Neural network agent model construction method for flow field prediction

Abstract

The invention discloses a neural network agent model construction method for flow field prediction. The method comprises the following steps: dividing a whole flow field into a plurality of sub-regions; respectively training different neural network sub-models for the sub-regions; the neural network sub-models are dynamically integrated through a hybrid expert system MoE technology, so that the overall efficient prediction of the flow field is realized; through the flow field division strategy, the model can divide the flow field into sub-regions with similar flow characteristics, it is ensured that the model more accurately adapts to a specific flow mode in each sub-region, and the accuracy of overall flow field prediction is improved; according to the method, transfer learning is introduced into each sub-domain model, the general characteristics of global flow are reserved, and efficient learning of a new flow mode is realized, so that the adaptability of the model to different flow modes is enhanced, and the dependence on a large amount of data is reduced; according to the method, the calculation and training cost can be reduced, high-precision prediction is ensured, and hardware resources are greatly saved.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method for constructing a neural network proxy model for flow field prediction. Background Art

[0002] Combining artificial intelligence with fluid dynamics calculations has become a major trend in future development. In particular, deep neural networks, with their flexible architecture and efficient computing power, have demonstrated strong application value in multiple complex tasks such as aerospace and mechanical engineering. Traditional numerical solution methods for fluid mechanics, such as numerical calculations of partial differential equations, often consume a lot of computing resources and time, while the introduction of neural network proxy models provides a new way to reduce computing costs, making efficient flow field simulation and optimization possible.

[0003] In specific applications, neural network proxy models have been widely used in the design and optimization of complex mechanical systems. For example, in flow field prediction tasks such as the flow field in liquid rocket engines, by training deep neural networks, new technologies can use design variables such as nozzle geometry and mass flow to quickly predict key performance indicators such as thrust and efficiency. Such models can not only significantly reduce the computational burden in the design iteration process, but also improve the efficiency of the optimization process. However, existing methods are often limited to the prediction of a single or a few performance indicators and lack the flexibility to adapt to different needs. When faced with changing design requirements or optimization goals, it is often necessary to rebuild the proxy model, which affects its versatility and scope of application.

[0004] To overcome these limitations, researchers are exploring a new type of proxy model that can not only output a single performance indicator, but also predict the complete flow field solution in all aspects. Such models directly output the distribution of physical quantities such as velocity and pressure in the flow field by accepting inputs such as initial conditions and boundary conditions. Deep learning architectures such as U-Net are particularly suitable for this task due to their multi-layer convolution and decoder structure, and can quickly capture and express multi-scale features in the flow field. In practical applications, such neural network models can generate prediction results for complex flow fields within milliseconds. In contrast, traditional solvers (such as OpenFOAM) may take tens of seconds to achieve similar accuracy under the same conditions. This progress not only speeds up the solution process of the flow field, but also provides a more effective tool for understanding complex flow patterns, which will help promote the intelligent development of fluid mechanics and its application in practical engineering.

[0005] Despite this, the flow patterns in real flow fields are still extremely complex, usually involving multiple areas with significantly different characteristics and that influence each other. Currently, many neural network-based flow field prediction methods still rely on a single model to uniformly model the entire flow field area. This method has certain difficulties in capturing the unique characteristics of different flow patterns and often requires a large number of model parameters, which not only increases the complexity of training, but may also lead to parameter redundancy and waste of computing resources. Therefore, how to effectively integrate and optimize the characteristics of different flow patterns has become an important issue that needs to be urgently addressed in fluid mechanics engineering applications. Summary of the invention

[0006] This paper aims to propose an efficient hybrid expert system MoE prediction method for multi-mode coupled flow field solutions to meet the prediction needs of multiple flow modes in complex flow fields. The specific objectives are as follows:

[0007] Accurate modeling of multiple flow modes: Modeling and training are performed separately for different sub-regions in the flow field to capture the unique flow characteristics of each region. This method aims to improve the prediction accuracy of the model in each sub-region, thereby achieving accurate prediction of the overall flow field.

