System simulation modeling method based on combination of neural potential field model and physical constraint

Abstract

The invention provides a system simulation modeling method based on a neural potential field model in combination with physical constraints, and belongs to the technical field of system simulation. The method for constructing a structured neural potential field model based on a deep neural network in combination with a potential field method comprises the following steps: mapping a low-dimensional input coordinate of a space-time coordinate of a query point to a high-dimensional feature space through a position coding technology to obtain a coordinate coding vector of the query point; performing local feature extraction on the coordinate coding vectors of the query points to generate local features of the query points, and further performing global information aggregation to obtain deep fusion features; mapping the deep fusion features into basic potential codes representing different attributes of the to-be-simulated system; splicing the basic potential codes of different attributes to obtain structured potential codes; using the trained structured neural potential field model to receive data of a system to be simulated, and outputting structured potential codes of the system to be simulated to realize system simulation. A single and continuous parameterization function is adopted to represent elements in the system, so that the fidelity of system simulation is improved.

Description

Technical Field

[0001] The present invention relates to the field of computer simulation technology, and in particular to a system simulation modeling method based on a neural potential field model combined with physical constraints. Background Art

[0002] In the production scheduling and equipment coordination of smart manufacturing workshops, it is often necessary to accurately simulate the thermal effects of processing equipment under long-term high-load operation (thermals), the mechanical structure deformation and material fatigue caused by thermal stress (mechanics), and the reverse impact of these deformations on process flow, production line rhythm and product accuracy (production process and logistics dynamics). Strong coupling effect.

[0003] Currently, system simulation relies primarily on a combination of specialized simulation technologies, including geometric and environmental modeling solutions. Solid geometry and environmental context are typically constructed based on computer-aided design models, geographic information system (GIS) data, or remote sensing data. Physical process modeling solutions, such as mechanics, thermals, fluid dynamics, and electromagnetics, typically simulate the physical processes involved using independent specialized solvers. For example, finite element analysis (FEA) software is used to simulate structural stress, computational fluid dynamics (CFD) software is used to simulate air or liquid flow, and multibody dynamics engines are used to simulate mechanical system motion. Independent simulation tools are connected through a high-level architecture. However, in actual production, simulators for different physical domains, such as FEA and CFD, often use mathematical models and discretization methods with vastly different principles. Collaborative simulation technology only exchanges discrete data at the macro level, which can lead to significant numerical errors, oscillations, and delays at interfaces, making it difficult to accurately simulate strong cross-domain coupling effects. While purely data-driven machine learning models can learn and generate complex behavioral patterns, their "black box" nature means their outputs may violate fundamental physical laws or domain constraints, rendering simulation results unreliable and unsuitable for analysis, planning, and decision support scenarios requiring high credibility. This creates a significant "credibility gap." In summary, current system simulation methods suffer from insufficient inter-model collaboration and physical constraints in their modeling processes, leading to insufficient simulation accuracy, poor adaptability, and insufficient physical credibility.

[0004] Therefore, a high-fidelity and highly adaptable system simulation modeling method is needed. Summary of the Invention

[0005] In view of this, the present invention provides a system simulation modeling method based on a neural potential field model combined with physical constraints. It uses a single, continuous parameterized function to characterize the geometric, appearance, physical and behavioral properties of each element in the system, replacing the traditional federal architecture composed of multiple discrete models to meet the high-fidelity requirements of simulation.

[0006] To this end, the present invention provides the following technical solutions: A system simulation modeling method based on a neural potential field model combined with physical constraints, including: Construct a structured neural potential field model based on deep neural network combined with potential field method; The structured neural potential field model includes: The low-dimensional input coordinates of the space-time coordinates of the query point are mapped to the high-dimensional feature space through the position encoding technology to obtain the coordinate encoding vector of the query point; Performing local feature extraction on the coordinate encoding vector of the query point to generate a local feature of the query point; Perform global information aggregation on the local features of each query point to obtain deep fusion features; Mapping the deep fusion features into basic potential codes that characterize different properties of the system to be simulated; splicing basic potential codes of different attributes to obtain structured potential codes; the attributes include: geometry, appearance, physics and behavior; Training the structured neural potential field model; using the trained structured neural potential field model to receive data of the system to be simulated, and outputting a structured potential code of the system to be simulated; The system simulation is realized by utilizing the structured potential coding of the system to be simulated.

