A Flow Field Video Generation Method Based on Policy Value Architecture and Online Physics Exploration

Abstract

This invention discloses a flow field video generation method based on a policy-value architecture and online physics exploration, belonging to the interdisciplinary field of artificial intelligence and computational fluid dynamics (CFD). It includes the following steps: Step 1, constructing a multimodal training dataset for unsteady flow fields; Step 2, constructing a generative network system based on an Actor-Critic architecture; Step 3, constructing an interactive environment and reward function based on physical equations; Step 4, supervised fine-tuning training in the first stage; Step 5, online physics exploration of the generator in the second stage; Step 6, feedback co-evolution of the physical encoder in the third stage; Step 7, inference generation of unsteady flow field videos. This method constructs a reinforcement learning architecture including a policy network (Actor) and a value network (Critic), encapsulates the physical equations into a digital environment, and employs a training strategy of basic supervised fine-tuning, online physics exploration of the generator, and feedback co-evolution of the encoder.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and computational fluid dynamics (CFD), and in particular to a method for constructing a digital environment based on physical equations, utilizing an Actor-Critic (strategy-value) architecture for online physics exploration, thereby generating high-fidelity flow field videos. Background Technology

[0002] Flow field evolution is highly nonlinear and chaotic. Traditional offline learning can only mimic existing data patterns in the training set. When faced with complex, unseen conditions, the model cannot proactively adjust its strategy, resulting in poor generalization ability. Furthermore, current technologies lack mechanisms for evaluating the long-term value of flow fields, often focusing only on the error of the current frame, leading to larger errors in the latter half of the flow field video.

[0003] Current physical alignment methods mostly employ static regularization or offline preference learning, resulting in models lacking proactive exploration capabilities and only able to optimize within existing data distributions. This invention introduces an online interaction mechanism from reinforcement learning, constructing an Actor-Critic architecture that allows the model to continuously trial and error and explore in the physical environment like an intelligent agent, thereby breaking through the quality ceiling of training data. Summary of the Invention

[0004] The purpose of this invention is to provide a flow field video generation method based on policy value architecture and online physical exploration to address the above-mentioned shortcomings. This method solves the problem that existing technologies often use static regularization or offline preference learning for physical alignment, resulting in models that lack active exploration capabilities and can only optimize within existing data distributions.

[0005] This invention is achieved through the following scheme:

[0006] A flow field video generation method based on policy value architecture and online physics exploration includes the following steps:

[0007] Step 1, construct an unsteady flow field multimodal training dataset: obtain a standardized dataset containing the physical state of the flow field, working condition description, and true values ​​of dynamic evolution;

[0008] Step 2, Construct a generative network system based on the Actor-Critic architecture: Construct the policy network and value network required for reinforcement learning. The policy network is used to generate the flow field video, and the value network is used to evaluate the state value.

[0009] Step 3, Construct an interactive environment and reward function based on physical equations: Construct a digital environment that requires no human intervention, which performs real-time numerical scoring on the generated flow field video based on fluid dynamics equations;

[0010] Step 4, the first stage of supervised fine-tuning training: Initialize the policy network to give it basic generative capabilities;

[0011] Step 5: Online physics exploration of the second-stage generator: A near-end strategy optimization algorithm is adopted to enable the generator to actively explore the optimal solution space under the guidance of physics rewards;

[0012] Step 6: Third-stage feedback co-evolution of the physical encoder: Utilizing the gradient penetration mechanism of reinforcement learning, the downstream physical feedback forces the upstream physical encoder to evolve.

[0013] Step 7: Inference and generation of unsteady flow field video.

[0014] Step 1 specifically includes the following steps:

[0015] Step 1.1: Use a high-precision CFD solver to perform unsteady solutions on the Navier-Stokes equations and obtain flow field simulation data under various geometric boundaries and Reynolds number conditions.

[0016] Step 1.2: Extract from the simulation data obtained in Step 1.1 The physical field data at time t is used as the initial flow field image. ;

[0017] Step 1.3: Extract the continuous data after the initial moment from the simulation data obtained in Step 1.1. Frame physical field data as target video sequence ;

[0018] Step 1.4: Generate the corresponding operating condition description text based on the simulation settings parameters in Step 1.1. ;

[0019] Step 1.5, take the result obtained in step 1.2 The results obtained in step 1.4 and the results obtained in step 1.3 Combine and construct a triplet sample set .

