A method for accuracy compensation in large eddy simulation of thruster wake.

Abstract

This invention discloses a method for compensating the accuracy of large eddy simulation (LES) for propeller wakes, belonging to the field of deep learning technology. Based on a coarse-resolution LES velocity field, this method selects multiple time snapshots near the prediction time to construct multi-time local input features. The input is a local deformation residual compensation network composed of a 3D encoding / decoding backbone network, displacement output branches, residual output branches, and a 3D differentiable deformation sampling module. After the two branches output local displacement fields and velocity residual fields, the coarse-resolution velocity field is sampled using the local displacement fields to obtain a 3D differentiable deformation sampled velocity field, resulting in an aligned velocity field. This aligned velocity field is then superimposed with the residual field to obtain the compensated wake velocity field. This invention is applicable to propeller wake engineering analysis scenarios using low-cost LES, filtered LES, or coarse-grid LES, and can obtain engineering evaluation results closer to high-fidelity LES simulations without significantly increasing computational costs.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of computational fluid dynamics and deep learning, specifically to a method for compensating for the accuracy of large eddy simulation of propeller wake. Background Technology

[0002] Thruster wakes exhibit strong vortices, shear layers, multi-scale turbulent structures, and significant unsteady evolution characteristics. Large eddy simulation (LES) can resolve large-scale turbulent structures at an engineering-acceptable computational cost, and is therefore widely used for thruster wake analysis. However, high-fidelity LES still requires high mesh resolution and time step costs. While low-cost LES, coarse-grid LES, or filtered LES can reduce computational load, they lose small-scale turbulent structures, causing key turbulence statistics such as high wavenumber bands of the energy spectrum, velocity increment statistics, turbulent kinetic energy, Reynolds stress, vorticity, and vorticity dissipation to deviate from the results of high-fidelity LES.

[0003] In the process of propulsion engineering design and scheme comparison, it is usually necessary to repeatedly evaluate wake performance for different geometric parameters, advance ratios, rotational speeds, and operating conditions. If each scheme is calculated using high-fidelity large eddy simulation (LES), it will result in high costs for mesh generation, computational resources, and time, making it difficult to meet the needs of rapid screening, rapid iteration, and multi-condition evaluation in the engineering design phase. Although low-cost LES or coarse-grid LES can improve computational efficiency, deviations in key turbulence statistics can affect the reliability of propulsion wake noise prediction, wake-induced vibration analysis, propulsion efficiency assessment, and flow mechanism judgment. Therefore, how to retain the advantages of low-cost computational efficiency while making the wake turbulence statistics close to the results of high-fidelity LES is an urgent problem to be solved in the rapid evaluation of propulsion engineering.

[0004] Existing neural network-based flow field super-resolution or reconstruction methods typically aim to minimize point-by-point errors in the instantaneous field or directly transfer image super-resolution networks to flow field data. However, these methods suffer from the following problems: First, even if coarse-scale eddy simulations (CFS) and high-fidelity CFS are at the same or similar evolutionary stages, their small-scale turbulence phases may not correspond strictly point-by-point, and forcibly pursuing point-by-point reconstruction weakens the statistical restoration objective. Second, full-field generative adversarial learning may alter the existing large-scale wake structure in CFS. Third, using only mean square error or pixel-level loss is insufficient to constrain turbulence statistical characteristics such as energy spectrum, structure function, two-point correlation, velocity increment, Reynolds stress, and vorticity. Fourth, a single U-shaped network or a single-path residual network cannot simultaneously accommodate local structure alignment, velocity amplitude compensation, and constraints on key engineering areas.

[0005] Furthermore, while existing flow field prediction methods based on physical information neural networks introduce Navier-Stokes equation residuals into the loss function to improve physical consistency, their optimization objectives are still dominated by point-by-point deviations in the instantaneous field. They do not specifically optimize for the consistency of statistical characteristics in the large eddy simulation of the propulsion system wake, and they do not consider the possible local positional shifts between the coarse-resolution wake structure and the high-fidelity wake structure.

[0006] The problems with existing large eddy simulations of propeller wakes can be summarized in the following aspects:

[0007] 1. The problem of missing high wavenumber energy and small-scale turbulence statistical characteristics in coarse eddy simulation or filtered large eddy simulation;

[0008] 2. The traditional point-by-point reconstruction loss is not suitable for compensating for the accuracy of weakly paired large eddy simulations;

[0009] 3. The neural network repair process is prone to damaging the existing large-scale wake structure of coarse eddy simulation;

[0010] 4. There is a local positional offset between the coarse-resolution wake structure and the high-fidelity wake structure. Compensation by velocity residual alone can easily produce excessive residuals.

[0011] 5. The repair results lack consistent constraints on turbulence statistical indices such as spectrum, structure function, velocity increment, turbulent kinetic energy, Reynolds stress, vorticity, and vorticity dissipation.

[0012] Therefore, a data-driven simulation accuracy compensation method is needed specifically for propeller wake large eddy simulation, so that low-cost or filtered large eddy simulation can compensate for the missing small-scale turbulence statistical features while retaining the existing large-scale wake structure, and make it approach high-fidelity large eddy simulation in key turbulence statistical indicators. Summary of the Invention

[0013] The purpose of this invention is to solve the above-mentioned technical problems in the simulation of large eddy currents in the wake of propellers, and to provide a method for compensating the accuracy of large eddy current simulation of propeller wakes.

[0014] To achieve the above-mentioned objectives, the technical solution adopted by this invention is as follows:

[0015] In a first aspect, the present invention provides a method for compensating for the accuracy of large eddy simulation of thruster wake, comprising:

[0016] S1. Based on the geometric characteristics and operating conditions of the propeller in the propulsion system, a simulation model of the flow field in an unsteady open water area is constructed. The three-dimensional time-varying large eddy simulation flow field data under the target operating conditions is obtained through computational fluid dynamics simulation. The coarse resolution velocity field at the predicted time, the set of coarse resolution velocity fields at neighboring times, the velocity residual field at the predicted time relative to the neighboring time, the velocity modulus field at the predicted time, the velocity gradient modulus field, the vorticity modulus field, the vortex modulus field, and the normalized spatial coordinate field are extracted from the wake region behind the propeller to form multi-time local input features.

[0017] S2. The multi-time local input features are input into a pre-trained local deformation residual compensation network. First, the joint features are extracted by the three-dimensional encoder-decoder backbone network. Then, the joint features are simultaneously input into the displacement output branch and the residual output branch to obtain the local displacement field and the velocity residual field, respectively. Then, three-dimensional differentiable deformation sampling is performed on the coarse resolution velocity field at the prediction time according to the local displacement field to obtain the local deformation aligned velocity field. Finally, the local deformation aligned velocity field is added to the velocity residual field to obtain the compensated high-fidelity velocity field for propulsion system wake prediction.

[0018] As a preferred embodiment of the first aspect above, the local deformation residual compensation network includes a three-dimensional encoding and decoding backbone network, a displacement output branch, a residual output branch, and a three-dimensional differentiable deformation sampling module.

