A reverse design method, apparatus, device, and storage medium for material structures

Abstract

This application provides a method, apparatus, device, and storage medium for reverse design of material structures, relating to the field of materials design. A three-dimensional feature vector matrix of the sample material is constructed; the three-dimensional feature vectors in the matrix are used to characterize the three-dimensional microstructure features of the sample material. Macroscopic performance parameters of the sample material are extracted and used as model condition information. The three-dimensional feature vectors and model condition information are used to form training data pairs. The conditional diffusion model is iteratively trained using the training data pairs until the training termination condition is met, resulting in the target conditional diffusion model. The target macroscopic performance parameters are input into the target conditional diffusion model to obtain the target three-dimensional microstructure that satisfies the target macroscopic performance parameter conditions. This method enables the reverse design of the three-dimensional microstructure of materials based on required performance, improving material development efficiency and reducing development costs; moreover, the designed structure can more accurately meet the required performance requirements and better fit the actual process.

Description

Technical Field

[0001] This application relates to the field of materials design, and in particular to a method, apparatus, device and storage medium for reverse design of material structures. Background Technology

[0002] With the rapid development of aerospace, automotive, and energy equipment industries, the performance requirements for various materials are becoming increasingly stringent. Traditional materials development follows a "trial and error" approach, which involves preparing a large number of candidate materials through experiments; determining the performance of each candidate material through experimental testing or the establishment of mathematical models; and selecting the material that meets the required performance from among the candidate materials. However, this method can only predict the performance of the prepared materials and cannot reverse engineer materials to meet the required performance, resulting in low development efficiency and high costs.

[0003] Therefore, there is a need to provide a reverse design method for material structures to improve material development efficiency and reduce development costs. Summary of the Invention

[0004] In view of this, this application provides a reverse design method for material structures, the method comprising:

[0005] A three-dimensional feature vector matrix of the sample material is constructed; the three-dimensional feature vectors in the three-dimensional feature vector matrix are used to characterize the three-dimensional microstructure features of the sample material.

[0006] Extract the macroscopic performance parameters of the sample material; and use the macroscopic performance parameters as model condition information;

[0007] The three-dimensional feature vector matrix and the model condition information are used to form training data pairs;

[0008] Using the training data pairs, the pre-constructed conditional diffusion model is iteratively trained until the training termination condition is met to obtain the target conditional diffusion model.

[0009] The target macroscopic performance parameters are input into the target conditional diffusion model to obtain the target three-dimensional microstructure that meets the requirements of the target macroscopic performance parameters.

[0010] Optionally, construct a three-dimensional feature vector matrix of the sample material, including:

[0011] Obtain target documents in the field of alloy materials;

[0012] Structural parameters of the sample material are extracted from the target document; the structural parameters include at least one of grain characteristic parameters and phase characteristic parameters.

[0013] Based on the structural parameters, a representative volume unit corresponding to the sample material is generated;

[0014] The representative volume unit is discretized to obtain a three-dimensional voxel structure composed of each three-dimensional voxel unit.

[0015] Based on the structural parameters, generate the three-dimensional feature vector corresponding to the three-dimensional voxel unit;

[0016] Based on the position of the three-dimensional voxel unit in the three-dimensional voxel structure, determine the target position of the corresponding three-dimensional feature vector in the three-dimensional feature vector matrix;

[0017] According to each target position, the three-dimensional feature vectors are combined to form the three-dimensional feature vector matrix.

[0018] Optionally, the structural parameters include the grain characteristic parameters and the phase characteristic parameters; the grain characteristic parameters include at least the grain orientation; the phase characteristic parameters include at least the phase type; and based on the structural parameters, a three-dimensional feature vector corresponding to the three-dimensional voxel unit is generated, including:

[0019] Determine the Euler angles corresponding to the grain orientation of the three-dimensional voxel unit; and convert the Euler angles into quaternions;

[0020] Obtain the pre-constructed mapping relationship; the mapping relationship is the correspondence between phase type and number;

[0021] Based on the mapping relationship, the target number corresponding to the three-dimensional voxel unit is determined;

[0022] The quaternion and the target number are combined to obtain the three-dimensional feature vector.

[0023] Optionally, macroscopic performance parameters of the sample material are extracted, including:

[0024] Extract the stress-strain curve of the sample material from the target document;

[0025] The constitutive equation of the sample material is determined by performing parameter fitting on the stress-strain curve data.

[0026] Based on the constitutive equation, the mechanical property parameters of the sample material are determined;

[0027] The mechanical performance parameters are used as the macroscopic performance parameters.

[0028] Optionally, based on the structural parameters, a representative volume unit corresponding to the sample material is generated, including:

[0029] The structural parameters are used as simulation constraints.

[0030] Based on the simulation constraints, a microstructure generation tool is used to generate representative volumetric units with three-dimensional microstructures that satisfy the conditions corresponding to the structural parameters. There are multiple representative volumetric units, and the three-dimensional microstructures of each representative volumetric unit are different from each other.

[0031] Optionally, the pre-built conditional diffusion model is trained using the training data pair until the training termination condition is met to obtain the target conditional diffusion model, including:

[0032] The three-dimensional feature vector matrix is ​​mapped to the latent space through a pre-built encoder to obtain the corresponding latent space representation;

[0033] Perform conditional forward diffusion to add real noise to the latent space representation, resulting in the first... The potential noise representation of the step;

[0034] The model condition information corresponding to the three-dimensional feature vector matrix is ​​input into the encoding network to obtain the conditional embedding vector;

[0035] The noise latent representation, diffusion steps The conditional embedding vector is input into a denoising network for inverse denoising to obtain the predicted noise.

[0036] Based on the predicted noise and the actual noise, a loss function is constructed;

[0037] With the goal of minimizing the loss function, the parameters of the encoder and the denoising network are updated to optimize the conditional diffusion model.

