Neural operator-based special small model rapid prediction method

Abstract

The invention relates to the field of active flow control, and particularly discloses a neural operator-based special small model rapid prediction method, which comprises the following steps of S1, performing data enhancement based on a generative adversarial network method; s2, constructing a special neural operator DeepMONet proxy model based on the multi-physical field characteristics of the valve; s3, based on the enhanced data generated in the step S1, substituting the enhanced data into the step S2 for training, and constructing an agent model based on a neural operator; and S4, learning framework reinforcement is carried out based on the agent model in the S3, and structural performance optimization and rapid prediction are carried out on the valve. On the basis of the generative adversarial network method, low-resolution input data is effectively enhanced through the network, the data volume requirement is reduced, the special DeepMONet model is constructed according to the multi-physical field characteristics of the valve, time and resources needed by traditional CFD simulation are saved, and rapid prediction is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of active flow control, and in particular to a special small model fast prediction method based on a neural operator. Background Art

[0002] Valves currently occupy an irreplaceable position on large nuclear-powered ships and submarines. Their dynamic performance directly impacts the combat capability and survivability of ships, making their research a hot topic and a challenge within the field of ship propulsion systems. Complex thermal-fluid-solid multi-physics coupling processes exist within this system. This coupling process simultaneously involves structural analysis of ship steam systems, involving fluid flow, heat transfer, phase change, transient heat conduction, structural response, and noise. It is a complex multidisciplinary problem, and the coordinated optimization design of these processes is crucial for improving ship performance. A key issue is how to efficiently and accurately calculate and simulate the coupled effects of multiple physical fields, such as fluid flow, pressure distribution, and temperature changes.

[0003] As a key component in fluid delivery systems, valve performance directly impacts the overall system's efficiency and stability. To accurately evaluate valve flow and heat transfer characteristics under various operating conditions, three-dimensional numerical simulation techniques, particularly computational fluid dynamics (CFD), are widely used. CFD technology provides highly accurate flow and temperature field simulations, and is therefore widely used in valve performance evaluation, design optimization, and heat transfer analysis.

[0004] However, while CFD methods offer high-precision simulations, they also come with significant computational costs. Especially when dealing with complex valve systems or varying valve openings, the computational time and hardware resources required increase exponentially, severely limiting their effectiveness in engineering practice. Traditional CFD methods struggle to meet the demands for rapid feedback and real-time decision-making, especially in scenarios requiring multiple iterations of design optimization.

[0005] In summary, while 3D CFD simulation methods offer excellent accuracy, their high computational complexity and resource consumption limit their widespread application in practical engineering. Reducing computational complexity and improving simulation efficiency while ensuring accuracy has become a critical issue in the current research of valves and related fluid control systems. Summary of the Invention

[0006] In response to the above-mentioned problems in the prior art, the present invention proposes a special small model fast prediction method based on neural operators, which can effectively reduce the computational complexity and improve the simulation efficiency while ensuring the computational accuracy.

[0007] To achieve the above objectives, the present invention proposes a fast prediction method of a dedicated small model based on a neural operator, comprising the following steps:

[0008] S1. Data enhancement based on generative adversarial network method;

[0009] S2. Based on the multi-physics characteristics of valves, a dedicated neural operator DeepMONet agent model is constructed;

[0010] S3, based on the enhanced data generated by S1, is brought into S2 for training to build a proxy model based on neural operators;

[0011] S4, based on the agent model of S3, strengthens the learning framework and optimizes and quickly predicts the structural performance of the valve.

[0012] Preferably, in S1, the data enhancement method is implemented using a generative adversarial network (PI-GAN), and the implementation process includes the following steps:

[0013] S11. Construct a generative adversarial network (PI-GAN) that generates high-resolution outputs by training low-resolution inputs, thereby building a PI-GAN model for multi-physics super-resolution reconstruction. The low-resolution input is a sparse grid, and the high-resolution output is a fine grid.

[0014] S12. According to the PI-GAN constructed in step S11, the low-resolution input data is enhanced through the network to generate high-resolution data and a large-scale fluid-thermal-solid coupling database, which are used as input data for training the neural operator agent model;

[0015] S13. Preprocess the training set consisting of input data for training the neural operator agent model, and determine the incoming flow conditions and valve information with different openings corresponding to each training data.

