Method for extracting evolution characteristics of time-varying three-dimensional flow field in situ

Abstract

This invention discloses a method for extracting evolutionary features of time-varying three-dimensional flow fields under in-situ computation, comprising: setting up an in-situ computation environment and allocating available computers; using one part of the computers to construct a time-varying three-dimensional flow field model and obtain the intermediate time-sequence derivation process of the time-varying three-dimensional flow field; and using another part of the computers to generate a time-varying three-dimensional flow field animation description and output the final flow field animation. This invention alleviates the limitations of computer input / output speed and storage during storage and analysis by utilizing an in-situ computation environment. Specifically, it constructs a time-varying three-dimensional flow field model and generates flow field transition times for time-varying three-dimensional flow fields with fewer time steps, thus solving the problem in existing technologies where time-varying three-dimensional flow fields with few time steps and low accuracy cannot accurately reconstruct the intermediate time-sequence derivation process of the flow field. Results show that this invention can generate flow field dynamics that roughly conform to the original data over time, and achieves a high degree of decompression of the original data.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of fluid mechanics analysis, and more particularly, relates to a method for extracting evolution characteristics of a time-varying three-dimensional flow field in situ under on-the-fly calculation. BACKGROUND

[0002] The time-varying three-dimensional flow field is used to record the overall situation change of the flow field in multiple time steps, and in recent years, the time-varying three-dimensional flow field is used to simulate more complex physical phenomena with larger scale, higher precision and longer time. The three-dimensional time-varying flow field data generated thereby usually consumes a lot of time overhead in generating and storing simulation data due to large single time step data files, multiple time steps and other reasons, and the large-scale data volume challenges the input / output speed and storage capacity of supercomputers. This results in that the time-varying three-dimensional flow field can only be sampled in time and stored at the cost of reducing precision.

[0003] Meanwhile, the change generated by the time-varying flow field over time is usually nonlinear, but the traditional method generates intermediate bodies by using standard linear interpolation, and these interpolations are only based on local information around the interpolation position, so they may not be able to capture complex nonlinear changes. Secondly, because the typical network architecture is usually not very concerned about the position of the key point and the change in dynamics when it is applied to the time-varying three-dimensional flow field data, the generated data has a high peak signal-to-noise ratio. Therefore, a time-varying flow field compression technology is needed when saving the time-varying three-dimensional flow field data, which can only retain a small amount of time step data to reduce the overhead on the memory, while effectively and accurately restoring the intermediate time step when used.

[0004] In summary, the large-scale data volume and time steps of the time-varying three-dimensional flow field not only challenge the input / output speed and storage capacity of the computer, but also demand compression and construction of the time-varying three-dimensional flow field. SUMMARY

[0005] The purpose of the present application is to provide a method for extracting evolution characteristics of a time-varying three-dimensional flow field in situ under on-the-fly calculation, which constructs a time-varying three-dimensional flow field model for a time-varying three-dimensional flow field with fewer time steps, and generates a flow field transition moment, to solve the problem that the time-varying three-dimensional flow field in the prior art cannot accurately restore the derivation process of the intermediate moment of the flow field under the condition of fewer time steps and low precision.

[0006] In a first aspect, a method for extracting evolution characteristics of a time-varying three-dimensional flow field in situ under on-the-fly calculation is provided, comprising:

[0007] dimensionally reducing the three-dimensional flow field data of M time steps to obtain three-dimensional flow field data of N time steps;

[0008] The time super-resolution network is trained using the three-dimensional flow field data of the N time steps, specifically including: inputting the three-dimensional flow field data of the N time steps into the time super-resolution network to be trained to obtain three-dimensional flow field estimation data of M-N intermediate time steps, the three-dimensional flow field estimation data of the M-N intermediate time steps being prediction data of the three-dimensional flow field data of the M-N intermediate time steps in the M time steps, and adjusting parameters of the time super-resolution network through a loss function.

[0009] For example, the M time steps can be time step 1, time step 1-1, …, time step 1-k, time step 2, time step 2-1, …, time step 2-k, …, time step N-1, time step N-1-1, …, time step N-1-k, and time step N. The time dimension of the M time steps is reduced to obtain time step 1, …, and time step N. Therefore, the time super-resolution network can be trained through the M time steps to predict time step 1-1, …, time step 1-k, time step 2-1, …, time step 2-k, …, time step N-1, time step N-1-1, …, and time step N-1-k.

