Physical field determination method, device, equipment and medium of engineering structure

Abstract

The application provides a physical field determination method, device, equipment and medium of an engineering structure. The method first acquires working condition parameters of a target entity structure, and inputs the working condition parameters to a physical field reconstruction model. The physical field reconstruction model comprises a coefficient prediction network and a basis generator. The basis generator is loaded from a trained basic modal model. The trained basic modal model is a neural network model that learns the mapping rule between the geometric characteristics of different engineering structures and the corresponding global response physical field. The basis generator is used to provide a target reduced-order basis that is adapted to the target entity structure. Then, the coefficient prediction network is used to extract the target basis coefficients from the working condition parameters, and the physical field reconstruction processing is performed based on the target basis coefficients and the target reduced-order basis to obtain a target global response physical field that reflects the structure state of the target entity structure under the working condition parameters. The embodiments of the application can improve the reconstruction accuracy of the physical field.

Description

Technical Field

[0001] This application relates to the field of engineering structure technology, and more specifically, to a method, apparatus, device, and storage medium for determining the physical field of an engineering structure. Background Technology

[0002] High-fidelity numerical simulation technology (with finite element analysis as a typical example) is one of the core technologies for accurately predicting the mechanical behavior of structures and assessing structural safety. To solve the computational efficiency bottleneck of high-fidelity numerical simulation, developing surrogate models that combine accuracy and efficiency has become the core research direction in this field. Among them, the reduced-order model (ROM) has become the mainstream technical path due to its core advantage of "reducing the dimensionality of high-dimensional problems".

[0003] In related technologies, orthogonal decomposition (POD), singular value decomposition (SVD), or data-driven reduction model (DDROM, also known as AI-ROM) are commonly used for reduction. However, these technologies share common inherent limitations in basis generation: both use traditional mathematical methods such as SVD / POD to extract the basis from simulation snapshot data of a specific problem. Although this basis generation mode has mathematical rigor, it suffers from the two core defects mentioned above, which fundamentally limit the expressive power and generalization potential of the reduction model, resulting in low accuracy of the physical representation of the structure and thus reducing the accuracy of the structural state assessment. Summary of the Invention

[0004] In view of this, this application provides a method, apparatus, device and storage medium for determining the physical field of an engineering structure, so as to at least solve the problems existing in the related art.

[0005] Specifically, this application is implemented through the following technical solution: This application provides a method for determining the physical field of an engineering structure, including: The working parameters of the target entity structure are obtained and input into the physical field reconstruction model. The physical field reconstruction model includes a coefficient prediction network and a basis generator. The basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping law between the geometric features of different engineering structures and the corresponding global response physical field. The basis generator is used to provide a target reduced-order basis that is adapted to the target entity structure. Using the coefficient prediction network, the basis coefficients of the operating conditions are extracted to obtain the target basis coefficients. Based on the target basis coefficients and the target reduced-order basis provided by the basis generator, the physical field is reconstructed to obtain the target global response physical field. The target global response physical field is used to reflect the structural state of the target entity structure under the operating conditions.

[0006] This application also provides a device for determining the physical field of an engineering structure, including: The parameter acquisition module is used to acquire the working condition parameters of the target entity structure and input the working condition parameters into the physical field reconstruction model. The physical field reconstruction model includes a coefficient prediction network and a basis generator. The basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping law between the geometric features of different engineering structures and the corresponding global response physical field. The basis generator is used to provide a target reduced-order basis that is adapted to the target entity structure. The physics field reconstruction module is used to extract the basis coefficients of the operating parameters using the coefficient prediction network to obtain the target basis coefficients, and to perform physics field reconstruction processing based on the target basis coefficients and the target reduced-order basis provided by the basis generator to obtain the target global response physics field; the target global response physics field is used to reflect the structural state of the target entity structure under the operating parameters.

[0007] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the physical field determination method for any of the engineering structures described in the foregoing embodiments.

[0008] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the physical field determination method for any of the engineering structures described in the foregoing embodiments.

[0009] This application also provides a computer program product, including a computer program that, when run by a processor, performs the steps of a physical field determination method for any of the possible engineering structures described above.

[0010] The technical solutions provided by the embodiments of this application may include the following beneficial effects: In this embodiment, the basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping rules between the geometric features of different engineering structures and the corresponding global response physical fields. This enables the basis generator to have strong nonlinear expression capabilities and cross-structure transfer capabilities, improving the generalization of the basis generator. The basis generator is then transferred to the physical field reconstruction model, enabling it to provide a target reduced-order basis that adapts to the target entity structure. Thus, the physical field is reconstructed based on the target basis coefficients and the target reduced-order basis, which can improve the accuracy of the target global response physical field and thereby improve the evaluation accuracy of the structural state of the target entity structure under the working parameters.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this specification. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating a method for determining the physical field of an engineering structure according to an exemplary embodiment of this application; Figure 2 This is a flowchart illustrating the training process of a base mode model according to an exemplary embodiment of this application; Figure 3 This is a flowchart illustrating the training process of a physical field reconstruction model according to an exemplary embodiment of this application; Figure 4 This is a complete flowchart illustrating a model training and physics prediction process according to an exemplary embodiment of this application; Figure 5 This is a schematic diagram of the physical field determination device for an engineering structure, as shown in an exemplary embodiment of this application; Figure 6 This is a schematic diagram of the structure of a physical field determination device for another engineering structure, as illustrated in an exemplary embodiment of this application. Figure 7 This is a hardware structure diagram of a computer device illustrated in an exemplary embodiment of this application. Detailed Implementation

[0013] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0014] 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.