[0038] Optionally, when there are multiple target three-dimensional microstructures, after determining the target three-dimensional microstructures that satisfy the target macroscopic performance parameters, the method further includes:

[0039] Macroscopic performance simulation of the target's three-dimensional microstructure is performed to determine the actual macroscopic performance parameters;

[0040] By comparing the target macroscopic performance parameters with the actual macroscopic performance parameters, the performance deviation value of the target three-dimensional microstructure is determined; the performance deviation value is used to characterize the degree of difference between the target macroscopic performance parameters and the actual macroscopic performance parameters; the performance deviation value is positively correlated with the degree of difference.

[0041] The target three-dimensional microstructure with the smallest performance deviation value is taken as the final target three-dimensional microstructure.

[0042] This application also provides a reverse design apparatus for material structures, the apparatus comprising:

[0043] A matrix construction module is used to construct a three-dimensional feature vector matrix of the sample material; the three-dimensional feature vectors in the three-dimensional feature vector matrix are used to characterize the three-dimensional microstructure features of the sample material.

[0044] The condition construction module is used to extract the macroscopic performance parameters of the sample material and use the macroscopic performance parameters as model condition information.

[0045] The training pair construction module is used to construct training data pairs from the three-dimensional feature vector matrix and the model condition information;

[0046] The model training module is used to iteratively train the pre-built conditional diffusion model using the training data pairs until the training termination condition is met to obtain the target conditional diffusion model.

[0047] The structural design module is used to input the target macroscopic performance parameters into the target conditional diffusion model to obtain the target three-dimensional microstructure that meets the requirements of the target macroscopic performance parameters.

[0048] This application also provides an apparatus, comprising:

[0049] Memory, used to store computer programs;

[0050] A processor, used to execute the computer program to implement the steps of the reverse design method for any of the above-mentioned material structures.

[0051] This application also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the reverse design method for any of the above-described material structures.

[0052] In summary, this application provides a method, apparatus, device, and storage medium for reverse design of material structures. A three-dimensional feature vector matrix of the sample material is constructed; the three-dimensional feature vectors in the matrix are used to characterize the three-dimensional microstructure features of the sample material. Macroscopic performance parameters of the sample material are extracted and used as model condition information. The three-dimensional feature vectors and model condition information are used to form training data pairs. The conditional diffusion model is iteratively trained using the training data pairs until the training termination condition is met to obtain the target conditional diffusion model. The target macroscopic performance parameters are input into the target conditional diffusion model to obtain the target three-dimensional microstructure that satisfies the target macroscopic performance parameter conditions.

[0053] As can be seen, this application utilizes a three-dimensional vector matrix that characterizes the three-dimensional microstructure of the sample material, along with the macroscopic performance parameters of the sample material, to train a conditional diffusion model. This enables the trained target conditional diffusion model to generate a target three-dimensional microstructure that satisfies both the parameter requirements and conforms to real-world processes, using the target macroscopic performance parameters as generation conditions. This achieves reverse engineering of the material's three-dimensional microstructure, improves material development efficiency, and reduces development costs, while ensuring that the designed material structure more accurately meets the required performance requirements and better conforms to real-world processes. Attached Figure Description

[0054] Figure 1 A first flowchart illustrating a reverse design method for a material structure provided in this application;

[0055] Figure 2 A schematic diagram of the second process of a reverse design method for a material structure provided in this application;

[0056] Figure 3 A schematic diagram of the third process of a reverse design method for a material structure provided in this application;

[0057] Figure 4(a) is a schematic diagram of a three-dimensional voxel structure provided in this application;

[0058] Figure 4(b) is a schematic diagram of the grain orientation of a three-dimensional voxel structure provided in this application;

[0059] Figure 4(c) is a schematic diagram of the phase type of a three-dimensional voxel structure provided in this application;

[0060] Figure 5(a) is a schematic diagram of the grain orientation of a target three-dimensional microstructure provided in this application;

[0061] Figure 5(b) is a schematic diagram of the phase type of a target three-dimensional microstructure provided in this application;

[0062] Figure 6 A schematic diagram of a reverse design device for a material structure provided in this application;

[0063] Figure 7 This is a schematic diagram of the structure of a device provided in this application. Detailed Implementation

[0064] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0065] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0066] Please refer to Figure 1 , Figure 1 A first flowchart illustrating a reverse design method for a material structure provided in this application, the method comprising:

[0067] S101. Construct a three-dimensional feature vector matrix for the sample material; the three-dimensional feature vectors in the three-dimensional feature vector matrix are used to characterize the three-dimensional microstructure features of the sample material.

[0068] S102. Extract the macroscopic performance parameters of the sample material; and use the macroscopic performance parameters as model condition information.

[0069] S103. Combine the three-dimensional feature vector matrix and model condition information to form training data pairs.

[0070] S104. Using the training data pairs, iteratively train the pre-built conditional diffusion model until the training termination condition is met to obtain the target conditional diffusion model.

[0071] S105. Input the target macroscopic performance parameters into the target conditional diffusion model to obtain the target three-dimensional microstructure that meets the requirements of the target macroscopic performance parameters.

[0072] To achieve reverse design of the material structure, this application pre-constructs a conditional diffusion model and iteratively trains it to obtain a target conditional diffusion model. Based on the required macroscopic performance parameters, the target conditional diffusion model generates a target three-dimensional microstructure that meets the parameter requirements, thus completing the reverse design of the material structure.

[0073] To achieve the above functions, it is first necessary to construct the training data pairs required for training the conditional diffusion model.

