Dynamic scene adaptive simulation generation method based on physical information neural network

Abstract

The application is suitable for the technical field of dynamic scene simulation and deep learning, and discloses a dynamic scene adaptive simulation generation method based on a physical information neural network, which reconstructs the spatio-temporal evolution of a dynamic scene into a continuous field learning task constrained by partial differential equations, constructs a generative simulation framework including an implicit neural representation module, a physical residual module and an adversarial discriminator module, adopts a three-stage progressive training strategy to complete parameter optimization, and realizes continuous adjustment of generation precision and calculation efficiency through an accuracy control scalar; the application can improve simulation generation efficiency while maintaining high physical consistency through the fusion of physical priori and generative priori, has certain generalization ability for unseen scene modes, and can be applied to the fields of automatic driving test, robot navigation simulation and the like.

Description

Technical Field

[0001] This invention belongs to the field of interdisciplinary technology of dynamic scene simulation and deep learning, and particularly relates to a dynamic scene adaptive simulation generation method based on physical information neural networks. Background Technology

[0002] In the field of dynamic scene simulation, existing technologies are mainly divided into two technical routes: physics engine-driven and data-driven. The invention patent with publication number CN119670820B, entitled "Generative Adversarial Network and Method for Generalizing Autonomous Driving Test Scenes," discloses a scene generalization method based on temporal generative adversarial networks. This method can generate diverse driving scene sequences, but it does not introduce physical constraints, and the generated results are prone to abnormal states that violate mechanical principles. Its generalization ability is weak in out-of-distribution scenes. Furthermore, this scheme relies on discrete grid temporal sampling, making it impossible to achieve arbitrary point query generation in continuous spatiotemporal domains.

[0003] Existing technologies generally suffer from the following four problems: Discretized simulation paradigms have computational efficiency bottlenecks: Traditional simulation discretizes time and space and solves physical equations through numerical iteration. High-frequency dynamic scenes require small solution step sizes, and computational overhead grows superlinearly with the number of entities, making it difficult to support the rapid generation of large-scale dynamic scenes. Although physics engine-type methods have high fidelity, the generation time per frame is long, which cannot meet the needs of rapid sampling of batch scenes, thus limiting the coverage scale and generation efficiency of simulation scenes. Pure data-driven methods suffer from a lack of physical consistency: Data-driven methods based on generative adversarial networks and diffusion models learn scene distribution directly from data without explicitly solving physical equations, resulting in high generation efficiency. However, they lack inductive bias based on physical priors, making it easy to generate results that do not conform to physical laws. Physical consistency decreases when no traffic flow pattern is observed, making it unable to support high-reliability simulation testing tasks. Generative networks suffer from insufficient spectral adaptability: Existing implicit neural representations mostly use sinusoidal activation networks with fixed frequency parameters, which cannot adaptively distinguish between static low-frequency components and dynamic high-frequency components in a scene. They have limited ability to characterize complex scenes including multi-scale dynamic entities, and it is difficult to simultaneously ensure background stability and motion detail accuracy. In scenes including entities with mixed fast and slow motion, detail blurring or artifacts are prone to occur. The lack of dynamic adjustment mechanism for accuracy and efficiency: Existing simulation methods usually adopt a fixed-precision solution configuration, which cannot dynamically balance fidelity and computation speed according to the resource conditions and task requirements of the application scenario. It is difficult to adapt to the needs of multiple scenarios from offline high-fidelity training to online real-time inference, and the same model cannot take into account the usage requirements of different hardware environments and task priorities.

[0004] Therefore, a dynamic scene adaptive simulation generation method based on physical information neural networks is needed to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide a dynamic scene adaptive simulation generation method based on physical information neural networks to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution: The dynamic scene adaptive simulation generation method based on physical information neural network includes the following steps: S1. Define the continuous field learning task under partial differential equation constraints for the spatiotemporal evolution task of dynamic scene, determine the scene state vector, flux tensor, diffusion coefficient, source term parameters, and configure spatial boundary conditions and initial time conditions. S2. Construct a generative simulation framework, which includes an implicit neural representation module, a physical residual module, and an adversarial discriminator module; the implicit neural representation module receives spatiotemporal coordinates, implicit coding, and precision control scalar inputs, and outputs single spatiotemporal point scene state data; S3. The parameters of all modules in the generative simulation framework are iteratively optimized through a three-stage progressive training process. S4. During the inference phase, the implicit neural representation module is invoked. By changing the precision control scalar value, the sequence of scene states within the specified spatiotemporal interval is output. This approach reconstructs dynamic scenes into continuous field learning tasks, constructs a generative framework comprising three major modules: implicit neural representation, physical residuals, and adversarial discrimination, and combines three-stage progressive training with adaptive precision adjustment to achieve physically consistent and efficiency-adjustable dynamic scene generation, alleviating the problem of traditional methods struggling to balance efficiency and fidelity.

