Porous structure cross-domain reverse design method based on property constraint cooperative training

Abstract

The invention relates to a porous structure cross-domain reverse design method based on property constraint cooperative training, and the method comprises the steps: taking mechanical property information and random noise as the input of a trained reverse generation model, obtaining a coding vector, and obtaining reconstructed porous structure data through a trained variational auto-encoder model. According to the method, in the training process of the variational auto-encoder, mechanical property prediction cooperative training is directly carried out on feature vector distribution in an encoding space, the feature distribution of the encoding space is improved, and a more direct relation between the encoding space and the property is established. By increasing the constraint of the forward relationship, in the training process of the reverse generation model, the establishment of'property-coding space 'is easier, and compared with a common reverse design scheme, the method can improve the accuracy of the prediction structure and support the accurate cross-domain reverse design of the porous structure. According to the method, the characteristic that cross-domain structures are different in distribution is fully considered, and cross-domain reverse design with high accuracy is achieved.

Description

Technical Field

[0001] The present invention relates to the field of material structure design, and in particular to a cross-domain inverse design method of porous structures based on property constraint collaborative training. Background Art

[0002] Porous structures are material structures that can be used in aerospace, medical and other fields. Designing a workpiece into a porous structure can reduce the weight of the workpiece as much as possible while retaining the required mechanical properties, such as strength and stiffness. In addition, through artificial design, porous materials can meet the mechanical properties required by people under specific conditions, such as specific buffer materials. Current porous materials are usually implemented using repeated porous units (generally cube units). Researchers design the geometric structure within the unit to meet the requirements of mechanical properties. Common structures include rod structures, shell structures, spinodal structures, etc. Researchers change the structural design to achieve specific performance of the workpiece with the least amount of material. This type of design method is a forward design, and the direction of the researcher's optimization is relatively blind.

[0003] Reverse design has also developed in recent years. Some solutions analyze the stress distribution and adjust the size of the workpiece at the stress concentration point to accurately improve the mechanical properties of the workpiece. However, this type of optimization method relies on finite element analysis to obtain stress distribution, which requires a lot of calculations. The current reverse design solution based on neural networks still has some limitations:

[0004] (1) It mainly supports structural design within a specific domain, such as rod structure, shell structure, random spinodal structure, etc., and cannot generate structures that meet the requirements across domains, which greatly limits the diversity of reverse design;

[0005] (2) Some neural network inverse design schemes based on generative model structures (generative adversarial models, diffusion models, etc.) support cross-domain generation in terms of data format. However, since the distribution of cross-domain data is usually very different, such schemes often have poor results.

[0006] In summary, there is currently a lack of a cross-domain reverse design method for porous structures to solve or partially solve the aforementioned problems. Summary of the invention

[0007] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a cross-domain reverse design method for porous structures based on collaborative training of property constraints, so as to solve the problem that cross-domain reverse design cannot be realized or the effect is poor.

[0008] The purpose of the present invention can be achieved by the following technical solutions:

[0009] In one aspect of the present invention, a cross-domain inverse design method for porous structures based on property constraint collaborative training is provided, wherein mechanical property information and random noise are used as inputs of a trained inverse generation model to obtain a coding vector, and the trained variational autoencoder model is used to obtain reconstructed porous structure data, wherein the training process of the inverse generation model and the variational autoencoder model comprises the following steps:

[0010] Step S1, constructing cross-domain porous structure-mechanical property data pairs as training samples;

[0011] Step S2, using the porous structure data in the training sample as the input of the variational autoencoder model to obtain an encoding vector, using an encoding space-mechanical property mapper to calculate the first property loss, and implementing the training of the variational autoencoder model based on the first property loss;

[0012] Step S3, using the noisy coding vector as the input of the reverse generation model to obtain the reconstructed coding vector, using the coding space-mechanical property mapper to calculate the second property loss, and realizing the training of the reverse generation model based on the second property loss.

[0013] As a preferred technical solution, in the step S1, the cross-domain porous structure-mechanical property data pairs include porous structure-mechanical property data pairs corresponding to the rod structure, the shell structure and the spinodal structure.

[0014] As a preferred technical solution, the first property loss, porous structure reconstruction loss and encoding space loss are combined to achieve the training of the variational autoencoder model.

