Solid mechanics calculation method and device based on double-grid dynamic weight optimization

Abstract

The invention particularly relates to a solid mechanical calculation method and device based on double-grid dynamic weight optimization, and the method comprises the steps: judging whether a solid mechanical calculation demand in a plane plate tension process exists or not; if so, establishing a solid mechanics problem in the surface plate tension process, and establishing a coarse mesh discrete model and a fine mesh discrete model; establishing a first nonlinear mapping relation and a second nonlinear mapping relation based on a preset deep neural network and the two models; and constructing a composite loss function, and based on a preset training strategy, performing staged cooperative training on the composite loss function according to the first nonlinear mapping relationship and the second nonlinear mapping relationship until a preset training round is reached or the trained composite loss function meets a preset precision requirement, and solving according to the trained composite loss function to obtain a solution of a solid mechanics problem displacement field. Therefore, the problem that the convergence speed of a fine mesh model is low in the prior art is solved, and the training efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computational mechanics and artificial intelligence, in particular to a solid mechanics calculation method and device based on double-grid dynamic weight optimization. BACKGROUND

[0002] The field of solid mechanics has long relied on traditional numerical methods such as the finite element method and the finite difference method. In recent years, with the rise of physical information deep neural networks, the effective integration of data-driven and physical laws has been achieved by embedding solid mechanics control equations into the loss function of neural networks. However, the training efficiency still faces challenges when dealing with specific complex problems. For example, for crack problems with high gradients and singularities (such as crack tip stress concentration), the training of physical information deep neural networks often requires more computational resources and longer convergence time. In addition, due to the dramatic changes in the solution near the crack tip, physical information deep neural networks may also require more refined parameter settings and longer training time to achieve satisfactory accuracy.

[0003] In related technologies, adaptive weight adjustment algorithms can dynamically balance the loss term weights to address the training efficiency issues of traditional physical information deep neural networks.

[0004] However, the adaptive weight adjustment algorithm requires additional calculation of weight gradients at each iteration, increasing the single iteration time. The transfer learning pre-training strategy often has poor transfer effect in solid mechanics problems due to the diversity of boundary conditions, which needs to be addressed. SUMMARY

[0005] The present application provides a solid mechanics calculation method and device based on double-grid dynamic weight optimization to solve the problem of slow convergence of fine grid models in related technologies.

[0006] The first aspect of the present application provides a solid mechanics calculation method based on double-grid dynamic weight optimization, comprising the following steps: determining whether there is a solid mechanics calculation demand for the tensile process of a plane plate; If there is a solid mechanics calculation demand for the tensile process of a plane plate, a solid mechanics problem for the tensile process of a plane plate is established, and a coarse grid discrete model and a fine grid discrete model are established for the solid mechanics problem; Based on a pre-set deep neural network, a first nonlinear mapping relationship and a second nonlinear mapping relationship are established according to the coarse grid discrete model and the fine grid discrete model; construct a composite loss function, and based on a preset training strategy, perform staged collaborative training on the composite loss function according to the first nonlinear mapping relationship and the second nonlinear mapping relationship until a preset training round is reached or the composite loss function after training meets a preset accuracy requirement, and solve the solid mechanics problem of the flat panel tension process according to the composite loss function after training to obtain a solution of the displacement field of the solid mechanics problem.

[0007] Optionally, the first nonlinear mapping relationship and the second nonlinear mapping relationship are established based on a preset deep neural network according to the coarse grid discrete model and the fine grid discrete model, including: obtaining a first node coordinate set discretized by the coarse grid discrete model and a second node coordinate set discretized by the fine grid discrete model; establishing a mapping relationship between the first node coordinate set and a first displacement field to obtain the first nonlinear mapping relationship, and establishing a mapping relationship between the second node coordinate set and a second displacement field to obtain the second nonlinear mapping relationship.

[0008] Optionally, the composite loss function is constructed, including: determining a first physical information constraint loss function of the coarse grid discrete model, a second physical information constraint loss function of the fine grid discrete model, a first weight corresponding to the first physical information constraint loss function, and a second weight corresponding to the second physical information constraint loss function; constructing the composite loss function according to the first physical information constraint loss function, the first weight, the second physical information constraint loss function, and the second weight.

[0009] Optionally, the staged collaborative training includes a pre-training stage, a migration stage, and a fine-tuning stage, and the composite loss function is trained according to the first nonlinear mapping relationship and the second nonlinear mapping relationship based on a preset training strategy, including: in the pre-training stage, setting the first weight to 1 and the second weight to 0, and initializing network parameters through the coarse grid discrete model; in the migration stage, controlling the first weight to decrease based on a preset decreasing strategy, and controlling the second weight to increase based on a preset increasing strategy, so as to perform knowledge migration from the coarse grid discrete model to the fine grid discrete model; in the fine-tuning stage, setting the first weight to 0 and the second weight to 1, and training the fine grid discrete model based on a preset accuracy training requirement.