[0007] A further technical solution is that the implicit neural representation module includes a SIREN network unit and a super network unit; the SIREN network unit is configured with six sinusoidal activation hidden layers and a single linear output layer, and the implicit code is connected to each layer of the SIREN network unit via a FiLM structure; the super network unit generates the corresponding level frequency parameters of the SIREN network unit channel by channel according to the input implicit code. This scheme uses the SIREN network as the backbone architecture of the implicit representation, and utilizes the high-order differentiability of the sinusoidal activation function to support the stable calculation of physical residuals. The implicit coding is injected into each layer of the network through the FiLM mechanism to realize conditional scene generation based on implicit coding and support fast sampling of different scene modes.

[0008] In a further technical solution, the supernetwork unit is a multilayer perceptron structure that matches the layers of the SIREN network unit. After receiving the hidden code, the supernetwork unit outputs the original parameters of the layer neuron frequency. The original parameters are then activated by Softplus to output positive frequency parameters. Different hidden code inputs correspond to different frequency distribution parameters output by the supernetwork unit. This solution uses a hypernetwork to dynamically generate the frequency parameters of neurons in each layer based on the hidden coding. Softplus activation ensures that the frequency is non-negative, enabling the network to adaptively adjust the spectral response according to the dynamic characteristics of the scene, adapting to entities with different movement speeds and scene regions with different rates of change.

[0009] In a further technical solution, the physical residual module includes an automatic differentiation unit and a loss calculation unit; the automatic differentiation unit solves the spatiotemporal partial derivatives of the output results of the implicit neural representation module, and substitutes them into a preset partial differential equation to obtain the physical residual; the loss calculation unit generates the physical loss term based on the weighted operation of the residual mean of the configuration point set, the initial condition constraint term, and the boundary condition constraint term. This scheme calculates the spatiotemporal partial derivatives of the scene state based on an automatic differentiation mechanism, substitutes them into the partial differential equation to obtain the physical residuals, and constructs a complete physical loss by combining initial conditions and boundary conditions. Physical laws are integrated into the training process in the form of soft constraints, and physical priors can be injected without an explicit numerical solver.

[0010] A further technical solution is that the adversarial discriminator module includes a feature fusion unit and a discriminant output unit; the feature fusion unit concatenates spatiotemporal coordinate parameters and scene state values ​​to form an input triplet; the discriminant output unit is a four-layer multilayer perceptron structure, which outputs the sample authenticity probability value; the adversarial loss is solved using the WGAN-GP calculation formula. The adversarial discriminator in this scheme simultaneously inputs spatiotemporal coordinates and state values, learns the spatiotemporal correlations implied by physical laws, distinguishes the physical rationality of states under different spatiotemporal locations, uses WGAN-GP loss to improve training stability, and further optimizes the distribution realism and perceived naturalness of the generated results.

[0011] A further technical solution involves a three-stage progressive training process executed by the training control module, with the following steps: Phase 1: Perform parameter training on the implicit neural representation module using only the physical loss term, initial condition loss term, and boundary condition loss term; Phase 2: Overlay adversarial loss terms and diversity loss terms, and jointly optimize the implicit neural representation module and the adversarial discriminator module; Phase 3: Adjust the sampling weight distribution of configuration points based on the spatial distribution results of physical residuals; This scheme adopts a three-stage progressive training strategy. First, it learns the basic physical manifold through physical loss. Then, it introduces adversarial loss to refine the distribution fitting. Finally, it optimizes the configuration point distribution through adaptive sampling, gradually improving the model performance and alleviating the mode collapse problem caused by direct joint training.

