatial sub-regions; multi-scale heterogeneity index i Used to characterize the width of grain size distribution. Therefore, this invention will simultaneously satisfy G. i ≥G0 and M i Microstructures with a value ≥ M0 are defined as doubly heterogeneous structures and are denoted as y. i =1; otherwise, record it as y. i =0.

[0040] G0 and M0 are the threshold values ​​for determining two-fold isomorphic tissues, which can be determined based on a homogeneous or weakly isomorphic control sample library. Specifically, G0 in the control sample is calculated. i and M i The mean values ​​μ_G^ref and μ_M^ref, and the standard deviations σ_G^ref and σ_M^ref are calculated, and G0 = μ_G^ref + κ_Gσ_G^ref and M0 = μ_M^ref + κ_Mσ_M^ref are set, where κ_G and κ_M are threshold adjustment coefficients. Alternatively, G0 and M0 can be determined statistically based on a set of positive samples that meet the target mechanical performance requirements. Through this method, the two heterogeneous tissue labels are automatically generated from measurable tissue parameters and preset thresholds.

[0041] , : Threshold for determining two heterogeneous tissues. i : The dual heterogeneous tissue label y of the i-th sample i =1; indicates that the requirement of dual heterogeneous organization is met, y i =0 indicates that the condition is not met.

[0042] Step 4: Print parameters - training the organization structure mapping model The tissue feature vector Z extracted in step 3 i and two heterogeneous tissue labels y i As supervisory information, a prediction model is established with printing parameters as input and tissue features and two heterogeneous discrimination results as output. The model is preferably structured as "parameter input layer + fully connected mapping layer + classification output layer", where the convolutional neural network is used for image feature encoding and the parameter mapping network is used to predict the tissue features from the printing parameters.

[0043] The mapping relationship between printing parameters and tissue characteristics is as follows: The probability of the formation of two heterogeneous tissues is: The organizational feature regression loss is: The classification loss for two heterogeneous tissues is: The total loss function is: in: After training is completed, the target dual heterogeneous organization feature vector is defined. The optimal combination of printing parameters is obtained through parameter space search: And satisfy: in: : A mapping model from printing parameters to organizational characteristics. : Mapping model parameters. The model predicts the first Each sample organizes a feature vector. : Two heterogeneous organization classifiers. : Classifier parameters. The model predicts the first The probability of a sample forming two heterogeneous organizations. Organizational feature regression loss. : L2 norm. Organizational classification loss. : The total loss function of the model. : The weighting coefficients of regression loss and classification loss. Regularization coefficient. The set of all parameters to be optimized in the model. : Target dual heterogeneous organization feature vector. The optimal combination of printing parameters is obtained by reverse calculation. : Print the feasible field of parameters. : The weighting coefficient of the probability constraint term for the formation of two heterogeneous organizations. : Weighting coefficient of the process stability constraint term. : Refer to the print parameter vector to limit the search results from deviating too much from the existing stable process window. : The lower and upper bounds of the values ​​for each printing parameter.

[0044] Step 5: Experimental Validation of Additive Manufacturing Guided by Model Predicted Parameters The optimal combination of printing parameters obtained in step 4 is used in reverse calculation. Additive manufacturing experiments were conducted to prepare new high-entropy alloy samples, and experimental microstructure images were obtained. The image is then input into the convolutional neural network encoder in step 3 to extract the feature vector of the experimental tissue. Simultaneously, the corresponding predicted tissue feature vector is obtained from the prediction model in step 4: The relative error between the experimental results and the model predictions is defined as: Calculate the dual isomeric tissue labels of the experimental samples according to the criteria in step 3. .when and If the selected combination of printing parameters is valid, then the parameter search range is corrected, and the parameter search and experimental verification are re-executed.