[0010] Optionally, the first weight is determined by a first dynamic weight function, and the second weight is determined by a second dynamic weight function, wherein, the first dynamic weight function is: ; the second dynamic weight function is: ; wherein, ; ; wherein, is a first dynamic weight function, is an unnormalized first dynamic weight function, is an unnormalized second dynamic weight function, is a second dynamic weight function, is a current training epoch, and is a training epoch.

[0011] The second aspect embodiment of the application provides a solid mechanics calculation device based on double-grid dynamic weight optimization, comprising: a judgment module configured to determine whether there is a solid mechanics calculation demand for a plane plate tension process; a modeling module configured to, if there is a solid mechanics calculation demand for a plane plate tension process, establish a solid mechanics problem of the plane plate tension process, and establish a coarse-grid discrete model and a fine-grid discrete model for the solid mechanics problem; a mapping module configured to, based on a preset deep neural network, establish a first nonlinear mapping relationship and a second nonlinear mapping relationship according to the coarse-grid discrete model and the fine-grid discrete model, respectively; a determination module configured to construct a composite loss function, and based on a preset training strategy, perform staged collaborative training on the composite loss function according to the first nonlinear mapping relationship and the second nonlinear mapping relationship until a preset training epoch is reached or the composite loss function after training meets a preset accuracy requirement, solve the solid mechanics problem of the plane plate tension process according to the composite loss function after training, and obtain a solution of a displacement field of the solid mechanics problem.

[0012] Optionally, the mapping module is specifically configured to: obtain a first node coordinate set generated by discretization of the coarse-grid discrete model and a second node coordinate set generated by discretization of the fine-grid discrete model; establish a mapping relationship between the first node coordinate set and a first displacement field to obtain the first nonlinear mapping relationship, and establish a mapping relationship between the second node coordinate set and a second displacement field to obtain the second nonlinear mapping relationship.

[0013] Optionally, the composite loss function is constructed, comprising: determining a first physical information constraint loss function of the coarse grid discrete model, a second physical information constraint loss function of the fine grid discrete model, a first weight corresponding to the first physical information constraint loss function, and a second weight corresponding to the second physical information constraint loss function; constructing the composite loss function according to the first physical information constraint loss function, the first weight, the second physical information constraint loss function, and the second weight.

[0014] Optionally, the staged co-training comprises a pre-training stage, a migration stage, and a fine-tuning stage, and the determining module is specifically configured to: in the pre-training stage, setting the first weight as 1 and the second weight as 0, and initializing network parameters through the coarse grid discrete model; in the migration stage, controlling the first weight to decrease based on a preset decreasing strategy, and controlling the second weight to increase based on a preset increasing strategy, so as to perform knowledge migration from the coarse grid discrete model to the fine grid discrete model; in the fine-tuning stage, setting the first weight as 0 and the second weight as 1, and training the fine grid discrete model based on a preset accuracy training requirement.

[0015] Optionally, the first weight is determined by a first dynamic weight function, and the second weight is determined by a second dynamic weight function, wherein, the first dynamic weight function is: ; the second dynamic weight function is: ; wherein, ; ; wherein, is the first dynamic weight function, is an unnormalized first dynamic weight function, is an unnormalized second dynamic weight function, is the second dynamic weight function, is a current training round, and is a training round.

[0016] An electronic device is provided in a third aspect of the present application, comprising: at least one processor; and a memory connected with the at least one processor in communication; wherein the memory stores instructions executable by the at least one processor, and the instructions are configured to execute the solid mechanics calculation method based on double-grid dynamic weight optimization as described in the above embodiments.

[0017] A computer readable storage medium is provided in a fourth aspect of the present application, and the computer readable storage medium stores a computer program, which is executed by a processor to implement the solid mechanics calculation method based on double-grid dynamic weight optimization as described in the above embodiments.

[0018] A computer program product is provided in a fifth aspect of the present application, and the computer program product stores a computer program, which is executed by a processor to implement the solid mechanics calculation method based on double-grid dynamic weight optimization as described in the above embodiments.