[0074] Specifically, this application first constructs a three-dimensional feature vector matrix based on the three-dimensional microstructure characteristics of the sample material. The three-dimensional feature vectors in the three-dimensional feature vector matrix are used to characterize the three-dimensional microstructure characteristics of the sample material. It can be understood that the aforementioned sample material is a pre-prepared material with known macroscopic performance parameters. Based on the three-dimensional microstructure characteristics, the conditional diffusion model is trained, enabling the reverse design of the material structure to reach a higher dimension, ensuring that the reverse-designed material structure better matches the required performance parameters and is closer to the actual process. The specific implementation method for constructing the three-dimensional feature vector matrix of the sample material will be explained in subsequent embodiments and will not be elaborated here.

[0075] Furthermore, to ensure that the 3D microstructure obtained through reverse engineering is controllable and customizable, this application also extracts the macroscopic performance parameters of the sample material and uses these parameters as model condition information. Utilizing this model condition information, the conditional diffusion model is guided to denoise, enabling the trained target conditional diffusion model to design 3D microstructures that meet the user's required macroscopic performance parameters.

[0076] Furthermore, considering the diverse macroscopic properties of materials, including but not limited to mechanical properties, corrosion resistance, oxidation resistance, and wear resistance, the corresponding macroscopic performance parameters can be extracted based on the target macroscopic properties required for reverse engineering of the material. This application does not impose any particular limitations on this. For example, when the target macroscopic property is mechanical property, the macroscopic performance parameters may include Young's modulus and yield strength; when the target macroscopic property is corrosion resistance, the macroscopic performance parameters may include corrosion rate, corrosion potential, and polarization resistance; when the target macroscopic property is oxidation resistance, the macroscopic performance parameters may include oxidation induction time, oxidation induction temperature, and oxidation stability index.

[0077] Next, training data pairs are constructed using the obtained three-dimensional feature vector matrix and model condition information. These training data pairs are then used to train the conditional diffusion model until the training termination condition is met, resulting in the target conditional diffusion model. The specific training process of the conditional diffusion model will be described in detail in subsequent embodiments and will not be elaborated here.

[0078] Based on the above, the obtained target macroscopic performance parameters are input into the target conditional diffusion model. The target conditional diffusion model, using the target macroscopic performance parameters as conditions, recovers the target three-dimensional microstructure that meets the requirements of the target macroscopic performance parameters from pure noise, thus realizing the reverse design of the material structure.

[0079] Thus, this application achieves reverse design from target macroscopic performance parameters to target three-dimensional microstructure, improving material development efficiency and reducing development costs. Simultaneously, it advances the reverse design of material structures to a higher three-dimensional dimension, ensuring that the reverse-designed material structure better matches the required performance parameters and more closely resembles actual processes, thus possessing greater practical value.

[0080] Based on the above embodiments:

[0081] The process of constructing the three-dimensional feature vector matrix of the sample material is explained in detail below.

[0082] Please refer to Figure 2 , Figure 2 This is a schematic diagram of the second process of a reverse design method for a material structure provided in this application. As an optional embodiment, constructing a three-dimensional feature vector matrix of the sample material includes:

[0083] S201. Obtain target documents in the field of alloy materials;

[0084] S202. Extract the structural parameters of the sample material from the target document; the structural parameters include at least one of the grain characteristic parameters and phase characteristic parameters;

[0085] S203. Based on structural parameters, generate representative volume units corresponding to the sample material;

[0086] S204. Discretize the representative volume unit to obtain a three-dimensional voxel structure composed of each three-dimensional voxel unit.

[0087] S205. Generate the three-dimensional feature vectors corresponding to the three-dimensional voxel units based on the structural parameters.

[0088] S206. Based on the position of the three-dimensional voxel unit in the three-dimensional voxel structure, determine the target position of the corresponding three-dimensional feature vector in the three-dimensional feature vector matrix;

[0089] S207. According to the target positions, the three-dimensional feature vectors are combined into a three-dimensional feature vector matrix.

[0090] Alloy materials are widely used in numerous fields such as aerospace, automotive, and energy equipment. To achieve reverse engineering of alloy materials, this embodiment uses target documents in the field of alloy materials as the data source for constructing a three-dimensional feature vector matrix. This embodiment does not impose any particular limitations on the method for determining the target documents. For example, documents published later than a preset time point and containing preset keywords can be extracted from a document database in the field of alloy materials as target documents. The preset keywords can be set according to actual needs. This ensures the cutting-edge nature and professionalism of the target documents, providing a data foundation for the subsequent training of the conditional diffusion model.

[0091] For alloy materials, their macroscopic properties include, but are not limited to, mechanical properties, corrosion resistance, oxidation resistance, and wear resistance. Among these macroscopic properties, mechanical properties are particularly important. The structure of an alloy material determines its mechanical properties. Based on this, in this embodiment, structural parameters of the material sample are extracted from the obtained target document. Structural parameters include at least one of grain characteristic parameters and phase characteristic parameters.

[0092] Among these, grain characteristic parameters include, but are not limited to, grain orientation, grain size, and grain size distribution. Phase characteristic parameters include, but are not limited to, phase type, phase size, and phase distribution. Both have a significant impact on the mechanical properties of alloy materials, such as strength, toughness, and plasticity.

[0093] As mentioned above, to ensure that the reverse-engineered material structure better matches the required performance parameters and more closely resembles the actual process, this application extends the reverse design of material structures to the three-dimensional dimension. Therefore, this embodiment generates representative volumetric units of the sample material based on structural parameters to determine the three-dimensional microstructural characteristics of the sample material, providing a foundation for the subsequent construction of a three-dimensional feature vector matrix.

[0094] Among them, the representative volume element can statistically represent the three-dimensional microstructure characteristics of the entire sample material, that is, the various phases, grain size distributions, and orientation distributions appearing within the representative volume element are consistent with the sample material. At the same time, the size of the representative volume element is much smaller than the scale of the engineering structure, which significantly reduces the computational cost of the three-dimensional feature vector matrix.