[0045] The correction of the parameter search range can be expressed as: in: Microscopic images of the experimentally prepared samples. Normalized experimental microscopic tissue images. : Tissue feature vectors extracted from experimental samples. The model prints the optimal combination of parameters. The predicted tissue feature vector. The relative error between experimental results and predicted results. : Allowable error threshold. : Two heterogeneous tissue labels of experimental samples. : No. In the first round of search The upper and lower boundaries of each printing parameter. The first in the optimal combination of printing parameters Each parameter value. : No. Wheel of Life The search half-width of each parameter. Iteration rounds. Print parameter dimension number.

[0046] Step 6: Experimental Data-Driven Iterative Update of Convolutional Neural Network The new experimental data obtained in step 5, including the optimal combination of printing parameters, corresponding microscopic tissue images, tissue feature vectors, and two heterogeneous tissue labels, are added to the original training database to construct the expanded database: The model parameters are retrained based on the expanded database, and the parameter update process can be represented as follows: To determine whether the closed-loop optimization process has converged, the rate of change of loss between two adjacent training rounds is defined as: When the following conditions are met: When the closed-loop iteration between the model and the experiment reaches a convergence state, the stable preparation of the dual heterogeneous high-entropy alloy microstructure is achieved.

[0047] in: : No. The training database corresponding to each iteration. : No. Print the parameter vector of newly added experimental samples. : No. Microscopic tissue images of newly added experimental samples. : No. The tissue feature vectors of newly added experimental samples. : No. Two heterogeneous tissue labels for newly added experimental samples. , : No. Wheel and First Wheel model parameters. Learning rate. Regarding model parameters The gradient operator. , : No. Wheel and First The loss function value obtained from the training round. : Rate of change of loss between two adjacent rounds. Loss convergence threshold. The relative error between the experiment and the prediction. Error tolerance threshold.

[0048] Step 7: Microstructure characterization and wide-temperature-range mechanical property testing Metallographic specimens, electron backscatter diffraction specimens, and transmission electron microscope specimens were prepared according to standards. The microstructural parameters of the specimens, such as grain morphology, precipitate characteristics, grain boundary / interface structure, and dislocation configuration, were systematically characterized. Standard tensile specimens were prepared using an electrical discharge wire cutting machine. Wide-temperature-range uniaxial tensile tests were conducted using an Instron 8801 servo hydraulic testing machine (equipped with a high and low temperature environment chamber) at three temperature gradients: -196℃, 25℃, and 900℃. Load-unload-reload (LUR) tests and nanoindentation tests were carried out simultaneously to obtain core mechanical property parameters such as yield strength, tensile strength, and uniform elongation.

[0049] Example: 1) Implementation Method In this embodiment, firstly, high-entropy alloy structure-process-performance data covering the NiCoCrFe system are collected to construct an integrated database of NiCoCrFe high-entropy alloy structure-process-performance with a sample size of ≥200 groups. Taking deformation / recrystallization zone ratio, precipitated phase and grain boundary / interface characteristics, and process parameters as inputs, and mechanical properties from -196 to 900℃ as outputs, a deep learning + genetic algorithm AI optimization model is built by integrating CALPHAD phase diagram calculations. After multiple rounds of iteration, the model prediction deviation is reduced to ≤1%. After inputting into the design space, two heterogeneous structure design schemes and full-process process parameters are output.

[0050] 15-53μm nickel-based high-entropy alloy spherical powder was added to a laser selective melting device. The substrate was preheated to 200℃ and the forming cavity was filled with high-purity argon. A 67° interlayer rotation scanning strategy was adopted, and two sets of parameters (250W / 1400 mm / s and 350 W / 1000 mm / s) were used alternately in every 10 layers as the control unit. The layer thickness was fixed at 0.05 mm and the scanning interval was 0.12 mm. The in-situ forming of the sample was completed according to the optimal parameters output by AI.

[0051] Real data from microstructure characterization and wide-temperature-range mechanical property testing of the samples were extracted as new samples, the AI ​​model training set was updated and retrained and optimized, and the deviation between the model prediction and the experimental results was further narrowed.