[0016] Preferably, in S2, a dedicated neural operator DeepMONet proxy model is constructed based on the multi-physical field characteristics of the valve, and the specific process includes:

[0017] S21. Construct a dedicated neural operator agent model based on the fluid mechanics characteristics, consider the components of the multi-physics field in the valve independently, and input them into the branch network Branch Net respectively;

[0018] S22. Calculate the parameters of the solid boundary, physical field, outlet, and inlet in the Branch net respectively, extract features through Max-pooling, embed them in the Trunk net respectively, and use a single linear layer with an activation function as a mapping to obtain a neural operator model.

[0019] Preferably, in S2, the dedicated neural operator DeepMONet proxy model constructed has multiple network functions and multiple stacked operator layers for quickly predicting unknown spatial variables.

[0020] Preferably, in S3, the specific process of constructing the agent model based on the neural operator is:

[0021] S31, constructing a training database using the high-resolution data and large-scale fluid-thermal-solid coupling data generated in S1;

[0022] S32. Based on the database built in S31, train a neural operator model.

[0023] Therefore, the present invention proposes a fast prediction method based on a dedicated small model of a neural operator, which has the following beneficial effects:

[0024] (1) The present invention trains proxy models based on high-resolution data generated by a generative adversarial network and a large-scale fluid-thermal-solid coupling database, effectively reducing the amount of calculation and improving the computational efficiency and accuracy of the system. It is particularly suitable for complex valve systems that require fast calculation and optimization.

[0025] (2) The method of the present invention integrates high-resolution simulation data generated by generative adversarial networks (GANs) and a large-scale fluid-heat-solid coupling database to train efficient proxy models and construct dedicated neural operator proxy models, achieving significant reductions in computational costs while maintaining high accuracy, thereby significantly improving prediction efficiency and optimization capabilities in complex environments.

[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 Schematic diagram of a data enhancement model of a neural operator-based fast prediction method using a dedicated small model according to the present invention;

[0028] Figure 2 It is a schematic diagram of a DeepMONet proxy model based on a special small model fast prediction method based on a neural operator of the present invention. DETAILED DESCRIPTION

[0029] To make the technical solutions, advantages, and purposes of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are part of the embodiments of the present invention, not all of them. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0030] Unless otherwise defined, technical or scientific terms used in the present invention shall have the same meaning as commonly understood by one of ordinary skill in the art to which the present invention belongs.

[0031] like Figure 1-Figure 2 As shown, a dedicated small model fast prediction method based on a neural operator provided by the present invention includes the following steps:

[0032] S1. Data enhancement based on generative adversarial network method;

[0033] S2. Based on the multi-physics characteristics of valves, a dedicated neural operator DeepMONet agent model is constructed;

[0034] S3, based on the enhanced data generated by S1, is brought into S2 for training to build a proxy model based on neural operators;

[0035] S4, based on the agent model of S3, strengthens the learning framework and optimizes and quickly predicts the structural performance of the valve.

[0036] Example

[0037] S1. Data enhancement based on generative adversarial network method.

[0038] like Figure 1 As shown in the figure, CFD simulation data and low-resolution data obtained from experiments are input into PI-GAN to generate corresponding high-resolution data, thus achieving multi-physics field super-resolution reconstruction.

[0039] S11. Construct a generative adversarial network (PI-GAN) that generates high-resolution outputs by training low-resolution inputs, thereby building a PI-GAN model for multi-physics super-resolution reconstruction. The low-resolution input is a sparse grid, and the high-resolution output is a fine grid.

[0040] S12. According to the PI-GAN constructed in step S11, the low-resolution input data is enhanced through the network to generate high-resolution data and a large-scale fluid-thermal-solid coupling database, which are used as input data for training the neural operator agent model;

[0041] S13. Preprocess the training set consisting of input data for training the neural operator agent model, and determine the incoming flow conditions and valve information with different openings corresponding to each training data.