[0019] The three-dimensional encoding and decoding backbone network adopts a U-Net architecture based on three-dimensional residual convolutional blocks. Its input is the multi-time local input features, and its output is a joint feature with six channels. The joint feature is decomposed into a first sub-feature with three channels and a second sub-feature with three channels along the channels.

[0020] The displacement output branch maps the first sub-feature through the first convolutional neural network using the tanh function, and then scales it using the first scaling factor to obtain the local displacement field.

[0021] The residual output branch passes the second sub-feature through the second convolutional neural network to obtain intermediate features. At the same time, it passes the second sub-feature through a spatial gating unit composed of a cascaded 3×3×3 convolutional layer, GELU activation function, 1×1×1 convolutional layer, and Sigmoid activation function to obtain spatial gating coefficients. Then, it multiplies the spatial gating coefficients with the intermediate features element by element and scales them with the second scaling factor to obtain the velocity residual field.

[0022] In the three-dimensional differentiable deformation sampling module, an initial sampling grid is generated for the propeller wake region, and the local displacement field is converted into a grid offset and applied to the initial sampling grid to obtain an offset sampling grid. This offset sampling grid is then used to perform differentiable sampling on the coarse-resolution velocity field in the multi-time local input features to obtain a local deformation aligned velocity field. This local deformation aligned velocity field is then added to the velocity residual field to obtain a predicted high-fidelity velocity field as the final output of the local deformation residual compensation network.

[0023] As a preferred embodiment of the first aspect, the training dataset used to train the local deformation residual compensation network is obtained by performing computational fluid dynamics simulations under different operating conditions. Each training sample consists of the multi-time local input features and high-fidelity velocity field corresponding to a local block in the wake region of the propeller under one operating condition. The local blocks included in the training samples in the wake region of the propeller are obtained by first calculating the sum of the velocity gradient modulus, vorticity modulus, and swirl modulus as the structural saliency score of each position in the region, and then selecting the positions with structural saliency scores exceeding the threshold as candidate sampling centers and sampling near each candidate sampling center.

[0024] Furthermore, in the training dataset, the coarse-resolution velocity field used to construct the multi-time local input features is obtained directly through coarse-grained computational fluid dynamics simulation, or by sequentially performing spatial filtering, downsampling, and upsampling on the high-fidelity velocity field.

[0025] As a preferred embodiment of the first aspect above, the loss function used to train the local deformation residual compensation network is a weighted sum of at least two of the following: velocity error, local correlation error, velocity gradient magnitude error, vorticity magnitude error, swirl magnitude error, key wake region error, displacement smoothing constraint, displacement amplitude constraint, and residual amplitude constraint; wherein the key wake region error includes at least two of the following: wake center region error, axial velocity deficit region error, swirl region error, high gradient region error, and high vorticity region error.

[0026] As a preferred embodiment of the first aspect above, the geometric features include propeller diameter, pitch ratio, disk area ratio, hub diameter ratio, and number of blades, and the operating conditions include feed rate ratio, rotational speed, and fluid medium density.

[0027] As a preferred embodiment of the first aspect, the local deformation residual compensation network performs overlapping local block sliding window prediction on the propeller wake region during the inference process, and after fusing the high-fidelity velocity fields of the overlapping regions of adjacent local blocks, it reassembles them to form a complete predicted high-fidelity velocity field for the propeller wake region.

[0028] In a second aspect, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the large eddy simulation accuracy compensation method for propeller wake as described in any of the first aspects above.

[0029] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, can implement the large eddy simulation accuracy compensation method for propeller wake as described in any of the first aspects above.

[0030] Fourthly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0031] The memory is used to store computer programs;

[0032] The processor is configured to, when executing the computer program, implement the large eddy simulation accuracy compensation method for propeller wake as described in any of the first aspects above.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] This invention does not aim for complete point-by-point reconstruction of the instantaneous flow field, but rather for improving the consistency of key turbulence statistical indicators and engineering evaluation indicators in large eddy simulation (LES) of propulsion system wakes. Therefore, it is more suitable for weakly paired scenarios where there is a small-scale phase difference between coarse eddy simulation (CES) and high-fidelity LES. This invention performs small-range deformation alignment of the existing wake structure from CES using a local displacement field, and combines this with amplitude compensation using a velocity residual field. This reduces the risk of excessive residuals when relying solely on velocity residuals to correct local structural offsets. This invention can also be extended to other wake LES data that include a three-dimensional velocity field.

[0035] In the local deformation residual compensation network constructed in this invention, a three-dimensional encoding and decoding backbone, displacement output branches, residual output branches, and a spatial gating module enable the network to simultaneously learn local structural position correction and velocity amplitude compensation. By introducing unsteady wake evolution information through multi-time local input features, the model's ability to determine the local movement direction of shear layers, vortex zones, and high vorticity structures is improved. Through compensation results constrained by local correlation errors, gradient modulus errors, vorticity modulus errors, vortex modulus errors, and errors in key wake regions, the output field more closely approximates high-fidelity large eddy simulations in terms of local morphology, shear structures, vortex structures, and vortex structures of engineering interest.

[0036] During the training process of the local deformation residual compensation network, this invention can maintain the continuity and controllability of local deformation and residual compensation by introducing constraints such as displacement smoothing constraints, displacement amplitude constraints, and residual amplitude constraints. It is suitable for capturing spatial region data in the wake portion of the propeller in the thin layer thickness direction. Attached Figure Description

[0037] Figure 1 A schematic diagram illustrating the steps of a large eddy simulation accuracy compensation method for thruster wake.

[0038] Figure 2 This is a schematic diagram of the overall structure of the local deformation residual compensation network of the present invention;

[0039] Figure 3 This is a schematic diagram of the superposition of local displacement field deformation sampling and velocity residuals according to the present invention;

[0040] Figure 4 This is a schematic diagram of the multi-time local input feature and structural saliency sampling local block of the present invention;

[0041] Figure 5 This is a schematic diagram of the structure of a computer electronic device;

[0042] Figure 6 This is a schematic diagram of a three-dimensional model of a marine propeller provided in this embodiment;

[0043] Figure 7 This is the result of performance evaluation on the test set according to an embodiment of the present invention. Detailed Implementation

[0044] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in various embodiments of the present invention can be combined accordingly without mutual conflict.

[0045] This invention provides a method for accuracy compensation in large eddy simulation (LES) of propeller wake. This method can perform accuracy compensation based on coarse-resolution velocity fields obtained from low-cost LES or coarse-grid LES, achieving accuracy close to that of high-fidelity LES. The primary optimization objective of this method is to enhance the consistency of local structural positions, velocity amplitudes, velocity gradients, vorticity, swirl, and key wake regions in the LES of the propulsion system wake, rather than solely focusing on the complete point-by-point reconstruction of the instantaneous flow field. Figure 1As shown, the large eddy simulation accuracy compensation method for propeller wake includes steps S1 and S2 during accuracy compensation inference. The specific implementation of the two steps is described in detail below.