[0012] In a further technical solution, the implicit neural representation module also integrates a precision control unit; the precision control unit retrieves a precision control scalar and performs a centering scaling operation on the penultimate layer output features of the implicit neural representation module. The operation method is to subtract the mean of the training set features from the penultimate layer output, multiply it by the precision control scalar, and then add the mean of the training set features. This scheme achieves precision adjustment by scaling the penultimate layer feature output with precision control scalar, centering it based on the mean of the training set, and then scaling it again. It utilizes the long-tail characteristic of feature contributions to truncate low-value components, ensuring the controllability of precision degradation.

[0013] A further technical solution is to set the precision control scalar value range to 0 to 1; when the scalar value is 1, the precision control unit does not perform feature scaling; when the scalar value is 0, the implicit neural representation module outputs a fixed constant value result. The precision control scalar value range of this solution is from 0 to 1 and can be continuously and steplessly adjusted. Different values ​​correspond to different computational loads and generation precisions. Users can flexibly choose according to hardware resources and task requirements and switch between different simulation levels without retraining the model.

[0014] In a further technical solution, the training control module is equipped with a sampling weight calculation unit; the sampling weight calculation unit uses the Softmax weighting formula to calculate the sampling weight of the configuration point, the weight value is positively correlated with the physical residual exponential function, and the convergence range of the weight distribution is limited by the temperature parameter; In the third stage of training, this scheme adaptively adjusts the sampling weights of the configuration points based on the spatial distribution of the physical residuals. Regions with high residuals receive more sampling points, and the concentration of the weight distribution is controlled by temperature parameters, thereby tilting computational resources toward regions with larger errors and improving training efficiency and accuracy.

[0015] In a further technical solution, in step S1, the partial differential equation is built based on traffic flow theory, and the equation includes convection terms, diffusion terms, and source terms; the spatial boundary conditions include Dirichlet boundary constraints at scene entrances and exits and Neumann boundary constraints at road edges; the initial time value is determined by scene sampling data or preset layout parameters. The partial differential equations of this scheme are constructed based on traffic flow theory and include three types of components: convection, diffusion, and source terms. The boundary conditions adopt a combination of Dirichlet and Neumann methods. The initial values ​​can be derived from real data or preset configurations, which can flexibly adapt to the simulation needs of different types of intersection traffic scenarios.

[0016] Compared with the prior art, the beneficial effects of the present invention are: This invention can improve the generation efficiency of large-scale dynamic scenes: it reconstructs scene evolution into a continuous field learning task, directly learns the parameterized representation of low-dimensional physical manifolds through implicit neural representation modules, bypasses the point-by-point iterative process of traditional numerical solutions, and the state query of a single spatiotemporal point only requires one network forward computation. It also supports fully parallel batch processing, which can reduce a certain proportion of computational overhead compared with traditional physics engines. It is suitable for batch generation of scenes of a certain scale and can support rapid sampling and testing of multiple scenes. This invention helps to enhance the physical consistency and generalization ability of the generated results: by transforming partial differential equation constraints into soft loss and injecting them into the training process through the physical residual module, it provides the generator network with physical prior inductive bias. At the same time, it combines the adversarial discriminator module to improve the data distribution fit. The introduction of physical constraints enables the model to maintain a high level of physical rationality even in traffic flow patterns that have not been observed. Compared with pure data-driven methods, it can reduce the performance degradation in out-of-distribution scenarios and support more reliable simulation tests. This invention achieves adaptive spectrum adaptation for multi-scale scenes: by generating channel-by-channel learnable frequency parameters through hypernetwork units, the frequency response of each channel is dynamically adjusted according to the input implicit coding. The weight of low-frequency channels corresponding to static backgrounds is increased, and the weight of high-frequency channels corresponding to fast-moving entities is increased. This mechanism realizes conditional frequency band allocation at the function space level, which can simultaneously take into account the stability of static regions and the ability to depict details in dynamic regions, reducing detail loss and artifacts in multi-scale scenes. This invention provides continuously adjustable accuracy and efficiency: by scaling the hidden layer output of the implicit neural representation module through an accuracy control unit and an accuracy control scalar, low-contribution components are safely truncated using the long-tail distribution characteristics of network feature contributions. The accuracy control scalar can be continuously adjusted in the range of 0 to 1, achieving a trade-off between different fidelity and speed without retraining the network. It is suitable for various application scenarios such as offline high-fidelity simulation and online real-time generation, improving the environmental adaptability of the model.