[0010] With reference to the first aspect, in some implementations of the first aspect, the time super-resolution network adopts a recurrent generative network (RGN) including a generator and a discriminator.

[0011] The generator is configured to generate twice intermediate prediction data according to input flow field velocity vector data of two time steps, and add the twice intermediate prediction data through weights to obtain final prediction data.

[0012] The discriminator is configured to match the prediction data and the original data to adjust parameters of the time super-resolution network.

[0013] With reference to the first aspect, in some implementations of the first aspect, the generator includes a feature extraction module, a time module, and an amplification module. The feature extraction module adopts four residual blocks to perform feature extraction and down-sampling operation on time-varying three-dimensional flow field data. The time module adopts a ConvLSTM architecture to learn time-varying trends and corresponding dynamic changes of data. The amplification module adopts a residual block and a voxel shuffle layer to amplify data.

[0014] With reference to the first aspect, in some implementations of the first aspect, in the feature extraction module, the convolution kernel size of the first residual block is 5x5x5, and the convolution kernel size of the other residual blocks is 3x3x3, in the first route, the step size of the first convolution layer is 1, the step size of the second convolution layer is 2, and the padding size is (number of convolution kernels-1) / / 2; in the second route, the convolution kernel size of the convolution layer is 1x1x1, the step size is 1, and no padding is performed, and then an average pooling is performed for down-sampling.

[0015] With reference to the first aspect, in some implementations of the first aspect, in the time module, the weight operation is performed through convolution operation, the convolution kernel size is 3x3x3, the step size is 1, and the padding size is 1.

[0016] With reference to the first aspect, in some implementations of the first aspect, in the amplification module, each amplification layer includes a residual block and a voxel shuffle layer, the voxel shuffle layer is used to amplify data, integrate data in a certain channel number, arrange the data in a certain order, and finally output data in one channel; the convolution kernel of the last residual block is 5x5x5, and the convolution kernel size of the other residual blocks is 3x3x3, the step size is 1, and the padding is (number of convolution kernels-1) / / 2.

[0017] With reference to the first aspect, in some implementations of the first aspect, in the discriminator, the convolution kernel size of each layer of the convolution layer except the last layer of the convolution layer is 4x4x4, the step size is 2, and the padding is 1, the step size of the last layer of the convolution layer is 0, and there is a spectral normalization layer SN and a leaky rectified linear unit LeakyReLU layer with leakage after each convolution layer.

[0018] With reference to the first aspect, in some implementations of the first aspect, the loss function of the time super-resolution network includes an adversarial loss function, and the adversarial loss function satisfies:

[0019]

[0020]

[0021] wherein, is the minimum value of the generator G loss, is the expectation operation, V T is a set of flow field data pairs, that is, D(·) is a function related to the discriminator D, and G(·) is a function related to the generator G, is the minimum value of the discriminator D loss, V I is V T is the intermediate time step data of the data pair in V

[0022] In conjunction with the first aspect, in some implementations of the first aspect, the loss function of the temporal super-resolution network includes a volumetric loss function, which satisfies:

[0023]

[0024] in, For volume loss, To obtain the expected operation, V′ is the data generated by the generator, V T A set of flow field data pairs, where V represents the original data, V I For V T The intermediate time step data of the data pair, G(·) is a function of the generator G, and ||·|| is the L2 normal form.

[0025] In conjunction with the first aspect, in some implementations of the first aspect, the loss function of the temporal super-resolution network includes a feature loss function, which satisfies:

[0026]

[0027] in, For feature loss, To obtain the expected operation, V′ is the data generated by the generator, V T A set of flow field data pairs, where V represents the original data, V I For V T The data in the middle time step of the data pair, where N is the total number of convolutional layers in the discriminator D, and k represents the k-th convolutional layer. k F represents the number of elements in the k-th convolutional layer. k G represents the features extracted by the k-th convolutional layer in the discriminator D, G(·) is a function of the generator G, and ||·|| is the L2 norm.