[0020] Step 2 specifically includes the following steps:

[0021] Step 2.1, Construct a text encoder: Use a pre-trained language model encoder and configure it to receive the text described in step 1.4. Output semantic embedding ;

[0022] Step 2.2, Construct a physical encoder: The image described in step 1.2... After segmentation and linear projection, global features are extracted using a Transformer encoder, outputting the physical embedding. ;

[0023] Step 2.3, Constructing the video generation decoder: Using a 3D denoising network based on a diffusion model, the video generated in step 2.1... And obtained in step 2.2 Videos are generated under dual guidance; during the reinforcement learning phase, the network is defined as a policy network. ;

[0024] Step 2.4, Construct the value network: Construct a regression network with a similar structure to the policy network described in Step 2.3, but with the output layer modified to have scalar output. This network is used to receive input status. It also predicts the expected physical reward value of the video generated in this state to help train stability.

[0025] Step 3 specifically includes the following steps:

[0026] Step 3.1, Environment Interface Definition: Define the environment to receive the video frame sequence generated in Step 2.3 as actions, and invert it into a physical velocity field. ;

[0027] Step 3.2, Calculate the divergence reward: based on the continuity equation for incompressible fluids. Calculate the full-field divergence error of the generated video and convert it into a positive reward;

[0028] Step 3.3, Calculate vorticity reward: Calculate the statistical deviation between the vorticity distribution of the generated flow field and the actual physical laws, and convert it into a positive reward;

[0029] Step 3.4, Synthesize Total Physical Reward: Define the total physical reward for a single generation. ;in The weighting coefficient for divergence rewards is used to adjust the importance of the mass conservation constraint in the overall objective. The weighting coefficient representing the vorticity bonus is used to adjust the importance of flow field structural constraints.

[0030] Step 4 specifically includes the following steps:

[0031] Step 4.1, take the sample from step 1.5 Input the network system constructed in step 2;

[0032] Step 4.2: Generate a predicted video using the policy network described in Step 2.3, and calculate its comparison with the ground truth video. The pixel mean square error loss and gradient loss;

[0033] Step 4.3: Update the parameters of the physical encoder and video generation decoder using the backpropagation algorithm;

[0034] Step 4.4, Value Network Warm-up: Generate a batch of data using the SFT-enhanced model, calculate its actual physical reward, and pre-train the value network described in Step 2.4 using mean squared error loss. .

[0035] Step 5 specifically includes the following steps:

[0036] Step 5.1, Environment Initialization: Freeze the physical encoder and text encoder; Unfreeze the video generation decoder and value network;

[0037] Step 5.2, Online Sampling: Utilizing the current policy network For input Perform online inference to generate a batch of flow field video samples. And record the logarithmic probability distribution of the generation process;

[0038] Step 5.3, using the methods from step 3.4, calculate the instantaneous physical reward for each sample. ;

[0039] Step 5.4: Calculate the state value using the value network from step 2.4. ;

[0040] Step 5.5: Based on reward and state value, calculate the advantage function using the generalized advantage estimation formula. ;

[0041] Step 5.6, Strategy Update: Construct the PPO cropping objective function and maximize the function to update the parameters of the video generation decoder;

[0042] Step 5.7, Value Network Update: Calculate the error between the value network prediction and the actual reward, and update the value network parameters.

[0043] Step 6 specifically includes the following steps:

[0044] Step 6.1, Network State Switching: Freeze the video generation decoder, value network, and text encoder; Unfreeze the physical encoder;

[0045] Step 6.2, Feature-driven sampling: Extracting features using a physical encoder. The policy network frozen in step 6.1 is then used to generate another batch of video samples online. ;

[0046] Step 6.3, Physical Feedback Calculation: Repeat steps 5.3 and 5.5 to calculate the physical reward for the new sample. With advantage function ;

[0047] Step 6.4, Gradient Penetration Backpropagation: Calculate the function described in Step 5.6 based on the results of Step 6.3;

[0048] Step 6.5, Co-evolution: Update the physical encoder parameters using the gradients returned from Step 6.4.