[0095] Next, the representative volume element is discretized to obtain a three-dimensional voxel structure composed of individual three-dimensional voxel elements. Please refer to Figure 4(a), which is a schematic diagram of a three-dimensional voxel structure provided in this application. Discretizing the representative volume element is equivalent to dividing it into a three-dimensional mesh, with each three-dimensional mesh being a three-dimensional voxel element. For each three-dimensional voxel element, a three-dimensional feature vector is generated based on the structural parameters of the sample material; the three-dimensional feature vector can characterize the three-dimensional microstructural features of the sample material at that three-dimensional voxel element.

[0096] Furthermore, this embodiment determines the target position of the three-dimensional feature vector corresponding to the three-dimensional voxel unit in the three-dimensional voxel structure based on the position of the three-dimensional voxel unit in the three-dimensional voxel structure, to ensure that the position of the three-dimensional voxel unit in the three-dimensional voxel structure is consistent with the position of the three-dimensional feature vector of the three-dimensional voxel unit in the three-dimensional feature vector matrix. According to the target position corresponding to each three-dimensional feature vector, the three-dimensional feature vectors corresponding to each three-dimensional voxel unit are combined to obtain the three-dimensional feature vector matrix of the sample material.

[0097] Taking the red three-dimensional voxel unit in Figure 4(a) as the target three-dimensional voxel unit as an example, the position of the target three-dimensional voxel unit in the three-dimensional voxel structure is: the first one in the horizontal direction (from left to right), the first one in the vertical direction (from back to front), and the fourth one in the depth direction (from bottom to top); then the position of the three-dimensional feature vector corresponding to this three-dimensional voxel unit in the three-dimensional feature vector matrix is ​​also: the first one in the horizontal direction (from left to right), the first one in the vertical direction (from back to front), and the fourth one in the depth direction (from bottom to top).

[0098] The process of generating representative volumetric elements is explained below.

[0099] As an optional embodiment, representative volumetric units corresponding to the sample material are generated based on structural parameters, including:

[0100] Use structural parameters as simulation constraints;

[0101] Based on simulation constraints, a microstructure generation tool is used to generate representative volumetric elements that satisfy the conditions corresponding to the structural parameters in three-dimensional microstructures. There are multiple representative volumetric elements, and the three-dimensional microstructures of each representative volumetric element are different.

[0102] In this embodiment, structural parameters are used as simulation constraints. A microstructure generation tool is used to generate representative volumetric units of a three-dimensional microstructure that satisfy the conditions corresponding to the structural parameters. The aforementioned microstructure generation tool can be selected according to actual needs, including but not limited to three-dimensional material microstructure data processing and visualization platforms such as Dream3D.

[0103] It is important to emphasize that this embodiment uses a microstructure generation tool to generate multiple representative volumetric units based on each set of structural parameters, and the three-dimensional microstructures of each representative volumetric unit are distinct. That is, one sample material corresponds to one structural parameter, and one structural parameter corresponds to multiple representative volumetric units. For each representative volumetric unit, a corresponding three-dimensional feature vector matrix can be obtained through steps such as discretization.

[0104] In addition, simulation programs can be used to perform macroscopic performance simulations on each representative volume element, further accurately determining the macroscopic performance parameters of each representative volume element, and using the newly determined macroscopic performance parameters as the corresponding model condition information.

[0105] In summary, in this embodiment, by using limited sample material, as many representative volumetric units as possible are generated, thereby constructing as many three-dimensional feature vector matrices as possible and expanding the number of training data pairs.

[0106] The process of generating three-dimensional feature vectors will be explained below.

[0107] Please refer to Figure 3 , Figure 3 This is a schematic diagram of the third process of a reverse design method for a material structure provided in this application. As an optional embodiment, the structural parameters include grain characteristic parameters and phase characteristic parameters; the grain characteristic parameters include at least the grain orientation; the phase characteristic parameters include at least the phase type; based on the structural parameters, a three-dimensional feature vector corresponding to a three-dimensional voxel unit is generated, including:

[0108] S301. Determine the Euler angles corresponding to the grain orientation of the three-dimensional voxel unit; and convert the Euler angles into quaternions;

[0109] S302. Obtain the pre-built mapping relationship; the mapping relationship is the correspondence between phase type and number;

[0110] S303. Determine the target number corresponding to the three-dimensional voxel unit based on the mapping relationship;

[0111] S304. Combine the position coordinates and quaternions to obtain the three-dimensional feature vector.

[0112] This embodiment determines the Euler angles corresponding to the grain orientation of the three-dimensional voxel unit. Euler angles are a set of parameters that describe the three-dimensional spatial orientation using three rotation angles. Furthermore, to eliminate gimbal lock and improve numerical compactness and stability, the Euler angles are converted into quaternions, which can still characterize the grain orientation at the three-dimensional voxel unit. The specific process of determining the Euler angles and quaternions of the three-dimensional voxel unit is not described in detail in this embodiment. Based on the quaternions corresponding to each three-dimensional voxel unit, the grain orientation of each three-dimensional voxel unit is visualized, resulting in the grain orientation schematic diagram of the three-dimensional voxel structure shown in Figure 4(b).

[0113] On the other hand, this embodiment also pre-constructs a mapping relationship, which is a correspondence between phase types and numbers. The phase types mentioned above can be set according to actual needs. For example, the phase types include martensite and austenite. Correspondingly, the mapping relationship can be a correspondence between martensite and the first number, and a correspondence between austenite and the second number. Based on the above mapping relationship and the phase types of the three-dimensional voxel units, the target number corresponding to the three-dimensional voxel unit can be determined. Based on the target number corresponding to each three-dimensional voxel unit, the phase type of each three-dimensional voxel unit is visualized, and the phase type schematic diagram of the three-dimensional voxel structure shown in Figure 4(c) can be obtained.