[0028] In conjunction with the first aspect, in some implementations of the first aspect, the loss function of the temporal super-resolution network includes a keypoint loss function, which satisfies:

[0029]

[0030]

[0031]

[0032] in, and The loss consists of two parts. To obtain the expected value, pos and pos′ represent the positions of key points in the original data and the generated data, respectively. OriginCp and pos FakeCp N represents the locations of all original data key points and generated data key points, respectively. c V represents the total number of keypoints in the original data, where i represents the keypoint in the i-th original data point. OriginCp V FakeCp V and V′ represent the regions surrounding all the original data keypoints and the generated data keypoints, respectively. λ represents the region surrounding a subset of the original data keypoints and the generated data keypoints, respectively. C1 and λ C2 Represent and The weight.

[0033] In conjunction with the first aspect, in some implementations of the first aspect, the loss function of the temporal super-resolution network satisfies:

[0034]

[0035] in, Let λ be the minimum value of the generator G loss, and λ1 be the weight corresponding to the first part of the loss. To obtain the expected operation, V represents the original data, V T Given the time-varying flow field data as input, D(·) is a function of the discriminator D, G(·) is a function of the generator G, and λ² is the weight corresponding to the volumetric loss. λ3 represents the volumetric loss, and λ3 represents the weights corresponding to the feature loss. λ4 represents the feature loss, and λ4 represents the weights corresponding to the keypoint loss. This represents a loss at a critical point.

[0036] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes:

[0037] Identify key regions in a three-dimensional flow field, wherein the key regions include all key points where the velocity vector is 0;

[0038] The information of the key points is input into the temporal super-resolution network to complete the training of the temporal super-resolution network.

[0039] In conjunction with the first aspect, in some implementations of the first aspect, the method further includes:

[0040] The computers are divided into a first computer cluster and a second computer cluster to build an in-situ computing environment. The first computer cluster is used to preprocess the data and train the time super-resolution network. The second computer cluster is used to generate intermediate time-stack three-dimensional flow field data through the trained time super-resolution network and perform visualization operations on the three-dimensional flow field data.

[0041] Secondly, a method for generating intermediate-time three-dimensional flow field data is provided, including:

[0042] Acquire three-dimensional flow field data at multiple time steps;

[0043] The three-dimensional flow field data of the plurality of time steps are input into the time super-resolution network as described in any of the implementations of the first aspect above, to obtain the three-dimensional flow field data of the intermediate time step of the plurality of time steps.

[0044] Perform visualization operations on the three-dimensional flow field data to obtain visualization results corresponding to the three-dimensional flow field data and the intermediate time step three-dimensional flow field data.

[0045] In conjunction with the second aspect, in some implementations of the second aspect, the method further includes:

[0046] The information of key points in the three-dimensional flow field data of multiple time steps is input into the time super-resolution network to obtain the three-dimensional flow field data of the intermediate time step. The key points are the points where the velocity vector is 0.

[0047] The advantages of this invention compared to the prior art are:

[0048] (1) This invention trains and visualizes time super-resolution networks based on an in-situ computing environment, which effectively alleviates the problem that flow field data cannot be visualized and stored in a timely manner due to the limitations of computer input / output speed and storage.

[0049] (2) By introducing a time super-resolution network, this invention can more accurately obtain the changes of time-varying flow field over time compared with the traditional standard linear interpolation to generate intermediates, thereby generating flow field animation at intermediate moments when only two time steps of flow field are stored.

[0050] (3) The present invention also uses key point information as input in the training of the temporal super-resolution network, so that the trained neural network parameters can better focus on the position and dynamic changes of key points, thereby reducing the peak signal-to-noise ratio of the data.

[0051] Other features and advantages of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0052] The above and other objects, features and advantages of the present invention will become more apparent from the more detailed description of exemplary embodiments of the invention in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments of the invention.

[0053] Figure 1 A schematic diagram of the overall process for extracting evolution features of a time-varying three-dimensional flow field under in-situ computation according to an embodiment of the present invention is shown.

[0054] Figure 2 A schematic diagram illustrating the in-situ computational application of a method for extracting evolution features of a time-varying three-dimensional flow field under in-situ computation according to an embodiment of the present invention is shown.

