An asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physics field characteristic consistency constraint

Abstract

The application relates to an asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physical field feature consistency constraint, relates to the field of fluid mechanics, can automatically correct the phase error of low-precision data, realizes high-quality flow field prediction and reconstruction, solves the problem that the existing data-driven flow field reconstruction method is difficult to process asynchronous (exists phase deviation) multi-precision CFD data, introduces physical consistency constraints containing velocity, pressure and vorticity features, automatically constructs a high-confidence training sample pair, and thus improves the accuracy of flow field time series prediction and the physical authenticity of high-resolution reconstruction.

Description

Technical Field

[0001] This invention relates to the field of fluid mechanics, and more specifically to an asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physics characteristic consistency constraints. Background Technology

[0002] Computational fluid dynamics (CFD) is an important tool for studying fluid flow patterns. In practical engineering, extremely high-resolution meshes (DNS or LES) are typically required to capture turbulent details (such as vortex shedding and boundary layer separation), but this incurs enormous computational costs. Therefore, industry often uses coarse meshes (RANS or coarsened LES) for rapid simulations, but this leads to the loss of flow field details and the accumulation of numerical errors.

[0003] In recent years, super-resolution reconstruction techniques for flow fields based on deep learning have become a research hotspot. However, existing techniques have significant limitations when processing real-world engineering data: 1. Existing technologies assume "strict time synchronization": The low-resolution input data of most methods is obtained by directly downsampling the high-resolution data, and the two are strictly corresponding in time.

[0004] 2. The "Asynchronous" Pain Point of Real Data: In real coarse-grid CFD simulations, due to numerical dissipation and truncation errors, the evolution rate of the flow field is often asynchronous with that of fine-grid simulations (there is phase lag or lead). For example, the vortex shedding period calculated by the coarse-grid may be slower than that calculated by the fine-grid.

[0005] 3. The dangers of direct pairing: If this physical "asynchronous" characteristic is ignored and low-precision data is forcibly paired with high-precision data for training based on timestamps, the neural network will learn incorrect mapping relationships, resulting in blurry reconstruction results, artifacts, and even violations of fluid dynamics laws. Summary of the Invention

[0006] In order to overcome the shortcomings of the above technologies, this invention provides a method that can automatically correct the phase error of low-precision data and achieve high-quality flow field prediction and reconstruction.

[0007] The technical solution adopted by this invention to overcome its technical problems is: An asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physics feature consistency constraints includes: S1. Obtain a high-precision dataset , , For time steps High-precision data, , The number of time steps; S2. Utilizing high-precision datasets Building a low-precision dataset ; S3. Calculate a metric function to quantify the difference in physical state between low-precision and high-precision data. ; S4. Using a metric function Data pairs were obtained through optimization using a sliding window. ; S5. Pair the data The Middle Low-precision data The input is fed into the flow field time series prediction model, and the output is a low-precision data prediction value; S6. Input the low-precision data predictions into the high-resolution reconstruction model, and output the reconstructed high-precision data. ; S7. Utilizing reconstructed high-precision data and data pairs Time step in high-precision data Training flow field time series prediction model and high-resolution reconstruction model.

[0008] Furthermore, in step S1, direct numerical simulation data of flow around a two-dimensional square cylinder is used as a high-precision dataset. Time step high-precision data The grid resolution is 128×256.

[0009] Furthermore, step S2 includes the following steps: S2-1. Transfer high-precision data The data is input into an average pooling layer with a pooling kernel of 16×16, resulting in low-precision data with a resolution of 8×16. ,all The dataset consists of several low-precision data points. , ; S2-2. From the dataset Randomly selected A number of low-precision data points constitute a low-precision dataset. .

[0010] In step S3, the formula is used. Calculate the metric function In the formula , , All are weighting coefficients. It is the L2 norm. For low-precision datasets The Middle Velocity field in low-precision data , For high-precision datasets Mid-time step high-precision data The velocity field in , For downsampling operators, For low-precision datasets The Middle Pressure field in low-precision data High-precision dataset Mid-time step high-precision data The pressure field in the middle, For curl operator.

[0011] Furthermore, step S4 includes the following steps: S4-1. Set the search range as follows , , Maximum phase tolerance; S4-2. Traversing time steps to All high-precision data within and the first For each low-precision data metric function, the time step corresponding to the metric function with the minimum value among all metric function values ​​is selected. As the optimal physical matching moment; S4-3. The first Low-precision data With time step high-precision data Constituting data pairs .

[0012] Preferred, The value is 10.

[0013] Preferably, the above-mentioned flow field time series prediction model is a ConvLSTM model.