[0015] 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."

[0016] In the field of structural engineering, high-fidelity numerical simulation technology (with finite element analysis as a typical example) is one of the core technologies for accurately predicting structural mechanical behavior and assessing structural safety. This technology, through refined discretization of the structure and solution of mechanical equations, can provide physical field data on the full-domain response, including stress, displacement, and strain, providing a reliable basis for engineering design optimization and risk assessment. However, high-fidelity numerical simulation has extremely high computational complexity; a single simulation often requires several hours to several days of computation time. This high computational cost makes it unsuitable for scenarios requiring numerous repetitive calculations, such as structural parameter sensitivity analysis, multi-condition optimization design, and real-time health monitoring—practical engineering needs.

[0017] To address the computational efficiency bottleneck in high-fidelity numerical simulations, developing surrogate models that combine accuracy and efficiency has become a core research direction in this field. Among them, the reduced-order model (ROM) has become the mainstream technical approach due to its core advantage of "reducing the dimensionality of high-dimensional problems".

[0018] In related technologies, orthogonal decomposition (POD), singular value decomposition (SVD), or data-driven AI-reduced order modeling (AI-ROM) are commonly used for order reduction. However, these technologies share common inherent limitations in the basis generation method: (1) Limitations and insufficient expressive power of the substrate: SVD / POD is essentially a linear decomposition technique, and the substrate extracted is linear. For complex physical behaviors with high nonlinearity such as structural cracking and material plasticity, a large number of linear substrates need to be superimposed to approximate the simulation. This not only weakens the effect of order reduction, but also often makes it difficult to accurately capture high-frequency and local fine physical features, resulting in an inherent accuracy bottleneck in the model.

[0019] (2) The "task specificity" and lack of generalization ability of the base: The traditional base is extracted from the simulation data of a specific structure under a specific working condition. Therefore, it is only effective for "this one" problem and does not have any knowledge transfer ability across structures or tasks. Once a new structure or working condition beyond the training range is encountered, the entire base will fail and new simulation data must be generated and the base must be recalculated.

[0020] (3) High cost and inefficiency of basis construction: Since the basis does not have generalization ability, every new reduced-order model project must repeat the complete process of "performing a large number of high-fidelity simulations to generate snapshot data and performing SVD decomposition on the snapshot matrix to extract the basis". The above process consumes a lot of time, thus increasing the application cost of the technology.

[0021] Based on the above analysis of the relevant technologies, it can be seen that the above technologies all adopt traditional mathematical methods such as orthogonal projection decomposition and singular value decomposition to extract the basis from the simulation snapshot data of specific problems. Although this basis generation mode has mathematical rigor, it fundamentally limits the expressive power and generalization potential of the reduced-order model.

[0022] Based on the above research, this disclosure provides a method for determining the physical field of an engineering structure. The method first obtains the operating parameters of the target entity structure and inputs these parameters into a physical field reconstruction model. The physical field reconstruction model includes a coefficient prediction network and a basis generator. The basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping relationship between the geometric features of different engineering structures and their corresponding global response physical fields. The basis generator provides a target reduced-order basis adapted to the target entity structure. Then, the coefficient prediction network is used to extract basis coefficients from the operating parameters to obtain target basis coefficients. Based on the target basis coefficients and the target reduced-order basis provided by the basis generator, physical field reconstruction is performed to obtain the target global response physical field. The target global response physical field reflects the structural state of the target entity structure under the operating parameters.

[0023] In this embodiment, the basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping rules between the geometric features of different engineering structures and the corresponding global response physical fields. This enables the basis generator to have strong nonlinear expression capabilities and cross-structure transfer capabilities, improving the generalization of the basis generator. The basis generator is then transferred to the physical field reconstruction model, enabling it to provide a target reduced-order basis that adapts to the target entity structure. Thus, the physical field is reconstructed based on the target basis coefficients and the target reduced-order basis, which can improve the accuracy of the target global response physical field and thereby improve the evaluation accuracy of the structural state of the target entity structure under the working parameters.

[0024] To facilitate understanding of this embodiment, a detailed description of the physical field determination method for an engineering structure disclosed in this disclosure is provided first. The execution entity of the physical field determination method for an engineering structure provided in this disclosure is generally a computer device. This computer device can be a server, which can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud storage, big data, and artificial intelligence platforms. In other embodiments, the computer device can also be a terminal device, which can be a mobile device, terminal, handheld device, computing device, vehicle-mounted device, etc.

[0025] In other embodiments, the method can also be applied to an implementation environment consisting of computer equipment and servers, or an implementation environment consisting of terminal equipment and servers. Furthermore, the method for determining the physical field of this engineering structure can also be implemented by a processor calling computer-readable instructions stored in memory.

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention.