[0052] 2) Experimental Results The prepared sample successfully formed a dual heterogeneous structure that highly matched the AI ​​design scheme. The micro parameters deviated very little from the design values. The proportion of deformed grain bands and recrystallized grain bands matched the design expectations. The key size parameters of the serrated grain boundaries met the design requirements. The proportion of precipitated phases was moderate. They were all L12 type structures and coherent with the face-centered cubic matrix. The interface mismatch was low. The composition of the sample micro-regions was uniform with no obvious segregation. The microstructure control effect of the in-situ process was significant.

[0053] The samples achieved a synergistic improvement in strength and plasticity across a wide temperature range from ultra-low temperatures to 900℃. Yield strength and uniform elongation at each temperature range met design specifications. High strength was maintained even at 900℃, while uniform elongation was effectively improved, demonstrating excellent synergistic strength and plasticity performance across a wide temperature range. Specifically, the yield strength was 1000-1500 MPa and the uniform elongation was 10-30% at -196℃; the yield strength was 800-1200 MPa and the uniform elongation was 10-20% at 25℃; and the strength-plasticity product in the mid-temperature range of 900℃ was increased by up to 200-500% compared to traditional homogeneous high-entropy alloys, demonstrating a synergistic improvement in strength and plasticity across a wide temperature range.

[0054] 3) Technical Effect Analysis This embodiment verifies the effectiveness of the closed-loop optimization procedure of "AI dual heterogeneous structure design - additive manufacturing in-situ forming - performance characterization - model feedback optimization". The prepared high-entropy alloy successfully solves the technical bottleneck of significant mid-temperature brittleness and difficulty in synergistic strength and plasticity in a wide temperature range of traditional high-entropy alloys. It achieves the transformation from brittleness to toughness in the mid-temperature range of 900℃, and maintains high strength and excellent plasticity in a wide temperature range of -196℃ to 900℃.

[0055] The deep integration of AI technology has enabled the precise design of high-entropy alloy structures and processes, replacing the traditional experience-based "trial and error" approach and significantly shortening the R&D cycle. Additive manufacturing avoids problems such as secondary deformation and compositional segregation in offline processing, ensuring the stable forming and consistent performance of the two heterogeneous structures.

[0056] The design and preparation method of this invention is not only applicable to NiCoCrFe nickel-based high-entropy alloys, but can also be extended to other precipitation-strengthened high-entropy alloy systems, providing an efficient and reliable new paradigm for the research and development of high-performance structural materials for extreme environments, and has both important scientific significance and engineering application value.

[0057] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for preparing a dual heterogeneous high-entropy alloy using artificial intelligence-driven additive manufacturing, characterized in that, The method specifically includes the following steps: S1. Construct an additive manufacturing experimental database, the database containing multiple samples, each sample including a set of printing parameters, a microstructure image of a high-entropy alloy sample prepared under the set of printing parameters, and microstructure information extracted from the microstructure image; S2. Construct a convolutional neural network model, use the convolutional neural network model to extract features from the microscopic tissue image to obtain tissue feature vectors, and determine whether the microscopic tissue image meets the preset requirements for two heterogeneous tissues based on the tissue feature vectors, and generate two heterogeneous tissue labels. S3. Establish and train a printing parameter-organizational structure mapping model. The model is trained with printing parameters as input and the organization feature vector and the two heterogeneous organization labels as supervision information. It is used to predict the organization features that can be obtained under given printing parameters and the probability of forming two heterogeneous organizations. S4. Set the target organization feature vector, and use the trained printing parameter-organization structure mapping model to perform a reverse search within the feasible domain of printing parameters to obtain the optimal combination of printing parameters that makes the predicted organization feature vector approach the target organization feature vector and maximizes the probability of forming two heterogeneous organizations. S5. Using the optimal combination of printing parameters, a high-entropy alloy sample is prepared by additive manufacturing process, and the prepared sample is characterized by microstructure to obtain experimental microstructure images and experimental structure feature vectors. S6. The optimal combination of printing parameters, experimental microscopic tissue images, experimental tissue feature vectors and corresponding dual heterogeneous tissue labels obtained in step 5 are added as new samples to the database of step 1, and the printing parameter-tissue structure mapping model in step 3 is iteratively updated and trained. S7. Repeat steps 4 to 6 until the error between the model prediction result and the experimental result converges to within the preset threshold, thereby achieving the stable preparation of the dual heterogeneous high-entropy alloy.