[0014] Preferably, the high-resolution reconstruction model mentioned above is the ESRGAN model.

[0015] Furthermore, step S7 includes the following steps: S7-1. Utilizing reconstructed high-precision data and data pairs Time step in high-precision data Calculate the RMSE loss function; S7-2. Use the Adam optimizer to train the flow field time series prediction model and the high-resolution reconstruction model using the RMSE loss function.

[0016] The beneficial effects of this invention are: 1. Significantly improves the detail of flow field reconstruction (solves ghosting problem) Due to the phase discrepancy (asynchronous) between low-precision input and high-precision labels, traditional methods, which force pairing training, struggle to converge to the true physical laws of the model. In the reconstructed results, the edges of the vortex structure are blurred, and significant ghosting artifacts appear in the central region of the vortex core. The asynchronous alignment mechanism of this invention significantly improves the quality of the reconstructed flow field. Vortex edges become sharp and clear, ghosting completely disappears, and the fine structure of the flow field (such as shear layers) is effectively restored.

[0017] 2. Extremely strong physical consistency Thanks to the "vorticity" constraint introduced in the metric function, the location and shape of vortex shedding in the reconstructed flow field closely match the high-precision ground truth. This demonstrates that the method not only recovers image pixels but also accurately recreates the fluid topology.

[0018] 3. Reconstruction error is significantly reduced. Through append Figure 2 As can be seen from the comparison error map, the reconstruction error of this invention in the vortex core and regions with drastic flow field changes is significantly reduced compared to traditional methods. This further proves that the physical alignment algorithm can effectively correct the phase lag problem caused by coarse mesh calculation, enabling the model to learn the correct mapping relationship. Attached Figure Description

[0019] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a comparison diagram of the reconstruction effects of the present invention. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1 Appendix Figure 2 The present invention will be further described below.

[0021] An asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physics feature consistency constraints includes: S1. Obtain a high-precision dataset , , For time steps High-precision data, , This represents the number of time steps.

[0022] S2. Utilizing high-precision datasets Building a low-precision dataset .

[0023] S3. Calculate a metric function to quantify the difference in physical state between low-precision and high-precision data. .

[0024] S4. Using a metric function Data pairs were obtained through optimization using a sliding window. .

[0025] S5. Pair the data The Middle Low-precision data The data is input into the flow field time series prediction model and outputs low-precision data prediction values.

[0026] S6. Input the low-precision data predictions into the high-resolution reconstruction model, and output the reconstructed high-precision data. .

[0027] S7. Utilizing reconstructed high-precision data and data pairs Time step in high-precision data Training flow field time series prediction model and high-resolution reconstruction model.

[0028] It can automatically correct phase errors in low-precision data, achieving high-quality flow field prediction and reconstruction. It solves the problem that existing data-driven flow field reconstruction methods cannot handle asynchronous (with phase deviation) multi-precision CFD data. By introducing physical consistency constraints that include velocity, pressure, and vorticity characteristics, it automatically constructs high-confidence training sample pairs, thereby improving the accuracy of flow field time series prediction and the physical realism of high-resolution reconstruction.

[0029] In one embodiment of the present invention, step S1 uses direct numerical simulation data of flow around a two-dimensional square cylinder as a high-precision dataset. Time step high-precision data The grid resolution is 128×256.

[0030] In one embodiment of the present invention, step S2 includes the following steps: S2-1. Transfer high-precision data The data is input into an average pooling layer with a pooling kernel of 16×16, resulting in low-precision data with a resolution of 8×16. Simulates the spatial fuzziness characteristics of coarse meshes, all The dataset consists of several low-precision data points. , .

[0031] S2-2. From the dataset Randomly selected A number of low-precision data points constitute a low-precision dataset. This operation simulates the phase evolution lag or lead phenomenon caused by numerical dissipation in coarse-grid calculations.

[0032] In one embodiment of the present invention, step S3 is performed using the formula Calculate the metric function In the formula , , These are all weighting coefficients because the numerical ranges of velocity, pressure, and vorticity are different (for example, pressure might be in the hundreds, while velocity is only a few), and directly adding them together would cause one of the terms to be ineffective. Therefore... , , It is used to adjust the proportions, ensuring that all three are equally important. The L2 norm is used to quantify the cumulative numerical error of all pixels in the field. For low-precision datasets The Middle Velocity field in low-precision data , For high-precision datasets Mid-time step high-precision data The velocity field in , This is a pooling operator used for downsampling because high-precision data has a very dense grid (e.g., 128×256), while low-precision data has a very sparse grid (e.g., 8×16), making direct subtraction impossible. The downsampling operator "fuzzifies" the high-precision data to the same size as the low-precision data, allowing for fair error calculation. For low-precision datasets The Middle Pressure field in low-precision data High-precision dataset Mid-time step high-precision data The pressure field in the middle, For curl operator. , All of these are vorticity. This item mandates that the vortex core position and vortex shape of the high-precision flow field must be precisely aligned with those of the low-precision flow field to prevent mismatches where the values ​​are similar but the structures are misaligned (such as vortices being reversed).