[0015] As a preferred technical solution, the calculation process of the porous structure reconstruction loss includes:

[0016] The porous structure data in the training sample is used as the input of the encoder in the variational autoencoder model to obtain an encoding vector, and the encoding vector is input into the decoder in the autoencoder model to obtain the reconstructed porous structure data, and the porous structure reconstruction loss is calculated based on the porous structure data in the training sample and the reconstructed porous structure data.

[0017] As a preferred technical solution, the calculation process of the coding space loss includes:

[0018] The loss between the encoding vector output by the variational autoencoder model and the preset standard distribution is calculated as the encoding space loss.

[0019] As a preferred technical solution, the process of calculating the first property loss and the second property loss using the coding space-mechanical property mapper includes the following steps:

[0020] The encoding vector obtained by the variational autoencoder model or the inverse generation model is used as the input of the encoding space-mechanical property mapper to obtain the predicted mechanical property information, and the first property loss or the second property loss is calculated based on the predicted mechanical property information and the mechanical property information in the training samples.

[0021] As a preferred technical solution, the training of the inverse generation model is achieved by combining the second property loss and the noise loss.

[0022] As a preferred technical solution, the calculation process of the noise loss includes the following steps:

[0023] The coding vector output by the variational autoencoder model is subjected to denoising to obtain the noisy coding vector and use it as the input of the inverse generation model to obtain predicted noise information, a reconstructed coding vector is obtained based on the noise information, and the noise loss is calculated based on the predicted noise information and the actual noise.

[0024] Another aspect of the present invention provides an electronic device, comprising: one or more processors and a memory, wherein the memory stores one or more programs, and the one or more programs include instructions for executing the aforementioned cross-domain inverse design method of porous structure based on collaborative training of property constraints.

[0025] Another aspect of the present invention provides a computer-readable storage medium comprising one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the aforementioned cross-domain inverse design method for porous structures based on collaborative training of property constraints.

[0026] Compared with the prior art, the present invention has at least one of the following beneficial effects:

[0027] (1) Taking into account the different characteristics of cross-domain structural distribution, a highly accurate cross-domain reverse design is achieved: During the training of the variational autoencoder, the distribution of the feature vectors in the encoding space is directly trained for mechanical property prediction, the feature distribution of the encoding space is improved, and a more direct relationship between "encoding space → property" is established. By adding constraints on the positive relationship, it is easier to establish "property → encoding space" during the training of the reverse generation model. Compared with ordinary reverse design schemes, the present invention can improve the accuracy of the predicted structure and support accurate cross-domain reverse design of porous structures.

[0028] (2) Low burden of computing resources and time cost, easy to promote and apply: Compared with the inverse design method that does not use property prediction collaborative training, the training process of the present invention only adds additional loss function calculation and a simple network structure (encoding space-mechanical property mapper), and almost does not increase the time cost of training. The reasoning process (reverse design process) does not increase any computing or time cost. No additional loss is added, which is conducive to the implementation and use of the reverse design algorithm.

[0029] (3) Wide application scenarios: Compared with the inverse design method that does not use property prediction collaborative training, the present invention adds a "coding space → property" mapper, so the model can be simultaneously applied to multi-task scenarios of forward property prediction and inverse design. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 A six-layer diagram of a cross-domain inverse design method for porous structures based on collaborative training of property constraints in an embodiment;

[0031] Figure 2 This is an architecture diagram of cross-domain reverse design of porous structures based on property constraint collaborative training in an embodiment;

[0032] Figure 3 is a schematic diagram of an electronic device in an embodiment;

[0033] Figure 4 This is an example diagram of the reverse design effect of the cross-domain reverse design of the porous structure based on the collaborative training of property constraints in the embodiment. DETAILED DESCRIPTION

[0034] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0035] Example 1

[0036] In view of the problem that the cross-domain reverse design of the aforementioned prior art cannot be realized or has poor effect, this embodiment provides a cross-domain reverse design method for porous structures based on property constraint collaborative training, which relies on variational autoencoders, inverse generation models and encoding space-mechanical property mappers. In this embodiment, cross-domain refers to cross-structural domains, and common structures include rod structures, shell structures, spinodal structures, etc.

[0037] Preferably, the reverse generation model in this embodiment is implemented using an implicit diffusion model or a generative adversarial model.

[0038] See also Figure 2 ,The variational autoencoder includes an encoder and a decoder mapper, and the encoding space - mechanical property mapper can be composed of several layers of fully connected layers.

[0039] See also Figure 1 and Figure 4 ,This method mainly includes the training data preparation stage, the ,variational autoencoder training stage, the inverse generative model training stage and the inverse ,design stage. Each stage will be described below.