[0027] Please see the appendix Figure 1 This is a flowchart illustrating a method for determining the physical field of an engineering structure, as shown in an exemplary embodiment of this application. Figure 1 As shown, the method for determining the physical field of the engineering structure in this embodiment may include the following steps S101~S102: S101: Obtain the working condition parameters of the target entity structure and input the working condition parameters into the physical field reconstruction model; the physical field reconstruction model includes a coefficient prediction network and a basis generator; the basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping law between the geometric features of different engineering structures and the corresponding global response physical field, and the basis generator is used to provide a target reduced-order basis that adapts to the target entity structure.

[0028] The target entity structure can be of any type, such as a flat plate, truss, polyhedron, bridge, or other topological structures, without any limitation.

[0029] The operating parameters may include at least one of the following: load parameters, physical constraint parameters, and environmental parameters. Load parameters are parameters that characterize the external forces (or energy) acting on the target structure, and may include load magnitude, load distribution, load application location, etc. Physical constraint parameters are used to constrain the displacement and rotation of the target structure in a specific direction. Specifically, they may include constraint location parameters (e.g., endpoints, edges, etc.), mechanical constraints of structural materials (e.g., strength, hardness, elastic modulus, etc.), and geometric constraints (e.g., deformation constraints and geometric correlation constraints; deformation constraints include the maximum allowable linear displacement of the structure, and geometric correlation constraints may include the geometric constraint relationships between multiple components). Environmental parameters may include temperature parameters, humidity parameters, wind force parameters, etc., and are not limited here.

[0030] Here, the trained base mode model is a neural network model that learns the mapping law between the geometric features of different engineering structures and the corresponding global response physical field. In this way, the basis generator can learn a set of nonlinear reduced-order basis with good orthogonality and strong expressive power.

[0031] In practical applications, the physical field reconstruction model is a lightweight surrogate model. This physical field reconstruction model is fine-tuned for the target entity structure. That is, the basis generator loaded from the trained base mode model does not have specificity for a certain type of structure. Therefore, the physical field reconstruction model can be fine-tuned so that the basis generator can generate a reduced-order basis that is adapted to a certain type of structure. In this embodiment, the basis generator can provide a target reduced-order basis that is adapted to the target entity structure. In this way, for different operating parameters of the target entity structure, it is not necessary to repeatedly generate the target reduced-order basis.

[0032] The fine-tuning process of the physics field reconstruction model will be described in detail later (see [link]). Figure 3 (This will not be elaborated upon here.) S102: Using the coefficient prediction network, the basis coefficients of the operating conditions are extracted to obtain the target basis coefficients. Based on the target basis coefficients and the target reduced-order basis provided by the basis generator, the physical field is reconstructed to obtain the target global response physical field. The target global response physical field is used to reflect the structural state of the target entity structure under the operating conditions.

[0033] Among them, the coefficient prediction network is a lightweight, well-trained neural network (also known as the task head), which is usually a multilayer perceptron network (MLP).

[0034] Here, the target global response physical field can refer to the physical distribution field that characterizes the structural mechanics of the target entity structure in the whole space under the action of the corresponding working parameters. Its core is the quantification process of the global response mapping of the structure to external forces. Specifically, the target global response physical field can include at least one of the stress field, strain field, displacement field or temperature field. The coupled physical field can be used to characterize the coupling relationship between any two of the above three physical fields.

[0035] Using a coefficient prediction network, the basis coefficients of the operating conditions are extracted to obtain the target basis coefficients. Based on the target basis coefficients and the target reduced-order basis, the physical field is reconstructed. Specifically, the target basis data and the target reduced-order basis are multiplied to obtain the target global response physical field. For example, the stress distribution, displacement, and strain distribution of the entire structure can be obtained. Based on the above information, the structural state of the target entity structure under the operating conditions can be reflected, such as whether stress concentration occurs, whether the structural deformation is within the preset range, and whether material damage affects the structural health.

[0036] In this embodiment, the basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping rules between the geometric features of different engineering structures and the corresponding global response physical fields. This enables the basis generator to have strong nonlinear expression capabilities and cross-structure transfer capabilities, improving the generalization of the basis generator. The basis generator is then transferred to the physical field reconstruction model, enabling it to provide a target reduced-order basis that adapts to the target entity structure. Thus, the physical field is reconstructed based on the target basis coefficients and the target reduced-order basis, which can improve the accuracy of the target global response physical field and thereby improve the evaluation accuracy of the structural state of the target entity structure under the working parameters.

[0037] The following is combined with Figure 2 The flowchart shown illustrates the training process of a basic modal model, and provides a detailed introduction to the training process of the trained basic modal model.

[0038] like Figure 2 As shown, the training process includes steps S201 to S204: S201: Obtain multiple training sample pairs, wherein the training sample pairs include geometric feature samples of the same engineering structure and corresponding masking response physical field samples, the training sample pairs have sample labels, the sample labels are used to indicate the global response physical field of the engineering structure, and the masking response physical field samples are obtained by masking the global response physical field.

[0039] Here, multiple training sample pairs can correspond to different engineering structures, which can be physical engineering structures or virtual engineering structures (such as virtual 3D models).

[0040] Geometric feature samples are used to characterize the geometric features of the corresponding engineering structure. Optionally, geometric features may include at least one of topological features, shape features, size features, and position and orientation features. Topological features can characterize the connection methods and hole distribution between components; shape features are used to characterize the structural outline and the outline shape of the components in the engineering structure; size features may include the dimensions of the engineering structure and the size information of the components within the engineering structure; and position and orientation features may include the spatial orientation of the components relative to the engineering structure and the relative positional relationships between multiple components. The above-described geometric features are merely exemplary; other geometric features may be included in other embodiments, which are not limited here.

