A method for constructing reduced-order manifold neural operators for non-equivalent domain mapping of spatiotemporal processes
By building a combination of non-peer domain encoding and decoding modules and homodomain mapping modules, the problem of poor prediction effect of non-peer domain mapping in the spatiotemporal process in the prior art is solved, and more efficient prediction and training efficiency is achieved.
Patent Information
- Application Number
- CN202411198809.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-29
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2044-08-29
AI Technical Summary
When handling non-peer domain mapping of spatiotemporal processes, the prior art requires additional domain expansion or reduction of the model, which is difficult to ensure prediction effect and has low computational efficiency.
By obtaining the non-public domain temporal basis function or spatial basis function of the input function and the output function, the non-peer domain encoding module and the decoding module are constructed, and combined with the same domain mapping module, the down-order manifold neural computing for space-time processes is constructed. , directly use the manifold neural operator model for prediction.
It realizes that without additional domain changes, the prediction effect and training efficiency of non-peer domain mapping are improved, with higher stability and faster training speed.
Smart Images

Figure CN118709724B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of machine learning technology, and in particular to a method for constructing a reduced-order manifold neural operator for non-equivalent domain mapping of spatiotemporal processes. Background Art
[0002] A space-time process refers to a process that evolves in time and space. Many processes in the fields of natural science and technical science belong to space-time processes, such as processes controlled by time-varying partial differential equations, climate change processes, aortic blood flow processes, and composite material curing processes. Space-time processes involve three types of functions: time function , spatial function and space-time function . Predictive learning of spatiotemporal processes refers to constructing operator mappings between these three types of functions. There is an important type of non-equivalent domain mapping in predictive learning of spatiotemporal processes: one of the input function and the output function is a spatiotemporal function, and the other is a time or space function. The time domain or space domain shared by the two is the same, which is called the common domain; the other additional domain is called the non-common domain. For example, for non-equivalent domain mapping , whose common domain is the spatial domain and the non-common domain is the temporal domain. Traditional numerical simulation methods have low computational efficiency in non-equivalent domain mapping for complex scenarios or repeated calculation scenarios, while neural operators can effectively learn the mapping between function spaces using deep learning methods, and can achieve fast predictions once training is completed.
[0003] For existing patents, please refer to the Chinese patent with announcement number CN116187386A, which discloses a method for constructing (manifold) neural operators for complex geometric shapes, which can learn the mapping from function space to function space defined on complex geometric shapes. Its method works well on the same domain mapping problem, but the above technology still has defects: it requires that the domains of the input function and the output function must be the same, and when processing non-equivalent domain mappings, the model needs to be additionally expanded or reduced in domain. These operations can only guarantee the feasibility of the model in form, but it is difficult to guarantee the prediction effect. In view of the above defects, it is necessary for the present application to provide a reduced-order manifold neural operator construction method for non-equivalent domain mapping of spatiotemporal processes. Summary of the invention
[0004] The purpose of the present invention is to provide a method for constructing a reduced-order manifold neural operator for non-equivalent domain mapping of spatiotemporal processes, so as to solve the problems existing in the above-mentioned background technology.
[0005] To achieve the above object, the present invention provides a method for constructing a reduced-order manifold neural operator for non-equivalent domain mapping of spatiotemporal processes, comprising the following steps:
[0006] S1. Obtain a set of time basis functions for the non-common domains of the input function and the output function or spatial basis functions ;
[0007] S2, time basis function obtained using S1 or spatial basis functions Constructing non-equivalent domain encoding modules and non-equivalent domain decoding modules ;
[0008] S3. For the common domain of input function and output function, a same-domain mapping module is constructed using the manifold neural operator model ;
[0009] S4. According to the type of non-equivalent domain mapping, the non-equivalent domain encoding module or non-equivalent domain decoding module With the same domain mapping module Combination, thereby constructing a reduced-order neural operator for non-equivalent domain mapping of spatiotemporal processes.
[0010] Preferably, in step S1, a set of time basis functions is obtained for the non-common domains of the input function and the output function. or spatial basis functions The specific methods include: using the modal decomposition method based on the training data, such as proper orthogonal decomposition and dynamic modal decomposition; directly using basis functions that are independent of the training data, such as the Laplace Beltrami operator basis function in the space domain and the Fourier basis function in the time domain.
[0011] Preferably, the non-equivalent domain encoding module in step S2 To use the time basis function or spatial basis functions The space-time function Reduced to spatial weight function Or time weight function .
[0012] Preferably, the non-equivalent domain decoding module in step S2 To use the time basis function or spatial basis functions The corresponding spatial weight function Or time weight function Reconstruction into space-time function .