[0040] Training data preparation stage: Construct a cross-domain porous structure-mechanical property data pair as a training sample set, which contains a total of N "porous structure data-property information" data pairs {Si, Pi} (i=0,1,...N) from k domains, where Si refers to the porous structure, expressed using a signed distance field, and the data format is a three-dimensional array of shape [r,r,r], where r represents the resolution. Pi refers to mechanical properties, which can be stiffness, yield strength, stress-strain curve, etc. In this embodiment, Pi specifically refers to the stiffness matrix, which can characterize the linear relationship between stress and strain under small deformation, and its data format is a [6, 6] matrix.

[0041] Variational autoencoder training phase: The variational autoencoder includes an encoder and a decoder. During training, the encoder maps the structure Si to the encoding space to obtain the encoding vector li, and the decoder restores the encoding vector li to the structure Si*. Three loss functions are constructed in the training phase, one is the reconstruction loss between the original structure Si and the reconstructed structure Si*, the other is the loss function between the distribution of the encoding vector li and the standard Gaussian distribution, and the encoding space loss. By setting the first two loss functions, an encoder that satisfies the "structure → encoding space" mapping relationship and a decoder that satisfies the "encoding space → structure" mapping relationship can be obtained. At the same time, in order to make the encoding space and the properties have a more direct connection, the third loss function is designed as the loss function between li and Pi - property loss. The property loss is calculated by the encoding space-mechanical property mapper, which inputs the encoding vector li and outputs the predicted property value Pi*. The loss function is the difference between Pi and Pi*.

[0042] Inverse generative model training phase: The input of the inverse generative model is the sample's encoding vector li, property condition Pi and random Gaussian noise n obtained by encoder mapping in the variational autoencoder training phase, and the output is the inversely generated encoding vector li*. The sample's encoding vector is first weighted and summed with random Gaussian noise to obtain a noisy encoding vector. The noisy encoding vector is input into the inverse generative model. The inverse generative model predicts the noise information in the noise encoding and reconstructs the sample's encoding vector based on the noise value. This phase contains two losses. The first loss is the difference between the noise information predicted by the inverse generative model and the actual noise, and the second loss is the property loss of the reconstructed encoding vector. The second loss is calculated by the encoding space-mechanical property mapper, which inputs the reconstructed li* and outputs the predicted property value Pi**. The loss function is the difference between Pi** and Pi.

[0043] Inverse design stage: The property condition information Pj and random Gaussian noise N are used as the input of the inverse generation model, and the inverse generation model reconstructs the encoding vector, and the reconstructed encoding vector is input into the decoder trained in the variational autoencoder training stage to obtain the reconstructed porous structure. The process of inverse design is as follows Figure 4 As shown, this scheme can generate porous structures of different domains, including rod structures, shell structures, etc., according to a control condition (i.e., stiffness matrix).

[0044] Specifically, the process of using the inverse generation model to predict the noise information in the noise coding and reconstructing the sample's code vector according to the noise value is as follows: during the noise addition process, the sample's code vector and random Gaussian noise are weighted and summed to obtain a noisy code vector (noise code vector = a*random Gaussian noise + b*code vector). The noise code vector is input into the inverse generation model, and the inverse generation model predicts the random Gaussian noise added to the code vector after several inference iterations. Under the premise of knowing the noise code vector and Gaussian noise, the initial sample code vector is obtained according to the weighted relationship (code vector = (noise code vector - a*random Gaussian noise) ÷ b), which is the reconstructed code vector.

[0045] Compared with training using the training mapping relationship of "structure → encoding space → structure", the encoding space usually has a more direct relationship with the structure. However, the relationship between structure and property itself is relatively hidden, so this method usually requires a lot of calculations to obtain. The connection between the "encoding space" and "property" obtained in this way is also relatively hidden. At the same time, in the encoding space, structures within the same domain are usually similar, while cross-domain structures are often in different distributions, which is not conducive to generating cross-domain structural encodings under the same property conditions.

[0046] Compared with the above method, the variational autoencoder training stage of this embodiment directly performs property prediction and collaborative training on the feature vector distribution in the encoding space, which can improve the feature distribution of the encoding space and establish a more direct relationship between "encoding space → property". By adding constraints on the positive relationship, it is easier to establish "property → encoding space" in the reverse generation model training stage. Compared with ordinary reverse design schemes, this method can improve the accuracy of the predicted structure and support accurate cross-domain reverse design of porous structures.