[0008] Improve training and reasoning efficiency: On the basis of ensuring high-resolution flow field prediction, optimize the training and reasoning process of the neural network. By dynamically selecting and weighting different subdomain models, redundant training and waste of computing resources are avoided, thus significantly improving training efficiency.

[0009] Ensure the accuracy of physical solutions: In the model integration stage, prior physical knowledge and dynamic gating mechanisms are combined to ensure that the prediction results conform to physical laws, thereby improving the credibility and scientificity of the flow field solution.

[0010] By achieving the above objectives, it is expected that the present invention will be able to construct an agent model that can not only accurately capture local flow characteristics but also have efficient global flow field prediction capabilities, providing solid technical support for the understanding and application of complex flow fields.

[0011] In addition, the flow field in the present invention includes various physical fields, including velocity field, temperature field, pressure field, etc.

[0012] To this end, the technical solution provided by the present invention is: a method for constructing a neural network proxy model for flow field prediction, comprising:

[0013] (1) The entire flow field is divided into multiple sub-regions, where the flow patterns in different sub-regions are not similar;

[0014] (2) Train different neural network sub-models for these sub-regions;

[0015] (3) The neural network sub-models are dynamically integrated through the hybrid expert system MoE technology to achieve efficient prediction of the overall flow field.

[0016] Preferably, the flow field division specifically includes the following steps:

[0017] (1.1) According to different flow characteristics in the flow field, the flow field is divided into several sub-areas according to a division strategy; the division strategy includes division experience, or using a clustering data algorithm to find the optimal division method, so that the flow characteristics in each sub-area are similar, so as to simplify the prediction difficulty;

[0018] (1.2) After the division, a dedicated neural network sub-model is constructed and trained for each sub-region so that each sub-model can minimize the prediction error in the corresponding region.

[0019] Preferably, the neural network sub-model training specifically includes the following steps:

[0020] (2.1) During the training process, a typical sub-region is first selected, and the key flow characteristics of the region are learned through a large number of data samples to train a basic model;

[0021] (2.2) Migrate this model to other sub-regions and fine-tune it to adapt to the flow characteristics of each region to ensure that the model can efficiently identify and apply the common characteristics of each region while capturing the unique flow patterns of each region.

[0022] Preferably, the dynamic integration of the neural network sub-models specifically comprises the following steps:

[0023] (3.1) During integration, the gating network of the hybrid expert system MoE dynamically selects the most suitable sub-model to predict the flow field according to the characteristics of the input data; the hybrid expert system MoE automatically selects one or more most suitable expert sub-models to participate in the prediction based on the characteristics of the data, ensuring that the data in different regions can call the best sub-model; the final output of the integrated model is the weighted sum of the outputs of each sub-model, and the weight is determined by the output of the gating network, thereby ensuring that the entire flow field obtains the best prediction effect in each sub-region.

[0024] (3.2) Model training and optimization

[0025] The integrated model generated by (3.1) is trained. When training the entire hybrid expert system MoE model, the parameters of the gating network and each sub-model are jointly optimized to minimize the error between the prediction result and the true solution of the flow field. Through dynamic expert selection, the hybrid expert system MoE model can quickly adapt to different flow patterns and improve the overall prediction accuracy. At the same time, since different sub-models can be calculated in parallel, the hybrid expert system MoE method can significantly reduce the consumption of computing resources while maintaining accuracy.

[0026] The efficient subdomain integrated model based on MoE (hybrid expert system) proposed in this paper focuses on solving the complex flow field prediction problem with multiple flow modes coexisting. Compared with a single model, this method has significant advantages:

[0027] (1) Improving the accuracy of flow field prediction: Through the flow field division strategy, the model can divide the flow field into sub-regions with similar flow characteristics, ensuring that the model can adapt to the specific flow pattern more accurately in each sub-region, thereby improving the accuracy of the overall flow field prediction.