[0013] Preferably, step S3 is used to construct a same domain mapping module The manifold neural operator model includes one or more kernel integration modules, wherein the selection of the kernel integration module is divided into: if the common domain of the input function and the output function is the time domain, the kernel integration module in the manifold neural operator model is constructed using the Fourier basis function; if the common domain of the input function and the output function is the space domain, the kernel integration module in the manifold neural operator model is constructed using the characteristic function of the Laplace Beltrami operator.
[0014] Preferably, in step S4, the non-equivalent domain encoding module is mapped according to the type of non-equivalent domain mapping. or non-equivalent domain decoding module With the same domain mapping module Combination, thus constructing a reduced-order neural operator for non-equivalent domain mapping of spatiotemporal processes, including two cases:
[0015] In the first case, if the input function is a spatiotemporal function, the architecture of the reduced-order neural operator is: Input function-non-equivalent domain encoding module —Same domain mapping module — output function;
[0016] In the second case, if the output function is a spatiotemporal function, the architecture of the reduced-order neural operator is: Input function-same domain mapping module —Non-equivalent domain decoding module —Output function.
[0017] Therefore, the present invention adopts the above-mentioned reduced-order manifold neural operator construction method for non-equivalent domain mapping of spatiotemporal processes, which has the following beneficial effects compared with the prior art:
[0018] The constructed non-equivalent domain encoding module or non-equivalent domain decoding module is used to reduce the order of non-equivalent domain mapping to equal-domain mapping, so that the manifold neural operator model can be directly applied without the need for additional domain expansion or domain reduction modifications to the model, and the prediction effect is more stable; and because of the introduction of the order reduction idea, the training efficiency is faster.
[0019] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0020] Figure 1 is a flow chart of the present invention;
[0021] Figure 2 A schematic diagram of a reduced-order neural operator including a same-domain mapping module and a non-equivalent-domain decoding module in an embodiment of the present invention;
[0022] Figure 3 1 is a comparison between the actual value at four typical moments of a set of test data and the predicted value of this method in an embodiment of the present invention. DETAILED DESCRIPTION
[0023] Example
[0024] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0025] This embodiment uses Figure 2 Take the solution of the two-dimensional Burgers equation as an example, the input is the initial flow velocity field , is a spatial function, and the output is Velocity field inside , is a space-time function, which is a typical non-equivalent domain mapping. Among them, the discrete quantity in the time domain is , the discrete quantity in the spatial domain is In this embodiment, 3000 groups of Function pairs are used as training data, and 500 function pairs are used as test data. Figure 1 As shown, a method for constructing a reduced-order manifold neural operator for non-equivalent domain mapping of spatiotemporal processes includes the following steps:
[0026] S1. Obtain a set of time basis functions for the non-common domains of the input function and the output function or spatial basis functions ;
[0027] In this embodiment, the non-common domain of the input function and the output function is the time domain, so a set of time basis functions needs to be obtained. Based on 3000 sets of training data, the first 32 modes in the time domain are obtained using intrinsic orthogonal decomposition calculation as a set of time basis functions. .
[0028] S2: Use the time basis function or space basis function obtained in S1 to construct a non-equivalent domain coding module and non-equivalent domain decoding modules ;
[0029] In this embodiment, a set of time basis functions is used The spatial function Transformed into space-time function The non-equivalent domain decoding module is denoted as .
[0030] S3. For the common domain of the input function and the output function, a same-domain mapping module is constructed using a manifold neural operator model containing one or more kernel integral modules. ;
[0031] In this embodiment, a manifold neural operator model including four kernel integration modules is used to construct a same domain mapping module. , and since the common domain of the input function and the output function is the spatial domain, the characteristic function of the Laplace Beltrami operator is used to construct the kernel integration module. Each kernel integration module includes a frequency domain transformation submodule, a linear transformation submodule and a nonlinear activation submodule, wherein the frequency domain transformation submodule includes: (1) using the characteristic functions of the first 128 Laplace Beltrami operators to reduce the dimension of the submodule input function to obtain a coefficient vector with a dimension of 128; (2) performing a linear transformation on the 128-dimensional coefficient vector; (3) using the characteristic functions of the first 128 Laplace Beltrami operators to increase the dimension of the coefficient vector after the linear transformation to obtain the submodule output function. The nonlinear activation submodule adopts the gelu activation function. In this embodiment, in order to enhance the learning ability of the manifold neural operator model, a layer of fully connected neural network with 64 neurons is set before the four kernel integration modules, and two layers of fully connected neural networks are set after the four kernel integration modules, with the numbers of neurons being 32 and 32 respectively, and the fully connected layer adopts the gelu activation function. Same domain mapping module Final output spatial weight function .