[0046] S1. Based on the geometric characteristics and operating conditions of the propeller in the propulsion system, a simulation model of the unsteady open water flow field is constructed. The three-dimensional time-varying large eddy simulation flow field data under the target operating conditions is obtained through computational fluid dynamics simulation. The coarse resolution velocity field at the prediction time, the set of coarse resolution velocity fields at the neighboring time, the velocity residual field at the prediction time relative to the neighboring time, the velocity modulus field at the prediction time, the velocity gradient modulus field at the prediction time, the vorticity modulus field at the prediction time, the swirl modulus field at the prediction time, and the normalized spatial coordinate field are extracted from the propeller wake region to form multi-time local input features.

[0047] It should be noted that the geometric features required for constructing the simulation model of unsteady open water flow field in this invention include propeller diameter, pitch ratio, disk area ratio, hub diameter ratio, and number of blades. Specific geometric features can also be adjusted according to actual modeling needs. Furthermore, the operating conditions required for performing the simulation include feed rate ratio, rotational speed, and fluid medium density. Specific operating condition parameters can be adjusted according to actual requirements.

[0048] It should be noted that the terms "coarse resolution" and "high fidelity" used to describe the velocity field in this invention are actually relative concepts of resolution. The resolution of the coarse-resolution velocity field is denoted as the first resolution, and the resolution of the high-fidelity velocity field is denoted as the second resolution. Theoretically, as long as the second resolution is higher than the first resolution, there is no restriction on their absolute resolution values. Since accuracy compensation is required subsequently in this invention, step S1 only needs to obtain the three-dimensional time-varying coarse-resolution velocity field; its accuracy does not need to be too high, and the specific accuracy can be adjusted according to actual needs. However, the specific resolution of the coarse-resolution velocity field in the inference stage should be as consistent as possible with the resolution of the coarse-resolution velocity field in the samples during the training stage.

[0049] Furthermore, the multi-time local input features used as model inputs in this invention are all extracted based on three-dimensional time-varying large eddy simulation flow field data. They consist of a series of sub-features, each a field data point, with the spatial dimension corresponding to the propeller wake region. The resolution is coarse, i.e., the aforementioned first resolution. The propeller wake region is a three-dimensional space. Therefore, for the coarse-resolution velocity field at different times in the multi-time local input features, each location in this three-dimensional space (i.e., the spatially discretized voxel) records velocity components in three directions. The velocity modulus field records the velocity modulus at each location in this three-dimensional space, the velocity gradient modulus field records the velocity gradient modulus at each location in this three-dimensional space, the vorticity modulus field records the vorticity modulus at each location in this three-dimensional space, the swirling modulus field records the swirling modulus at each location in this three-dimensional space, and the normalized spatial coordinate field records the normalized three-dimensional coordinates at each location in this three-dimensional space.

[0050] Taking any prediction time t as the current time, the multi-time local input features at time t can be expressed by the following formula:

[0051]

[0052] In the formula, To predict the local input features at multiple time points t, To predict the coarse-resolution velocity field at time t, This is the set of coarse-resolution velocity fields in the neighborhood. To predict the velocity residual field at a given time relative to neighboring times, To predict the velocity modulus field at time t, To predict the velocity gradient magnitude field at time t, To predict the vorticity modulus long field at time t, To predict the swirl mode length field at time t, For a normalized spatial coordinate field, This indicates splicing along the channel dimension.

[0053] It should be noted that since the coarse-resolution velocity field is time-domain variable, the set of coarse-resolution velocity fields at the aforementioned neighborhood time points is actually a set of velocity fields from adjacent time snapshots. Assuming the neighborhood radius in the time dimension is... Then the neighborhood time containing the center time itself can be represented as The neighborhood time, which does not include the central time itself, can be represented as Therefore, the set of coarse-resolution velocity fields at neighboring time points is: The velocity residual field at the predicted time relative to the neighboring time is ,in Indicates all velocity field Find the mean.

[0054] Additionally, it should be noted that if the three velocity components of the current coarse-resolution velocity field are, in order... , and Then the swirl modulus length Calculated based on two velocity components perpendicular to the main flow direction, satisfying... , where ε is a stabilizing term to prevent the value from being zero.

[0055] It should be noted that when the above-mentioned multi-time local input features are input into the network, the sub-features can be concatenated along the channel dimension and input into the network in the form of concatenated features to participate in the forward computation.

[0056] S2. The multi-time local input features are input into a pre-trained local deformation residual compensation network. First, the joint features are extracted by the three-dimensional encoder-decoder backbone network. Then, the joint features are simultaneously input into the displacement output branch and the residual output branch to obtain the local displacement field and the velocity residual field, respectively. Then, three-dimensional differentiable deformation sampling is performed on the coarse resolution velocity field at the prediction time according to the local displacement field to obtain the local deformation aligned velocity field. Finally, the local deformation aligned velocity field is added to the velocity residual field to obtain the compensated high-fidelity velocity field for propulsion system wake prediction.

[0057] In embodiments of the present invention, the local deformation residual compensation network includes a three-dimensional encoding and decoding backbone network, a displacement output branch, a residual output branch, and a three-dimensional differentiable deformation sampling module. The three-dimensional encoding and decoding backbone can extract joint features from the multi-time-time local input features, including the multi-time-time velocity field, time residual, velocity modulus, velocity gradient modulus, vortex modulus, swirl modulus, and spatial coordinates. Then, the two branches generate the local displacement field and the velocity residual field, respectively. The local displacement field obtained by the displacement output branch is used to describe the spatial offset of the local wake structure relative to the high-fidelity velocity field in the current coarse-resolution velocity field, while the velocity residual field obtained by the residual output branch is used to compensate for the velocity differences that still exist after deformation alignment. The specific network structure of this local deformation residual compensation network is described in detail below.

[0058] like Figure 2 The diagram illustrates the forward computation flow between network modules in a local deformation residual compensation network. The dimension is P. C ×P D ×P H ×P W Multi-time local input features After being spliced, it is input into the 3D encoding and decoding backbone network, P C P represents the number of feature channels. D P H PW These represent the number of voxels along the depth, height, and width directions after the three-dimensional spatial discretization of the propeller wake region. The three-dimensional encoder-decoder backbone network adopts a U-Net architecture based on three-dimensional residual convolutional blocks. The encoder is composed of multiple three-dimensional residual convolutional blocks cascaded through downsampling blocks. The encoder and decoder are connected through a bottleneck layer and have skip connections. The decoder is composed of multiple three-dimensional residual convolutional blocks cascaded through upsampling blocks, and finally outputs a six-channel joint feature through a 1×1×1 three-dimensional convolution. This six-channel joint feature is decomposed into two three-channel features along the channels, with the first three channels corresponding to the first sub-features. Input displacement output branch, the second sub-feature corresponding to the last three channels Input residual output branch. The 3D residual convolutional block is an existing technology, consisting of a Conv3d (3×3×3) layer, a GroupNorm layer, a ReLU layer, a Dropout3d layer, a Conv3d (3×3×3) layer, a GroupNorm layer, a residual summation layer, and a ReLU layer cascaded together.