[0055] Figure 3 A schematic diagram of the construction of a time-varying three-dimensional flow field is shown, illustrating a method for extracting evolutionary features of a time-varying three-dimensional flow field under in-situ calculation according to an embodiment of the present invention.

[0056] Figure 4 A schematic diagram of the time super-resolution network for a method of extracting evolution features of time-varying three-dimensional flow fields under in-situ computation according to an embodiment of the present invention is shown.

[0057] Figure 5 The diagram illustrates the global network architecture and IB block schematic of a time super-resolution network for a method of extracting evolution features of a time-varying three-dimensional flow field under in-situ computation according to an embodiment of the present invention.

[0058] Figure 6 The parameters of each layer of the generator of the time super-resolution network for the method of extracting evolution features of time-varying three-dimensional flow fields under in-situ computation according to an embodiment of the present invention are shown.

[0059] Figure 7 A schematic diagram of the discriminator architecture of a time super-resolution network for a method of extracting evolution features of a time-varying three-dimensional flow field under in-situ computation according to an embodiment of the present invention is shown.

[0060] Figure 8 The paper presents a comparison of cloud map results at the same time step for different network architectures of a method for extracting evolution features of time-varying three-dimensional flow fields under in-situ computation according to an embodiment of the present invention. Detailed Implementation

[0061] Preferred embodiments of the present invention will now be described in more detail.

[0062] The method for extracting evolution features of time-varying three-dimensional flow fields under in-situ calculation according to the present invention, such as... Figure 1 As shown, it includes:

[0063] Set up an in-situ computing environment while en route and allocate the available computers;

[0064] A time-varying three-dimensional flow field model was constructed using a time super-resolution network.

[0065] The intermediate time derivation process of the time-varying three-dimensional flow field is obtained by training on the partitioned time-varying three-dimensional flow field dataset;

[0066] Flow field animation descriptions are generated based on neural network parameterization models.

[0067] This embodiment mainly describes the method from four aspects: building an in-situ computing environment, constructing a time-varying three-dimensional flow field model, obtaining the intermediate time derivation process of the time-varying flow field, and generating flow field animation descriptions.

[0068] 1. Set up an in-situ computing environment

[0069] In one example, the analysis and calculation of time-varying three-dimensional flow fields in an in-situ environment can be achieved through the following steps:

[0070] like Figure 2 As shown, the number of computers was divided into an 8:2 ratio to build an in-situ computing environment. The cluster of 80% of the computers was used to preprocess the data and construct a time-varying 3D flow field model to obtain the intermediate time-derivation process of the time-varying 3D flow field. The other 20% of the computers used the parameters trained by the time-varying 3D flow field model from memory to perform in-situ analysis and calculation, and output the final visualized flow field animation description. The visualization analysis of several time steps output by the time super-resolution network was performed to reduce the time and storage costs required for experimental analysis.

[0071] 2. Construction of Time-Varying Three-Dimensional Flow Field Model

[0072] Based on 80% of the computer in the in-situ computing environment described in section 1, the construction of the time-varying three-dimensional flow field is mainly accomplished through the following steps, such as... Figure 3 As shown, firstly, key regions are automatically identified in the three-dimensional flow field. Then, high-resolution data in the time dimension is compressed into low-resolution data. Finally, a time super-resolution network is constructed to establish a time-varying flow field model that can describe the flow field animation.

[0073] Step 1: In one example, the following steps can be used to automatically identify key regions in a time-varying 3D flow field:

[0074] In a vector field, points where the velocity vector is 0 are called key points. In the time-varying 3D flow field data in VTK format used in this invention, the velocity vectors are u, v, and w, which are parallel to the x-axis, y-axis, and z-axis, respectively. All key point information in the time-varying 3D flow field can be obtained and saved by iterating through the velocity vectors pointed to by u, v, and w in the VTK file to see if they are 0.

[0075] Step 2: In one example, high-resolution data of time-varying flow fields can be compressed into low-resolution data using the following steps:

[0076] Dimensionality reduction is achieved by downsampling high-resolution data in the time dimension. Specifically, time-varying 3D flow fields are extracted at specific step sizes, and the unextracted moments are used as the ground truth for training the temporal super-resolution network to determine the effects derived from intermediate moments in the flow field. The corresponding time step information and low-resolution data are saved, and the saved data is normalized to construct training and testing datasets.