[0041] Here, if the engineering structure is a physical engineering structure, the global response physical field can be obtained through actual measurement; if the engineering structure is a virtual engineering structure, the global response physical field can be calculated or obtained.

[0042] The content of the global response physics field is similar to that of the target global response physics field described above, and will not be repeated here.

[0043] Here, for any engineering structure, the geometric space can be transformed into a set of spatial grid cells. That is, the geometric space can be divided into multiple grid cells, each of which has a corresponding physical quantity (such as stress or displacement). This constitutes a global response physical field. Based on this, a part of the global response physical field can be masked. Specifically, some grid cells can be masked to obtain the masked response physical field.

[0044] S202: For each training sample pair, the masking response physical field sample is input to the coefficient extractor to obtain the basis coefficients, and the geometric feature sample is input to the basis generator to obtain the reduced basis.

[0045] In this embodiment, the base mode model includes a coefficient extractor (Transformer encoder), an information aggregation layer, and a base generator (Transformer decoder).

[0046] In this way, for each training sample pair, the mask response physical field sample can be input into the coefficient extractor to obtain the basis coefficients, where the basis coefficients are a coefficient vector including multidimensional coefficients.

[0047] Simultaneously, geometric feature samples are input into the basis generator to obtain a reduced-order basis, wherein the reduced-order basis includes a basis tensor with multiple channels.

[0048] It should be noted that each coefficient corresponds to a basis tensor for one channel. The basis tensors of each channel are used to characterize a distribution pattern or physical mode of the target global response physical field (at least one of stress field, strain field, displacement field, and temperature field). The target global response physical field is reconstructed by weighted superposition of the basis tensors of the multiple channels according to their respective basis coefficients. Specifically, the basis tensors of the multiple channels can be divided according to the type of physical field. The basis tensors of at least one channel corresponding to the same physical field type are used to jointly characterize the corresponding physical field. For example, the basis tensors of at least one channel corresponding to the stress field type are used to jointly characterize the stress field; similarly, the basis tensors of at least one channel corresponding to the strain field type are used to jointly characterize the strain field.

[0049] In some implementations, when the geometric feature samples are input into the basis generator to obtain the reduced-order basis, the initial learnable modality query parameters of the basis generator can be obtained, and the cross-attention mechanism can be used to fuse the initial learnable modality query parameters with the geometric feature samples to obtain the reduced-order basis.

[0050] Here, the initial learnable modality query parameters can be obtained through random initialization (such as Gaussian distribution initialization or normal distribution (Xavier) initialization).

[0051] The initial learnable modality query parameters are a set of feature vector matrices that exist as parameters of the neural network. ,in K The number of base channels is the preset number (i.e., the number of potential modal prototypes). D For feature dimension, this K Each query vector corresponds to K A general physical deformation pattern to be learned.

[0052] During the training of the base modality model, the initial learnable modality query parameters do not depend on any specific structural input, but rather participate as globally shared parameters in the backpropagation of all training samples. Specifically, in the basis generator, this... K Each query vector is used as a query parameter Query, and cross-attention is performed with the geometric feature samples to obtain a reduced-order basis.

[0053] S203: Perform physical field reconstruction processing on the basis coefficients and the reduced-order basis to obtain the reconstructed global response physical field.

[0054] Here, the information aggregation layer can be used to reconstruct the physical field of the basis coefficients and the reduced basis. The reconstruction process refers to multiplying the basis coefficients and the reduced basis to obtain the global response physical field.

[0055] Specifically, the coefficients of each dimension can be multiplied by the basis tensors of the corresponding channels, and the results of each multiplication can be summed to obtain the reconstructed global response physical field.

[0056] In other words, the information aggregation layer performs the following operation: converting the coefficient vector... c Each coefficient The basis tensor corresponding to the channel of the basis tensor Perform scalar multiplication, and then sum along the channel dimension to obtain the reconstructed global response physics field, as shown in formula (1): (1) S204: Based on the reconstructed global response physical field and the global response physical field indicated by the label, determine the model loss, and adjust the parameters of the coefficient extractor and the basis generator based on the model loss until the preset convergence condition is reached to obtain the trained basic mode model.

[0057] It is understandable that after determining the reconstructed global response physics field, the model loss can be determined based on the reconstructed global response physics field and the global response physics field indicated by the label. The parameters of the coefficient extractor and the basis generator can be adjusted by backpropagation based on the model loss until the preset convergence condition is reached, and the trained basic mode model is obtained.

[0058] Specifically, when adjusting the parameters of the basis generator, one can adjust the initial learnable mode query parameters. Specifically, the gradient descent algorithm continuously adjusts these initial learnable mode query parameters based on the model loss. K The values ​​of the query vectors are gradually converged to the optimal state that can most effectively detect various orthogonal physical characteristics in the structure (such as overall bending, local torsion, high-frequency oscillation, etc.).

[0059] The preset convergence condition can be that the model loss is less than a preset loss threshold.

[0060] Thus, the sample structure of the above linear combination forces the basis generator to learn a set of nonlinear "reduced-order basis" with good orthogonality, strong expressiveness and generalization, while the coefficient extractor learns how to extract the corresponding basis coefficients.

