A method for constructing spatiotemporal manifold neural operators for complex geometries
By constructing the combination of Laplace operator feature function and Fourier nested kernel integral module, the problem that is difficult to deal with in the spatial-temporal function mapping of complex geometric shapes is solved, and high efficiency independent of discrete resolution is achieved Spatial-temporal function prediction, suitable for spatial-temporal dynamic modeling of complex geometric shapes.
Patent Information
- Application Number
- CN202411198804.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2044-08-29
AI Technical Summary
When existing neural operators deal with the spatial and temporal function mapping of complex geometric shapes, it is difficult to accurately process the spatial and temporal dimensions at the same time. The complexity of the model is related to the discrete resolution of the spatial and temporal dimensions, and it is not effectively applicable to the spatial and temporal and temporal geometric shapes. Dynamic modeling.
By constructing the Laplace operator feature function as a basis function, a coding module for spatial and temporal dimensions is constructed, combined with the Laplace-Fourier nested kernel integral module, a space-time manifold neural computing for complex geometric shapes is formed Sub, independent of the discrete resolution of the input domain, can learn mappings between spatiotemporal functions on complex geometries.
A model structure independent of the space-time dimension discrete resolution in complex geometric shapes is realized, which improves the accuracy and applicability of predictions, and can handle uneven discrete grid node data, which is lower than the previous operator network. model complexity and higher prediction accuracy.
Smart Images

Figure CN118690794B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of machine learning, and in particular to a method for constructing a spatiotemporal manifold neural operator for complex geometric shapes. Background Art
[0002] The mapping model between function spaces constructed using deep neural networks is called a neural operator (NO) or operator network. Neural operator networks can construct prediction models that are independent of the discretization resolution of input features, can be applied to high-dimensional mapping problems, and have potential advantages in spatiotemporal dynamics modeling. It has the ability to transfer problems between grids, and only needs to train neural operators once for different grids of the same problem. For the parameters of the new instance, only the forward pass of the network is required to obtain the solution, thereby alleviating the main computational problems arising from the neural finite element method.
[0003] Among the existing neural operators, the DeepONet-like neural operator framework uses the universal approximation theorem to learn function mapping through the encoded information from the trunk and branch networks. DeepONet can be regarded as a finite-dimensional operator model that parameterizes the input data point by point, which means that the number of network parameters still depends on the discretization resolution of the input domain. When faced with complex spatiotemporal functions, its branch networks have to process complex spatial coordinates, and the trunk network has to extract the spatiotemporal attributes of each spatial node at all times, making it difficult to achieve accurate learning.
[0004] The Fourier neural operator (FNO) transforms complex function mapping into a low-dimensional mapping in the frequency domain, greatly reducing the model complexity. Subsequently, variant operator networks such as the wavelet neural operator and the U-type neural operator were proposed and widely used in spatiotemporal dynamics modeling. Although these operator networks have been successfully applied in spatiotemporal field prediction problems, the Fourier transform or wavelet transform requires uniform grid data defined on a rectangular domain, which makes them unsuitable for mapping spatiotemporal functions on complex geometries.
[0005] The recently proposed Neural Operator on Riemannian Manifold (NORM) uses the eigenfunction of the Laplacian operator to encode complex geometric information, thereby learning the mapping between function spaces defined on arbitrarily complex geometric domains. Although NORM can effectively handle the spatial dimension of the function, it is still high-dimensionally discretized in the time dimension. Summary of the invention
[0006] The purpose of the present invention is to provide a method for constructing a spatiotemporal manifold neural operator for complex geometric shapes, by constructing a parameterized model to represent the mapping between two spatiotemporal functions defined on complex geometric bodies, and the structural complexity of the model is independent of the discretization resolution of the spatiotemporal dimension.
[0007] To achieve the above purpose, the present invention provides a method for constructing a spatiotemporal manifold neural operator for complex geometric shapes, the steps of which are as follows:
[0008] A method for constructing a spatiotemporal manifold neural operator for complex geometric shapes, characterized by the following steps:
[0009] S1. The geometric shapes of the input space-time function and the output space-time function , solve a set of Laplace operator characteristic functions as basis functions, and then construct a spatial dimension encoding module and decoding module ;
[0010] S2. Solve a set of time basis functions and then construct a time dimension encoding module and decoding module ;
[0011] S3, spatial dimension encoding module and decoding module With the time dimension encoding module and decoding module Nested combination, constructing Laplace-Fourier nested kernel integration modules, multiple kernel integration modules are connected in series to construct space-time manifold neural operators for complex geometric shapes.