[0028] (2) Improving model adaptability and reducing data requirements: Introducing transfer learning in each subdomain model to retain the common characteristics of global flow while achieving efficient learning of new flow patterns, thereby enhancing the model's adaptability to different flow patterns and reducing dependence on large amounts of data.

[0029] (3) Saving computing resources: This method allows different sub-regions to be processed in parallel, reducing computing and training costs, ensuring high-precision predictions while also significantly saving hardware resources.

[0030] In summary, this project not only effectively solved the prediction problems caused by the differences in various flow patterns, but also reduced the calculation and resource requirements of the model, demonstrating strong technical advantages and application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of a method for constructing a neural network proxy model for flow field prediction according to the present invention. DETAILED DESCRIPTION

[0032] 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.

[0033] In the description of the embodiments of the present invention, it should be noted that if the terms "center", "upper", "lower", "left", "right", "vertical", "horizontal", "inner", "outer", etc. appear, the orientation or position relationship indicated is based on the orientation or position relationship shown in the drawings, or the orientation or position relationship in which the invented product is usually placed when used. It is only for the convenience of describing the present invention and simplifying the description, and does not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operate in a specific orientation, and therefore cannot be understood as a limitation on the present invention. In addition, the terms "first", "second", "third", etc. are only used to distinguish the description, and cannot be understood as indicating or implying relative importance.

[0034] In addition, the terms "horizontal", "vertical", "overhanging" and the like do not mean that the components must be absolutely horizontal or overhanging, but can be slightly tilted. For example, "horizontal" only means that its direction is more horizontal than "vertical", and does not mean that the structure must be completely horizontal, but can be slightly tilted.

[0035] In the description of the embodiments of the present invention, "plurality" means at least 2.

[0036] In the description of the embodiments of the present invention, it is also necessary to explain that, unless otherwise clearly specified and limited, the terms "set", "install", "connect", and "connect" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal connection of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0037] In recent years, the successful application of machine learning in many fields, especially in natural language processing and computer vision, has attracted the attention of the scientific community to its potential in the field of fluid mechanics. Among them, the Mixture of Experts (MoE) technology has become an important research direction in flow field prediction due to its excellent performance in handling diverse tasks.

[0038] The core of MoE technology is to decompose complex problems into multiple sub-models, each of which focuses on a specific task or field. Through the routing network, the system can dynamically select the most appropriate sub-model for prediction at each input stage. This method improves processing efficiency and maintains high accuracy in different fields. For example, in natural language processing tasks, the MoE model achieves better performance by selecting different language experts to process specific inputs.

[0039] In fluid mechanics applications, MoE technology also shows great potential. Faced with diverse flow patterns in complex flow fields, MoE can efficiently integrate different sub-models to accurately capture flow field characteristics. This method makes the learning process more flexible and efficient, thereby improving model performance overall.

[0040] This research plan is to design an efficient large model of full-flow coupled neural network based on the underlying architecture of various neural networks. The model integrates different sub-models and makes full use of MoE technology to ensure accurate prediction of each flow mode. Each sub-model will focus on a specific flow area, such as turbulent or laminar, and adopt a shared parameter strategy to improve learning efficiency. The sharing mechanism not only reduces the consumption of computing resources, but also accelerates the training and reasoning process of the model. By sharing the underlying architecture parameters, the system can efficiently learn the common characteristics of the flow field while maintaining sensitivity to the unique characteristics of each sub-area.

[0041] In order to fully capture the characteristics of the entire flow field, the present invention will also ensure information interaction between the various sub-models through inter-regional coupling strategies. This coupling mechanism enables the model to establish connections between different flow patterns, thereby more comprehensively understanding the overall behavior of the flow field. By dynamically adjusting the weights and parameters of the sub-models, the system can adaptively optimize the prediction results and improve the overall prediction accuracy.

[0042] In summary, this paper combines the advantages of MoE technology and deep learning to establish an efficient and accurate full flow field prediction model. By dynamically selecting and coupling each sub-model, it can achieve efficient prediction of diversified flow patterns in complex flow fields, providing new ideas and methods for fluid mechanics research and application.