[0059] Furthermore, the displacement output branch passes the first sub-feature through a first convolutional neural network (a 3D convolutional CNN network), maps it using the tanh function, and then scales it using the first scaling factor to obtain the local displacement field. The residual output branch passes the second sub-feature through a second convolutional neural network (a 3D convolutional CNN network) to obtain intermediate features. Simultaneously, it passes the second sub-feature through a spatial gating unit consisting of a cascaded 3×3×3 convolutional layer, a GELU activation function, a 1×1×1 convolutional layer, and a Sigmoid activation function to obtain spatial gating coefficients. These spatial gating coefficients are then multiplied element-wise with the intermediate features and scaled using the second scaling factor to obtain the velocity residual field. The spatial gating unit can generate spatial gating coefficients based on the decoded joint features to adjust the output intensity of the velocity residual field in the high gradient, high vorticity, and swirling regions of the wake.

[0060] The process of calculating the local displacement field and velocity residual field in the above displacement output branch and residual output branch can be expressed by the following formula:

[0061]

[0062]

[0063] In the formula, This represents the local displacement field of the three channels. This represents the maximum displacement magnitude as the first scaling factor. This is the first convolutional neural network in the shift output branch. The first three channels are the joint features extracted from the multi-time local input features after passing through a 3D encoding and decoding backbone network. The last three channels are the joint features extracted from multi-time local input features after passing through a 3D encoding and decoding backbone network. The velocity residual field is a three-channel field. This represents the residual scaling factor, which serves as the second scaling factor. For spatial gating units, ⊙ represents element-wise multiplication. This is the second convolutional neural network in the residual output branch.

[0064] See Figure 3 As shown, in the above-mentioned three-dimensional differentiable deformation sampling module, an initial sampling grid needs to be generated for the propeller wake region. The local displacement field obtained by the displacement output branch. Convert to grid offset Then applied to the initial sampling grid The offset sampling grid is obtained and used for coarse-resolution velocity fields in the multi-time local input features. Differentiable sampling is performed to obtain the local deformation-aligned velocity field and then compared with the velocity residual field. Adding them together, we obtain the predicted high-fidelity velocity field. This is the final output of the local deformation residual compensation network. The process can be expressed by the formula:

[0065]

[0066] In the formula, To predict the high-fidelity velocity field, i.e. the compensated large eddy simulation velocity field of the propulsion system wake, This represents three-dimensional differentiable interpolation sampling. This is the coarse-resolution velocity field at the current predicted time t. The initial sampling grid after normalization. For the local displacement field The normalized grid offset obtained from the conversion, The velocity residual field is obtained from the residual output branch.

[0067] It should be noted that if there are many voxels in the propeller wake region, resulting in excessive computation, the region can be divided into blocks. Each local block undergoes forward inference independently, and the results are then concatenated. In the embodiments of this invention, the local deformation residual compensation network performs overlapping local block sliding window prediction on the propeller wake region during inference. That is, the entire propeller wake region is extracted using a three-dimensional sliding window with overlapping sliding. The local blocks extracted within each window are used to construct multi-time local input features independently. Input the data into the local deformation residual compensation network to obtain the predicted high-fidelity velocity field corresponding to the local block. After obtaining the predicted high-fidelity velocity field corresponding to each local block. Then, the velocity fields of the overlapping areas of adjacent local blocks are fused using methods such as averaging, weighted averaging, or window functions, and then reassembled to form a complete high-fidelity predicted velocity field of the propeller wake region.

[0068] Additionally, it should be noted that the above Figure 2 This describes the forward inference process of a local deformation residual compensation network. However, this network needs to be supervisedly trained on a training dataset before it can be used for inference. In an embodiment of the invention, the training dataset used to train the local deformation residual compensation network is obtained through computational fluid dynamics simulations under different operating conditions. Each training sample consists of multi-time-time local input features corresponding to a local block in the propeller wake region under one operating condition, as well as a high-fidelity velocity field. Since the flow in most areas of the propeller wake is a smooth and regular background region, only the shear layer, high vorticity region, and swirling region are the most physically complex and difficult to simulate areas. Therefore, if random sampling is used, most of the local blocks corresponding to the training samples will come from the background region, causing the model to fail to learn key physical features. In embodiments of the present invention, local blocks included in the training samples within the wake region of the propeller need to be sampled based on structural saliency scores. Specifically, the sampling method is as follows: first, the sum of the velocity gradient magnitude, vorticity magnitude, and swirl magnitude of each voxel within the region is calculated as the structural saliency score of the corresponding voxel. Voxels with structural saliency scores exceeding a threshold are selected as candidate sampling centers, and sampling is performed near each candidate sampling center. Specifically, when sampling each candidate sampling center, local training blocks can be cropped from the vicinity of the candidate sampling center according to a preset structural sampling probability (in specific sampling, a random offset can be applied to each candidate sampling center, and then a batch of offset sampling centers can be finally selected as the centers of the local blocks for actual cropping sampling according to the preset structural sampling probability), to increase the probability of the wake shear layer, high vorticity region, and swirl region entering the training samples. See details below. Figure 4 As shown, the process of selecting local blocks and constructing training samples based on structural saliency scores is illustrated, where the neighborhood radius is set to two simulation time steps, i.e. .

[0069] It is also important to note that during the construction of the training dataset, the sample input is based on a coarse-resolution velocity field, while the ground truth labels require a high-fidelity velocity field. Therefore, theoretically, the optimal approach is to perform hydrodynamic simulations at two different resolutions to obtain the velocity fields for each resolution. In other words, the coarse-resolution velocity field used to construct multi-time-step local input features is directly obtained through coarse-grained computational hydrodynamic simulation, while the high-fidelity velocity field used to construct sample labels is obtained through high-fidelity fine-grained computational hydrodynamic simulation. However, in some other embodiments, only high-fidelity fine-grained computational hydrodynamic simulation can be performed, and then the coarse-resolution velocity field can be obtained by sequentially performing spatial filtering, downsampling, and upsampling on the high-fidelity velocity field. Specifically, spatial filtering, downsampling, and upsampling are performed by applying three-dimensional average pooling for low-pass filtering and downsampling on the high-fidelity velocity field, followed by trilinear interpolation to restore the spatial dimensions of the high-fidelity velocity field. The latter method, using a combination of spatial filtering, downsampling, and upsampling to obtain the coarse-resolution velocity field, reduces the simulation time and cost.

[0070] Furthermore, the loss function used in training the aforementioned local deformation residual compensation network in this invention can comprehensively consider loss terms from different dimensions, particularly focusing on velocity error, local correlation error, velocity gradient magnitude error, vorticity magnitude error, vortex magnitude error, key wake region error, displacement smoothing constraint, displacement amplitude constraint, and residual amplitude constraint. This ensures that the compensated propulsion system wake large eddy simulation approximates a high-fidelity large eddy simulation in terms of local structural positions, velocity amplitudes, and engineering-sensitive regions. In the actual construction of the total loss function, a weighted sum of at least two of the following can be used as the total loss function form: velocity error, local correlation error, velocity gradient magnitude error, vorticity magnitude error, vortex magnitude error, key wake region error, displacement smoothing constraint, displacement amplitude constraint, and residual amplitude constraint. The key wake region error includes at least two of the following: wake center region error, axial velocity deficit region error, vortex region error, high gradient region error, and high vorticity region error.