[0017] To more clearly illustrate the structural features and effects of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0020] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0021] Example 1 like Figure 1 As shown, this embodiment of the invention provides a dynamic scene adaptive generative simulation system based on a physical information neural network, including an implicit neural representation module, a physical residual module, an adversarial discriminator module, and a training control module.

[0022] The implicit neural representation module includes a SIREN network unit, a super network unit, and a precision control unit. The SIREN network unit is a network structure with six sinusoidal activation hidden layers and one linear output layer. The input is the spatiotemporal coordinate (x,t) and the implicit code z, and the output is the scene state vector u(x,t) at the corresponding spatiotemporal point. The super network unit is a small multilayer perceptron corresponding to each hidden layer. After receiving the implicit code z, it outputs the frequency parameters and FiLM modulation parameters of all neurons in that layer. The precision control unit is located between the penultimate layer and the output layer and is used to control the scaling of the feature output by a scalar α according to the precision.

[0023] The physical residual module includes an automatic differentiation unit and a loss calculation unit. The automatic differentiation unit calculates the spatiotemporal partial derivatives of the output of the SIREN network unit based on the automatic differentiation mechanism of the deep learning framework. The loss calculation unit substitutes the partial derivatives into the preset partial differential equation to obtain the physical residuals, and calculates the total physical loss by combining the initial condition loss and the boundary condition loss.

[0024] The adversarial discriminator module includes a feature fusion unit and a discriminant output unit. The feature fusion unit concatenates the spatiotemporal coordinates with the corresponding state values ​​to form an input triple. The discriminant output unit is a four-layer multilayer perceptron that uses the LeakyReLU activation function to output the authenticity probability of the corresponding state.

[0025] The training control module includes a phase switching unit and a sampling weight calculation unit; the phase switching unit is used to control the switching of training phases and the adjustment of loss weights; the sampling weight calculation unit is used to calculate the sampling weights of the configuration points based on the spatial distribution of the physical residuals.

[0026] The core physical residual calculation formula in this embodiment is as follows:

[0027] in: For spacetime points The scene state vector, including traffic flow density, velocity field, position field, and orientation field, is calculated and output by the SIREN network unit. The first-order partial derivative of the state vector with respect to time is calculated by the automatic differentiation unit. Gradient tracking is enabled in the time dimension during the calculation, and the solution is obtained through a single backpropagation. For the flux tensor, it describes the spatial transport of state variables. Its form is determined based on traffic flow theory and satisfies the convective transport relationship. Its functional form is preset based on domain priors before training. The flux divergence is obtained by calculating the spatial partial derivatives using the automatic differentiating unit, reflecting the rate of change of the state quantity in the spatial dimension. The diffusion coefficient characterizes the effective diffusion caused by the randomness of driving behavior. It is calibrated through a data-driven approach and optimized synchronously with the network parameters during training. The Laplace operator is obtained by calculating the second-order spatial partial derivatives from the automatic differentiating unit, and it characterizes the spatial diffusion effect of the state variables. The source terms include control inputs such as vehicle acceleration / deceleration and steering, as well as external constraints such as lane boundaries and traffic signals, which are preset according to scenario rules. The physical residual is the value of the physical residual. The smaller the residual value, the better the fit between the state at that point and the physical equation.

[0028] The formula as a whole is used to quantify the degree of deviation between the generated state and the physical equation, serving as the basis for calculating the physical loss. It transforms physical laws into training signals that can be backpropagated, thereby injecting physical priors into the generator network.

[0029] The working principle of this embodiment is as follows: During the training phase, the phase switching unit first starts the first phase of training, using only physical loss, initial condition loss, and boundary condition loss to train the implicit neural representation module, enabling the network to learn the basic evolutionary manifold that conforms to physical laws. After completing the first phase of training, it switches to the second phase, introducing an adversarial discriminator module and using a weighted joint optimization of physical loss, adversarial loss, and diversity loss. After the second phase is completed, it switches to the third phase, where the sampling weight calculation unit dynamically adjusts the sampling weights of the configuration points according to the physical residual distribution, achieving adaptive allocation of computing resources.

[0030] During the inference phase, only the implicit neural representation module is retained. The user inputs the spatiotemporal coordinate range, implicit encoding, and precision control scalar. The network forward computes and outputs the scene state sequence within the target spatiotemporal range.