[0061] After training, the adjusted initial learnable modal query parameters are fixed, thus solidifying into a universal "physical modal dictionary." For any new structure, the basis generator uses the adjusted initial learnable modal query parameters to "scan" the structure's geometry and topology, thereby generating a geometry adapted to that structure. K Given a basis tensor, we obtain the reduced-order basis corresponding to the structure.

[0062] The training process of the physics field reconstruction model is described in detail below. In this embodiment, the physics field reconstruction model is obtained by supervised training based on multiple working condition parameter training samples of the target structure and the geometric features of the target structure. The working condition parameter training samples correspond to the first global response physics field label. The multiple working condition parameter training samples include the working condition parameters of the target structure under different working conditions. The similarity between the target structure and the target entity structure is greater than the preset similarity threshold.

[0063] As described above, for any new structure, the basis generator uses adjusted initial learnable modal query parameters to "scan" the structure's geometric topology, thereby generating a geometry that fits the structure. K Given a basis tensor, we obtain the reduced-order basis corresponding to the structure. Therefore, the physics reconstruction model refers to the fine-tuned model for a specific structure. That is, for any type of structure, we need to perform targeted model fine-tuning so that the physics reconstruction model can adapt to this structure. Therefore, during the training process, the similarity between the target structure and the target entity structure is greater than the preset similarity threshold. For example, the target structure and the target entity structure are both truss structures, or the target structure and the target entity structure are both octahedral hollow structures, etc.

[0064] The preset similarity threshold can be set according to actual needs, and there is no limitation here. For example, it can be set to 85% or 90%.

[0065] Figure 3 A flowchart illustrating the training process of a physics field reconstruction model provided as an exemplary embodiment of this application.

[0066] like Figure 3 As shown, the training process of the physics field reconstruction model includes the following steps S301~S304: S301: Load the basis generator from the trained base modality model and fix the parameters of the basis generator.

[0067] Here, the parameters of the fixed basis generator can refer to the learnable modal query parameters of the fixed basis generator (i.e., the adjusted initial learnable modal query parameters mentioned above), which enable the physical field reconstruction model to scan the geometry of the target structure based on the learnable modal query parameters.

[0068] S302: Input the training samples of the working condition parameters into the coefficient prediction network to be trained to obtain the predicted basis coefficients, and input the geometric features of the target structure into the basis generator with fixed parameters to obtain the target reduced basis corresponding to the target structure.

[0069] The training samples for operating parameters include the operating parameters of the target structure.

[0070] In this step, the training samples of the working condition parameters are input into the coefficient prediction network to be trained to obtain the predicted basis coefficients, and the geometric features of the target structure are input into the basis generator with fixed parameters to obtain the target reduced basis corresponding to the target structure. Here, the basis generator can scan the geometric features of the target structure based on the learnable modality query parameters to obtain the target reduced basis corresponding to the target structure, that is, to generate a target reduced basis that is adapted to the target structure.

[0071] S303: Based on the predicted basis coefficients and the target reduced-order basis, the response physical field is reconstructed to obtain the first reconstructed global response physical field.

[0072] After generating a target reduced-order basis that fits the target structure, the response physical field can be reconstructed based on the predicted basis coefficients and the target reduced-order basis to obtain the first reconstructed global response physical field.

[0073] S304: Based on the difference between the first reconstructed global response physical field and the first global response physical field, update the network parameters of the coefficient prediction network to be trained until the preset training requirements are met, and obtain the physical field reconstruction model.

[0074] In this way, based on the difference between the first reconstructed global response physical field and the first global response physical field, the network parameters of the coefficient prediction network to be trained can be updated by backpropagation until the preset training requirements are met, and the physical field reconstruction model is obtained.

[0075] The preset training requirement can be that the difference between the first reconstructed global response physical field and the first global response physical field is less than a preset difference threshold.

[0076] Through the above training process, the model can be fine-tuned to obtain a physical field reconstruction model for the target structure.

[0077] Similarly, for different types of structures, a "one-time training (basic mode model training) and multiple fine-tuning (fine-tuning of the physical field reconstruction model for different structures)" mode can be adopted. For example, fine-tuning can be done once for truss structures and once for octahedral structures. Through the above mode, the training cost can be distributed, thereby reducing the construction cost and construction time of high-precision reduced-order models for each structure.

[0078] The following is combined with Figure 4 The training process of the aforementioned fundamental mode model, the training process of the physical field reconstruction model, and the online prediction process are explained in detail.

[0079] Figure 4 This is a complete flowchart illustrating model training and physics prediction as an exemplary embodiment of this application. Figure 4 As shown, training sample pairs are first generated, which include masking response physics field samples and geometric feature samples. The basis mode model is trained based on the training sample pairs. Specifically, during the training process, the masking response physics field samples are input into the coefficient extractor to obtain the basis coefficients, and the geometric feature samples are input into the basis generator to obtain the reduced-order basis. Based on the basis coefficients and the reduced-order basis, reconstruction processing is performed to obtain the reconstructed global response physics field. The model loss is determined based on the reconstructed global response physics field and the global response physics field indicated by the label. The model convergence is judged based on the model loss. If it does not converge, the training of the basis mode model is returned.