[0071] It should be noted that the method steps shown in S1~S2 above can essentially be implemented in the form of computer programs or software functional modules.

[0072] Therefore, based on the same inventive concept, such as Figure 5 As shown, the present invention also provides a computer electronic device corresponding to the large eddy simulation accuracy compensation method for propeller wake provided in the above embodiments, which includes a memory and a processor;

[0073] The memory is used to store computer programs;

[0074] The processor is used to implement the large eddy simulation accuracy compensation method for propeller wake as described above when executing the computer program.

[0075] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0076] Therefore, based on the same inventive concept, this invention provides a computer-readable storage medium corresponding to the large eddy simulation accuracy compensation method for propeller wake, wherein the storage medium stores a computer program, and when the computer program is executed by a processor, it can realize the large eddy simulation accuracy compensation method for propeller wake as described above.

[0077] Therefore, based on the same inventive concept, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can realize the large eddy simulation accuracy compensation method for propeller wake as described above.

[0078] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S2.

[0079] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0080] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0081] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0082] The present invention will further demonstrate the detailed implementation process and technical effects of the large eddy simulation accuracy compensation method for propeller wake shown in steps S1 to S2 above on a specific dataset through a specific embodiment, so as to facilitate understanding of the essence of the present invention.

[0083] Example

[0084] The steps in this embodiment are consistent with the large eddy simulation accuracy compensation method for thruster wake shown in steps S1 to S2 above. The following mainly demonstrates in detail the construction of the training sample set, network construction, and training evaluation of this method in this embodiment.

[0085] I. High-fidelity data acquisition and sample construction

[0086] In this embodiment, the geometric parameters of the propeller design specifically include the propeller diameter, pitch ratio, disk area ratio, hub diameter ratio, and number of blades. As a specific embodiment, the values ​​of these geometric parameters are shown in Table 1. A three-dimensional propeller geometric model is constructed using the geometric parameters shown in Table 1. Figure 6 As shown, the structural components in the three-dimensional propeller geometry model include the propeller shaft 1, the propeller blade 2, the propeller hub 3, and the propeller cap 4.

[0087] Table 1 Geometric parameters of the propeller in the embodiment

[0088] parameter unit numerical values propeller diameter m 0.25 Pitch ratio … 1.635 plate ratio … 0.77896 Hub diameter ratio … 0.300 Number of leaves … 5

[0089] Based on the aforementioned geometric characteristics and different operating conditions (including advance ratio, rotational speed, and fluid medium density), a simulation model of the unsteady open water flow field was established. Computational fluid dynamics was used to solve the three-dimensional time-varying flow field of the propeller under different advance ratios. This three-dimensional time-varying flow field is a high-fidelity large eddy simulation result. The obtained flow field data was preprocessed, and spatiotemporal local flow field samples were extracted as network input. The target flow field at the corresponding time point was used as the ground truth for supervision, constructing a pairwise training dataset.

[0090] The input high-fidelity large eddy simulation data is organized as a five-dimensional tensor, with dimensions representing time, channel, first spatial direction, second spatial direction, and third spatial direction, respectively. The time dimension represents the number of time snapshots, the channel dimension represents the number of physical channels, and flow thickness, radial height, and radial width represent the three-dimensional spatial dimensions. For the high-fidelity velocity field of the thruster wake, the data in the channel dimension is selected as the first three channels as the three-dimensional velocity components in a Cartesian coordinate system, corresponding to the velocities in the first coordinate direction. Velocity in the second coordinate direction and velocity in the third coordinate direction .

[0091] To construct a coarse-resolution velocity field, the high-fidelity velocity field is processed as follows:

[0092] (1) Convert the velocity field into a floating-point tensor, perform three-dimensional average pooling for each time snapshot and each velocity component, keep the resolution unchanged in the third spatial direction, and perform downsampling in the first and second spatial directions;

[0093] (2) The pooled low-resolution velocity field is restored to the original spatial size through trilinear interpolation and used as the input for the filtered large eddy simulation, with the original high-fidelity velocity field as the target.

[0094] (3) Normalize the standard scores of the input field and the target field;

[0095] (4) Extract a portion of the spatial region from the wake of the propeller to construct training samples.

[0096] This method can construct input / target pairs from a single set of high-fidelity large eddy simulation (LES) data, making it suitable for situations where strictly paired real coarse-grid LES and high-fidelity LES data are lacking. Compared to traditional pairing methods that require running high-fidelity and coarse-grid simulations simultaneously and strictly aligning time steps, this embodiment only requires a single set of high-fidelity data to generate a large number of training samples, significantly reducing data acquisition costs.

[0097] Once coarse-resolution and high-fidelity three-dimensional time-varying large eddy current simulation flow field data under different operating conditions are obtained, multi-time local input features can be constructed around different moments. In this embodiment, the time offset set is taken as... The system clips local velocity field blocks at the same spatial location from the current time snapshot and its adjacent time snapshots. Then, it concatenates the current coarse-resolution velocity field, the velocity fields from adjacent time snapshots, time residual information, velocity magnitude, velocity gradient magnitude, vorticity magnitude, vortex magnitude, swirl magnitude, and normalized spatial coordinates along the channel dimension to obtain the multi-time local input features corresponding to any prediction time t. The format is as follows:

[0098]

[0099] In the formula: To predict the local input features at multiple time points t, To predict the coarse-resolution velocity field at time t, This is the set of coarse-resolution velocity fields in the neighborhood. To predict the velocity residual field at a given time relative to neighboring times, To predict the velocity modulus field at time t, To predict the velocity gradient magnitude field at time t, To predict the vorticity modulus long field at time t, To predict the swirl mode length field at time t, For a normalized spatial coordinate field, This indicates splicing along the channel dimension. Among them, , , , , This represents the aforementioned averaging operation. The three velocity components of the current coarse-resolution velocity field are, in order: , and Then calculate the swirl modulus. When this is necessary, a stabilizing term ε needs to be added to prevent the value from being zero. The calculation formula is as follows: .

[0100] Furthermore, during the training sample pruning stage, a structural saliency score is calculated based on the current coarse-resolution velocity field, and this score is used to increase the probability of key wake structures entering the training samples. The formula for calculating the structural saliency score for each voxel is:

[0101]

[0102] In the formula, The structural significance score, For the velocity gradient magnitude, For vortex modulus, The length of the swirl modulus. According to... A preset quantile threshold is used to select candidate sampling centers. In this embodiment, the quantile threshold is set to 0.85, which is the entire wake region behind the propeller. Voxels with a value greater than 0.85 can be used as candidate sampling centers. During training, local training blocks are clipped from the vicinity of the candidate sampling centers (the specific clipped local block center is randomly offset relative to the candidate sampling center) with a preset structure sampling probability (0.65 in this embodiment), and the remaining probability is used for uniform random clipping throughout the entire wake region. This sampling method makes it easier for local structures near the shear layer, high vorticity region, swirling region, and wake center to be included in the training batch.