[0032] S4. According to the type of non-equivalent domain mapping, the non-equivalent domain encoding module or non-equivalent domain decoding module With the same domain mapping module Combining them to construct a reduced-order neural operator for mapping non-equivalent domains of spatiotemporal processes;
[0033] In this embodiment, since the output function is a spatiotemporal function, the non-equivalent domain decoding module constructed in step S2 and the same domain mapping module constructed in step S3 are used to construct a reduced-order neural operator. The specific architecture from the input function to the output function is as follows: Input function —Same domain mapping module —Spatial weight function —Non-equivalent domain decoding module — Output function .
[0034] In this embodiment, 3000 groups of The function pairs are used as training data, and 500 sets of function pairs are used as test data. The classic neural network training method is used, the batch size is 50, the learning rate is 0.01 (every 100 iterations, the learning rate decays by 50%), the optimizer is Adam, and the loss function is defined as follows:
[0035] ;
[0036] in, is the number of training data, represents the true value of the velocity field, represents the predicted value of the velocity field. After 500 iterations of training, the test error on 500 sets of test data is 4.356%. Figure 3 As shown, a comparison of the true value of a set of test data at four typical moments and the predicted value of this method is given. It can be seen that the method according to the present invention can construct a reduced-order neural operator for non-equidomain mapping of spatiotemporal processes, with higher prediction accuracy.
[0037] Therefore, the present invention adopts the above-mentioned reduced-order manifold neural operator construction method for non-equivalent domain mapping of spatiotemporal processes, which includes a reduced-order neural operator architecture for non-equivalent domain mapping of spatiotemporal processes, including a non-equivalent domain encoding module, a non-equivalent domain decoding module and a same-domain mapping module. Compared with the existing technology, it has higher prediction accuracy and can also take into account training efficiency.
[0038] 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 reduced-order manifold neural operator for non-equivalent domain mapping of spatiotemporal processes, characterized in that: The following steps are involved: S1. Obtain a set of time basis functions for the non-common domains of the input function and the output function or spatial basis functions ; S2, time basis function obtained using S1 or spatial basis functions Constructing non-equivalent domain encoding modules and non-equivalent domain decoding modules ; S3. For the common domain of input function and output function, a same-domain mapping module is constructed using the manifold neural operator model ; S4. According to the type of non-equivalent domain mapping, the non-equivalent domain encoding module or non-equivalent domain decoding module With the same domain mapping module Combining them to construct a reduced-order neural operator for mapping non-equivalent domains of spatiotemporal processes; The non-equivalent domain encoding module in step S2 To use the time basis function or spatial basis functions The space-time function Reduced to spatial weight function Or time weight function ; The non-equivalent domain decoding module in step S2 To use the time basis function or spatial basis functions The corresponding spatial weight function Or time weight function Reconstruction into space-time function ; Step S3 is used to construct the same domain mapping module The manifold neural operator model includes one or more kernel integration modules, wherein the selection of the kernel integration module is divided into: if the common domain of the input function and the output function is the time domain, the kernel integration module in the manifold neural operator model is constructed using the Fourier basis function; if the common domain of the input function and the output function is the space domain, the kernel integration module in the manifold neural operator model is constructed using the characteristic function of the Laplace Beltrami operator; In step S4, the non-equivalent domain encoding module is mapped according to the type of non-equivalent domain mapping. or non-equivalent domain decoding module With the same domain mapping module Combination, thus constructing a reduced-order neural operator for non-equivalent domain mapping of spatiotemporal processes, including two cases: In the first case, if the input function is a spatiotemporal function, the architecture of the reduced-order neural operator is: Input function-non-equivalent domain encoding module —Same domain mapping module — output function; In the second case, if the output function is a spatiotemporal function, the architecture of the reduced-order neural operator is: Input function-same domain mapping module —Non-equivalent domain decoding module —Output function.
2. According to claim 1, a method for constructing a reduced-order manifold neural operator for non-equivalent domain mapping of spatiotemporal processes is characterized in that: In step S1, a set of time basis functions is obtained for the non-common domains of the input function and the output function. or spatial basis functions The specific methods include: using the modal decomposition method based on the training data, including proper orthogonal decomposition and dynamic modal decomposition; directly using basis functions that are independent of the training data, including the Laplace Beltrami operator eigenfunction in the spatial domain and the Fourier basis function and wavelet basis function in the time domain.