[0019] Therefore, when it is determined that there is a solid mechanics calculation demand for the tensile process of the plane plate, the solid mechanics problem of the tensile process of the plane plate is established, and the coarse-grid discrete model and the fine-grid discrete model are established. The first nonlinear mapping relationship and the second nonlinear mapping relationship are established based on the preset deep neural network and the two models. The composite loss function is constructed, and based on the preset training strategy, the composite loss function is trained in stages according to the first nonlinear mapping relationship and the second nonlinear mapping relationship, until the preset training round is reached or the composite loss function after training meets the preset accuracy requirement, and the solution of the displacement field of the solid mechanics problem is obtained according to the composite loss function after training. Therefore, the problem of slow convergence of the fine-grid model in the related art is solved, and the training efficiency is improved.

[0020] Additional aspects and advantages of the present application will be in part apparent and in part pointed out hereinafter. BRIEF DESCRIPTION OF DRAWINGS

[0021] The above and / or additional aspects and advantages of the present application will become apparent and be readily appreciated from the following description, taken in conjunction with the accompanying drawings, in which: Figure 1 A flowchart of a solid mechanics calculation method based on double-grid dynamic weight optimization according to an embodiment of the present application is provided. Figure 2 A schematic diagram of different element discrete forms of a solid mechanics problem of a solid mechanics calculation method based on double-grid dynamic weight optimization according to an embodiment of the present application is provided. Figure 3A weight function schematic diagram of a solid mechanics calculation method based on double-grid dynamic weight optimization according to an embodiment of the present application is provided. Figure 4 A physical information deep neural network calculation flowchart of a solid mechanics calculation method based on double-grid dynamic weight optimization according to an embodiment of the present application is provided. Figure 5 A schematic diagram of a solid mechanics calculation device based on double-grid dynamic weight optimization according to an embodiment of the present application is provided. Figure 6 A structural schematic diagram of an electronic device according to an embodiment of the present application is provided. DETAILED DESCRIPTION

[0022] The embodiments of the present application are described in detail below, examples of which are shown in the accompanying drawings, in which the same or similar notations represent the same or similar elements or elements having the same or similar functions throughout. The embodiments described below by reference to the drawings are exemplary and are intended to explain the present application, and cannot be understood as limiting the present application.

[0023] The solid mechanics calculation method and device based on double-grid dynamic weight optimization of the embodiments of the present application are described below with reference to the accompanying drawings. In view of the slow convergence speed of the fine grid model in the related technology mentioned in the above background technology, the present application provides a solid mechanics calculation method based on double-grid dynamic weight optimization. In the method, when it is judged that there is a solid mechanics calculation demand for the tensile process of a plane plate, a solid mechanics problem of the tensile process of the plane plate is established, and a coarse grid discrete model and a fine grid discrete model are established. Based on a preset deep neural network and the two models mentioned above, a first nonlinear mapping relationship and a second nonlinear mapping relationship are established. A composite loss function is constructed, and based on a preset training strategy, the composite loss function is trained in stages according to the first nonlinear mapping relationship and the second nonlinear mapping relationship until a preset training round is reached or the composite loss function after training meets a preset accuracy requirement. The solution of the displacement field of the solid mechanics problem is obtained according to the composite loss function after training. Thus, the slow convergence speed of the fine grid model in the related technology is solved, and the training efficiency is improved.

[0024] Specifically, Figure 1 A flowchart of a solid mechanics calculation method based on double-grid dynamic weight optimization according to an embodiment of the present application is provided.

[0025] As Figure 1 shown, the solid mechanics calculation method based on double-grid dynamic weight optimization includes the following steps: In step S101, it is judged whether there is a solid mechanics calculation demand for the tensile process of a plane plate. The solid mechanics calculation requirement is a technical requirement for quantitatively predicting and evaluating the mechanical behavior of a solid structure or material under external load by using a numerical method or theoretical analysis.

[0026] In step S102, if there is a solid mechanics calculation requirement for the tensile process of the plane plate, a solid mechanics problem of the tensile process of the plane plate is established, and a coarse-grid discrete model and a fine-grid discrete model of the solid mechanics problem are established.

[0027] The solid mechanics problem refers to a scientific problem of studying the mechanical behavior of a solid material or structure, such as deformation, stress, stability, and dynamic response, under the action of external load.

[0028] Specifically, as shown in Figure 2 (a), as shown in Figure 2 (a) is a schematic diagram of a solid mechanics calculation method based on double-grid dynamic weight optimization for a plane plate tensile problem according to an embodiment of the present application. When there is a solid mechanics calculation requirement for the tensile process of the plane plate, a double-grid discretization strategy is used to construct a model for this typical solid mechanics problem. The coarse-grid model (model 1), i.e., the coarse-grid discrete model of the embodiment of the present application, is shown in Figure 2 (b), as shown in Figure 2 (b) is a schematic diagram of a solid mechanics problem coarse-grid element discretization form of a solid mechanics calculation method based on double-grid dynamic weight optimization according to an embodiment of the present application. A larger element size is used for discretization modeling. This model has the characteristics of fewer elements and high computational efficiency, and can achieve fast convergence during deep neural network training. The fine-grid model (model 2), i.e., the fine-grid discrete model of the embodiment of the present application, is shown in Figure 2 (c), as shown in Figure 2 (c) is a schematic diagram of a solid mechanics problem fine-grid element discretization form of a solid mechanics calculation method based on double-grid dynamic weight optimization according to an embodiment of the present application. A smaller element size is used for fine modeling. The increase in the number of elements reduces the convergence speed, but ensures that the solution meets the engineering precision requirements. It should be noted that the coarse-grid discrete model and the fine-grid discrete model are both discretizations of the same solid mechanics problem.