[0049] Step 7 specifically includes the following steps:

[0050] Step 7.1, Receive user input and ;

[0051] Step 7.2: Extract physical features using the optimized physical encoder from step 6.5;

[0052] Step 7.3: Use the optimized generator decoder from step 5.6 to output a high-fidelity flow field video with physical conservation.

[0053] In step 3.2, the calculation formula is as follows:

[0054]

[0055] in This represents the mass conservation reward value calculated based on divergence. Represents the natural exponential function; This represents the divergence sensitivity coefficient, used to control the sensitivity of the reward function to divergence error; This indicates the total number of frames contained in the generated video sequence; Represents the pixel height of a video frame; Indicates the pixel width of a video frame image; Indicates the first In the frame at pixel coordinates The horizontal velocity component at that location; Indicates the first In the frame at pixel coordinates The vertical velocity component at that location; This represents the absolute value operation;

[0056] In step 3.3, the calculation formula is as follows:

[0057]

[0058] in This represents the flow structure consistency bonus value calculated based on the statistical characteristics of vorticity. Indicates the vorticity sensitivity coefficient; Denotes the L2 norm of a matrix; Indicates the generation of video number 1 The vorticity field matrix of the frame; Indicates the first The reference vorticity field matrix corresponding to the frame.

[0059] In step 5.6, the calculation formula is as follows:

[0060]

[0061] in Let represent the policy optimization objective function of the PPO algorithm, and the training objective is to maximize this value; This represents the set of network parameters for the current policy to be optimized. This represents the empirical expectation operator, used to calculate the mean of the sampled data within the current batch. Representing the probability ratio, defined as ; Indicates the new policy network in state Take action below The probability density of generating the current video frame; This indicates the state of the network before the old policy was updated. Take action below The probability density; This represents the advantage function value calculated in step 5.5, which is used to measure the superiority or inferiority of the current action relative to the average level. This represents a function that takes the minimum value. This represents a truncation function that takes the input value as input. Limited to the range Inside; This represents the pruning hyperparameter, used to limit the magnitude of policy updates and prevent drastic oscillations during training.

[0062] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0063] 1. Strong proactive exploration capability and higher physical upper limit: Compared with offline methods, the online reinforcement learning method of this invention allows the model to actively explore the solution space. Guided by physical rewards, the model can attempt to generate fine flow field structures with smaller physical errors that have not been seen in the supervised fine-tuning stage, thereby breaking through the quality upper limit of the training data.

[0064] 2. Improved robustness and stability: The introduction of a value network to predict the value of the current flow field state enables the model to not only focus on the physical error of the current frame, but also to predict the long-term physical compliance of the generation trend, which significantly improves the stability of the model when facing unseen Reynolds numbers or geometric boundaries.

[0065] 3. The encoder achieves true physical intelligence: Through gradient penetration in step 6, the physical encoder is no longer a static image feature extractor, but has evolved into an intelligent perception module that can understand "what features can lead to results that conform to physical equations", thus achieving deep collaboration between perception and generation. Attached Figure Description

[0066] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0067] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0068] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0069] In the description of this invention, it should be understood that the terms "upper," "lower," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a predetermined orientation, or be constructed and operated in a predetermined orientation. Therefore, they should not be construed as limitations on this invention.

[0070] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the quantity indicated. Therefore, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature.

[0071] Example 1

[0072] like Figure 1 As shown, this invention provides a technical solution: the method constructs a reinforcement learning architecture including a policy network (Actor) and a value network (Critic), encapsulates the physical equations into a digital environment, and adopts a training strategy of basic supervised fine-tuning, online physics exploration by the generator, and co-evolutionary learning by the encoder feedback; the specific implementation steps are as follows:

[0073] Step 1, Constructing a Multimodal Training Dataset for Unsteady Flow Fields: This step aims to obtain a standardized dataset containing the physical states of the flow field, descriptions of operating conditions, and true values ​​of dynamic evolution, providing a foundation for model initialization; it specifically includes the following steps:

[0074] Step 1.1: Use a high-precision CFD solver to perform unsteady solutions on the Navier-Stokes equations and obtain flow field simulation data under various geometric boundaries and Reynolds number conditions.

[0075] Step 1.2: Extract from the simulation data obtained in Step 1.1 The physical field data at time t is used as the initial flow field image. .