[0031] This embodiment can be applied to two typical scenarios: one is a complex intersection simulation scenario for autonomous driving, which is used to generate intersection interaction scenarios with different traffic flow densities and different driving styles to support batch testing of autonomous driving algorithms; the other is an indoor navigation simulation scenario for robots, which is used to generate indoor environment evolution sequences including dynamic obstacles to support the training and verification of robot navigation algorithms.

[0032] In this embodiment, four modules work collaboratively: the implicit neural representation module is responsible for backbone generation; the physical residual module injects physical priors to ensure rationality; the adversarial discriminator module optimizes distribution realism; and the training control module ensures training stability and efficiency. Three core technologies advance layer by layer: the first layer, continuous field reconstruction and implicit representation, lays the foundation for efficient generation; the second layer, physical constraints and adversarial refinement, improve fidelity and generalization; and the third layer, adaptive sampling and precision adjustment, achieves dynamic controllability of efficiency and precision. These three elements work together to form a complete technical closed loop of "efficient generation - physical fidelity - dynamic adjustability."

[0033] Example 2 The difference between this embodiment and Embodiment 1 is that it further defines the specific structure of the supernetwork unit and the control logic for frequency adjustment.

[0034] In this embodiment, the hypernetwork unit is a two-layer multilayer perceptron with a hidden layer dimension of 64. The input is a 128-dimensional hidden code, and the output dimension is consistent with the number of neurons in the corresponding SIREN layer. The hypernetwork unit simultaneously outputs the scaling parameter γ and offset parameter β required by the FiLM mechanism, sharing the same set of hidden layers as the frequency parameters, with only the output layer being separate. After the frequency parameters are output, they are activated by Softplus to ensure that all frequency values ​​are positive, avoiding numerical instability caused by negative frequencies.

[0035] The channel-by-channel frequency adjustment formula corresponding to this embodiment is as follows:

[0036] in: For the first Layer Frequency parameters of each neuron; For the first The hypernetwork unit corresponding to the layer has implicit coding as its input. The output is the original frequency values ​​of all neurons in this layer; As an activation function, ensure that the frequency parameter is positive and constrain the frequency range to positive real numbers to avoid numerical oscillations; The latent coding vector has a dimension of 128 and is obtained by random sampling or encoding of real scene to represent different scene patterns. For the first Layer The activation output of each neuron; , The first Layer The weights and biases of each neuron are the trainable parameters of the network, which are updated synchronously during training. For the first The output features of the current layer are used as the input of the current layer.

[0037] During training, the supernetwork unit and the SIREN network unit jointly optimize the process. When the input hidden code corresponds to a high-speed motion scene, the weights of the high-frequency channels in the supernetwork output are increased, improving the ability to capture details of fast motion. When the input hidden code corresponds to a static scene, the weights of the low-frequency channels are dominant, ensuring the smoothness and stability of the background region. This mechanism does not require manual setting of frequency parameters and can automatically adapt to the frequency band requirements of different scenes through data-driven methods, improving the reconstruction accuracy of dynamic regions compared to the fixed-frequency SIREN network.

[0038] Example 3 The difference between this embodiment and Embodiment 1 is that it further defines the automatic differentiation implementation method and configuration point sampling rules of the physical residual module.

[0039] In this embodiment, the automatic differentiation unit is implemented based on the native automatic differentiation mechanism of the deep learning framework. By setting gradient tracking for the input spatiotemporal coordinates, the first-order spatiotemporal partial derivative can be calculated through backpropagation after a single forward propagation, and the second-order partial derivative is calculated through a second backpropagation. The configuration points are generated by a combination of random sampling and boundary fixed-point sampling. Internal configuration points are randomly sampled within the spatial domain, and boundary configuration points are sampled separately at the boundaries. Initial configuration points are sampled separately at the initial time step.

[0040] The loss calculation unit calculates the PDE residual loss of internal placement points, the boundary condition loss of boundary placement points, and the initial condition loss at the initial time, respectively. These three losses are weighted according to preset weights to obtain the total physical loss. The boundary condition loss distinguishes between Dirichlet boundaries and Neumann boundaries; Dirichlet boundaries directly constrain the state values, while Neumann boundaries constrain the normal derivative to zero. This design can flexibly adapt to scene boundaries of different shapes, ensuring the integrity of physical constraints.