[0114] Finally, the position coordinates and quaternions are combined to obtain the three-dimensional feature vector. The quaternion can be represented as... Target number is Then the three-dimensional eigenvector can be represented as .

[0115] In summary, in the three-dimensional eigenvector matrix, the position of the three-dimensional eigenvectors represents the position of the three-dimensional voxel unit in the three-dimensional voxel structure; the value of the three-dimensional eigenvectors represents the grain orientation and phase type of the three-dimensional voxel unit at the three-dimensional voxel structure, so as to jointly characterize the three-dimensional microstructure features of the sample material and provide a basis for subsequent reverse design of material structure from three dimensions.

[0116] The process of extracting macroscopic performance parameters from sample materials is described below.

[0117] As an optional embodiment, macroscopic performance parameters of the sample material are extracted, including:

[0118] Extract the stress-strain curve of the sample material from the target document;

[0119] The constitutive equation of the sample material is determined by performing parameter fitting on the stress-strain curve data.

[0120] Based on the constitutive equation, the mechanical property parameters of the sample material are determined;

[0121] Mechanical performance parameters are used as macroscopic performance parameters.

[0122] In order to achieve the goal of reverse designing the three-dimensional microstructure of alloy materials based on the required mechanical performance parameters, the mechanical performance parameters of the alloy materials are used as macroscopic performance parameters in this embodiment.

[0123] Specifically, the stress-strain curves of the sample material are first extracted from target documents in the field of alloy materials. Stress-strain curves describe the deformation behavior of alloy materials under external forces, representing the relationship between stress and strain in the form of a coordinate graph.

[0124] Next, parameter fitting is performed on the stress-strain curve to determine the constitutive equation of the sample material. The constitutive equation is a mathematical model used to describe the relationship between stress and strain in alloy materials under external loads. This embodiment does not specifically limit the specific implementation of parameter fitting. For example, nonlinear least squares methods or global optimization methods can be used for parameter fitting.

[0125] Finally, the mechanical property parameters of the sample material can be determined based on the constitutive equation. These mechanical property parameters include, but are not limited to, elastic modulus, yield strength, ultimate strength, and elongation after fracture; this embodiment does not impose any particular limitations on these parameters. These mechanical property parameters are used as macroscopic performance parameters so that they can subsequently be used as model condition information to guide the denoising process of the conditional diffusion model. This allows the trained target conditional diffusion model to design a three-dimensional microstructure that meets the user's required macroscopic performance parameters.

[0126] The training process of the conditional diffusion model will be explained in detail below.

[0127] As an optional embodiment, a pre-built conditional diffusion model is trained using training data pairs until the training termination condition is met to obtain the target conditional diffusion model, including:

[0128] By using a pre-built encoder, the three-dimensional feature vector matrix is ​​mapped to the latent space to obtain the corresponding latent space representation;

[0129] Perform conditional forward diffusion to add real noise to the latent space representation, obtaining the first... The potential noise representation of the step;

[0130] The model condition information corresponding to the three-dimensional feature vector matrix is ​​input into the encoding network to obtain the conditional embedding vector;

[0131] Representing noise potential and the number of diffusion steps The conditional embedding vector is input into the denoising network for inverse denoising to obtain the predicted noise.

[0132] Construct a loss function based on the predicted noise and the actual noise;

[0133] With the goal of minimizing the loss function, the parameters of the encoder and denoising network are updated to optimize the conditional diffusion model.

[0134] First, a high-dimensional 3D feature vector matrix is ​​mapped to a low-dimensional latent space using a pre-built encoder (e.g., a 3D convolutional neural network) to obtain a latent space representation, thereby reducing computational complexity.

[0135] Subsequently, forward noise addition is performed. Specifically, conditional forward diffusion is executed to add realistic noise to the latent space representation, resulting in the... The latent noise representation is obtained by gradually adding real noise to the latent space over a preset number of diffusion steps, eventually transforming the latent space representation into pure noise.

[0136] Then, inverse denoising is performed. The inverse denoising process involves training a denoising network to predict the actual noise added in each diffusion step; this is the core technology the conditional diffusion model needs to learn. Specifically, the model conditional information corresponding to the three-dimensional feature vectors is first input into the encoding network to obtain the conditional embedding vector. Then, the noise latent representation and the number of diffusion steps are... The conditional embedding vector is input into the denoising network for inverse denoising to obtain the predicted noise.

[0137] The denoising network can be a U-net network. Conditional embedding vectors are injected into the denoising network through a cross-attention mechanism. Specifically, the conditional embedding vectors are first projected onto a fully connected layer as keys and values; the intermediate feature maps of the U-net network are then projected onto a query through convolution; the attention weights between the query and the key are calculated; and the values ​​are weighted and summed using these attention weights, ensuring that the denoising process is guided by model conditional information, i.e., macroscopic performance parameters.

[0138] The conditional diffusion model also includes a decoder symmetric to the encoder, used to reconstruct the denoised latent space representation back into three-dimensional space, obtaining the corresponding three-dimensional eigenvector matrix. Based on the three-dimensional eigenvector matrix, the three-dimensional microstructure can be determined, enabling reverse design of material structures.

[0139] Finally, a loss function is constructed based on the predicted noise and the actual noise; with the goal of minimizing the loss function, the parameters of the encoder and the denoising network are updated to optimize the conditional diffusion model.