[0080] If convergence is achieved, a basis generator is loaded from the fundamental modal model, and a physical field reconstruction model is constructed based on the basis generator with fixed parameters and the coefficient prediction network. Then, for a specific structure (the target structure in this application), the physical field reconstruction model is fine-tuned (i.e., the parameters of the physical field reconstruction model are adjusted). After fine-tuning, the physical field reconstruction model is used to reconstruct the physical field based on the operating parameters of the target structure (or other structures of the same type as the target structure, such as the target entity structure) to obtain the target global response physical field.

[0081] Corresponding to the embodiments of the aforementioned method for determining the physical field of an engineering structure, this application also provides embodiments of a device for determining the physical field of an engineering structure.

[0082] Please see Figure 5 This is a schematic diagram illustrating the structure of a physical field determination device for an engineering structure, as shown in an exemplary embodiment of this application. Figure 5 As shown, the physical field determination device 500 for this engineering structure includes: The parameter acquisition module 510 is used to acquire the working condition parameters of the target entity structure and input the working condition parameters into the physical field reconstruction model; the physical field reconstruction model includes a coefficient prediction network and a basis generator; the basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping law between the geometric features of different engineering structures and the corresponding global response physical field; the basis generator is used to provide a target reduced-order basis that adapts to the target entity structure. The physics field reconstruction module 520 is used to extract the basis coefficients of the operating parameters using the coefficient prediction network to obtain the target basis coefficients, and to perform physics field reconstruction processing based on the target basis coefficients and the target reduced-order basis provided by the basis generator to obtain the target global response physics field; the target global response physics field is used to reflect the structural state of the target entity structure under the operating parameters.

[0083] Please see Figure 6 This is a schematic diagram illustrating the structure of a physical field determination device for another engineering structure, as shown in an exemplary embodiment of this application. Figure 6 As shown, the physical field determination device 500 of the engineering structure also includes a model training module 530; The base modal model further includes a coefficient extractor; the model training module 530 is used for: Multiple training sample pairs are obtained. Each training sample pair includes geometric feature samples of the same engineering structure and corresponding masking response physical field samples. Each training sample pair has a sample label, which is used to indicate the global response physical field corresponding to the engineering structure. The masking response physical field samples are obtained by masking the global response physical field. For each training sample pair, the masking response physical field sample is input into the coefficient extractor to obtain the basis coefficients, and the geometric feature sample is input into the basis generator to obtain the reduced basis; The basis coefficients and the reduced-order basis are subjected to physical field reconstruction processing to obtain the reconstructed global response physical field; Based on the reconstructed global response physics field and the global response physics field indicated by the label, the model loss is determined, and the parameters of the coefficient extractor and the basis generator are adjusted based on the model loss until the preset convergence condition is reached, thereby obtaining the trained basic mode model.

[0084] In some implementations, the model training module 530 is specifically used for: Obtain the initial learnable modality query parameters of the basis generator; The initial learnable modality query parameters and the geometric feature samples are fused using a cross-attention mechanism to obtain the reduced-order basis; When the model training module 530 adjusts the parameters of the coefficient extractor and the basis generator based on the model loss, it is specifically used for: The parameters of the coefficient extractor are adjusted based on the model loss, and the initial learnable modality query parameters of the basis generator are adjusted based on the model loss.

[0085] In some embodiments, the basis coefficients are coefficient vectors including multidimensional coefficients, and the reduced-order basis includes a basis tensor with multiple channels, wherein each dimension of the coefficient corresponds to a basis tensor with one channel; the model training module 530 is specifically used for: The coefficients of each dimension are multiplied by the basis tensors of the corresponding channels, and the results of each multiplication are summed to obtain the reconstructed global response physical field.

[0086] In some implementations, the physical field reconstruction model is obtained through supervised training based on multiple working condition parameter training samples for the target structure and the geometric features of the target structure. The working condition parameter training samples correspond to a first global response physical field label. The multiple working condition parameter training samples include the working condition parameters of the target structure under different working conditions. The similarity between the target structure and the target entity structure is greater than a preset threshold.

[0087] In some embodiments, the model training module 530 is further configured to: Load the basis generator from the trained base modality model and fix the parameters of the basis generator; The training samples of the operating condition parameters are input into the coefficient prediction network to be trained to obtain the predicted basis coefficients, and the geometric features of the target structure are input into the basis generator with fixed parameters to obtain the target reduced basis corresponding to the target structure. Based on the predicted basis coefficients and the target reduced basis, the response physical field is reconstructed to obtain the first reconstructed global response physical field. Based on the difference between the first reconstructed global response physical field and the first global response physical field, the network parameters of the coefficient prediction network to be trained are updated until the preset training requirements are met, and the physical field reconstruction model is obtained.

[0088] In some embodiments, the operating parameters include at least one of the load parameters, physical constraint parameters, and environmental parameters of the target physical structure; the target global response physical field includes at least one of the stress field, strain field, displacement field, or temperature field.

[0089] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0090] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0091] Corresponding to the above-described method for determining the physical field of engineering structures, this disclosure also provides a computer device, such as... Figure 7 The diagram shown is a structural schematic of a computer device provided in an embodiment of this disclosure. Figure 7 As shown, the computer device 700 includes a processor 710, an internal bus 720, memory 730, a network interface 740, and non-volatile memory 750, and may also include other hardware required for its functions. One or more embodiments of this specification can be implemented in software, for example, the processor 710 reads the corresponding computer program from the non-volatile memory 750 into the memory 730 and then runs it. Of course, besides software implementation, one or more embodiments of this specification do not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0092] The memory 730, also known as internal memory, is used to temporarily store the computational data in the processor 710, as well as the data exchanged with non-volatile memory 750 such as hard disk. The processor 710 exchanges data with the non-volatile memory 750 through the memory 730.