[0076] Step 1.3: Extract the continuous data after the initial moment from the simulation data obtained in Step 1.1. Frame physical field data as target video sequence .

[0077] Step 1.4: Generate the corresponding operating condition description text based on the simulation settings parameters in Step 1.1. .

[0078] Step 1.5, take the result obtained in step 1.2 The results obtained in step 1.4 and the results obtained in step 1.3 Combine and construct a triplet sample set .

[0079] Step 2, Constructing a Generative Network System Based on the Actor-Critic Architecture: This step constructs the policy network and value network required for reinforcement learning. The policy network is used to generate the flow field video, and the value network is used to evaluate the state value. Specifically, it includes the following steps:

[0080] Step 2.1, Construct a text encoder: Use a pre-trained language model encoder and configure it to receive the text described in step 1.4. Output semantic embedding .

[0081] Step 2.2, Construct a physical encoder: The image described in step 1.2... After segmentation and linear projection, global features are extracted using a Transformer encoder, outputting the physical embedding. .

[0082] Step 2.3, Constructing the video generation decoder: Using a 3D denoising network based on a diffusion model, the video generated in step 2.1... And obtained in step 2.2 Videos are generated under dual guidance. During the reinforcement learning phase, this network is defined as a policy network. .

[0083] Step 2.4, Construct the value network: Construct a regression network with a similar structure to the policy network described in Step 2.3, but with the output layer modified to have scalar output. This network is used to receive input status. It also predicts the expected physical reward value of the video generated in this state to help train stability.

[0084] Step 3, Constructing an interactive environment and reward function based on physical equations: This step constructs a digital environment that requires no human intervention. This environment performs real-time numerical scoring on the generated flow field video based on fluid dynamics equations. Specifically, it includes the following steps:

[0085] Step 3.1, Environment Interface Definition: Define the environment to receive the video frame sequence generated in Step 2.3 as actions, and invert it into a physical velocity field. .

[0086] Step 3.2, Calculate the divergence reward: based on the continuity equation for incompressible fluids. The full-field divergence error of the generated video is calculated and converted into a positive reward. The calculation formula is as follows:

[0087]

[0088] in This represents the mass conservation reward value calculated based on divergence. Represents the natural exponential function; This represents the divergence sensitivity coefficient, used to control the sensitivity of the reward function to divergence error; This indicates the total number of frames contained in the generated video sequence; Represents the pixel height of a video frame; Indicates the pixel width of a video frame image; Indicates the first In the frame at pixel coordinates The horizontal velocity component at that location; Indicates the first In the frame at pixel coordinates The vertical velocity component at that location; This represents absolute value operations.

[0089] Step 3.3, Calculate vorticity reward: Calculate the statistical deviation between the vorticity distribution of the generated flow field and the actual physical laws, and convert it into a positive reward. The calculation formula is as follows:

[0090]

[0091] in This represents the flow structure consistency bonus value calculated based on the statistical characteristics of vorticity. Indicates the vorticity sensitivity coefficient; Denotes the L2 norm of a matrix; Indicates the generation of video number 1 The vorticity field matrix of the frame; Indicates the first The reference vorticity field matrix corresponding to the frame.

[0092] Step 3.4, Synthesize Total Physical Reward: Define the total physical reward for a single generation. .in The weighting coefficient for divergence rewards is used to adjust the importance of the mass conservation constraint in the overall objective. The weighting coefficient representing the vorticity bonus is used to adjust the importance of flow field structural constraints.

[0093] Step 4, First Stage Supervised Fine-Tuning Training: This step initializes the policy network, giving it basic generative capabilities and avoiding the cold start difficulties of reinforcement learning. It specifically includes the following steps:

[0094] Step 4.1, take the sample from step 1.5 Input the network system constructed in step 2.

[0095] Step 4.2: Generate a predicted video using the policy network described in Step 2.3, and calculate its comparison with the ground truth video. The pixel mean square error loss and gradient loss.

[0096] Step 4.3: Update the parameters of the physical encoder and video generator decoder using the backpropagation algorithm.

[0097] Step 4.4, Value Network Warm-up: Generate a batch of data using the SFT-enhanced model, calculate its actual physical reward, and pre-train the value network described in Step 2.4 using mean squared error loss. .