[0140] After obtaining the target conditional diffusion model, the target macroscopic performance parameters input by the user are acquired. These parameters are used as the target model's conditional information. The target model's conditional information is input into the encoding network to obtain the target conditional embedding vector c_target. The target conditional diffusion model starts with pure noise z_T and performs a T-step denoising process. In each diffusion step, the denoising network determines the predicted noise based on the target conditional embedding vector c_target, the current noise latent representation, and the current diffusion step number. Based on the predicted noise, the noise z_T-1 for the next diffusion step is determined. This process is repeated until a clean latent space representation z_0 is obtained. The latent space representation z_0 is input into the decoder to obtain the target three-dimensional feature vector matrix. Based on the target three-dimensional feature vector matrix, the target three-dimensional microstructure can be determined. Because the above denoising process is guided by the target macroscopic performance parameters, the final target three-dimensional microstructure meets the requirements of the target macroscopic performance parameters, realizing the reverse design of the material structure and more closely resembling real-world processes.

[0141] Furthermore, after obtaining the target three-dimensional microstructure, the target three-dimensional microstructure can be discretized to obtain the target three-dimensional voxel structure; based on the quaternion corresponding to each three-dimensional voxel unit in the target three-dimensional voxel structure, the grain orientation of the target three-dimensional microstructure is visualized, resulting in the grain orientation schematic diagram shown in Figure 5(a); based on the target number corresponding to each three-dimensional voxel unit in the target three-dimensional voxel structure, the phase type of the target three-dimensional microstructure is visualized, resulting in the phase type schematic diagram shown in Figure 5(b).

[0142] As an optional embodiment, when there are multiple target three-dimensional microstructures, after determining the target three-dimensional microstructures that satisfy the target macroscopic performance parameters, the method further includes:

[0143] Perform macroscopic performance simulation on the three-dimensional microstructure of the target to determine the actual macroscopic performance parameters;

[0144] By comparing the target macroscopic performance parameters with the actual macroscopic performance parameters, the performance deviation value of the target's three-dimensional microstructure is determined; the performance deviation value is used to characterize the degree of difference between the target macroscopic performance parameters and the actual macroscopic performance parameters; the performance deviation value is positively correlated with the degree of difference.

[0145] The target three-dimensional microstructure with the smallest performance deviation value is taken as the final target three-dimensional microstructure.

[0146] To ensure the effectiveness and reliability of the target three-dimensional microstructures obtained through reverse engineering, closed-loop verification was also performed on each target three-dimensional microstructure in this embodiment.

[0147] Specifically, a pre-built simulation program can be used to perform macroscopic performance simulations on the three-dimensional microstructures of each target generated by the target condition diffusion model, determining the actual macroscopic performance parameters. The performance deviation value is determined by comparing the target macroscopic performance parameters with the actual macroscopic performance parameters. As an optional embodiment, the macroscopic performance parameters include yield strength and Young's modulus under mechanical properties; a first deviation value is determined between the actual yield strength and the target yield strength; a second deviation value is determined between the actual Young's modulus and the target Young's modulus; corresponding preset weights can be set according to the degree (or importance) of each of the yield strength and Young's modulus in determining the performance deviation value; the greater the contribution, the greater the preset weight; using each preset weight, the first deviation value and the second deviation value are weighted and summed to determine the performance deviation value.

[0148] Since the performance deviation value is positively correlated with the degree of difference between macroscopic performance parameters, the target three-dimensional microstructure with the smallest performance deviation value is taken as the final target three-dimensional microstructure, which further ensures the effectiveness and reliability of the target three-dimensional microstructure obtained by reverse design.

[0149] Please refer to Figure 6 , Figure 6 A schematic diagram of a reverse design apparatus for a material structure provided in this application is provided. The apparatus includes:

[0150] The matrix construction module 601 is used to construct a three-dimensional feature vector matrix of the sample material; the three-dimensional feature vectors in the three-dimensional feature vector matrix are used to characterize the three-dimensional microstructure features of the sample material.

[0151] The condition construction module 602 is used to extract the macroscopic performance parameters of the sample material and use the macroscopic performance parameters as model condition information.

[0152] Training pair construction module 603 is used to construct training data pairs from the three-dimensional feature vector matrix and model condition information;

[0153] The model training module 604 is used to iteratively train a pre-built conditional diffusion model using training data pairs until the training termination condition is met to obtain the target conditional diffusion model.

[0154] The structural design module 605 is used to input the target macroscopic performance parameters into the target conditional diffusion model to obtain the target three-dimensional microstructure that meets the requirements of the target macroscopic performance parameters.

[0155] For a detailed description of the reverse design apparatus for material structures provided in this application, please refer to the embodiments of the reverse design method for material structures described above; this application will not elaborate further here.

[0156] Based on the above embodiments:

[0157] As an optional embodiment, the matrix construction module 601 includes:

[0158] The document acquisition module is used to acquire target documents in the field of alloy materials.

[0159] The structural parameter extraction module is used to extract the structural parameters of the sample material from the target document; the structural parameters include at least one of grain feature parameters and phase feature parameters;

[0160] The volume element generation module is used to generate representative volume elements corresponding to the sample material based on structural parameters.

[0161] The discrete module is used to discretize representative volume units to obtain a three-dimensional voxel structure composed of each three-dimensional voxel unit.

[0162] The vector generation module is used to generate a three-dimensional feature vector corresponding to the three-dimensional voxel unit based on the structural parameters.

[0163] The target position determination module is used to determine the target position of the corresponding three-dimensional feature vector in the three-dimensional feature vector matrix based on the position of the three-dimensional voxel unit in the three-dimensional voxel structure.

[0164] The matrix generation module is used to assemble the three-dimensional feature vector matrix from the three-dimensional feature vectors according to the target positions.

[0165] As an optional embodiment, the structural parameters include grain characteristic parameters and phase characteristic parameters; the grain characteristic parameters include at least the grain orientation; the phase characteristic parameters include at least the phase type; the vector generation module includes:

[0166] The quaternion determination module is used to determine the Euler angles corresponding to the grain orientation of the three-dimensional voxel unit and convert the Euler angles into quaternions.