[0012] Preferably, in S1, the process of solving the basis function includes:
[0013] The corresponding Laplacian operator is defined for complex geometric shapes as follows:
[0014] (1)
[0015] Then the characteristic equation of the Laplace operator is constructed as follows:
[0016] (2)
[0017] The Laplace operator characteristic function that reflects the frequency domain information of the geometric shape is solved as follows:
[0018] (3)
[0019] Preferably, in S1, a spatial dimension encoding module is constructed The specific process includes: using the Laplace operator characteristic function to project the geometric space where the space-time function is located to the frequency domain space spanned by the Laplace operator characteristic function, and then Defined as the geometric shape of the space-time function in the Laplace operator characteristic function Spectral decomposition of .
[0020] Preferably, in S1, a spatial dimension decoding module is constructed The specific process includes: using the Laplace operator characteristic function to decode the encoded spatial frequency domain function from the frequency domain space spanned by the Laplace operator characteristic function to the geometric space where the output function is located, and then Defined as the Laplace operator characteristic function exist Spectral reconstruction on .
[0021] Preferably, in S2, a time dimension encoding module is constructed The specific process includes: using Fourier basis functions to perform fast Fourier transform, wavelet transform or Laplace transform on the time domain of the space-time function.
[0022] Preferably, in S2, a time dimension decoding module is constructed The specific process includes: performing inverse fast Fourier transform, inverse wavelet transform or inverse Laplace transform on the encoded time-frequency domain function.
[0023] Preferably, in S3, the Laplace-Fourier nested kernel integration module includes an encoding module for sequentially connecting input space-time function and spatial dimension mapping , encoding module for time dimension mapping , parameterization module, decoding module for time dimension mapping , decoding module for spatial dimension mapping , output space-time function.
[0024] Preferably, in S3, the Laplace-Fourier nested kernel integration module includes an encoding module for sequentially connecting an input space-time function and a time dimension mapping. , encoding module for spatial dimension mapping , parameterization module, decoding module of spatial dimension mapping , decoding module for time dimension mapping , output space-time function.
[0025] Therefore, the present invention adopts a method for constructing a space-time manifold neural operator for complex geometric shapes using the above steps, which can learn a parameterized model through paired training data to represent the mapping between two space-time functions defined on complex geometric bodies, and the structural complexity of the model is independent of the discretization resolution of the space-time dimension. Compared with DeepONet-type neural operators, the network architecture and parameter complexity of the present invention are independent of the discrete resolution of the input domain, and more accurate predictions can be achieved; compared with Fourier neural operators (FNO), the present invention can extract and process non-uniformly discrete grid node data, which is suitable for mapping space-time functions on complex geometries; compared with neural operators on Riemann manifolds (NORM), the present invention can simultaneously encode spatial dimensions and time dimensions to achieve learning of mappings between space-time functions.
[0026] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0027] Figure 1 A schematic diagram of a flow chart of an embodiment of a method for constructing a spatiotemporal manifold neural operator for complex geometric shapes according to the present invention;
[0028] Figure 2 It is a parameter process diagram of a composite part according to an embodiment of the present invention;
[0029] Figure 3 Schematic diagram of the structure of a spatiotemporal manifold neural operator according to an embodiment of the present invention;
[0030] Figure 4 is a temperature process curve at a single node in an embodiment of the present invention;
[0031] Figure 5 A comparison diagram of the predicted solidification state field distribution and the actual solidification state field distribution according to an embodiment of the present invention;
[0032] Figure 6 A statistical diagram of prediction errors according to an embodiment of the present invention;
[0033] Figure 7 It is the space-time function of the two-dimensional example and the three-dimensional example of the embodiment of the present invention. DETAILED DESCRIPTION
[0034] The technical solution of the present invention is further described below through the accompanying drawings and embodiments.
[0035] Example 1
[0036] Neural operator networks can construct prediction models that are independent of the resolution of the discretization of spatiotemporal features, and therefore can learn the mapping of high-dimensional spatiotemporal process data to high-dimensional spatiotemporal curing state fields. In the aerospace field, composite parts are large in size, complex in shape, and require high precision. It is crucial to control the deformation of composite materials during the curing process. The use of multi-zone self-resistance electric heating is an effective method to control the curing deformation of composite materials by applying different temperature fields in different areas to compensate for the curing stress. Figure 2 Taking the composite material shown as an example, the component is a complex closed rotating structure with multiple curved surfaces, and it deforms greatly after high-temperature curing.