[0103] Finally, apart from the ground truth labels required for some auxiliary loss terms, the remaining individual training samples in the training sample set can be represented as: To reduce the amount of computation required for a single calculation, Each corresponds to a local block within the wake region of the propeller.

[0104] II. Construction of Local Deformation Residual Compensation Network

[0105] In this embodiment, the local deformation residual compensation network adopts Figure 2 The U-Net architecture shown is built based on 3D residual convolutional blocks, and its input is the multi-time local input features corresponding to a single local block. The output is the compensated large eddy simulation velocity field of the propulsion system wake, that is, the predicted high-fidelity velocity field corresponding to a single local block in the propulsion system wake. In this 3D encoder-decoder backbone, the encoder consists of multiple 3D residual convolutional blocks and downsampling blocks connected in series, forming a bottleneck feature at the end. The decoder receives the bottleneck feature and skip connection features from the encoder at the corresponding scale, forming the 3D encoder-decoder backbone. The U-Net architecture of this 3D encoder-decoder backbone differs from the traditional U-Net architecture in that it uses 3D residual convolutional blocks instead of ordinary convolutional modules. Each 3D residual convolutional block is composed of a Conv3d (3×3×3) layer, a GroupNorm layer, a ReLU layer, a Dropout3d layer, a Conv3d (3×3×3) layer, a GroupNorm layer, a residual summation layer, and a ReLU layer, cascaded sequentially. Therefore, this 3D encoder-decoder backbone performs joint feature extraction on multi-time velocity fields, time residuals, velocity magnitudes, gradient magnitudes, vorticity magnitudes, vortex magnitudes, vortex magnitudes, and normalized spatial coordinates. The final extracted joint features are 6-channel features, with the first three channels taken as the first sub-feature. Input displacement output branch, take the last three channels as the second sub-feature Input residuals and output branches.

[0106] The process of calculating the local displacement field in the displacement output branch is represented as follows:

[0107]

[0108] The process of calculating the velocity residual field in the residual output branch is expressed as follows:

[0109]

[0110] In the formula, This represents the local displacement field of the three channels. The maximum displacement amplitude is represented by the first scaling factor (in this embodiment, the maximum displacement amplitude is taken as 2 voxels). It is the hyperbolic tangent function. This is a 3D convolutional neural network in the displacement output branch. The velocity residual field is a three-channel field. This represents the residual scaling factor, which serves as the second scaling factor. This represents a spatial gating unit; ⊙ indicates element-wise multiplication. For another 3D convolutional neural network.

[0111] The aforementioned local displacement field is obtained by the hyperbolic tangent function tanh and the maximum displacement amplitude. The scaled output ensures that the constraint primarily uses local deformation to correct small-scale structural offsets rather than performing large-scale flow field shifts. Similarly, the velocity residual field is scaled by the residual scaling factor. The output, after adjustment with the spatial gating coefficient, can be used to compensate for the remaining velocity differences after deformation alignment. The spatial gating unit described above is represented as follows: The gating coefficient, by multiplying the velocity residual field element by element, enables the network to suppress unnecessary residual outputs in the background and weak structure regions, while retaining stronger compensation capabilities in the wake shear layer, high vorticity region, and swirling region.

[0112] Finally, the local displacement field is obtained. and velocity residual field Then, it is necessary to consider the local displacement field. Coarse-resolution velocity field at predicted time t Perform three-dimensional differentiable deformation sampling and compare it with the velocity residual field. By summing the residuals, the final output of the network satisfies the following deformation residual combination relationship:

[0113]

[0114] In the formula, To predict a high-fidelity velocity field in the wake of the compensated propulsion system, This represents three-dimensional differentiable interpolation sampling. The initial sampling grid is used to represent the normalized spatial coordinates corresponding to the positions of each voxel in the coarse-resolution center frame velocity field; This is the normalized mesh offset obtained from the local displacement field transformation.

[0115] In the above deformation residual combination formula, three-dimensional differentiable interpolation sampling The operation can be implemented using the `grid_sample` function in the PyTorch framework. The initial sampling grid generated for the propeller wake region can be a base normalized sampling grid, a 3D coordinate grid perfectly consistent with the target output spatial resolution, providing a reference sampling position for deformation sampling without any offset. When no deformation is applied, [the following text is incomplete and likely refers to a different operation:] ... As sampling coordinate pairs conduct The operation will yield the same result as The spatially aligned original velocity field undergoes no geometric deformation. In contrast, in this invention, the initial sampling grid... Need to go through The mesh is offset to produce a deformation that conforms to the flow field pattern. Local displacement field. The data records the offsets of various locations in space along three spatial directions, due to the basic normalized sampling grid. Each coordinate value is normalized to an interval Therefore, the local displacement field The offset in the data also needs to be converted using a scaling function. Corresponding conversion mapping to Normalize the coordinate space so that it can be compared with the basic normalized sampling grid. Direct addition makes the distorted mesh more aligned with the target flow field on a large-scale structure, thus allowing the distorted mesh to be used... right conduct The operation yields the local deformation alignment velocity field, and finally, the local deformation alignment velocity field is combined with the velocity residual field. Adding them together, we obtain the compensated high-fidelity velocity field for predicting the propulsion system wake. Through the above two steps, the network finally outputs... It retains the original coarse-resolution velocity field The large-scale flow characteristics are further refined by two branches: displacement deformation and residual compensation, to achieve dual fine correction of the spatial structure and amplitude distribution of the wake field, thereby obtaining a high-fidelity predicted velocity field.

[0116] III. Model Training and Performance Evaluation

[0117] In this embodiment, during the training phase, a composite physical statistical loss function is formed as the total loss function by calculating multiple losses between the compensated velocity field and the target high-fidelity large eddy simulation velocity field. The total loss function consists of velocity data fidelity terms, local correlation terms, local structure terms, key wake region terms, displacement constraint terms, and residual constraint terms, used to simultaneously constrain the continuity of velocity amplitude, local morphology, vortex structure, vortex structure, key engineering regions, and local deformation output. The formula for calculating the total loss function is:

[0118]

[0119] In the formula, For the total loss function, The mean absolute error of speed, For the mean square error of velocity, This is a local correlation error. For the velocity gradient magnitude error, This is the error in vorticity modulus length. For the error of the swirl modulus, For displacement smoothing constraints, For displacement amplitude constraints, For residual amplitude constraints, For key wake region errors, to The corresponding loss weights.