[0167] The mapping relationship acquisition module is used to obtain pre-built mapping relationships; the mapping relationship is the correspondence between phase type and number.

[0168] The target number determination module is used to determine the target number corresponding to the three-dimensional voxel unit based on the mapping relationship;

[0169] The combination module is used to combine position coordinates and quaternions to obtain a three-dimensional feature vector.

[0170] As an optional embodiment, the condition construction module 602 includes:

[0171] The curve extraction module is used to extract the stress-strain curve of the sample material from the target document;

[0172] The constitutive equation determination module is used to perform parameter fitting on stress-strain curve data to determine the constitutive equation of the sample material.

[0173] The mechanical parameter determination module is used to determine the mechanical property parameters of the sample material based on the constitutive equation.

[0174] The macroscopic parameter determination module is used to use mechanical performance parameters as macroscopic performance parameters.

[0175] The condition construction submodule is used to use macroscopic performance parameters as model condition information.

[0176] As an optional embodiment, the volume unit generation module includes:

[0177] The constraint determination module is used to use structural parameters as simulation constraints.

[0178] The volume element generation submodule is used to generate representative volume elements whose three-dimensional microstructures satisfy the conditions corresponding to the structural parameters based on simulation constraints and using microstructure generation tools. There are multiple representative volume elements, and the three-dimensional microstructures of each representative volume element are different.

[0179] As an optional embodiment, the model training module 604 includes:

[0180] The latent space representation determination module is used to map the three-dimensional feature vector matrix to the latent space through a pre-built encoder to obtain the corresponding latent space representation;

[0181] The noise latent representation determination module performs conditional forward diffusion to add real noise to the latent space representation, obtaining the first... The potential noise representation of the step;

[0182] The conditional embedding vector determination module is used to input the model condition information corresponding to the three-dimensional feature vector matrix into the encoding network to obtain the conditional embedding vector.

[0183] The noise prediction module is used to represent the potential noise and the number of diffusion steps. The conditional embedding vector is input into the denoising network for inverse denoising to obtain the predicted noise.

[0184] The loss function construction module is used to construct the loss function based on the predicted noise and the actual noise;

[0185] The training submodule is used to update the parameters of the encoder and denoising network with the goal of minimizing the loss function in order to optimize the conditional diffusion model.

[0186] As an optional embodiment, when the number of target three-dimensional microstructures is multiple, the device further includes:

[0187] The performance simulation module is used to perform macroscopic performance simulation on the target three-dimensional microstructure after determining the target macroscopic performance parameters, and to determine the actual macroscopic performance parameters.

[0188] The performance deviation determination module is used to compare the target macroscopic performance parameters with the actual macroscopic performance parameters to determine the performance deviation value of the target's three-dimensional microstructure. The performance deviation value is used to characterize the degree of difference between the target macroscopic performance parameters and the actual macroscopic performance parameters. The performance deviation value is positively correlated with the degree of difference.

[0189] The final structure determination module is used to select the target three-dimensional microstructure with the smallest performance deviation value as the final target three-dimensional microstructure.

[0190] Please refer to Figure 7 , Figure 7 A schematic diagram of a device provided in this application, the device comprising:

[0191] Memory 701 is used to store computer programs;

[0192] Processor 702 is used to implement the steps of the reverse design method for any of the above-mentioned material structures when executing a computer program.

[0193] For a detailed description of the equipment provided in this application, please refer to the embodiments of the reverse design method for the material structure described above; this application will not elaborate further here.

[0194] This application also provides a storage medium storing a computer program, which, when executed by a processor, implements the steps of the reverse design method for any of the aforementioned material structures.

[0195] The aforementioned storage media include all forms of non-volatile memory, media, and memory devices, such as semiconductor memory devices (e.g., EPROM, EEPROM, and flash memory devices), magnetic disks (e.g., internal hard disks or removable disks), magneto-optical disks, and CD-ROMs and DVD-ROMs. Processors and memory may be supplemented by or integrated into dedicated logic circuitry.

[0196] For a detailed description of the storage medium provided in this application, please refer to the embodiments of the reverse design method for the above-described material structure; this application will not repeat the details here.

[0197] While this specification contains numerous specific implementation details, these should not be construed as limiting the scope of any invention or the scope of the claims, but rather are primarily intended to describe features of specific embodiments of a particular invention. Certain features described in the various embodiments herein may also be implemented in combination in a single embodiment. Conversely, various features described in a single embodiment may also be implemented separately in various embodiments or in any suitable sub-combination. Furthermore, while features may function in certain combinations as described above and even initially claimed in this way, one or more features from a claimed combination may be removed from that combination in some cases, and a claimed combination may refer to a sub-combination or a variation thereof.

[0198] Similarly, although the operations are depicted in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order shown or sequentially, or requiring all illustrated operations to be performed to achieve the desired result. In some cases, multitasking and parallel processing may be advantageous. Furthermore, the separation of various system modules and components in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0199] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A reverse design method for material structures, characterized in that, The method includes: A three-dimensional feature vector matrix of the sample material is constructed; the three-dimensional feature vectors in the three-dimensional feature vector matrix are used to characterize the three-dimensional microstructure features of the sample material. Extract the macroscopic performance parameters of the sample material; and use the macroscopic performance parameters as model condition information; The three-dimensional feature vector matrix and the model condition information are used to form training data pairs; Using the training data pairs, the pre-constructed conditional diffusion model is iteratively trained until the training termination condition is met to obtain the target conditional diffusion model. The target macroscopic performance parameters are input into the target conditional diffusion model to obtain the target three-dimensional microstructure that meets the requirements of the target macroscopic performance parameters. Constructing the three-dimensional feature vector matrix of the sample material includes: Obtain target documents in the field of alloy materials; extract structural parameters of the sample material from the target documents; the structural parameters include at least one of grain characteristic parameters and phase characteristic parameters; Based on the structural parameters, a representative volume unit corresponding to the sample material is generated; the representative volume unit is discretized to obtain a three-dimensional voxel structure composed of each three-dimensional voxel unit. Based on the structural parameters, generate the three-dimensional feature vector corresponding to the three-dimensional voxel unit; Based on the position of the three-dimensional voxel unit in the three-dimensional voxel structure, determine the target position of the corresponding three-dimensional feature vector in the three-dimensional feature vector matrix; According to each target position, the three-dimensional feature vectors are combined to form the three-dimensional feature vector matrix.