[0033] In one embodiment of the present invention, step S4 includes the following steps: S4-1. Set the search range as follows , , This represents the maximum phase tolerance.

[0034] S4-2. Traversing time steps to All high-precision data within and the first For each low-precision data metric function, the time step corresponding to the metric function with the minimum value among all metric function values ​​is selected. As the optimal physical matching moment.

[0035] S4-3. The first Low-precision data With time step high-precision data Constituting data pairs This step completes a "phase correction".

[0036] In this embodiment, preferably, The value is 10.

[0037] In one embodiment of the present invention, the flow field temporal prediction model is a ConvLSTM model (Convolutional Long Short-Term Memory Network), and the high-resolution reconstruction model is an ESRGAN model (Enhanced Super-Resolution Generative Adversarial Network).

[0038] In one embodiment of the present invention, step S7 includes the following steps: S7-1. Utilizing reconstructed high-precision data and data pairs Time step in high-precision data Calculate the RMSE loss function.

[0039] S7-2. Use the Adam optimizer to train the flow field time series prediction model and the high-resolution reconstruction model using the RMSE loss function.

[0040] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for asynchronous flow field data alignment and high-resolution reconstruction based on multi-physics feature consistency constraints, characterized in that, include: S1. Obtain a high-precision dataset , , For time step High-precision data, , The number of time steps; S2. Utilizing high-precision datasets Building a low-precision dataset ; S3. Calculate a metric function to quantify the difference in physical state between low-precision and high-precision data. ; S4. Using a metric function Data pairs were obtained through optimization using a sliding window. ; S5. Pair the data The Middle Low-precision data The input is given to the flow field time series prediction model, and the output is a low-precision data prediction value. S6. Input the low-precision data predictions into the high-resolution reconstruction model, and output the reconstructed high-precision data. ; S7. Utilizing reconstructed high-precision data and data pairs Time step in high-precision data Training flow field time-series prediction model and high-resolution reconstruction model; In step S1, direct numerical simulation data of flow around a two-dimensional square cylinder is used as a high-precision dataset. Time step high-precision data The grid resolution is 128×256; In step S3, the formula is used. Calculate the metric function In the formula , , All are weighting coefficients. It is the L2 norm. For low-precision datasets The Middle Velocity field in low-precision data , For high-precision datasets Mid-time step high-precision data The velocity field in , For downsampling operators, For low-precision datasets The Middle Pressure field in low-precision data High-precision dataset Mid-time step high-precision data The pressure field in the middle, For curl operator; Step S4 includes the following steps: S4-1. Set the search range as follows , , Maximum phase tolerance; S4-2. Traversing time steps to All high-precision data within and the first For each low-precision data metric function, the time step corresponding to the metric function with the minimum value among all metric function values ​​is selected. As the optimal physical matching moment; S4-3. The first Low-precision data With time step high-precision data Constituting data pairs .

2. The asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physics feature consistency constraints according to claim 1, characterized in that, Step S2 includes the following steps: S2-1. Transfer high-precision data The data is input into an average pooling layer with a pooling kernel of 16×16, resulting in low-precision data with a resolution of 8×16. ,all The dataset consists of several low-precision data points. , ; S2-2. From the dataset Randomly selected A number of low-precision data points constitute a low-precision dataset. .

3. The asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physics feature consistency constraints according to claim 1, characterized in that: The value is 10.

4. The asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physics feature consistency constraints according to claim 1, characterized in that: The flow field time series prediction model is a ConvLSTM model.

5. The asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physics feature consistency constraints according to claim 1, characterized in that: The high-resolution reconstruction model is the ESRGAN model.

6. The asynchronous flow field data alignment and high-resolution reconstruction method based on multi-physics feature consistency constraints according to claim 1, characterized in that, Step S7 includes the following steps: S7-1. Utilizing reconstructed high-precision data and data pairs Time step in high-precision data Calculate the RMSE loss function; S7-2. Use the Adam optimizer to train the flow field time series prediction model and high-resolution reconstruction model using the RMSE loss function.

Images