[0077] Specifically, the input data for the temporal super-resolution network is first preprocessed, which involves truncating the original data and extracting an image at every three time steps to obtain time-varying three-dimensional flow field data that is upsampled and downsampled in the time dimension. The intermediate time steps are used as ground truth values ​​for comparison with the data generated by the temporal super-resolution network.

[0078] Step 3: In one example, a time-varying 3D flow field temporal super-resolution network is constructed as follows:

[0079] This invention employs a recursive generative network (RGN) with input data as a vector field. It combines recurrent neural networks and generative adversarial networks, and its architecture actually uses a convolutional long short-term memory (ConvLSTM) artificial neural network structure. The flow field velocity vector data from two selected time steps are input into the generator. The generator generates two prediction data points based on the input pair of time steps, both of which can be predictions for the same intermediate time step. These two prediction data points are summed using weights to obtain the final output. Finally, the predicted data and the original data are input into a discriminator to obtain the final discrimination result. The overall architecture of the temporal super-resolution network is as follows: Figure 4 As shown.

[0080] First, in one example, the generator for the temporal super-resolution network is constructed as follows:

[0081] The generator G is mainly divided into three modules: a feature extraction module, a temporal module, and an amplification module. The feature extraction module primarily uses four residual blocks for feature extraction and downsampling of the time-varying 3D flow field data. The temporal module employs a ConvLSTM architecture to learn the time-varying trends and corresponding dynamic changes of the data. The amplification module uses residual blocks and voxel shuffling layers to amplify the data. The generator architecture of the temporal super-resolution network is as follows: Figure 5 As shown in (a).

[0082] In the feature extraction module, the kernel size of the first residual block is 5×5×5, and the kernel size of the remaining residual blocks is 3×3×3. In route P1, the stride of the first convolutional layer is 1, the stride of the second convolutional layer is 2, and the padding size is (number of kernels - 1) / / 2, so that the length, width, and height of the data become half of their original values ​​after passing through P1. In route P2, the kernel size of the convolutional layer is 1×1×1, the stride is 1, no padding is applied, and then it undergoes downsampling through average pooling, making the results of P1 and P2 the same size, allowing for summation at corresponding positions.

[0083] In the aforementioned time module, a Convolutional Long Short-Term Memory (ConvLSTM) artificial neural network is employed. The weight operations are performed through convolution operations, with a kernel size of 3×3×3, a stride of 1, and padding of 1, ensuring that the data size remains unchanged before and after convolution. By using the ConvLSTM artificial neural network, the network can predict the data for the next time step based on the previous input data, thus enabling the temporal super-resolution network to effectively learn the dynamic changes of time-varying 3D flow field data over time.

[0084] In the aforementioned amplification module, each amplification layer consists of a residual block and a voxel shuffling layer. The last residual block has a 5×5×5 convolutional kernel, while the remaining residual blocks all have 3×3×3 convolutional kernels. The stride is 1, and the padding is (number of kernels - 1) / / 2. The specific architecture of the residual block is as follows: Figure 5 As shown in (b), after the first convolutional layer of P1 is a voxel shuffling layer. This layer amplifies the data, combines data within a certain number of channels, and arranges these data in a specific order, ultimately summing them into a single channel for output. This means the output channel number cout of the residual block is equal to the input channel number cin × n × n × n. In this invention, n is set to 2, and then input to the voxel shuffling layer. The network parameters of each layer in the generator of the temporal super-resolution network are as follows: Figure 6 As shown.

[0085] Then, in one example, the discriminator of the temporal super-resolution network is constructed as follows:

[0086] In the discriminator network architecture, except for the last convolutional layer, the kernel size of each convolutional layer is 4×4×4, the stride is 2, and the padding is 1. The stride of the last convolutional layer is 0, resulting in a final output data size of 1×1×1. Each convolutional layer is followed by a spectral normalization layer (SN) and a leaky corrected linear unit (LeakyReLU) layer. The discriminator architecture of the temporal super-resolution network is as follows: Figure 7 As shown.

[0087] Finally, in one example, the loss function of the temporal super-resolution network is composed as follows:

[0088] The loss functions for temporal super-resolution can be categorized into adversarial loss, volumetric loss, feature loss, and keypoint loss.