2. The method for preparing a dual heterogeneous high-entropy alloy using artificial intelligence-driven additive manufacturing according to claim 1, characterized in that, In step S1, each set of printing parameters includes at least laser power, scanning speed, scanning spacing, powder layer thickness, and scanning path mode; the microstructure image undergoes grayscale normalization preprocessing; the tissue structure information includes at least the grain size d calculated from the grain boundary mask and the spatial gradient heterogeneity index G calculated from the average grain size of the sub-region. i The multi-scale heterogeneity index M, calculated from the percentile value of grain size distribution. i The phase area fraction f calculated from the phase category mask. i k, and the microstructure heterogeneity index U calculated from the grain size or phase area fraction variation coefficient of the sub-region. i .

3. The method for preparing a dual heterogeneous high-entropy alloy using artificial intelligence-driven additive manufacturing according to claim 2, characterized in that, In step S1, establishing the additive manufacturing experimental database of "printing parameters - microstructure images - structural features" specifically includes: The print parameter vector of the i-th sample is represented as: The scan path is represented using one-hot encoding as follows: Based on this, construct the original database: To eliminate the influence of dimensions, the printing parameters are normalized: Gray-level normalization processing was performed on the microscopic tissue images: in: For the first Print parameter vectors for each sample. For laser power, For scanning speed, For scanning spacing, To achieve a thicker powder layer, Encode vectors for scanning path methods. This represents the total number of categories for scanning path methods. For the first Does the _ sample belong to the _ ... Class scan path marker variable, For the original training database, For the first Microscopic tissue images corresponding to each sample The total number of samples, For the first In the nth sample One printing parameter, These are the normalized printing parameters. For the first Each sample image at pixel coordinates grayscale value at that location These are the minimum and maximum grayscale values ​​of the image. The normalized image grayscale values. These are the pixel coordinates of the image.

4. The method for preparing a dual heterogeneous high-entropy alloy using artificial intelligence-driven additive manufacturing according to claim 3, characterized in that, In step S2, the method for determining the requirement of dual heterogeneous structure includes: dividing the microstructure image into multiple sub-regions, calculating the average grain size of each sub-region and the overall average grain size, and constructing a spatial gradient heterogeneity index accordingly; simultaneously, calculating the ratio of the 90th percentile to the 10th percentile of the grain size distribution as a multi-scale heterogeneity index; when both the spatial gradient heterogeneity index and the multi-scale heterogeneity index exceed a preset threshold, it is determined that the requirement of dual heterogeneous structure is met.