[0043] Embodiment 1:

[0044] Combination Figure 1 , a neural network proxy model construction method for flow field prediction, comprising:

[0045] (1) The entire flow field is divided into multiple sub-regions, where the flow patterns in different sub-regions are not similar;

[0046] (1.1) According to different flow characteristics in the flow field, the flow field is divided into several sub-areas according to a division strategy; the division strategy includes division experience, or using a clustering data algorithm to find the optimal division method, so that the flow characteristics in each sub-area are similar, so as to simplify the prediction difficulty;

[0047] (1.2) After the division, a dedicated neural network sub-model is constructed and trained for each sub-region so that each sub-model can minimize the prediction error in the corresponding region.

[0048] (2) Train different neural network sub-models for these sub-regions;

[0049] (2.1) During the training process, a typical sub-region is first selected, and the key flow characteristics of the region are learned through a large number of data samples to train a basic model;

[0050] (2.2) Migrate this model to other sub-regions and fine-tune it to adapt to the flow characteristics of each region to ensure that the model can efficiently identify and apply the common characteristics of each region while capturing the unique flow patterns of each region.

[0051] (3) The neural network sub-models are dynamically integrated through the hybrid expert system MoE technology to achieve efficient prediction of the overall flow field.

[0052] (3.1) During integration, the gating network of the hybrid expert system MoE dynamically selects the most suitable sub-model to predict the flow field according to the characteristics of the input data; the hybrid expert system MoE automatically selects one or more most suitable expert sub-models to participate in the prediction based on the characteristics of the data, ensuring that the data in different regions can call the best sub-model; the final output of the integrated model is the weighted sum of the outputs of each sub-model, and the weight is determined by the output of the gating network, thereby ensuring that the entire flow field obtains the best prediction effect in each sub-region.

[0053] (3.2) Model training and optimization

[0054] The integrated model generated by (3.1) is trained. When training the entire hybrid expert system MoE model, the parameters of the gating network and each sub-model are jointly optimized to minimize the error between the prediction result and the true solution of the flow field. Through dynamic expert selection, the hybrid expert system MoE model can quickly adapt to different flow patterns and improve the overall prediction accuracy. At the same time, since different sub-models can be calculated in parallel, the hybrid expert system MoE method can significantly reduce the consumption of computing resources while maintaining accuracy.

[0055] 1. Taking the flow field prediction model of the liquid rocket engine as an example, the following briefly describes the embodiment:

[0056] The method for constructing a flow field prediction model for the internal flow field of a liquid rocket engine comprises the following steps:

[0057] (1) Flow field division and sub-model training

[0058] The combustion and flow process of the propellant in the engine is an extremely complex physical and chemical process. To facilitate analysis, the flow field in the engine is usually divided into multiple flow sub-regions without obvious boundaries based on the aerodynamic point of view, such as the atomization mixing region, the combustion region, the cooling film mixing region, the nozzle flow region, etc.; This embodiment first trains the U-netViT sub-model for each sub-region separately, and then integrates them through the MoE technology.

[0059] Assume that the entire flow field F can be divided into N sub-regions F i (i=1,2,…,N). Each sub-region Fi contains a specific flow feature φ i , the goal of flow field division is to find the optimal division strategy S so that the flow characteristics of each sub-region are as consistent as possible, that is:

[0060]

[0061] in, represents the expected value, and ||·|| represents a norm. The flow field partition can be defined directly according to the spatial dimension through expert experience; or try to use algorithms such as clustering to dynamically find the optimal partition strategy.