[0093] In this embodiment, memory 730 is specifically used to store application code that executes the solution of this application, and its execution is controlled by processor 710. That is, when the computer device is running, processor 710 communicates with network interface 740, memory 730 and non-volatile memory 750 through internal bus 720, so that processor 710 executes the application code stored in memory 730 and non-volatile memory 750, thereby executing the physical field determination method of the engineering structure described in the above method embodiment.

[0094] Processor 710 may be an integrated circuit chip with signal processing capabilities. The aforementioned processor can be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it can also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware microservices. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this invention. The general-purpose processor can be a microprocessor or any conventional processor.

[0095] It is understood that the structures illustrated in the embodiments of this application do not constitute a specific limitation on the computer device 700. In other embodiments of this application, the computer device 700 may include more or fewer components than illustrated, or combine some components, or split some components, or have different component arrangements. The illustrated components may be implemented in hardware, software, or a combination of software and hardware.

[0096] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the physical field determination method for the engineering structure described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0097] This disclosure also provides a computer program product carrying program code. The program code includes instructions that can be used to execute the steps of the physical field determination method for engineering structures in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0098] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0099] The embodiments of the subject matter and functional operation described in this specification can be implemented in the following ways: digital electronic circuits, tangibly embodied computer software or firmware, computer hardware including the structures disclosed in this specification and their structural equivalents, or combinations thereof. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions encoded on a tangible, non-transitory program carrier for execution by a data processing apparatus or for controlling the operation of a data processing apparatus. Alternatively or additionally, the program instructions may be encoded on artificially generated propagation signals, such as machine-generated electrical, optical, or electromagnetic signals, which are generated to encode information and transmit it to a suitable receiving device for execution by the data processing apparatus. The computer storage medium may be a machine-readable storage device, a machine-readable storage substrate, a random or serial access memory device, or combinations thereof.

[0100] The processing and logic flow described in this specification can be executed by one or more programmable computers that execute one or more computer programs to perform corresponding functions by operating on input data and generating output. The processing and logic flow can also be executed by dedicated logic circuitry—such as FPGAs (Field-Programmable Gate Arrays) or ASICs (Application-Specific Integrated Circuits), and the device can also be implemented as dedicated logic circuitry.

[0101] Computers suitable for executing computer programs include, for example, general-purpose and / or special-purpose microprocessors, or any other type of central processing unit. Typically, the central processing unit receives instructions and data from read-only memory and / or random access memory. Basic computer microservices include a central processing unit for implementing or executing instructions and one or more memory devices for storing instructions and data. Typically, a computer will also include one or more mass storage devices for storing data, such as disks, magneto-optical disks, or optical disks, or the computer will be operatively coupled to such mass storage devices to receive data from or transfer data to them, or both. However, a computer is not required to have such devices. Furthermore, a computer can be embedded in another device, such as a mobile phone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a global positioning system (GPS) receiver, or a portable storage device such as a universal serial bus (USB) flash drive, to name a few.

[0102] Computer-readable media suitable for storing computer program instructions and data 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-ROM and DVD-ROM disks. Processors and memory may be supplemented by or incorporated into dedicated logic circuitry.

[0103] 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.

[0104] 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 microservices in the above embodiments should not be construed as requiring such separation in all embodiments, and it should be understood that the described program microservices and systems can generally be integrated together in a single software product or packaged into multiple software products.

[0105] Thus, specific embodiments of the subject matter have been described. Other embodiments are within the scope of the appended claims. In some cases, the actions recited in the claims may be performed in a different order and still achieve the desired result. Furthermore, the processes depicted in the drawings are not necessarily shown in a specific order or sequence to achieve the desired result. In some implementations, multitasking and parallel processing may be advantageous.