[0047] Example 2

[0048] This embodiment provides Figure 3 The one shown corresponds to Figure 1 A schematic diagram of the electronic device, such as Figure 3 As mentioned above, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 The method of the cross-domain reverse design method of porous structure based on property constraint collaborative training. Of course, in addition to the software implementation, the present invention does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc., that is, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0049] Example 3

[0050] This embodiment provides a computer-readable storage medium, including one or more programs for execution by one or more processors of an electronic device, wherein the one or more programs include instructions for executing the cross-domain inverse design method of porous structures based on collaborative training of property constraints as described in Example 1.

[0051] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0052] Computer readable media include permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. Information can be computer readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer readable media does not include temporary computer readable media (transitory media), such as modulated data signals and carrier waves.

[0053] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0054] The above is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and these modifications or replacements should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be based on the protection scope of the claims.

Claims

1. A cross-domain inverse design method for porous structures based on property constraint collaborative training, characterized in that: The mechanical property information and random noise are used as inputs of the trained inverse generation model to obtain a coding vector, and the reconstructed porous structure data is obtained using the trained variational autoencoder model, wherein the training process of the inverse generation model and the variational autoencoder model includes the following steps: Step S1, constructing cross-domain porous structure-mechanical property data pairs as training samples; Step S2, using the porous structure data in the training sample as the input of the variational autoencoder model to obtain an encoding vector, using an encoding space-mechanical property mapper to calculate the first property loss, and implementing the training of the variational autoencoder model based on the first property loss; Step S3, using the noisy coding vector as the input of the reverse generation model to obtain the reconstructed coding vector, using the coding space-mechanical property mapper to calculate the second property loss, and realizing the training of the reverse generation model based on the second property loss.

2. According to claim 1, a porous structure cross-domain inverse design method based on property constraint collaborative training is characterized in that: In the step S1, the cross-domain porous structure-mechanical property data pairs include porous structure-mechanical property data pairs corresponding to the rod structure, the shell structure and the spinodal structure.

3. According to claim 1, a porous structure cross-domain inverse design method based on property constraint collaborative training is characterized in that: The first property loss, porous structure reconstruction loss and encoding space loss are combined to achieve the training of the variational autoencoder model.

4. According to claim 3, a porous structure cross-domain inverse design method based on property constraint collaborative training is characterized in that: The calculation process of the porous structure reconstruction loss includes: The porous structure data in the training sample is used as the input of the encoder in the variational autoencoder model to obtain an encoding vector, and the encoding vector is input into the decoder in the autoencoder model to obtain the reconstructed porous structure data, and the porous structure reconstruction loss is calculated based on the porous structure data in the training sample and the reconstructed porous structure data.

5. According to claim 3, a porous structure cross-domain inverse design method based on property constraint collaborative training is characterized in that: The calculation process of the coding space loss includes: The loss between the encoding vector output by the variational autoencoder model and the preset standard distribution is calculated as the encoding space loss.

6. According to claim 1, a porous structure cross-domain inverse design method based on property constraint collaborative training is characterized in that: The process of calculating the primary property loss and the secondary property loss using the encoding space-mechanical property mapper includes the following steps: The encoding vector obtained by the variational autoencoder model or the inverse generation model is used as the input of the encoding space-mechanical property mapper to obtain the predicted mechanical property information, and the first property loss or the second property loss is calculated based on the predicted mechanical property information and the mechanical property information in the training samples.

7. The method for cross-domain inverse design of porous structures based on property constraint collaborative training according to claim 1 is characterized in that: The second property loss and the noise loss are combined to realize the training of the reverse generation model.

8. The cross-domain inverse design method of porous structure based on property constraint collaborative training according to claim 7, characterized in that: The calculation process of the noise loss includes the following steps: The coding vector output by the variational autoencoder model is subjected to denoising to obtain the noisy coding vector and use it as the input of the inverse generation model to obtain predicted noise information, a reconstructed coding vector is obtained based on the noise information, and the noise loss is calculated based on the predicted noise information and the actual noise.

9. An electronic device, characterized in that: include: One or more processors and a memory, wherein the memory stores one or more programs, wherein the one or more programs include instructions for executing the cross-domain inverse design method for porous structures based on property constraint collaborative training as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that: It includes one or more programs for execution by one or more processors of an electronic device, and the one or more programs include instructions for executing the cross-domain inverse design method of porous structure based on property constraint collaborative training as described in any one of claims 1-8.

Images