[0062] For each sub-region F i , train a sub-model M i , whose training goal is to minimize the prediction error:

[0063]

[0064] Among them, θ i It is the sub-model M i Parameters, is the loss function, N' is the number of training samples, U tar,i It is the real flow field solution of each subdomain, obtained by numerical simulation. In the process of training the subdomain model, the project plans to first train a basic flow model for a representative or critical subregion, and deeply learn the flow characteristics of the region such as flow velocity and pressure distribution through a large number of data samples. Subsequently, the basic model is migrated to other subregions using transfer learning technology, and the flow pattern of the new region is learned by fine-tuning the strategy of a specific decoder layer. In this process, the shared attention layer and other parameters in U-net ViT enable the model to retain and utilize the learned general flow features, so that while quickly adapting to the unique characteristics of different subregions, it maintains the understanding and utilization of global flow knowledge, laying the foundation for the accurate prediction of the coupled full flow field after model integration.

[0065] (2) Model integration and tuning

[0066] This embodiment uses a hybrid expert system to integrate multiple sub-models. Each sub-domain model is regarded as an expert. The core of MoE is the gating network G, which dynamically selects the most appropriate sub-model M according to the characteristics of the input data. i G consists of a multi-layer neural network to learn the association between input data and each sub-model. In the forward propagation, the gating network outputs a probability distribution indicating the contribution of each sub-model to the current input. The integration process can be expressed as:

[0067]

[0068] Among them, U out is the output of the integrated model, G i is the gating weight and ψ is the learnable gating parameter.

[0069] By training and fine-tuning the parameters ψ of the gating mechanism and the sub-model parameters θ i , the prediction accuracy of the entire flow field can be further improved. This can be achieved by the gradient descent method:

[0070]

[0071] Among them, U tar The target sample for coupling the full flow field is obtained by sorting out numerical simulation results or experimental data.

[0072] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A method for constructing a neural network proxy model for flow field prediction, characterized in that: include: (1) The entire flow field is divided into multiple sub-regions, where the flow patterns in different sub-regions are not similar; (2) Train different neural network sub-models for these sub-regions; (3) The neural network sub-models are dynamically integrated through the hybrid expert system MoE technology to achieve efficient prediction of the overall flow field.

2. A method for constructing a neural network proxy model for flow field prediction according to claim 1, characterized in that: The flow field division specifically comprises the following steps: (1.1) According to different flow characteristics in the flow field, the flow field is divided into several sub-areas according to a division strategy; the division strategy includes division experience, or using a clustering data algorithm to find the optimal division method, so that the flow characteristics in each sub-area are similar, so as to simplify the prediction difficulty; (1.2) After the division, a dedicated neural network sub-model is constructed and trained for each sub-region so that each sub-model can minimize the prediction error in the corresponding region.

3. A method for constructing a neural network proxy model for flow field prediction according to claim 1, characterized in that: The neural network sub-model training specifically includes the following steps: (2.1) During the training process, a typical sub-region is first selected, and the key flow characteristics of the region are learned through a large number of data samples to train a basic model; (2.2) Migrate this model to other sub-regions and fine-tune it to adapt to the flow characteristics of each region to ensure that the model can efficiently identify and apply the common characteristics of each region while capturing the unique flow patterns of each region.

4. A method for constructing a neural network proxy model for flow field prediction according to claim 1, characterized in that: The dynamic integration of the neural network sub-models specifically includes the following steps: (3.1) During integration, the gated network of the hybrid expert system MoE dynamically selects the most appropriate sub-model to predict the flow field according to the characteristics of the input data; the hybrid expert system MoE automatically selects one or more most suitable expert sub-models to participate in the prediction based on the characteristics of the data, ensuring that the data in different regions can call the best sub-model; the final output of the integrated model is the weighted sum of the outputs of each sub-model, and the weight is determined by the output of the gated network, thereby ensuring that the entire flow field obtains the best prediction effect in each sub-region; (3.2) Model training and optimization The integrated model generated by (3.1) is trained. When training the entire hybrid expert system MoE model, the parameters of the gating network and each sub-model are jointly optimized to minimize the error between the prediction result and the true solution of the flow field. Through dynamic expert selection, the hybrid expert system MoE model can quickly adapt to different flow patterns and improve the overall prediction accuracy. At the same time, since different sub-models can be calculated in parallel, the hybrid expert system MoE method can significantly reduce the consumption of computing resources while maintaining accuracy.

Images