[0089] The adversarial loss makes the intermediate time-step data generated by the generator G more similar to the corresponding time-step data in the original data, allowing the discriminator D to better identify whether the data comes from the generator G or from the original data. Based on the definition of an adversarial generative network, the adversarial loss formula can be obtained as follows:

[0090]

[0091]

[0092] in, To minimize the loss of generator G, To obtain the expected operation, V T A set of flow field data pairs, i.e. D(·) is a function of the discriminator D, and G(·) is a function of the generator G. V is the minimum value of the discriminator D loss. I For V T The intermediate time step data of the data pair, i.e.

[0093] Volumetric loss is used to ensure that the data generated by the generator has a low mean square error, making the obtained data closer to the original data. The formula is as follows:

[0094]

[0095] in, For volume loss, To obtain the expected operation, V′ is the data generated by the generator, V T A set of flow field data pairs, where V represents the original data, V I For V T The intermediate time step data of the data pair, G(·) is a function of the generator G, and ||·|| is the L2 normal form.

[0096] Feature loss is used to ensure that the features of the data generated by the generator are similar to those of the corresponding original data as a whole.

[0097]

[0098] in, For feature loss, To obtain the expected operation, V′ is the data generated by the generator, V TA set of flow field data pairs, where V represents the original data, V I For V T The data in the middle time step of the data pair, where N is the total number of convolutional layers in the discriminator D, and k represents the k-th convolutional layer. k F represents the number of elements in the k-th convolutional layer. k G represents the features extracted by the k-th convolutional layer in the discriminator D, G(·) is a function of the generator G, and ||·|| is the L2 norm.

[0099] The loss function of keypoint loss consists of two parts, namely: and The calculation process involves using the key points in the original data as a basis, sequentially traversing the key points in the original data, and then searching for the point in the generated data whose distance is less than a threshold that is closest to a key point in the original data. When a key point in the generated data is closest to a key point in the original data and its distance is less than a threshold, the two key points are considered to be corresponding, recorded, and the distance between that point and the corresponding key point in the original data is calculated. The calculation is performed using MSE. Finally, settings are configured separately. and The weights are added together according to the weight ratio to obtain the final keypoint loss.

[0100]

[0101]

[0102]

[0103] in, and The loss consists of two parts. To obtain the expected value, pos and pos′ represent the positions of key points in the original data and the generated data, respectively. OriginCp and pos FakeCp N represents the locations of all original data key points and generated data key points, respectively. c V represents the total number of key points in the original data, where i represents the key point in the i-th original data. OriginCp V FakeCp V and V′ represent the regions surrounding all the original data keypoints and the generated data keypoints, respectively. λ represents the region surrounding a subset of the original data keypoints and the generated data keypoints, respectively. C1 and λ C2 Represent and The weight.

[0104] By multiplying and summing the aforementioned loss components according to their respective weights, the final loss function of the generator can be obtained as follows:

[0105]

[0106] in, Let λ be the minimum value of the generator G loss, and λ1 be the weight corresponding to the first part of the loss. To obtain the expected operation, V represents the original data, V T Given the time-varying flow field data as input, D(·) is a function of the discriminator D, G(·) is a function of the generator G, and λ² is the weight corresponding to the volumetric loss. λ3 represents the volumetric loss, and λ3 represents the weights corresponding to the feature loss. λ4 represents the feature loss, and λ4 represents the weights corresponding to the keypoint loss. This represents a loss at a critical point.

[0107] 3. Obtaining the intermediate time derivation process of time-varying flow fields

[0108] Based on the time-varying 3D flow field model in section 2, key points in the flow field and low-resolution time-varying 3D flow field data are used as inputs to a time super-resolution network. The network is trained in the manner described above to construct flow field animations derived from intermediate moments. Different loss functions are calculated using unsampled time step data from the time-varying 3D flow field data as the true values ​​of the derived flow field at intermediate moments, resulting in a set of neural network parameterization models with optimal performance.

[0109] Specifically, in this invention, the input data size is 3×64×64×64, that is, the data contains three channels u, v, and w, and the length, width, and height of the data itself are all 64. The training data time step is 100, the test data time step is 30, and the temporal super-resolution network without keypoint loss and with keypoint loss are trained respectively, with 100 training rounds for each.