[0029] In step S103, based on the preset deep neural network, a first nonlinear mapping relationship and a second nonlinear mapping relationship are established according to the coarse-grid discrete model and the fine-grid discrete model.

[0030] Optionally, in some embodiments, the first nonlinear mapping relationship and the second nonlinear mapping relationship are established according to the coarse-grid discrete model and the fine-grid discrete model based on the preset deep neural network, including: obtaining a first node coordinate set discretized by the coarse-grid discrete model and a second node coordinate set discretized by the fine-grid discrete model; establishing a mapping relationship between the first node coordinate set and the first displacement field to obtain the first nonlinear mapping relationship, and establishing a mapping relationship between the second node coordinate set and the second displacement field to obtain the second nonlinear mapping relationship.

[0031] It can be understood that, as Figure 3 shown, Figure 3 a weight function diagram of a solid mechanics calculation method based on double-grid dynamic weight optimization according to an embodiment of the present application, a mapping relationship between a spatial coordinate and a displacement field is established through a deep neural network. Specifically, a first node coordinate set and a second node coordinate set discretized by a coarse-grid discrete model and a fine-grid discrete model are respectively constructed and nonlinear mapping relationship with the corresponding displacement field, that is, the first nonlinear mapping relationship and the second nonlinear mapping relationship of the embodiment of the present application.

[0032] In step S104, a composite loss function is constructed, and based on a preset training strategy, the composite loss function is trained in stages in cooperation according to the first nonlinear mapping relationship and the second nonlinear mapping relationship until a preset training round is reached or the composite loss function after training meets a preset accuracy requirement, and a solid mechanics problem of a plane panel tension process is solved according to the composite loss function after training to obtain a solution of the displacement field of the solid mechanics problem. The preset accuracy requirement can be a threshold value preset by a user, can be a threshold value obtained through a limited number of experiments, or can be a threshold value obtained through a limited number of computer simulations, which is not specifically limited here.

[0033] Specifically, the embodiment of the present application guarantees the accuracy of the calculation result through the stage-by-stage cooperative training mechanism, and significantly improves the solving speed; when the value of the function of the embodiment of the present application falls below the preset threshold value and the error on the verification set meets the preset accuracy requirement, it is considered that the neural network training is completed, and the displacement field predicted by the neural network is output as the solution of the displacement field of the solid mechanics problem. Otherwise, the gradient of the composite loss function with respect to the weight W and the bias b of the deep neural network is calculated through an optimization algorithm (gradient descent algorithm), and the weight W and the bias b of the deep neural network are updated based on the gradient. Finally, the above steps are repeated until the convergence condition is met or the maximum iteration number is reached, and the solution of the displacement field of the solid mechanics problem is obtained.

[0034] Optionally, in some embodiments, constructing a composite loss function includes: determining a first physical information constraint loss function for the coarse-grid discrete model, a second physical information constraint loss function for the fine-grid discrete model, a first weight corresponding to the first physical information constraint loss function, and a second weight corresponding to the second physical information constraint loss function; and constructing a composite loss function based on the first physical information constraint loss function, the first weight, the second physical information constraint loss function, and the second weight.

[0035] Optionally, in some embodiments, the first weight is determined by a first dynamic weighting function, and the second weight is determined by a second dynamic weighting function, wherein... The first dynamic weighting function is: ; The second dynamic weighting function is: ; in, ; ; in, The first dynamic weight function, For the unnormalized first dynamic weight function, For the unnormalized second dynamic weight function, This is the second dynamic weight function. For the current training round, and For training rounds.