[0007] Furthermore, training the structured neural potential field model includes: Using an information maximization generative adversarial network to perform self-supervised learning training on the structured neural potential field model; The structured neural potential field model is trained with physical laws as constraints and structured true value data.

[0008] Furthermore, it also includes determining an element configuration scheme based on the structured potential coding: Constructing a reinforcement learning framework based on structured potential coding; Using the optimization objective as a reward function, an optimized structured potential encoding is obtained through the reinforcement learning framework; The optimized structured potential coding is used as the element configuration scheme of the simulation system.

[0009] Furthermore, the self-supervised learning training of the structured neural potential field model using an information maximization generative adversarial network includes: An information maximization generative adversarial network is used to maximize the mutual information between each basic potential code and the corresponding attribute. The loss function of the information maximization generative adversarial network is:

[0010] in, is the standard adversarial loss; is the mutual information term, It is the structured potential code to be decoupled; the decoder generates simulation element samples , is a preset weight coefficient greater than zero, used to adjust the strength of the mutual information regularization term.

[0011] Furthermore, the training of the structured neural potential field model based on physical laws and structured true value data includes: Through physical information loss, physical laws are used as hard constraints, and a composite loss function is constructed by combining structured ground truth data supervision loss terms. Training the structured neural potential field model with minimizing the composite loss function as the optimization goal; The composite loss function:

[0012] in, is the physical information loss, is the supervision loss term for structured ground truth data; Represents the weights and sums of the structured truth data supervision loss terms Represents the physical information loss weight, which are all positive hyperparameters.

[0013] Furthermore, the reinforcement learning framework includes: Using structured neural potential field models as a strategy; The potential generated by the structured neural potential field model is used to encode the action; The current moment, coordinates, and macro-context are structured representations as the state of the environment.

[0014] Furthermore, mapping the low-dimensional input coordinates of the space-time coordinates of the query point to the high-dimensional feature space by using the position encoding technology to obtain the coordinate encoding vector of the query point includes:

[0015]

[0016] in, and Apply by element; is the space-time coordinate, are the three-dimensional space coordinates, It's time, is a random Gaussian matrix, is the number of Fourier features.

[0017] Furthermore, performing local feature extraction on the coordinate encoding vector of the query point to generate local features of the query point includes: Coordinate Encoded Vector and macro context vector , extract local features through multi-layer perceptron; The forward propagation process of each layer of the multilayer perceptron is expressed as:

[0018] in, It is Hidden representation of the layer, It is Hidden representation of the layer; is the weight matrix, is the bias vector, is the total number of layers of the multilayer perceptron; Is the activation function; the output is the local feature , is the dimension of local features.

[0019] Furthermore, performing global information aggregation on the local features of each query point to obtain deep fusion features includes: The local features of all query points of the system to be simulated are used as the initial local feature matrix of the system to be simulated; The initial local feature matrix is ​​hierarchically fused through several global dependency aggregation modules to obtain deep fusion features, including:

[0020] in, , m is the index of the global dependency aggregation module; The global information aggregation feature matrix output by the previous global dependency aggregation module; Indicates the A global dependency aggregation module, For the The global information aggregation feature matrix output by the global dependency aggregation module; The global information aggregation feature matrix output by the last global dependency aggregation module is used as the deep fusion feature.

[0021] Furthermore, the data processing process of the global dependency aggregation module includes:

[0022]

[0023] in, Represents the attention feature, MHSA represents the multi-head attention mechanism, FFN represents the feedforward network, represents a normalization layer.

[0024] Advantages and positive effects of the present invention: This method uses a structured neural potential field model to simulate the system: a single and continuous parameterized function is used to represent the geometric, physical and behavioral properties of each element in the system to output a semantically decoupled structured potential code, solving the problem of insufficient system simulation fidelity caused by the existing system simulation through the combination of multiple discrete models.