[0106] 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 method for determining the physical field of an engineering structure, characterized in that, include: The working parameters of the target entity structure are obtained and input into the physical field reconstruction model. The physical field reconstruction model includes a coefficient prediction network and a basis generator. The basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping law between the geometric features of different engineering structures and the corresponding global response physical field. The basis generator is used to provide a target reduced-order basis that is adapted to the target entity structure. The physical field reconstruction model is obtained by supervised training based on multiple working condition parameter training samples for the target structure and the geometric features of the target structure. The working condition parameter training samples correspond to a first global response physical field label. The multiple working condition parameter training samples include the working condition parameters of the target structure under different working conditions. The similarity between the target structure and the target entity structure is greater than a preset threshold. Using the coefficient prediction network, the basis coefficients of the operating conditions are extracted to obtain the target basis coefficients. Based on the target basis coefficients and the target reduced-order basis provided by the basis generator, physical field reconstruction is performed to obtain the target global response physical field. The target global response physical field is used to reflect the structural state of the target entity structure under the operating conditions. The base mode model also includes a coefficient extractor; the trained base mode model is trained through the following training steps: obtaining multiple training sample pairs, the training sample pairs including geometric feature samples of the same engineering structure and corresponding masked response physical field samples, the training sample pairs having sample labels, the sample labels being used to indicate the global response physical field corresponding to the engineering structure, and the masked response physical field samples being obtained by masking the global response physical field. For each training sample pair, the mask response physical field sample is input into the coefficient extractor to obtain the basis coefficients, and the initial learnable modality query parameters of the basis generator are obtained; the initial learnable modality query parameters and the geometric feature sample are fused using a cross-attention mechanism to obtain a reduced-order basis; The basis coefficients and the reduced-order basis are subjected to physical field reconstruction processing to obtain the reconstructed global response physical field; based on the reconstructed global response physical field and the global response physical field indicated by the label, the model loss is determined, and the parameters of the coefficient extractor and the basis generator are adjusted based on the model loss until the preset convergence condition is reached to obtain the trained basic mode model. Wherein, the basis coefficients are coefficient vectors including multidimensional coefficients, and the reduced-order basis includes a basis tensor with multiple channels, with each dimension of coefficient corresponding to a basis tensor of one channel; the physical field reconstruction processing of the basis coefficients and the reduced-order basis to obtain the reconstructed global response physical field includes: performing a dot product of each dimension of coefficient with the basis tensor of the corresponding channel, and summing the results of each dot product to obtain the reconstructed global response physical field.

2. The method according to claim 1, characterized in that, The adjustment of the parameters of the coefficient extractor and the basis generator based on the model loss includes: The parameters of the coefficient extractor are adjusted based on the model loss, and the initial learnable modality query parameters of the basis generator are adjusted based on the model loss.

3. The method according to claim 1, characterized in that, The physical field reconstruction model is trained through the following steps: Load the basis generator from the trained base modality model and fix the parameters of the basis generator; The training samples of the operating condition parameters are input into the coefficient prediction network to be trained to obtain the predicted basis coefficients, and the geometric features of the target structure are input into the basis generator with fixed parameters to obtain the target reduced basis corresponding to the target structure. Based on the predicted basis coefficients and the target reduced basis, the response physical field is reconstructed to obtain the first reconstructed global response physical field. Based on the difference between the first reconstructed global response physical field and the first global response physical field, the network parameters of the coefficient prediction network to be trained are updated until the preset training requirements are met, and the physical field reconstruction model is obtained.

4. The method according to any one of claims 1-3, characterized in that, The operating parameters include at least one of the load parameters, physical constraint parameters, and environmental parameters of the target entity structure; the target global response physical field includes at least one of the stress field, strain field, displacement field, or temperature field.

5. A device for determining the physical field of an engineering structure, characterized in that, include: The parameter acquisition module is used to acquire the working condition parameters of the target entity structure and input the working condition parameters into the physical field reconstruction model. The physical field reconstruction model includes a coefficient prediction network and a basis generator. The basis generator is loaded from a trained base mode model, which is a neural network model that learns the mapping law between the geometric features of different engineering structures and the corresponding global response physical field. The basis generator is used to provide a target reduced-order basis that is adapted to the target entity structure. The physical field reconstruction model is obtained by supervised training based on multiple working condition parameter training samples for the target structure and the geometric features of the target structure. The working condition parameter training samples correspond to a first global response physical field label. The multiple working condition parameter training samples include the working condition parameters of the target structure under different working conditions. The similarity between the target structure and the target entity structure is greater than a preset threshold. The physics field reconstruction module is used to extract the basis coefficients from the operating parameters using the coefficient prediction network to obtain the target basis coefficients, and to perform physics field reconstruction processing based on the target basis coefficients and the target reduced-order basis provided by the basis generator to obtain the target global response physics field; the target global response physics field is used to reflect the structural state of the target entity structure under the operating parameters. The base mode model also includes a coefficient extractor; a model training module is used to train the trained base mode model through the following training steps: acquiring multiple training sample pairs, the training sample pairs including geometric feature samples of the same engineering structure and corresponding masked response physical field samples, the training sample pairs having sample labels, the sample labels being used to indicate the global response physical field corresponding to the engineering structure, and the masked response physical field samples being obtained by masking the global response physical field; For each training sample pair, the mask response physical field sample is input into the coefficient extractor to obtain the basis coefficients, and the initial learnable modality query parameters of the basis generator are obtained; the initial learnable modality query parameters and the geometric feature sample are fused using a cross-attention mechanism to obtain a reduced-order basis; The basis coefficients and the reduced-order basis are subjected to physical field reconstruction processing to obtain the reconstructed global response physical field; based on the reconstructed global response physical field and the global response physical field indicated by the label, the model loss is determined, and the parameters of the coefficient extractor and the basis generator are adjusted based on the model loss until the preset convergence condition is reached to obtain the trained basic mode model. Wherein, the basis coefficients are coefficient vectors including multidimensional coefficients, and the reduced-order basis includes a basis tensor with multiple channels, with each dimension of coefficient corresponding to a basis tensor with one channel. When the model training module performs physical field reconstruction processing on the basis coefficients and the reduced-order basis to obtain the reconstructed global response physical field, it specifically performs the following: multiplying each dimension coefficient with the basis tensor of the corresponding channel, and summing the results of each multiplication to obtain the reconstructed global response physical field.

6. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the physical field determination method for the engineering structure according to any one of claims 1-4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the physical field determination method for the engineering structure according to any one of claims 1-4.

Images