[0036] It is understandable that, such as Figure 4 As shown, Figure 4 This is a flowchart of a deep neural network computation method for solid mechanics based on dual-grid dynamic weight optimization, according to one embodiment of this application. Model 1 (i.e., the coarse-grid discrete model of this embodiment) is based on a larger cell size, and Model 2 (i.e., the fine-grid discrete model of this embodiment) is based on a smaller cell size. In the initial training phase, when the number of training epochs is less than... At that time, only Model 1 participated in the training of the deep neural network. When Model 1 was trained to... By the end of the round, the training had reached the predetermined convergence state. Then, in arrive Between rounds ( ), model 2 is gradually introduced into the training of the deep neural network, the parameters of model 1 are used as the network initialization values of model 2, and the process of dynamic optimization learning is adopted through the weight function, the weight function gradually reduces the loss function weight of model 1 and gradually increases the loss function weight of model 2, so as to utilize the prior knowledge provided by model 1 to accelerate the convergence of model 2. Finally, when the training round is greater than After the training round, only model 2 is trained in the deep neural network until the solid mechanics problem converges. The stage division parameter satisfies , wherein, represents the training round of transfer learning. In the physical constraint construction stage, the first physical information constraint loss function of the coarse grid discrete model and the second physical information constraint loss function of the fine grid discrete model respectively contain the sum of the residual terms of the boundary conditions, initial conditions and physical functions in each model. Based on the training displacement of the two models, the initial and boundary value condition constraint terms and the physical residual terms of the control equation are constructed, and the composite loss function is constructed through the weighted fusion strategy. The dynamic weight coefficient of the composite loss function of the embodiment of the application follows an adaptive adjustment mechanism, and the specific formula is as follows: ; , wherein, is the composite loss function, is the first dynamic weight function, is the first physical information constraint loss function of the coarse grid discrete model, is the second dynamic weight function, is the second physical information constraint loss function of the fine grid discrete model.

[0037] It should be noted that the embodiments of the present application can also construct more similar learning methods, such as continuing to add a grid discrete model with smaller grid size, and using the model 2 to accelerate the training process of the grid discrete model with smaller grid size according to the method described above. The division of the model unit can not only be quadrilateral, but also polygon; the embodiments of the present application are not only suitable for solving the plane problem of solid mechanics, but also can be used to solve the three-dimensional solid mechanics problem; the division of the model can adopt the unit form, or the point unit form, that is, the model with less number of points and the model with more number of points are used to realize. Moreover, the smaller the unit size is, the more the number of units in the model is, and the more accurate the model calculation is, and the more computer resources are required for each learning; on the contrary, the larger the unit size is, the less the number of units in the model is, and the lower the model calculation accuracy is, so that the computer resources required for each learning are less. Therefore, the embodiments of the present application define a new weight function suitable for learning of solid mechanics problems, in the early stage of neural network training, a large-size grid is used to make the model converge rapidly, and in the later stage of training, the converged large-size grid model is used to assist the small-size model to rapidly converge to the solution meeting the accuracy requirement; the simple model is used to guide the complex model to accelerate the convergence, thereby effectively improving the training efficiency.

[0038] Optionally, in some embodiments, the staged co-training includes a pre-training stage, a migration stage and a fine-tuning stage, based on a preset training strategy, the composite loss function is co-trained in stages according to the first non-linear mapping relationship and the second non-linear mapping relationship, including: in the pre-training stage, setting the first weight to 1 and the second weight to 0, and initializing the network parameters by the coarse grid discrete model; in the migration stage, controlling the first weight to decrease based on a preset decreasing strategy, and controlling the second weight to increase based on a preset increasing strategy, to perform knowledge migration from the coarse grid discrete model to the fine grid discrete model; in the fine-tuning stage, setting the first weight to 0 and the second weight to 1, and training the fine grid discrete model based on a preset accuracy training requirement.

[0039] It can be understood that the coarse grid discrete model of the embodiments of the present application is pre-trained only by the coarse grid model in the pre-training stage (when not converged), , , in the migration stage (after converged), based on a preset decreasing strategy from 1 to 0, based on a preset increasing strategy from 0 to 1, to realize knowledge migration learning from the coarse grid discrete model to the fine grid discrete model. In the fine-tuning stage, , , the fine grid discrete model is trained based on a preset accuracy training requirement.

[0040] Therefore, this application establishes a two-layer discrete system with a coarse-grid model and a fine-grid model, respectively calculating the initial boundary conditions and the residuals of the governing equations; it designs a dynamic weight function with cosine-sine square modulation and constructs a composite loss function, wherein the weight coefficients are dynamically adjusted according to the training stage. In the pre-training stage, only the coarse-grid discrete model is activated to achieve fast pre-convergence, and in the transfer stage, knowledge transfer learning is achieved through weight function adjustment; in the fine-tuning stage, the fine-grid discrete model is activated to achieve convergence to the solution accuracy; by providing high-quality initial parameter values ​​through pre-training of the coarse-grid discrete model, combined with the adaptive weight adjustment mechanism, the training efficiency of the deep neural network of the fine-grid discrete model can be greatly improved, which solves the contradiction between convergence speed and computational accuracy in solving solid mechanics problems by traditional physical information neural networks, and provides a new paradigm of high efficiency and intelligence for solving complex solid mechanics problems. Combined with the progressive weight adjustment strategy, the training efficiency is improved while ensuring computational accuracy.