[0041] In this embodiment, the physical loss is calculated using the mean squared error method, averaging the squared residuals of all placement points to ensure the stability of the loss gradient. The initial and boundary conditions also use mean squared error loss, which is superimposed on the PDE residual loss with preset weights to guide the network in learning a solution space that conforms to physical constraints. During training, placement points are resampled in each training round to ensure the network encounters constraints at different locations within the spatial domain, improving the global consistency of the solution.

[0042] Example 4 The difference between this embodiment and Embodiment 1 is that the spatiotemporal joint input structure and discrimination logic of the adversarial discriminator module are further defined.

[0043] In this embodiment, the feature fusion unit directly concatenates the spatiotemporal coordinates and the state vector along the channel dimension to form a joint input vector of "coordinates + state," which is then input to the discriminant output unit. The discriminant output unit is a four-layer fully connected network with 128 neurons per layer, employing the LeakyReLU activation function. The last layer outputs a single scalar, representing the authenticity score of the input sample.

[0044] Compared to traditional discriminators that only input state values, spatiotemporal joint input enables the discriminator to learn the spatiotemporal correlations implicit in physical laws. The same state value will receive different plausibility scores in different spatiotemporal locations; for example, a high-speed state appearing at the center of an intersection has different physical plausibility than appearing at the edge of a road. The adversarial loss adopts the WGAN-GP form, incorporating a gradient penalty term to constrain the discriminator's Lipschitz continuity, improving the stability of the training process and reducing the risk of pattern collapse.

[0045] In this embodiment, the discriminator and generator are trained alternately. The discriminator is trained for a fixed number of steps before the generator is trained for one step, ensuring a balanced training pace and preventing the generator's gradient from vanishing due to excessive discriminator performance. Realistic samples are obtained by sampling the state values ​​of corresponding spatiotemporal points from real-world scene datasets, while generated samples are obtained by forward computation from the implicit neural representation module. Both are then mixed and input into the discriminator for discrimination training.

[0046] Example 5 The difference between this embodiment and Embodiment 1 is that it further limits the weight parameter settings and stage switching logic of the three-stage progressive training.

[0047] In this embodiment, the training rounds in Phase 1 are 5000 rounds, and the loss weights are set to physical loss weight 10.0, initial condition loss weight 1.0, and boundary condition loss weight 1.0. In this phase, the adversarial discriminator module is not enabled, and only the implicit neural representation module is trained, so that the network first learns the basic solution manifold that conforms to the physical laws.

[0048] Phase 2 consists of 10,000 training rounds. In addition to the loss from Phase 1, adversarial loss and diversity loss are added with weights of 0.1 and 0.01, respectively. During this phase, the implicit neural representation module and the adversarial discriminator module are trained alternately. For every 5 steps the discriminator trains, the generator trains for 1 step to ensure a balanced training rhythm between the two.

[0049] Phase 3 consists of 5000 training epochs, with the loss weights remaining consistent with Phase 2. The configuration point sampling weights are updated every 100 epochs, adjusting the sampling density based on the spatial distribution of the current physical residuals. The phase switching unit automatically switches training phases according to a preset number of epochs, requiring no manual intervention and ensuring the automation and stability of the training process.

[0050] In this embodiment, the three-stage training gradually improves the model performance from easy to difficult. The first stage lays the physical foundation, the second stage optimizes the distribution fitting, and the third stage refines and improves the accuracy. The step-by-step training method can alleviate the training instability and pattern collapse problems caused by direct joint training, and improve the generation quality and physical consistency of the final model.

[0051] Example 6 The difference between this embodiment and Embodiment 1 is that the residual weighting mechanism of adaptive sampling and the temperature parameter adjustment logic are further defined.

[0052] In this embodiment, the sampling weight calculation unit uses the Softmax weighting mechanism to calculate the sampling weight of the configuration points. The calculation formula is as follows:

[0053] in For the first The sampling weights of each configuration point This represents the absolute value of the physical residual at that point. This is a temperature parameter, with a value of 2.0. This represents the total number of configuration points.

[0054] The temperature parameter β controls the sharpness of the weight distribution. A larger β results in higher weights for regions with high residuals, leading to more concentrated sampling; a smaller β results in a more uniform weight distribution. During training, the temperature parameter can be dynamically adjusted based on the residual distribution. When the overall residual is large, the temperature can be appropriately lowered to avoid a decrease in generalization caused by overly concentrated sampling. This mechanism directs computational resources towards regions with larger physical errors, achieving higher solution accuracy with the same computational load.