2. The reverse design method for material structures as described in claim 1, characterized in that, The structural parameters include the grain characteristic parameters and the phase characteristic parameters; the grain characteristic parameters include at least the grain orientation; the phase characteristic parameters include at least the phase type; based on the structural parameters, a three-dimensional feature vector corresponding to the three-dimensional voxel unit is generated, including: Determine the Euler angles corresponding to the grain orientation of the three-dimensional voxel unit; and convert the Euler angles into quaternions; Obtain the pre-constructed mapping relationship; the mapping relationship is the correspondence between phase type and number; Based on the mapping relationship, the target number corresponding to the three-dimensional voxel unit is determined; The quaternion and the target number are combined to obtain the three-dimensional feature vector.

3. The reverse design method for material structures as described in claim 1, characterized in that, Extract the macroscopic performance parameters of the sample material, including: Extract the stress-strain curve of the sample material from the target document; The constitutive equation of the sample material is determined by performing parameter fitting on the stress-strain curve data. Based on the constitutive equation, the mechanical property parameters of the sample material are determined; The mechanical performance parameters are used as the macroscopic performance parameters.

4. The reverse design method for material structures as described in claim 1, characterized in that, Based on the structural parameters, representative volume units corresponding to the sample material are generated, including: The structural parameters are used as simulation constraints. Based on the simulation constraints, a microstructure generation tool is used to generate representative volumetric units with three-dimensional microstructures that satisfy the conditions corresponding to the structural parameters. There are multiple representative volumetric units, and the three-dimensional microstructures of each representative volumetric unit are different from each other.

5. The reverse design method for material structures as described in claim 1, characterized in that, Using the aforementioned training data pairs, a pre-constructed conditional diffusion model is trained until the training termination condition is met to obtain the target conditional diffusion model, including: The three-dimensional feature vector matrix is ​​mapped to the latent space through a pre-built encoder to obtain the corresponding latent space representation; Perform conditional forward diffusion to add real noise to the latent space representation, resulting in the first... The potential noise representation of the step; The model condition information corresponding to the three-dimensional feature vector matrix is ​​input into the encoding network to obtain the conditional embedding vector; The noise latent representation, diffusion steps The conditional embedding vector is input into a denoising network for inverse denoising to obtain the predicted noise. Based on the predicted noise and the actual noise, a loss function is constructed; With the goal of minimizing the loss function, the parameters of the encoder and the denoising network are updated to optimize the conditional diffusion model.

6. The reverse design method for material structures as described in claim 1, characterized in that, When there are multiple target three-dimensional microstructures, after determining the target three-dimensional microstructures that satisfy the target macroscopic performance parameters, the method further includes: Macroscopic performance simulation of the target's three-dimensional microstructure is performed to determine the actual macroscopic performance parameters; By comparing the target macroscopic performance parameters with the actual macroscopic performance parameters, the performance deviation value of the target three-dimensional microstructure is determined; the performance deviation value is used to characterize the degree of difference between the target macroscopic performance parameters and the actual macroscopic performance parameters; the performance deviation value is positively correlated with the degree of difference. The target three-dimensional microstructure with the smallest performance deviation value is taken as the final target three-dimensional microstructure.

7. A reverse design device for material structures, characterized in that, The device includes: A matrix construction module is used to construct a three-dimensional feature vector matrix of the sample material; the three-dimensional feature vectors in the three-dimensional feature vector matrix are used to characterize the three-dimensional microstructure features of the sample material. The condition construction module is used to extract the macroscopic performance parameters of the sample material and use the macroscopic performance parameters as model condition information. The training pair construction module is used to construct training data pairs from the three-dimensional feature vector matrix and the model condition information; The model training module is used to iteratively train the pre-built conditional diffusion model using the training data pairs until the training termination condition is met to obtain the target conditional diffusion model. The structural design module is used to input the target macroscopic performance parameters into the target conditional diffusion model to obtain the target three-dimensional microstructure that meets the requirements of the target macroscopic performance parameters. The matrix construction module includes: The document acquisition module is used to acquire target documents in the field of alloy materials. The structural parameter extraction module is used to extract the structural parameters of the sample material from the target document; the structural parameters include at least one of grain feature parameters and phase feature parameters; The volume element generation module is used to generate representative volume elements corresponding to the sample material based on structural parameters. The discrete module is used to discretize representative volume units to obtain a three-dimensional voxel structure composed of each three-dimensional voxel unit. The vector generation module is used to generate a three-dimensional feature vector corresponding to the three-dimensional voxel unit based on the structural parameters. The target position determination module is used to determine the target position of the corresponding three-dimensional feature vector in the three-dimensional feature vector matrix based on the position of the three-dimensional voxel unit in the three-dimensional voxel structure. The matrix generation module is used to assemble the three-dimensional feature vectors into a three-dimensional feature vector matrix according to each target position.

8. A device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the reverse design method for the material structure as described in any one of claims 1 to 6.

9. A storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the steps of the reverse design method for the material structure as described in any one of claims 1 to 6.

Images