[0041] According to the solid mechanics calculation method based on dual-grid dynamic weight optimization proposed in this application, when it is determined that there is a solid mechanics calculation requirement for the tensile process of a planar plate, a solid mechanics problem of the tensile process of the planar plate is established, and a coarse-grid discrete model and a fine-grid discrete model are established. Based on a preset deep neural network and the above two models, a first nonlinear mapping relationship and a second nonlinear mapping relationship are established. A composite loss function is constructed, and based on a preset training strategy, the composite loss function is trained in stages according to the first nonlinear mapping relationship and the second nonlinear mapping relationship until a preset number of training rounds are reached or the composite loss function after training meets the preset accuracy requirements. The solution of the displacement field of the solid mechanics problem is obtained by solving the composite loss function after training. Thus, the problem of slow convergence speed of fine-grid models in related technologies is solved, and the training efficiency is improved.

[0042] Next, referring to the accompanying drawings, a solid mechanics calculation device based on dual-grid dynamic weight optimization proposed according to an embodiment of this application is described.

[0043] Figure 5 This is a block diagram of a solid mechanics calculation device based on dual-grid dynamic weight optimization according to an embodiment of this application.

[0044] like Figure 5 As shown, the solid mechanics calculation device 10 based on dual-grid dynamic weight optimization includes: a judgment module 100, a modeling module 200, a mapping module 300, and a determination module 400.

[0045] Among them, the judgment module 100 is used to determine whether there is a solid mechanics calculation requirement for the tensile process of the planar plate; The modeling module 200 is configured to establish a solid mechanics problem of the plane plate tension process if there is a solid mechanics calculation requirement of the plane plate tension process, and to establish a coarse-grid discrete model and a fine-grid discrete model for the solid mechanics problem; The mapping module 300 is configured to establish a first nonlinear mapping relationship and a second nonlinear mapping relationship according to the coarse-grid discrete model and the fine-grid discrete model based on a preset deep neural network. The determining module 400 is configured to construct a composite loss function, and perform staged collaborative training on the composite loss function according to the first nonlinear mapping relationship and the second nonlinear mapping relationship based on a preset training strategy until a preset training round is reached or the composite loss function after training meets a preset accuracy requirement, and to solve the solid mechanics problem of the plane plate tension process according to the composite loss function after training to obtain a solution of a displacement field of the solid mechanics problem.

[0046] Optionally, the mapping module 300 is specifically configured to: obtain a first node coordinate set generated by the coarse-grid discrete model and a second node coordinate set generated by the fine-grid discrete model; and establish a mapping relationship between the first node coordinate set and the first displacement field to obtain the first nonlinear mapping relationship, and establish a mapping relationship between the second node coordinate set and the second displacement field to obtain the second nonlinear mapping relationship.

[0047] Optionally, the composite loss function is constructed by: determining a first physical information constraint loss function of the coarse-grid discrete model, a second physical information constraint loss function of the fine-grid discrete model, a first weight corresponding to the first physical information constraint loss function, and a second weight corresponding to the second physical information constraint loss function; and constructing the composite loss function according to the first physical information constraint loss function, the first weight, the second physical information constraint loss function, and the second weight.

[0048] Optionally, the staged collaborative training includes a pre-training stage, a migration stage, and a fine-tuning stage, and the determining module 400 is specifically configured to: in the pre-training stage, set the first weight to 1 and the second weight to 0, and initialize network parameters through the coarse-grid discrete model; in the migration stage, control the first weight to decrease based on a preset decreasing strategy, and control the second weight to increase based on a preset increasing strategy, so as to perform knowledge migration from the coarse-grid discrete model to the fine-grid discrete model; and in the fine-tuning stage, set the first weight to 0 and the second weight to 1, and train the fine-grid discrete model based on a preset accuracy training requirement.

[0049] Optionally, the first weight is determined by a first dynamic weight function, and the second weight is determined by a second dynamic weight function, wherein, The first dynamic weight function is: ; The second dynamic weight function is: ; wherein, ; ; wherein, is a first dynamic weight function, is an unnormalized first dynamic weight function, is an unnormalized second dynamic weight function, is a second dynamic weight function, is a current training epoch, and is a training epoch.

[0050] It should be noted that the foregoing explanation and description of the embodiment of the solid mechanics calculation method based on double-grid dynamic weight optimization also applies to the embodiment of the solid mechanics calculation device based on double-grid dynamic weight optimization, which will not be described here.