[0098] Step 5, Online Physics Exploration of the Generator in the Second Stage: This step employs the Proximal Policy Optimization (PPO) algorithm, enabling the generator to actively explore the optimal solution space under the guidance of physics rewards. Specifically, it includes the following steps:

[0099] Step 5.1, Environment Initialization: Freeze the physical encoder and text encoder; unfreeze the video generation decoder (policy network) and value network.

[0100] Step 5.2, Online Sampling: Utilizing the current policy network For input Perform online inference to generate a batch of flow field video samples. And record the log probability distribution of the generation process.

[0101] Step 5.3, using the methods from step 3.4, calculate the instantaneous physical reward for each sample. .

[0102] Step 5.4: Calculate the state value using the value network from step 2.4. .

[0103] Step 5.5: Based on reward and state value, calculate the advantage function using the Generalized Advantage Estimation (GAE) formula. This value quantifies "how much better the currently generated video is than the average level".

[0104] Step 5.6, Strategy Update: Construct the PPO cropping objective function and maximize this function to update the parameters of the video generation decoder. The calculation formula is as follows:

[0105]

[0106] in Let represent the policy optimization objective function of the PPO algorithm, and the training objective is to maximize this value; This represents the set of network parameters for the current policy to be optimized. This represents the empirical expectation operator, used to calculate the mean of the sampled data within the current batch. Representing the probability ratio, defined as ; Indicates the new policy network in state Take action below The probability density of generating the current video frame; This indicates the state of the network before the old policy was updated. Take action below The probability density; This represents the advantage function value calculated in step 5.5, which is used to measure the superiority or inferiority of the current action relative to the average level. This represents a function that takes the minimum value. This represents a truncation function that takes the input value as input. Limited to the range Inside; This represents the pruning hyperparameter, used to limit the magnitude of policy updates and prevent drastic oscillations during training.

[0107] Step 5.7, Value Network Update: Calculate the error between the value network prediction and the actual reward, and update the value network parameters.

[0108] Step 6, Third-stage feedback co-evolution of the physical encoder: This step utilizes the gradient penetration mechanism of reinforcement learning to drive the evolution of the upstream physical encoder through downstream physical feedback. Specifically, it includes the following steps:

[0109] Step 6.1, Network State Switching: Freeze the video generation decoder (policy network), value network, and text encoder; unfreeze the physical encoder.

[0110] Step 6.2, Feature-driven sampling: Extracting features using a physical encoder. The policy network frozen in step 6.1 is then used to generate another batch of video samples online. .

[0111] Step 6.3, Physical Feedback Calculation: Repeat steps 5.3 and 5.5 to calculate the physical reward for the new sample. With advantage function .

[0112] Step 6.4, Gradient Penetration Backpropagation: Calculate the function described in Step 5.6 based on the results of Step 6.3.

[0113] Step 6.5, Co-evolution. The physical encoder parameters are updated using the gradients returned from Step 6.4. This step enables the physical encoder to learn to adaptively adjust its region of interest in the initial flow field image based on the "scoring" from the physical equations, thereby providing features that can induce the generation of higher physical quality videos.

[0114] Step 7, inference generation of unsteady flow field video; specifically including the following steps:

[0115] Step 7.1, Receive user input and .

[0116] Step 7.2: Extract physical features using the physical encoder optimized in step 6.5.

[0117] Step 7.3: Use the optimized generator decoder from step 5.6 to output a high-fidelity flow field video with physical conservation.

[0118] One of the key aspects of this solution is the adoption of a video generation architecture based on online physical feedback reinforcement learning, which includes a policy network and a value network. The policy network does not rely on a static dataset, but instead generates flow field video samples online and interacts directly with the digital environment based on the Navier-Stokes equations, using the PPO algorithm for closed-loop optimization.

[0119] Meanwhile, one of the key points of this scheme is the adoption of a real-time scalar reward mechanism based on fluid equations: a physical reward module that does not require manual annotation is constructed, which analyzes the generated flow field video in real time during the training process, calculates the whole field divergence and vorticity error, and transforms the physical residual into a positive scalar reward through an exponential decay function, which serves as the sole optimization objective of reinforcement learning.