[0042] S2. Based on the multi-physics characteristics of the valve, a dedicated neural operator DeepMONet proxy model is constructed. In this stage, a multi-physics neural operator proxy model is established for the fluid mechanics characteristics of the valve. Figure 2 The specific process is as follows:

[0043] S21. Construct a dedicated neural operator agent model based on the fluid mechanics characteristics, consider the components of the multi-physics field in the valve independently, and input them into the branch network Branch Net respectively;

[0044] S22. Calculate the parameters of the solid boundary, physical field, outlet, and inlet in the Branch net respectively, extract features through Max-pooling, embed them in the Trunk net respectively, and use a single linear layer with an activation function as a mapping to obtain a neural operator model.

[0045] In S2, a dedicated neural operator DeepMONet proxy model is constructed with multiple network functions and multiple stacked operator layers for quickly predicting unknown spatial variables.

[0046] S3, based on the enhanced data generated by S1, is brought into S2 for training to build a proxy model based on neural operators;

[0047] The high-resolution data obtained in step S1 and the large-scale flow-heat-solid coupling data are used to build a database. The DeepMONet constructed in step S2 is trained using this database. The specific implementation process of building a proxy model based on neural operators is as follows:

[0048] S31, constructing a training database using the high-resolution data and large-scale fluid-thermal-solid coupling data generated in S1;

[0049] S32. Based on the database built in S31, train a neural operator model.

[0050] S4, based on the agent model of S3, strengthens the learning framework and optimizes and quickly predicts the structural performance of the valve.

[0051] Therefore, the present invention provides a special small model fast prediction method based on neural operators, enhances low-resolution data based on generative adversarial networks, generates high-resolution data and large-scale flow-thermal-solid coupling data, and constructs a special DeepMONet model based on the multi-physical field characteristics of the valve, which solves the time and resources required for traditional CFD simulations, and achieves a significant reduction in computing costs while maintaining high accuracy, thereby greatly improving the prediction efficiency and optimization capabilities in complex environments.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. 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 solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A fast prediction method based on a dedicated small model of a neural operator, characterized in that: The following steps are involved: S1. Data enhancement based on generative adversarial network method; S2. Based on the multi-physics characteristics of valves, a dedicated neural operator DeepMONet agent model is constructed; S3, based on the enhanced data generated by S1, is brought into S2 for training to build a proxy model based on neural operators; S4, based on the agent model of S3, strengthens the learning framework and optimizes and quickly predicts the structural performance of the valve.

2. A fast prediction method based on a neural operator using a dedicated small model according to claim 1, characterized in that: In S1, the data enhancement method is implemented using the generative adversarial network PI-GAN. The implementation process includes the following steps: S11. Construct a generative adversarial network (PI-GAN) that generates high-resolution outputs by training low-resolution inputs, thereby building a PI-GAN model for multi-physics super-resolution reconstruction. The low-resolution input is a sparse grid, and the high-resolution output is a fine grid. S12. According to the PI-GAN constructed in step S11, the low-resolution input data is enhanced through the network to generate high-resolution data and a large-scale fluid-thermal-solid coupling database, which are used as input data for training the neural operator agent model; S13. Preprocess the training set consisting of input data for training the neural operator agent model, and determine the incoming flow conditions and valve information with different openings corresponding to each training data.

3. The neural operator-based fast prediction method of a dedicated small model according to claim 1, characterized in that: In S2, a dedicated neural operator DeepMONet proxy model is constructed based on the multi-physics field characteristics of the valve. The specific process includes: S21. Construct a dedicated neural operator agent model based on the fluid mechanics characteristics, consider the components of the multi-physics field in the valve independently, and input them into the branch network Branch Net respectively; S22. Calculate the parameters of the solid boundary, physical field, outlet, and inlet in the Branch net respectively, extract features through Max-pooling, embed them in the Trunk net respectively, and use a single linear layer with an activation function as a mapping to obtain a neural operator model.

4. The neural operator-based fast prediction method of a dedicated small model according to claim 1, characterized in that: In S2, a dedicated neural operator DeepMONet proxy model is constructed with multiple network functions and multiple stacked operator layers for quickly predicting unknown spatial variables.

5. The neural operator-based fast prediction method of a dedicated small model according to claim 1, characterized in that: In S3, the specific process of building a proxy model based on neural operators is as follows: S31, constructing a training database using the high-resolution data and large-scale fluid-thermal-solid coupling data generated in S1; S32. Based on the database built in S31, train a neural operator model.

Images