[0051] According to the solid mechanics calculation device based on double-grid dynamic weight optimization provided in the embodiments of the present application, when it is judged that there is a solid mechanics calculation demand for the tensile process of a plane plate, a solid mechanics problem of the tensile process of the plane plate is established, and a coarse-grid discrete model and a fine-grid discrete model are established; a first nonlinear mapping relationship and a second nonlinear mapping relationship are established based on a preset deep neural network and the two models; a composite loss function is constructed, and based on a preset training strategy, the composite loss function is trained in stages and cooperatively according to the first nonlinear mapping relationship and the second nonlinear mapping relationship until a preset training epoch is reached or the composite loss function after training meets a preset accuracy requirement, and a solution of the displacement field of the solid mechanics problem is obtained according to the composite loss function after training. Thus, the problem of slow convergence of the fine-grid model in the related art is solved, and the training efficiency is improved.

[0052] Figure 6 The structure schematic diagram of the electronic device provided in the embodiments of the present application is shown. The electronic device can include: a memory 601, a processor 602, and a computer program stored in the memory 601 and executable on the processor 602.

[0053] The processor 602 implements the solid mechanics calculation method based on double-grid dynamic weight optimization provided in the above embodiments when executing the program.

[0054] Further, the electronic device further includes: a communication interface 603 for communication between the memory 601 and the processor 602.

[0055] The memory 601 is used to store the computer program executable on the processor 602.

[0056] The memory 601 can include a high-speed RAM memory and can also include a non-volatile memory, such as at least one disk memory.

[0057] If the memory 601, the processor 602 and the communication interface 603 are implemented independently, the communication interface 603, the memory 601 and the processor 602 can be connected to each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, Figure 6 In the figure, only one thick line is used to represent, but it does not mean that there is only one bus or one type of bus.

[0058] Optionally, in a specific implementation, if the memory 601, the processor 602 and the communication interface 603 are integrated on a chip, the memory 601, the processor 602 and the communication interface 603 can complete communication between each other through an internal interface.

[0059] The processor 602 can be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0060] The embodiments of the present application also provide a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the solid mechanics calculation method based on double-grid dynamic weight optimization as above.

[0061] The embodiments of the present application also provide a computer program product, which stores a computer program, and the program is executed by a processor to implement the solid mechanics calculation method based on double-grid dynamic weight optimization as above.

[0062] In the description of the application, reference to "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" etc. means that a particular feature, structure, material or characteristic described in connection with the embodiment or example is included in at least one embodiment or example of the application. The illustrative appearances of the above-mentioned terms in various places in the specification are not necessarily referred to the same embodiment or example. Moreover, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples. Furthermore, in non-contradictory cases, those skilled in the art can combine and combine the features of different embodiments or examples described in the specification and the features of different embodiments or examples.

[0063] In addition, the terms "first", "second" are only for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly indicating the number of the technical features indicated. Therefore, the features defined with "first", "second" can explicitly or implicitly include at least one of the features. In the description of the application, the meaning of "N" is at least two, such as two, three, etc., unless otherwise specifically limited.

[0064] Any process or method descriptions or descriptions of the flow diagrams described herein or otherwise described in this application can be understood as representing the steps of the code modules, segments, or portions for implementing the custom logic using the executable instructions of the preferred embodiments of the application, and the various embodiments of the preferred embodiments of the application can include additional or fewer steps or methods, and the method can be implemented by a hardware logic circuit, software, firmware, or a combination thereof, and the order of the steps or methods can be changed, and some steps or methods can be performed in parallel, and some steps or methods can be omitted, and some steps or methods can be performed simultaneously, and some steps or methods can be performed in reverse order, and some steps or methods can be performed in substantially simultaneous, and the skilled in the art should understand that the scope of the application is not limited to the order of the steps or methods described in the preferred embodiments of the application.

[0065] It should be understood that parts of the application can be implemented in hardware, software, firmware, or a combination thereof. In the above-described embodiments, the N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. As in another embodiment, if implemented in hardware, any of the following technologies known in the art or their combinations can be used: discrete logic circuit with logic gate circuit for implementing logic functions on data signals, application specific integrated circuit with suitable combination logic gate circuit, programmable gate array (PGA), field programmable gate array (FPGA), etc.

[0066] Those skilled in the art can understand that all or part of the steps carried out by the above-mentioned embodiment method can be completed by a program instructing the relevant hardware, and the program can be stored in a computer readable storage medium, and the program includes one or a combination of the steps of the method embodiment when executed.