[0120] The aforementioned mean absolute error of velocity is used to constrain and compensate for the overall difference between the velocity field and the high-fidelity target velocity field in the three velocity components. Its calculation formula is as follows:

[0121]

[0122]

[0123] In the formula, To predict a high-fidelity velocity field in the wake of the compensated propulsion system, To simulate the velocity field of the target target in a high-fidelity large eddy simulation, which serves as the ground truth label in the training samples. This represents averaging across the channel and spatial position. The average absolute error of velocity is more robust to overall deviations, while the mean square error of velocity is more sensitive to large local errors. Combining the two can balance the consistency of overall velocity amplitude and the suppression of large local errors.

[0124] The aforementioned local correlation error is based on the local normalized correlation calculation of the velocity modulus, used to constrain the consistency of the local morphology between the compensation velocity field and the high-fidelity target velocity field. Let... To compensate for the velocity modulus of the velocity field, Given the velocity modulus of the target velocity field, the local correlation error within the local window W satisfies the calculation formula as follows:

[0125]

[0126]

[0127] In the formula, The normalized correlation coefficient within the local window W. Represents local covariance. Let represent the local variance, and ε be the numerically stable term. In this embodiment, the local window size is 5×5×5. By introducing local correlation error, the problem of overly smooth local structures caused by relying solely on point-by-point velocity error can be reduced.

[0128] The aforementioned local structural terms include velocity gradient modulus error, vorticity modulus error, and swirl modulus error, and their calculation formulas are as follows:

[0129]

[0130]

[0131]

[0132] In the formula, and The magnitudes of the velocity gradients in the compensation velocity field and the target velocity field are respectively. and These are the vortex moduli of the compensating velocity field and the target velocity field, respectively. and These are the swirling moduli for the compensated velocity field and the target velocity field, respectively. The velocity gradient modulus is used to characterize the wake shear layer and the intensity of velocity changes, the vorticity modulus is used to characterize the local vortex structure, and the swirling modulus is used to characterize the transverse rotation-induced velocity structure.

[0133] The displacement constraint term mentioned above is used to limit the spatial continuity and magnitude of the local displacement field, and the residual constraint term is used to suppress unnecessary large-amplitude velocity corrections. Its calculation formula is as follows:

[0134]

[0135]

[0136]

[0137] In the formula, For local displacement field, For the velocity residual field, This represents the difference of the local displacement field in three spatial directions. By constraining displacement smoothing, displacement amplitude, and residual amplitude, the local deformation and residual compensation can be kept continuous, local, and controllable, reducing the risk of damage to the existing large-scale wake structure of the coarse-resolution input field.

[0138] The aforementioned key wake region error consists of the weighted error corresponding to multiple region masks, and its calculation formula is as follows:

[0139]

[0140]

[0141] In the formula, For the regional category index, any one of the following can be selected: wake center region, axial velocity deficit region, swirling region, high gradient region, and high vorticity region; For the first Regional masks for key wake regions, The mean absolute error of velocity calculated based on the mask in this region is denoted as . Specifically, the regional errors corresponding to the wake center region, axial velocity deficit region, swirling region, high gradient region, and high vorticity region are respectively denoted as . , , , and . , , , and The corresponding region weights are defined based on the cross-sectional center radius, axial velocity deficit threshold, vortex modulus quantile, velocity gradient modulus quantile, and vorticity modulus quantile.

[0142] In this embodiment, after optimization, the values ​​of the above loss weights are: , , , , , , , , In the weighting of key wake regions, the weight of the wake center region is... The weight of the axial velocity deficit zone is 0.08. The weight of the swirl region is 0.10. The weight is 0.06, representing the weight in the high gradient region. The weight is 0.14 for high vorticity regions. The value is 0.14. The wake center region radius ratio is set to 0.28, the axial velocity deficit threshold ratio is set to 0.05, and the quantile thresholds for the swirling region, high gradient region, and high vorticity region are all set to 0.85. However, the above weights and thresholds are only specific training configurations for this embodiment, and in practice, each parameter can be adjusted according to the wake structure scale, training stability, and over-fixing degree.

[0143] In this embodiment, all acquired training samples are divided into training and validation sets to avoid data leakage. During training, an adaptive optimizer is used to optimize parameters based on the aforementioned total loss function, setting learning rate preheating, plateau-style learning rate descent based on validation metrics, and gradient pruning to prevent training instability. An early stopping strategy is implemented based on comprehensive statistical metrics of the validation set; training is terminated and the parameters are rolled back to optimal when the validation error fails to decrease over multiple consecutive rounds. When evaluating performance, multiple statistical metrics of the filtered large eddy simulation input field and the repaired field relative to the high-fidelity target field are calculated. By comparing the changes in these metrics between the input and repaired fields, the simulation accuracy compensation effect is quantitatively determined.

[0144] Specifically, in this embodiment, the input sample data includes 360 time snapshots, 3 velocity channels, and a propeller wake velocity field with a spatial size of 80×80×160. Training, validation, and testing are divided according to time snapshots, with 324 training time snapshots, 36 validation time snapshots, and the last 54 time snapshots used for testing. The local block size is 32×32×64. During inference, overlapping local block sliding window prediction is performed on the propeller wake region with a sliding window step size of 16×16×32. The preset structure sampling probability is 0.65, the structure significance quantile threshold is 0.85, and the maximum displacement amplitude is 2 voxels. The number of trainable parameters for the entire local deformation residual compensation network is approximately 2.28×10^6.

[0145] Finally, after performance evaluation on the test set, the error comparison between the coarse-resolution input field and the compensated velocity field relative to the high-fidelity target field in this embodiment is as follows: Figure 7As shown, the evaluation metrics include overall mean absolute error, overall root mean square error, root mean square error of velocity modulus, root mean square error of wake center region, root mean square error of axial velocity deficit region, root mean square error of vortex region, root mean square error of high gradient region, root mean square error of high vorticity region, root mean square error of high frequency structural residual, root mean square error of velocity gradient modulus, root mean square error of vorticity modulus, root mean square error of vortex modulus, and potential structural feature distance. The column lengths in the figure represent the error reduction rate of the compensated velocity field relative to the coarse resolution input field. The data is obtained through sliding window fusion evaluation of the last 54 time snapshots. The results of this embodiment show that the average absolute displacement of the local displacement field is 0.585 voxels, and the maximum absolute displacement is 2 voxels, which does not exceed the preset maximum displacement amplitude, indicating that the local deformation is confined to a small, controllable region. Test results show that the compensated velocity field significantly improves upon the coarse-resolution input field in terms of overall velocity error, velocity modulus error, wake center region, axial velocity deficit region, swirling region, high gradient region, and high vorticity region. Simultaneously, the high-frequency structural residual error is reduced by approximately 13.468%, the swirling modulus error by approximately 14.452%, and the potential structural feature distance by approximately 10.132%. This indicates that the local deformation residual compensation network can improve the consistency between the wake local structure, swirling structure, and high-frequency details and the high-fidelity target field while maintaining controllable local deformation amplitude.

[0146] Therefore, as verified by the above indicators, the compensated velocity field obtained by the method of this invention, compared with the coarse-resolution input field, can improve the consistency with high-fidelity large eddy simulation in terms of local velocity errors, key wake region errors, high-frequency structural errors, and potential structural feature spaces, under controlled deformation and controlled residual output conditions. The above verification indicators are used to illustrate the technical effects of this invention and do not constitute a limitation on specific training data, specific loss weights, or specific improvement magnitudes.