[0110] 4. Generation of flow field animation descriptions

[0111] Using the remaining 20% ​​of the computers in the in-situ computing environment built in section 1, and based on the neural network parameterization model obtained in section 3, the aforementioned time super-resolution network is run with the neural network parameterization model, time-varying 3D flow field data at two time points, and key point information as inputs. The result is the flow field at the transition time between the two time points of the time-varying 3D flow field. The final transition time flow field image is then output on this part of the computer.

[0112] The inference process of the temporal super-resolution network is performed according to the above procedure, and the result is as follows: Figure 8As shown, the times from top to bottom in the row are t=0.71, t=0.72, t=0.73, t=0.74. Figure 8 (a) represents the original data. Figure 8 (b) represents the prediction data generated after training the model without keypoint loss. Figure 8 (c) represents the predicted data generated after training the model with keypoint loss. The colors of the flow field contour maps are mapped according to the velocity modulus. The results show that regardless of whether keypoint loss is included, the temporal super-resolution network can generate predictions that roughly match the dynamic changes of the flow field over time in the original data. For some relatively smooth and obvious local regions in the original flow field, the predicted data generated by the model with keypoint loss can better reproduce these regions.

[0113] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments.

Claims

1. A method for extracting evolution features of a time-varying three-dimensional flow field under in-situ calculation, characterized in that, include: Dimensionality reduction in the time dimension is performed on the three-dimensional flow field data of M time steps to obtain three-dimensional flow field data of N time steps; this includes truncating the original data and extracting the image of one time step every 3 time steps to obtain time-varying three-dimensional flow field data with upsampling and downsampling in the time dimension. The time super-resolution network is trained using the three-dimensional flow field data from the N time steps. Specifically, this includes inputting the three-dimensional flow field data from the N time steps into the time super-resolution network to be trained, identifying key regions in the three-dimensional flow field, wherein the key regions include all key points with a velocity vector of 0; inputting the information of the key points into the time super-resolution network to complete the training of the time super-resolution network, obtaining three-dimensional flow field estimation data for MN intermediate time steps, wherein the three-dimensional flow field estimation data for MN intermediate time steps is the prediction data for the three-dimensional flow field data of MN intermediate time steps out of the M time steps, and adjusting the parameters of the time super-resolution network through a loss function. Based on the time-varying three-dimensional flow field model, the key points in the flow field and low-resolution time-varying three-dimensional flow field data are used as inputs to the time super-resolution network. The network is trained to construct the flow field animation derived from intermediate moments. The time step data that is not selected in the time-varying three-dimensional flow field data is used as the true value of the flow field derived from intermediate moments to calculate different loss functions, and the best set of neural network parameterization models is obtained. Based on the obtained neural network parameterization model, for the above-mentioned time super-resolution network, the neural network parameterization model, the time-varying three-dimensional flow field data at two time points and the key point information are taken as input, the time super-resolution network is run, and the result is the flow field at the transition time between the two time points of the time-varying three-dimensional flow field, and the final flow field image at the transition time is output. The temporal super-resolution network employs a recursive generator network (RGN), which includes a generator and a discriminator. The generator is used to generate two intermediate prediction data based on the flow field velocity vector data of the two input time steps, and add the two intermediate prediction data by weight to obtain the final prediction data. The discriminator is used to match the predicted data and the original data to adjust the parameters of the temporal super-resolution network; The generator includes a feature extraction module, a temporal module, and an amplification module. The feature extraction module uses four residual blocks for feature extraction and downsampling of time-varying 3D flow field data. The temporal module uses a ConvLSTM architecture to learn the time-varying trend and corresponding dynamic changes of the data. The amplification module uses residual blocks and voxel shuffling layers to amplify the data. In the feature extraction module, the kernel size of the first residual block is 5×5×5, and the kernel size of the remaining residual blocks is 3×3×3. In the first path, the stride of the first convolutional layer is 1, the stride of the second convolutional layer is 2, and the padding size is (number of kernels - 1) / / 2. In the second path, the kernel size of the convolutional layer is 1×1×1, the stride is 1, no padding is performed, and then it is downsampled by an average pooling. In the time module, the weighting operation is performed through convolution, with a kernel size of 3×3×3, a stride of 1, and a padding size of 1. In the amplification module, each amplification layer includes a residual block and a voxel shuffling layer. The voxel shuffling layer is used to amplify the data, combine the data within a certain number of channels, and arrange the data in a certain order, finally summarizing it into the data of one channel for output. The convolution kernel of the last residual block is 5×5×5, the convolution kernel size of the remaining residual blocks is 3×3×3, the stride is 1, and the padding is (number of convolution kernels - 1) / / 2. In the discriminator, except for the last convolutional layer, the kernel size of each other convolutional layer is 4×4×4, the stride is 2, and the padding is 1. The stride of the last convolutional layer is 0. Each convolutional layer is followed by a spectral normalization layer SN and a LeakyReLU layer with leakage correction linear unit.