5. The method for preparing a dual heterogeneous high-entropy alloy using artificial intelligence-driven additive manufacturing according to claim 4, characterized in that, In step S2, the obtained microscopic tissue image is input into a convolutional neural network. The convolutional neural network includes a shared encoder, a grain size regression branch, a phase distribution segmentation branch, a tissue feature vector regression branch, and a dual heterogeneous classification branch. Specifically, the grain size regression branch outputs a pixel-level grain size map or the average grain size of a sub-region; the phase distribution segmentation branch outputs a pixel-level phase category mask; and the tissue feature vector regression branch outputs a G... i M i f i ,k and U i The organizational feature vector, and the two heterogeneous classification branches are used to output the probability of the formation of two heterogeneous tissues; Layer-by-layer feature extraction of images is performed using convolutional layers, pooling layers, and nonlinear activation functions to obtain quantitative expressions of organizational features such as grain size, tissue gradient, multi-scale structure, phase distribution, and tissue inhomogeneity, including: The feature extraction process of a convolutional layer can be represented as follows: The output of the pooling layer is: After multiple convolutions and pooling, a high-dimensional feature vector of the microscopic tissue image is obtained: To achieve the discrimination of two heterogeneous tissues, spatial gradient heterogeneity index and multi-scale heterogeneity index are further constructed to divide the microscopic tissue image into... Each sub-region has an average grain size of [number] grains. The overall average grain size is: The spatial gradient heterogeneity index is defined as: Further define the two heterogeneous organization tags: in: : No. The first in the layer Feature maps output by each convolutional kernel; : No. The first in the layer One input feature map; : No. Layer connection The input channel and the first The convolutional kernel weights for each output channel; : No. Layer The bias term corresponding to each output channel; : Non-linear activation function; : No. Number of input channels in the layer; : No. Layer Each pooled output feature map; Pooling function; Convolutional neural network encoder; Model parameters of a convolutional neural network encoder; : No. The tissue feature vector of each sample; : The normalized first Microscopic tissue images of individual samples; The number of sub-regions in the image; : No. The sample at the th Average grain size in each sub-region; : No. Overall average grain size of the sample; : No. Spatial gradient heterogeneity index for each sample; : No. The 90th percentile value in the grain size distribution of each sample; : No. The 10th percentile value in the grain size distribution of each sample; : No. Multiscale heterogeneous indicators for a sample; : Threshold for determining two heterogeneous tissues; : No. Two heterogeneous tissue labels for a single sample; This indicates that the requirement of two heterogeneous organizations is met. This indicates dissatisfaction.

6. The method for preparing a dual heterogeneous high-entropy alloy using artificial intelligence-driven additive manufacturing according to claim 5, characterized in that, Steps S3 and S4 specifically include: Extracted tissue feature vectors and two heterogeneous tissue labels As supervisory information, a prediction model is established with printing parameters as input and tissue features and two heterogeneous discrimination results as output. The model is preferably structured as "parameter input layer + fully connected mapping layer + classification output layer," where a convolutional neural network is used for image feature encoding, and a parameter mapping network is used to predict the tissue features from printing parameters. Specifically, this includes: The mapping relationship between printing parameters and tissue characteristics is as follows: The probability of the formation of two heterogeneous tissues is: The organizational feature regression loss is: The classification loss for two heterogeneous tissues is: The total loss function is: in: After training is completed, the target dual heterogeneous organization feature vector is defined. The optimal combination of printing parameters is obtained through parameter space search: And satisfy: in: : A mapping model from printing parameters to organizational features; : Mapping model parameters; The model predicts the first Each sample organization feature vector; Two-dimensional heterogeneous tissue classifier; Classifier parameters; The model predicts the first The probability that a sample forms two heterogeneous organizations; Organizational feature regression loss; : L2 norm; Organizational classification loss; The model's total loss function; The weighting coefficients for regression loss and classification loss; Regularization coefficient; The complete set of parameters to be optimized in the model; : Feature vector of the target's two heterogeneous organizations; The optimal combination of printing parameters obtained by reverse calculation; : Print the feasible field of parameters; : Weighting coefficients of the probability constraint term for the formation of two heterogeneous organizations; Weighting coefficients for process stability constraints; : Refer to the printing parameter vector to limit the search results from deviating too much from the existing stable process window; : The lower and upper bounds of the values ​​for each printing parameter.