[0147] Furthermore, comparative results show that while excessively strengthening the constraint of a single deficit index can improve that individual index, it may lead to a decrease in related indices at the wake diffusion scale. When the relevant constraints are appropriately weakened, the total deficit deviation can be alleviated to some extent, while maintaining the positive improvement trend of the wake half-width. Therefore, this invention does not pursue the local optimum of a single deficit index or point-by-point error index, but rather achieves a better balance among wake half-width, axial deficit profile, axial velocity profile, mean total deficit, and local vortex structure through the synergistic design of multi-time local inputs, structural saliency sampling, local deformation alignment, residual compensation, and constraints in key wake regions. The synergistic constraint of multiple engineering quantities is an important condition for achieving accuracy compensation in the simulation of large eddy currents in the propeller wake.

[0148] In summary, the method of this invention can be used for large eddy simulation data augmentation of propulsion system wakes, ship propeller wakes, rotor wakes, jets, or other turbulent statistical features with multi-scale turbulence. While reducing reliance on high-fidelity calculations, it provides data augmentation means for propulsion hydrodynamic performance evaluation, wake noise prediction, wake-induced vibration analysis, flow mechanism research, and rapid selection of engineering solutions.

[0149] The embodiments described above are merely some preferred implementations of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for compensating the accuracy of large eddy simulation of propeller wake, characterized in that, include: S1. Based on the geometric characteristics and operating conditions of the propeller in the propulsion system, a simulation model of the flow field in an unsteady open water area is constructed. The three-dimensional time-varying large eddy simulation flow field data under the target operating conditions is obtained through computational fluid dynamics simulation. The coarse resolution velocity field at the predicted time, the set of coarse resolution velocity fields at neighboring times, the velocity residual field at the predicted time relative to the neighboring time, the velocity modulus field at the predicted time, the velocity gradient modulus field, the vorticity modulus field, the vortex modulus field, and the normalized spatial coordinate field are extracted from the wake region behind the propeller to form multi-time local input features. S2. The multi-time local input features are input into a pre-trained local deformation residual compensation network. First, the joint features are extracted by the three-dimensional encoder-decoder backbone network. Then, the joint features are simultaneously input into the displacement output branch and the residual output branch to obtain the local displacement field and the velocity residual field, respectively. Then, three-dimensional differentiable deformation sampling is performed on the coarse resolution velocity field at the prediction time according to the local displacement field to obtain the local deformation aligned velocity field. Finally, the local deformation aligned velocity field is added to the velocity residual field to obtain the compensated high-fidelity velocity field for propulsion system wake prediction.

2. The method for compensating for the accuracy of large eddy simulation of propeller wake as described in claim 1, characterized in that, The local deformation residual compensation network includes a three-dimensional encoding and decoding backbone network, a displacement output branch, a residual output branch, and a three-dimensional differentiable deformation sampling module; The three-dimensional encoding and decoding backbone network adopts a U-Net architecture based on three-dimensional residual convolutional blocks. Its input is the multi-time local input features, and its output is a joint feature with six channels. The joint feature is decomposed into a first sub-feature with three channels and a second sub-feature with three channels along the channels. The displacement output branch maps the first sub-feature through the first convolutional neural network using the tanh function, and then scales it using the first scaling factor to obtain the local displacement field. The residual output branch passes the second sub-feature through the second convolutional neural network to obtain intermediate features. At the same time, it passes the second sub-feature through a spatial gating unit composed of a cascaded 3×3×3 convolutional layer, GELU activation function, 1×1×1 convolutional layer, and Sigmoid activation function to obtain spatial gating coefficients. Then, it multiplies the spatial gating coefficients with the intermediate features element by element and scales them with the second scaling factor to obtain the velocity residual field. In the three-dimensional differentiable deformation sampling module, an initial sampling grid is generated for the propeller wake region, and the local displacement field is converted into a grid offset and applied to the initial sampling grid to obtain an offset sampling grid. This offset sampling grid is then used to perform differentiable sampling on the coarse-resolution velocity field in the multi-time local input features to obtain a local deformation aligned velocity field. This local deformation aligned velocity field is then added to the velocity residual field to obtain a predicted high-fidelity velocity field as the final output of the local deformation residual compensation network.

3. The method for compensating for the accuracy of large eddy simulation of propeller wake as described in claim 1, characterized in that, The training dataset used to train the local deformation residual compensation network is obtained by performing computational fluid dynamics simulations under different operating conditions. Each training sample consists of the multi-time local input features and high-fidelity velocity field corresponding to a local block in the wake region of the propeller under one operating condition. The local blocks in the wake region included in the training samples are obtained by first calculating the sum of the velocity gradient modulus, vorticity modulus, and swirl modulus as the structural saliency score of each position in the region, and then selecting the positions with structural saliency scores exceeding the threshold as candidate sampling centers and sampling in the vicinity of each candidate sampling center.

4. The method for compensating for the accuracy of large eddy simulation of propeller wake as described in claim 3, characterized in that, In the training dataset, the coarse-resolution velocity field used to construct the multi-time local input features is obtained directly through coarse-grained computational fluid dynamics simulation, or by sequentially performing spatial filtering, downsampling, and upsampling on the high-fidelity velocity field.

5. The method for compensating for the accuracy of large eddy simulation of propeller wake as described in claim 1, characterized in that, The loss function used to train the local deformation residual compensation network is a weighted sum of at least two of the following: velocity error, local correlation error, velocity gradient magnitude error, vorticity magnitude error, swirl magnitude error, key wake region error, displacement smoothing constraint, displacement amplitude constraint, and residual amplitude constraint; wherein the key wake region error includes at least two of the following: wake center region error, axial velocity deficit region error, swirl region error, high gradient region error, and high vorticity region error.

6. The method for compensating for the accuracy of large eddy simulation of propeller wake as described in claim 1, characterized in that, The geometric features include propeller diameter, pitch ratio, disk area ratio, hub diameter ratio, and number of blades. The operating conditions include feed rate ratio, rotational speed, and fluid medium density.

7. The method for compensating for the accuracy of large eddy simulation of propeller wake as described in claim 1, characterized in that, During the inference process, the local deformation residual compensation network performs overlapping local block sliding window prediction on the wake region behind the propeller, and after fusing the high-fidelity velocity fields of the overlapping regions of adjacent local blocks, it reassembles them to form a complete predicted high-fidelity velocity field for the wake region behind the propeller.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can realize the large eddy simulation accuracy compensation method for propeller wake as described in any one of claims 1 to 7.

9. A computer-readable storage medium, characterized in that, The storage medium stores a computer program, which, when executed by a processor, implements the large eddy simulation accuracy compensation method for propeller wake as described in any one of claims 1 to 7.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to implement, when executing the computer program, the large eddy simulation accuracy compensation method for propeller wake as described in any one of claims 1 to 7.

Images