[0037] This embodiment aims to quickly and accurately predict the spatiotemporal stress field during the curing process of a component to support a large number of iterative optimizations. The method of predicting the spatiotemporal curing state field based on finite element simulation modeling has low computational efficiency and cannot meet the needs of a large number of iterative optimizations. The data-driven prediction method is difficult to predict the high-dimensional spatiotemporal curing state field. Therefore, this embodiment establishes Figure 3 The spatiotemporal process data to spatiotemporal curing state field prediction model shown in the figure has the spatiotemporal process data of a composite material part with a complex shape as input and the spatiotemporal curing state field calculated by the spatiotemporal neural operator model as output. The establishment process is as follows:
[0038] S1. For input space-time function and output space-time function The geometric space ,in represents the position vector, Represents time and solves a set of Laplace operator eigenfunctions as basis functions.
[0039] like Figure 2 As shown in (a), 2149 mesh nodes are divided on the composite part, and the corresponding Laplace operator is defined according to the divided mesh type (regular tetrahedral mesh):
[0040] (1)
[0041] Then the characteristic equation of the Laplace operator is constructed as follows:
[0042] (2)
[0043] The Laplace operator characteristic function that reflects the frequency domain information of the geometric shape is solved as follows:
[0044] (3)
[0045] like Figure 2 As shown in (b), the time-space process parameters are input, specifically, the temperature process curves of the composite parts in different areas are randomly set. Figure 4 Represents the temperature process curve at a single node. The insulation temperature of all areas is randomly generated between 413 and 433 K, and other parameters remain unchanged. The solidification state on the divided grid nodes is sampled over the time process to obtain a set of frequency domain orthogonal basis functions in the Fourier domain:
[0046]
[0047] Sampling to obtain 1000 sets of spatiotemporal curing process data within the curing cycle , and each The spatiotemporal curing state field data reflecting the evolution of the curing state during the curing process of the part , forming N sets of spatiotemporal data pairs .
[0048] The Laplace operator characteristic function matrix is used to project the spatial domain features of the space-time function into the frequency domain space spanned by the Laplace operator characteristic function, and then the encoding module The spatial domain characteristics defined as space-time functions are in the Laplace operator characteristic function The spectral decomposition of , we get:
[0049] (4)
[0050] In the formula is the space where the geometric features of the complex shape parts are input. is the space where the time dimension feature is located, is the frequency domain space spanned by the characteristic function of the Laplace operator, It can be regarded as a vector of time attributes on a spatial node. is the spatial domain characteristic of the space-time function, is the function range space.
[0051] The Laplace operator characteristic function is used to decode the encoded spatial frequency domain function from the frequency domain space spanned by the Laplace operator characteristic function to a complex geometric space, and then the decoding module Defined as the Laplace operator characteristic function Spectral reconstruction on complex geometry yields:
[0052] (5)
[0053] In the formula, is the encoded spatial frequency domain function.
[0054] S2, using the Fourier basis function to project the time domain features of the space-time function into the Fourier domain spanned by the Fourier basis function, and then Defined as the spectral decomposition of the time domain characteristics of the space-time function on the Fourier basis function, we get:
[0055] (6)
[0056] in is the frequency domain space spanned by Fourier basis functions, input is the time domain characteristic of the space-time function, is the function range space, is the Fourier basis function.
[0057] The encoded time-frequency domain function is decoded from the frequency domain space spanned by the Fourier basis function to the time domain, and then the decoding module Defined as Fourier basis function in the time domain Reconstructing the spectrum on , we get:
[0058] (7)
[0059] in is the encoded time-frequency domain function.
[0060] Using spatiotemporal data samples The model parameters are trained to obtain the spatiotemporal solidification state field prediction model of composite parts. The predicted solidification state field distribution is compared with the actual solidification state field distribution. Figure 5 The resulting spatiotemporal manifold neural operator structure including the Laplace kernel integration module and the Fourier kernel integration module is shown in Figure 3 shown.