[0120] Meanwhile, one of the key points of this scheme is to adopt an end-to-end co-evolutionary method for the physical encoder: when the parameters of the generator network are frozen, the physical advantage function is calculated by online sampling, and the policy gradient is backpropagated through the frozen generator network to the front-end physical encoder, so as to realize the encoder's adaptive feature learning based on downstream physical feedback.

[0121] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A flow field video generation method based on strategy value architecture and online physics exploration, characterized by: Includes the following steps: Step 1, construct an unsteady flow field multimodal training dataset: obtain a standardized dataset containing the physical state of the flow field, working condition description, and true values ​​of dynamic evolution; Step 2, Construct a generative network system based on the Actor-Critic architecture: Construct the policy network and value network required for reinforcement learning. The policy network is used to generate the flow field video, and the value network is used to evaluate the state value. Step 2 specifically includes the following steps: Step 2.1, Construct a text encoder: Use a pre-trained language model encoder and configure it to receive text. Output semantic embedding ; Step 2.2, Construct a physical encoder: Transform the image After segmentation and linear projection, global features are extracted using a Transformer encoder, outputting the physical embedding. ; Step 2.3, Constructing the video generation decoder: Using a 3D denoising network based on a diffusion model, the video generated in step 2.1... And obtained in step 2.2 Videos are generated under dual guidance; during the reinforcement learning phase, the network is defined as a policy network. ; Step 2.4, Construct the value network: Construct a regression network with a similar structure to the policy network described in Step 2.3, but with the output layer modified to have scalar output. This network is used to receive input status. It also predicts the expected physical reward value of the video generated in this state to help train stability; Step 3, Construct an interactive environment and reward function based on physical equations: Construct a digital environment that requires no human intervention, which performs real-time numerical scoring on the generated flow field video based on fluid dynamics equations; Step 4, the first stage of supervised fine-tuning training: Initialize the policy network to give it basic generative capabilities; Step 5: Online physics exploration of the second-stage generator: A near-end strategy optimization algorithm is adopted to enable the generator to actively explore the optimal solution space under the guidance of physics rewards; Step 5 specifically includes the following steps: Step 5.1, Environment Initialization: Freeze the physical encoder and text encoder; Unfreeze the video generation decoder and value network; Step 5.2, Online Sampling: Utilizing the current policy network For input Perform online inference to generate a batch of flow field video samples. And record the logarithmic probability distribution of the generation process; Step 5.3, calculate the instantaneous physical reward for each sample. ; Step 5.4: Calculate state value using a value network. ; Step 5.5: Based on reward and state value, calculate the advantage function using the generalized advantage estimation formula. ; Step 5.6, Strategy Update: Construct the PPO cropping objective function and maximize the function to update the parameters of the video generation decoder; Step 5.7, Value Network Update: Calculate the error between the value network prediction and the actual reward, and update the value network parameters; Step 6: Third-stage feedback co-evolution of the physical encoder: Utilizing the gradient penetration mechanism of reinforcement learning, the downstream physical feedback forces the upstream physical encoder to evolve. Step 7: Inference and generation of unsteady flow field video.

2. The flow field video generation method based on strategy value architecture and online physics exploration as described in claim 1, characterized in that: Step 1 specifically includes the following steps: Step 1.1: Use a high-precision CFD solver to perform unsteady solutions on the Navier-Stokes equations and obtain flow field simulation data under various geometric boundaries and Reynolds number conditions. Step 1.2: Extract from the simulation data obtained in Step 1.1 The physical field data at time t is used as the initial flow field image. ; Step 1.3: Extract the continuous data after the initial moment from the simulation data obtained in Step 1.

1. Frame physical field data as target video sequence ; Step 1.4: Generate the corresponding operating condition description text based on the simulation settings parameters in Step 1.

1. ; Step 1.5, take the result obtained in step 1.2 The result obtained in step 1.4 and the results obtained in step 1.3 Combine and construct a triplet sample set .