2. The method according to claim 1, characterized in that, The loss function of the temporal super-resolution network includes an adversarial loss function, which satisfies: in, To minimize the loss of generator G, To obtain the expected operation, A set of flow field data pairs, i.e. , Let D be a function of the discriminator. A function relating to the generator G, To minimize the loss of the discriminator D, for The intermediate time step data of the data pair, i.e. .

3. The method according to claim 1, characterized in that, The loss function of the temporal super-resolution network includes a volumetric loss function, which satisfies the following: in, For volume loss, To obtain the expected operation, Data generated by the generator It is a set of flow field data pairs, where V represents the original data. for The intermediate time step data of the data pair. A function relating to the generator G, It is an L2 paradigm.

4. The method according to claim 1, characterized in that, The loss function of the temporal super-resolution network includes a feature loss function, which satisfies: in, For feature loss, To obtain the expected operation, Data generated by the generator It is a set of flow field data pairs, where V represents the original data. for The data in the middle time step of the data pair, where N is the total number of convolutional layers in the discriminator D, and k represents the k-th convolutional layer. k F represents the number of elements in the k-th convolutional layer. k This represents the features extracted by the k-th convolutional layer in discriminator D. A function relating to the generator G, It is an L2 paradigm.

5. The method according to claim 1, characterized in that, The loss function of the temporal super-resolution network includes a keypoint loss function, which satisfies the following: in, and The loss consists of two parts. To obtain the expected operation, and This represents the location of key points in the original data and key points in the generated data. and N represents the locations of all original data key points and generated data key points, respectively. c This represents the total number of key points in the original data, where i represents the i-th key point in the original data. , These represent the areas surrounding all the original data key points and the generated data key points, respectively. and These represent the areas surrounding key points in the original data and key points in the generated data, respectively. and Represent and The weight.

6. The method according to claim 1, characterized in that, The loss function of the temporal super-resolution network satisfies: in, To minimize the loss of generator G, The weights corresponding to the first part of the loss are... To obtain the expected operation, V represents the original data. For the input time-varying flow field data, Let D be a function of the discriminator. A function relating to the generator G, The weights corresponding to the volume loss. For volume loss, The weights corresponding to the feature loss. For feature loss, The weights corresponding to the key point loss are... This represents a loss at a critical point.

7. The method according to claim 1, characterized in that, The method further includes: The computers are divided into a first computer cluster and a second computer cluster to build an in-situ computing environment. The first computer cluster is used to preprocess the data and train the time super-resolution network. The second computer cluster is used to generate intermediate time-stack three-dimensional flow field data through the trained time super-resolution network and perform visualization operations on the three-dimensional flow field data.

8. A method for generating three-dimensional flow field data at intermediate time points, characterized in that, The method includes: Acquire three-dimensional flow field data at multiple time steps; The three-dimensional flow field data of the plurality of time steps are input into the time super-resolution network as described in any one of claims 1 to 7 to obtain the three-dimensional flow field data of the intermediate time step of the plurality of time steps; Perform visualization operations on the three-dimensional flow field data to obtain visualization results corresponding to the three-dimensional flow field data and the intermediate time step three-dimensional flow field data.

9. The method according to claim 8, characterized in that, The method further includes: The information of key points in the three-dimensional flow field data of multiple time steps is input into the time super-resolution network to obtain the three-dimensional flow field data of the intermediate time step. The key points are the points where the velocity vector is 0.

Images