[0055] In this embodiment, the configuration point sampling adopts a weighted random sampling method. Training samples are drawn from the candidate point set according to the calculated weights, and regions with high residuals have a higher probability of being selected. Before each update of the sampling weights, a batch of candidate points is uniformly sampled in the entire spatial domain to calculate the residuals, and then the weights are calculated based on the residuals. In subsequent training cycles, configuration points are sampled according to these weights, and the results are periodically updated to adapt to changes in the residual distribution during training.

[0056] Example 7 The difference between this embodiment and Embodiment 1 is that it further defines the feature scaling implementation and efficiency trade-off logic of the precision control unit.

[0057] In this embodiment, the precision control unit is embedded between the penultimate layer and the output layer of the SIREN network unit. During the training phase, the precision control scalar α is fixed at 1 to ensure complete training precision; during the inference phase, the value of α can be freely adjusted. After training is completed, the mean μ_h of the features of the penultimate layer of all samples in the training set is calculated and saved. During inference, this mean is directly called for scaling calculation.

[0058] The precision control scaling formula corresponding to this embodiment is as follows:

[0059] in: The scaled second-to-last layer features are input to the output layer for final state calculation. The original penultimate layer output features are obtained from the sixth layer output of the SIREN network unit; It is the mean of the second-to-last layer features on the training set. It is fixed and saved after training and can be directly called during inference without real-time calculation. For precise control of scalars, the value range is: It can be set by the user according to their needs; when When the feature is not scaled, the output is of full precision; when When the feature degenerates into the mean, the output is a constant function.

[0060] This formula leverages the long-tailed distribution of network feature contributions, achieving a continuous trade-off between accuracy and efficiency by scaling low-contribution components without modifying the network structure and parameters. When α is 0.5, some low-contribution feature components are truncated, improving inference speed but slightly reducing physical consistency. When α is 0.3, inference speed is further improved, but the decrease in physical consistency is more significant. Users can choose an appropriate α value based on task requirements; for example, higher α values ​​can be used for offline batch testing to ensure accuracy, while lower α values ​​can be used for online real-time simulation to ensure speed.

[0061] Example 8 The difference between this embodiment and Embodiment 1 is that it provides a specific application example of autonomous driving in complex intersection scenarios.

[0062] In this embodiment, the simulation object is a four-road intersection in an urban area, with the spatial scope being road sections of predetermined lengths in both the east-west and north-south directions, and the time range being 10 seconds. The state vector includes four dimensions: traffic flow density, velocity field, vehicle position, and orientation. The partial differential equations are based on an extension of the LWR traffic flow model, incorporating diffusion and source terms to describe the randomness of driving behavior and the influence of signal control.

[0063] The training data consisted of 160 intersection scenes from a publicly available motion dataset, while the test set included 40 unseen intersection scenes. After training, the single-frame generation time was low and the physical consistency was high in full-precision mode. When the precision was adjusted to 0.3, the single-frame generation time further decreased, while the physical consistency remained within a reasonable range. Even in out-of-distribution scenes with a 50% increase in density, the physical consistency remained high, outperforming purely data-driven diffusion model methods. This embodiment verifies the effectiveness and practicality of the proposed method in real-world autonomous driving scenarios.

[0064] In this embodiment, the boundary conditions are set as follows: Dirichlet boundaries are used for scene entrances and exits; traffic flow states are input statistically; Neumann boundaries are used for road edges; and zero-flux conditions are set to conform to the traffic flow constraints of real intersections. The initial state is extracted from the initial frame of the real scene to ensure the initial rationality of the generated sequence. The generated result can output the state value at any spatiotemporal point, or it can generate a scene sequence by sampling at a fixed frame rate to adapt to different downstream usage requirements.