[0061] In this embodiment, 1000 sets of spatiotemporal data obtained by sampling are used as training sets, and 200 sets of spatiotemporal data are used as test sets. The model parameters of the neural operator are learned by minimizing the loss function through the gradient descent method. The loss function is defined as follows:
[0062] (8)
[0063] In this embodiment, the learning rate is set to 0.01, decayed by 0.5 every 200 iterations, the number of iterations is 1000, the batch size is 50, and the Adam iterative optimizer is used. Figure 6 (a) is a statistical diagram of the errors of all grid nodes at different times. It can be seen from the figure that the average relative error at each time is less than 0.5%.t. Figure 6 (b) is the error distribution diagram of all nodes in all samples in the test set. The prediction error distribution of this method presents a Gaussian distribution with a mean of 0, and the prediction standard deviation is 0.0242. It can be seen that the method of this embodiment can more accurately realize the prediction of the spatiotemporal solidification state field of composite parts.
[0064] The space-time manifold neural operator proposed in this application can be applied to the space-time function evolution of many complex geometric shapes, such as the flow velocity field of the fluid in the curved pipe, the deformation field during the processing of aircraft structural parts, etc. Figure 7 (a) in the figure represents the space-time function of a two-dimensional example, which can be expressed as ; Figure 7 (b) in the figure represents the space-time function of the three-dimensional example, which can be expressed as For most engineering problems, the space-time function It is a scalar function with specific physical meaning.
[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.
Claims
1. A method for constructing a spatiotemporal manifold neural operator for complex geometric shapes, characterized by: Here are the steps: S1. The geometric shapes of the input space-time function and the output space-time function Solve a set of Laplace operator characteristic functions as basis functions, and then construct the spatial dimension encoding module ε and decoding module S2. Solve a set of time basis functions and then construct a time dimension encoding module and decoding module S3, the spatial dimension encoding module ε and decoding module With the time dimension encoding module and decoding module Nested combination, constructing a Laplace-Fourier nested kernel integration module, multiple kernel integration modules are serially connected to construct a spatiotemporal manifold neural operator for complex geometric shapes, wherein the complex geometric shapes are specifically composite material parts with complex shapes; In S1, the process of solving the basis function includes: The corresponding Laplacian operator is defined for complex geometric shapes as follows: Then the characteristic equation of the Laplace operator is constructed as follows: The Laplace operator characteristic function that reflects the frequency domain information of the geometric shape is solved as follows: In S1, the specific process of constructing the spatial dimension coding module ε includes: using the Laplace operator characteristic function to project the geometric space where the space-time function is located to the frequency domain space spanned by the Laplace operator characteristic function, and then defining the coding module ε as the spectral decomposition of the geometric shape where the space-time function is located on the Laplace operator characteristic function φi(x); In S1, build a spatial dimension decoding module The specific process includes: using the Laplace operator characteristic function to decode the encoded spatial frequency domain function from the frequency domain space spanned by the Laplace operator characteristic function to the geometric space where the output function is located, and then Defined as the Laplace operator characteristic function φi(x) in Spectral reconstruction on ; In S2, construct the time dimension encoding module The specific process includes: using Fourier basis function to perform fast Fourier transform, wavelet transform or Laplace transform on the time domain of space-time function; In S2, construct the time dimension decoding module The specific process includes: performing inverse fast Fourier transform, inverse wavelet transform or inverse Laplace transform on the encoded time-frequency domain function; In S3, the Laplace-Fourier nested kernel integration module includes the input space-time function, the encoding module ε of the spatial dimension mapping, and the encoding module of the time dimension mapping connected in sequence. Parameterization module, decoding module for time dimension mapping Decoding module for spatial dimension mapping Output space-time function.
2. The method for constructing a spatiotemporal manifold neural operator for complex geometric shapes according to claim 1, characterized in that: In S3, the Laplace-Fourier nested kernel integration module includes the input space-time function and the encoding module of the time dimension mapping connected in sequence. Encoding module ε of spatial dimension mapping, parameterization module, decoding module of spatial dimension mapping Decoding module for time dimension mapping Output space-time function.
Citation Information
Patent Citations
Magnetic resonance image reconstruction method based on regularized depth image prior method
CN110490832A
Composite material curing process modeling method
CN114065578A
Graph data compression method based on two-dimensional graph fractional order Fourier transform
CN114972552A
Traffic flow prediction method based on multi-dimensional time and space dependence mining
CN115953900A
Neural operator construction method for complex geometrical shape
CN116187386A