3. The flow field video generation method based on strategy value architecture and online physics exploration as described in claim 2, characterized in that: Step 3 specifically includes the following steps: Step 3.1, Environment Interface Definition: Define the environment to receive the video frame sequence generated in Step 2.3 as actions, and invert it into a physical velocity field. ; Step 3.2, Calculate the divergence reward: based on the continuity equation for incompressible fluids. Calculate the full-field divergence error of the generated video and convert it into a positive reward; Step 3.3, Calculate vorticity reward: Calculate the statistical deviation between the vorticity distribution of the generated flow field and the actual physical laws, and convert it into a positive reward; Step 3.4, Synthesize Total Physical Reward: Define the total physical reward for a single generation. ;in The weighting coefficient for divergence rewards is used to adjust the importance of the mass conservation constraint in the overall objective. The weighting coefficient representing the vorticity reward is used to adjust the importance of flow field structural constraints. This represents the mass conservation reward value calculated based on divergence. This represents the flow structure consistency bonus value calculated based on the statistical characteristics of vorticity.

4. The flow field video generation method based on strategy value architecture and online physics exploration as described in claim 3, characterized in that: Step 4 specifically includes the following steps: Step 4.1, take the sample from step 1.5 Input the network system constructed in step 2; Step 4.2: Generate a predicted video using the policy network described in Step 2.3, and calculate its comparison with the ground truth video. The pixel mean square error loss and gradient loss; Step 4.3: Update the parameters of the physical encoder and video generation decoder using the backpropagation algorithm; Step 4.4, Value Network Warm-up: Generate a batch of data using the SFT-enhanced model, calculate its actual physical reward, and pre-train the value network described in Step 2.4 using mean squared error loss. .

5. The flow field video generation method based on strategy value architecture and online physics exploration as described in claim 4, characterized in that: Step 6 specifically includes the following steps: Step 6.1, Network State Switching: Freeze the video generation decoder, value network, and text encoder; Unfreeze the physical encoder; Step 6.2, Feature-driven sampling: Extracting features using a physical encoder. The policy network frozen in step 6.1 is then used to generate another batch of video samples online. ; Step 6.3, Physical Feedback Calculation: Repeat steps 5.3 and 5.5 to calculate the physical reward for the new sample. With advantage function ; Step 6.4, Gradient Penetration Backpropagation: Calculate the function described in Step 5.6 based on the results of Step 6.3; Step 6.5, Co-evolution: Update the physical encoder parameters using the gradients returned from Step 6.

4.

6. The flow field video generation method based on strategy value architecture and online physics exploration as described in claim 5, characterized in that: Step 7 specifically includes the following steps: Step 7.1, Receive user input and ; Step 7.2: Extract physical features using the optimized physical encoder from step 6.5; Step 7.3: Use the optimized generator decoder from step 5.6 to output a high-fidelity flow field video with physical conservation.

7. The flow field video generation method based on strategy value architecture and online physics exploration as described in claim 5, characterized in that: In step 3.2, the calculation formula is as follows: in This represents the mass conservation reward value calculated based on divergence. Represents the natural exponential function; This represents the divergence sensitivity coefficient, used to control the sensitivity of the reward function to divergence error; This indicates the total number of frames contained in the generated video sequence; Represents the pixel height of a video frame; Indicates the pixel width of a video frame image; Indicates the first In the frame at pixel coordinates The horizontal velocity component at that location; Indicates the first In the frame at pixel coordinates The vertical velocity component at that location; This represents the absolute value operation; In step 3.3, the calculation formula is as follows: in This represents the flow structure consistency bonus value calculated based on the statistical characteristics of vorticity. Indicates the vorticity sensitivity coefficient; Denotes the L2 norm of a matrix; Indicates the generation of video number 1 The vorticity field matrix of the frame; Indicates the first The reference vorticity field matrix corresponding to the frame.

8. The flow field video generation method based on strategy value architecture and online physics exploration as described in claim 1, characterized in that: In step 5.6, the calculation formula is as follows: in Let represent the policy optimization objective function of the PPO algorithm, and the training objective is to maximize this value; This represents the set of network parameters for the current policy to be optimized. This represents the empirical expectation operator, used to calculate the mean of the sampled data within the current batch. Representing the probability ratio, defined as ; Indicates the new policy network in state Take action below The probability density of generating the current video frame; This indicates the state of the network before the old policy was updated. Take action below The probability density; This represents the advantage function value calculated in step 5.5, which is used to measure the superiority or inferiority of the current action relative to the average level. This represents a function that takes the minimum value. This represents a truncation function that takes the input value as input. Limited to the range Inside; This represents the pruning hyperparameter, used to limit the magnitude of policy updates and prevent drastic oscillations during training.

Images