[0065] The above description is merely 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 dynamic scene adaptive simulation generation method based on physical information neural networks, characterized in that, Includes the following steps: S1. Define the continuous field learning task under partial differential equation constraints for the spatiotemporal evolution task of dynamic scene, determine the scene state vector, flux tensor, diffusion coefficient, source term parameters, and configure spatial boundary conditions and initial time conditions. S2. Construct a generative simulation framework, which includes an implicit neural representation module, a physical residual module, and an adversarial discriminator module; the implicit neural representation module receives spatiotemporal coordinates, implicit coding, and precision control scalar inputs, and outputs single spatiotemporal point scene state data; S3. The parameters of all modules in the generative simulation framework are iteratively optimized through a three-stage progressive training process. S4. During the inference phase, the implicit neural representation module is invoked. By changing the precision control scalar value, the sequence of scene states within the specified spatiotemporal interval is output.

2. The dynamic scene adaptive simulation generation method based on physical information neural network according to claim 1, characterized in that, The implicit neural representation module includes a SIREN network unit and a super network unit. The SIREN network unit is configured with six sinusoidal activation hidden layers and a single linear output layer. The implicit code is connected to each layer of the SIREN network unit via a FiLM structure. The super network unit generates the corresponding level frequency parameters of the SIREN network unit channel by channel based on the input implicit code.

3. The dynamic scene adaptive simulation generation method based on physical information neural network according to claim 2, characterized in that, The supernetwork unit is a multilayer perceptron structure that matches the layers of the SIREN network unit. After receiving the hidden code, the supernetwork unit outputs the original parameters of the layer neuron frequency. The original parameters are then activated by Softplus to output positive frequency parameters. Different hidden code inputs correspond to different frequency distribution parameters output by the supernetwork unit.

4. The dynamic scene adaptive simulation generation method based on physical information neural network according to claim 1, characterized in that, The physical residual module includes an automatic differentiation unit and a loss calculation unit. The automatic differentiation unit solves the spatiotemporal partial derivatives of the output results of the implicit neural representation module and substitutes them into a preset partial differential equation to obtain the physical residual. The loss calculation unit generates the physical loss term by weighting the residual mean of the configuration point set, the initial condition constraint term, and the boundary condition constraint term.

5. The dynamic scene adaptive simulation generation method based on physical information neural network according to claim 1, characterized in that, The adversarial discriminator module includes a feature fusion unit and a discriminant output unit; The feature fusion unit concatenates spatiotemporal coordinate parameters and scene state values ​​to form an input triple; the discriminant output unit is a four-layer multilayer perceptron structure that outputs the probability value of sample authenticity; the adversarial loss is solved using the WGAN-GP calculation formula.

6. The dynamic scene adaptive simulation generation method based on physical information neural network according to claim 1, characterized in that, The training control module executes a three-stage progressive training process, which proceeds as follows: Phase 1: Perform parameter training on the implicit neural representation module using only the physical loss term, initial condition loss term, and boundary condition loss term; Phase 2: Overlay adversarial loss terms and diversity loss terms, and jointly optimize the implicit neural representation module and the adversarial discriminator module; Phase 3: Adjust the sampling weight distribution of configuration points based on the results of the physical residual spatial distribution.

7. The dynamic scene adaptive simulation generation method based on physical information neural network according to claim 1, characterized in that, The implicit neural representation module also integrates a precision control unit; the precision control unit retrieves a precision control scalar and performs a centering scaling operation on the penultimate layer output features of the implicit neural representation module. The operation method is to subtract the mean of the training set features from the penultimate layer output, multiply it by the precision control scalar, and then add the mean of the training set features.

8. The dynamic scene adaptive simulation generation method based on physical information neural network according to claim 7, characterized in that, The precision control scalar value range is set to 0 to 1; when the scalar value is 1, the precision control unit does not perform feature scaling. When the scalar value is 0, the implicit neural representation module outputs a fixed constant result.

9. The dynamic scene adaptive simulation generation method based on physical information neural network according to claim 6, characterized in that, The training control module is equipped with a sampling weight calculation unit; the sampling weight calculation unit uses the Softmax weighted formula to calculate the sampling weight of the configuration point. The weight value is positively correlated with the physical residual exponential function, and the convergence range of the weight distribution is limited by the temperature parameter.

10. The dynamic scene adaptive simulation generation method based on physical information neural network according to claim 1, characterized in that, In step S1, the partial differential equations are built based on traffic flow theory, and the equations include convection terms, diffusion terms, and source terms; the spatial boundary conditions include Dirichlet boundary constraints at scene entrances and exits and Neumann boundary constraints at road edges; the initial time values ​​are determined by scene sampling data or preset layout parameters.

Images