7. The method for preparing a dual heterogeneous high-entropy alloy using artificial intelligence-driven additive manufacturing according to claim 6, characterized in that, In step S5, the obtained optimal combination of printing parameters is used. Additive manufacturing experiments were conducted to prepare new high-entropy alloy samples, and experimental microstructure images were obtained. This includes: inputting the image into a convolutional neural network encoder to extract the feature vector of the experimental tissue. Simultaneously, the corresponding predicted tissue feature vector is obtained from the prediction model: The relative error between the experimental results and the model predictions is defined as: The dual isomeric tissue labels of the experimental samples were calculated according to the aforementioned criteria. ;when and If the selected combination of printing parameters is valid, then the parameter search range is corrected, and the parameter search and experimental verification are re-executed. The correction for the parameter search range is expressed as follows: in: Microscopic images of the experimentally prepared samples; Normalized experimental microscopic tissue images; : Tissue feature vectors extracted from experimental samples; The model prints the optimal combination of parameters. The predicted tissue feature vector; The relative error between experimental results and predicted results; : Allowable error threshold; : The dual isomeric tissue label of the experimental sample; : No. In the first round of search The upper and lower boundaries of each printing parameter; The first in the optimal combination of printing parameters Each parameter value; : No. Wheel of Life The search half-width for each parameter; Iteration rounds; Print parameter dimension number.

8. The method for preparing a dual heterogeneous high-entropy alloy using artificial intelligence-driven additive manufacturing according to claim 7, characterized in that, In step S6, the newly obtained experimental data, including the optimal combination of printing parameters, corresponding microscopic tissue images, tissue feature vectors, and two heterogeneous tissue labels, are added to the original training database to construct the expanded database. The model parameters are retrained based on the expanded database, and the parameter update process is represented as follows: To determine whether the closed-loop optimization process has converged, the rate of change of loss between two adjacent training rounds is defined as: When the following conditions are met: When the closed-loop iteration between the model and the experiment reaches a convergence state, the stable preparation of the dual heterogeneous high-entropy alloy microstructure is achieved. in: : No. The training database corresponding to each iteration; : No. Print the parameter vector of each newly added experimental sample; : No. Microscopic tissue images of newly added experimental samples; : No. The tissue feature vector of the newly added experimental samples in the round; : No. Two heterogeneous tissue labels for newly added experimental samples; , : No. Wheel and First Wheel model parameters; Learning rate; Regarding model parameters The gradient operator; , : No. Wheel and First The loss function value obtained from the first round of training; : Rate of change of loss between two adjacent rounds; Loss convergence threshold; The relative error between the experiment and the prediction; Error tolerance threshold.

9. An artificial intelligence-driven additive manufacturing system for preparing dual heterogeneous high-entropy alloys, used to perform the method according to any one of claims 1 to 8, characterized in that, The system includes: The database construction module is used to collect and store printing parameters, microscopic images and tissue structure information to build an additive manufacturing experimental database. The image feature extraction module has a built-in convolutional neural network model for extracting features from microscopic tissue images and outputting tissue feature vectors and the results of two heterogeneous tissue discrimination. The mapping model module is used to establish and store the mapping relationship between printing parameters, tissue feature vectors, and labels of two heterogeneous tissues, and can predict the output tissue features and the probability of forming two heterogeneous tissues based on the input printing parameters. The parameter optimization module is used to organize feature vectors according to the set target and use the mapping model module to search for the optimal combination of printing parameters in reverse. An additive manufacturing execution module is used to prepare high-entropy alloy samples using laser powder bed melting technology based on the optimal combination of printing parameters output by the parameter optimization module. The feedback update module is used to supplement the database construction module with newly added experimental data and trigger the iterative retraining of the mapping model module until the prediction error converges.

10. The AI-driven additive manufacturing system for preparing dual heterogeneous high-entropy alloys according to claim 9, characterized in that, The convolutional neural network model in the image feature extraction module includes multiple alternating stacked convolutional and pooling layers, as well as a fully connected encoder. The convolutional layers are used to extract local features from microscopic tissue images, the pooling layers are used to reduce the feature dimension, and the fully connected encoder is used to map the extracted multidimensional features into a fixed-length tissue feature vector.

Images