Claims

1. A method for solid mechanics calculation based on double-grid dynamic weight optimization, characterized in that, The method comprises the following steps: determining whether there is a solid mechanics calculation requirement for a flat plate tension process; if there is a solid mechanics calculation requirement for the flat plate tension process, establishing a solid mechanics problem of the flat plate tension process, and establishing a coarse-grid discrete model and a fine-grid discrete model for the solid mechanics problem; based on a preset deep neural network, establishing a first nonlinear mapping relationship and a second nonlinear mapping relationship according to the coarse-grid discrete model and the fine-grid discrete model; constructing a composite loss function, and based on a preset training strategy, performing staged collaborative training on the composite loss function according to the first nonlinear mapping relationship and the second nonlinear mapping relationship until a preset training round is reached or the composite loss function after training meets a preset accuracy requirement, and solving the solid mechanics problem of the flat plate tension process according to the composite loss function after training to obtain a solution of a displacement field of the solid mechanics problem.

2. The method of claim 1, wherein, The method comprises the following steps: obtaining a first node coordinate set generated by the coarse-grid discrete model and a second node coordinate set generated by the fine-grid discrete model; establishing a mapping relationship between the first node coordinate set and a first displacement field to obtain the first nonlinear mapping relationship, and establishing a mapping relationship between the second node coordinate set and a second displacement field to obtain the second nonlinear mapping relationship.

3. The method of claim 1, wherein, The method comprises the following steps: determining a first physical information constraint loss function of the coarse-grid discrete model, a second physical information constraint loss function of the fine-grid discrete model, a first weight corresponding to the first physical information constraint loss function, and a second weight corresponding to the second physical information constraint loss function; constructing the composite loss function according to the first physical information constraint loss function, the first weight, the second physical information constraint loss function, and the second weight.

4. The method of claim 3, wherein, The staged collaborative training comprises a pre-training stage, a migration stage, and a fine-tuning stage, and the method comprises the following steps: in the pre-training stage, setting the first weight to 1 and the second weight to 0, and initializing network parameters through the coarse-grid discrete model; in the migration stage, controlling the first weight to decrease based on a preset decreasing strategy, and controlling the second weight to increase based on a preset increasing strategy, so as to perform knowledge migration from the coarse-grid discrete model to the fine-grid discrete model; in the fine-tuning stage, setting the first weight to 0 and the second weight to 1, and training the fine-grid discrete model based on a preset accuracy training requirement.

5. The method of claim 3, wherein, The first weight is determined by a first dynamic weight function, and the second weight is determined by a second dynamic weight function, wherein the first dynamic weight function is: ； the second dynamic weight function is: ； wherein ； ； wherein, is a first dynamic weight function, is an unnormalized first dynamic weight function, is an unnormalized second dynamic weight function, is a second dynamic weight function, is a current training epoch, and is a training epoch.

6. A solid mechanics computing device based on double grid dynamic weight optimization, characterized in that, The method comprises the following steps: a determination module for determining whether there is a solid mechanics calculation requirement for a flat plate tension process; The modeling module is configured to, if there is a solid mechanics calculation requirement for the plane plate tension process, establish a solid mechanics problem of the plane plate tension process, and establish a coarse-grid discrete model and a fine-grid discrete model for the solid mechanics problem; The mapping module is configured to, based on a preset deep neural network, establish a first nonlinear mapping relationship and a second nonlinear mapping relationship according to the coarse-grid discrete model and the fine-grid discrete model; The determining module is configured to construct a composite loss function, and based on a preset training strategy, perform staged collaborative training on the composite loss function according to the first nonlinear mapping relationship and the second nonlinear mapping relationship until a preset training round is reached or the composite loss function after training meets a preset accuracy requirement, solve the solid mechanics problem of the plane plate tension process according to the composite loss function after training, and obtain a solution of a displacement field of the solid mechanics problem.

7. The apparatus of claim 6, wherein, The mapping module is specifically configured to: obtain a first node coordinate set generated by the coarse-grid discrete model and a second node coordinate set generated by the fine-grid discrete model; establish a mapping relationship between the first node coordinate set and a first displacement field to obtain the first nonlinear mapping relationship, and establish a mapping relationship between the second node coordinate set and a second displacement field to obtain the second nonlinear mapping relationship.

8. An electronic device, comprising: The computer program is executed by the processor to implement the solid mechanics calculation method based on the double-grid dynamic weight optimization according to any one of claims 1-5. The computer program is executed by the processor to implement the solid mechanics calculation method based on the double-grid dynamic weight optimization according to any one of claims 1-5.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the solid mechanics calculation method based on the double-grid dynamic weight optimization according to any one of claims 1-5.

10. A computer program product comprising a computer program, characterized in that, ​

Images