[0025] 1) The structured neural potential field model of this method encodes three-dimensional spatial coordinates through Fourier coding and periodic activation functions to ensure the ability to fit continuous fields; the non-local interactions between points in the field are captured through the self-attention mechanism; the hierarchical structure and independent mapping heads ensure the structured decoupling of the model's expressiveness and output; and the physical laws, operating specifications, and safety regulations are integrated into the structured neural potential field model through the physical information loss term to ensure the physical consistency of the system simulation.

[0026] 2) This method uses a reinforcement learning framework: structured potential coding based on system simulation obtains the element configuration scheme of the preset target through a preset reward function. For example, it generates the structure, material, behavior and other element configuration schemes that meet specific engineering goals, realizing the functional expansion from "forward simulation analysis" to "reverse optimization design". BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0028] Figure 1 This is a flow chart of the system simulation modeling method based on the neural potential field model combined with physical constraints; Figure 2 This is the structural diagram of the structured neural potential field model; Figure 3 Schematic diagram for integrating decoders and external application adapters; Figure 4 Schematic diagram of the structured potential space based on self-supervised decoupling; Figure 5 Schematic diagram of collaborative supervision of physical information and structured data; Figure 6 Schematic diagram of the online reinforcement learning optimization process driven by model performance requirements. DETAILED DESCRIPTION

[0029] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0030] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0031] The present invention provides a system simulation modeling method based on a neural potential field model combined with physical constraints, and constructs a structured neural potential field model based on a deep neural network combined with a potential field method, including: mapping the low-dimensional input coordinates of the spatiotemporal coordinates to a high-dimensional feature space through position coding technology to obtain a coordinate coding vector; performing local feature extraction based on the coordinate coding vector to generate local features of the query point; performing global information aggregation on the local features of each query point to obtain deep fusion features; mapping the deep fusion features into structured potential codes that characterize different attributes; splicing the structured potential codes of different attributes to obtain structured potential codes; using the trained structured neural potential field model to receive the system data to be simulated, and outputting the structured potential codes of the system to be simulated to realize system simulation.

[0032] Example 1 Combine Figure 1 , a system simulation modeling method based on the neural potential field model combined with physical constraints, including: S1. Build a structured neural potential field model, receive the data to be simulated, and output the structured potential code for a single query point.

[0033] S11. In this embodiment, a parameterized structured neural potential field model function is constructed based on a deep neural network combined with a potential field method. , to achieve a mapping of spatiotemporal and contextual coordinates into a structured potential encoding.

[0034] Structured neural potential field model:

[0035] in, Represents all trainable parameters of the structured neural potential field model; is the three-dimensional space coordinate of the query point; is the timestamp of the query point; is the macro context vector, is a preset positive integer; Used to represent the global state that affects the entire simulation system, that is, a fixed-length vector data structure composed of multiple named context variables. is the structured potential code of the output, and d is a preset integer. It is composed of multiple semantically decoupled basic potential codes that represent different attributes.

[0036] S12. Determine the structure of the structured neural potential field model: In order to achieve the goal of unifying the elements of the entire system into a single, continuous function, It is a hierarchical integrated network architecture with the following structure: Figure 2 shown. The overall trainable parameters of Is the collection of all its submodule parameters.

[0037] The submodules of the structured neural potential field model include: Input encoding module, local feature extraction module, hierarchical fusion module and mapping module.

[0038] 1. Input encoding module, which maps low-dimensional input coordinates such as spatiotemporal coordinates to high-dimensional feature space to enhance the network's ability to capture high-frequency detail information and achieve accurate representation of continuous fields.

[0039] In this embodiment, the input encoding module is implemented based on position encoding technology: Using Fourier eigenmap based on periodic function, the space-time coordinates of the query point are transformed into Mapping to a higher dimensional vector To preserve the original information, the encoding vector The space-time coordinates of the query point With its high-dimensional mapping Spliced ​​together.

[0040] The formula is:

[0041]

[0042] in, and Apply by element; represents the space-time coordinates of the query point, are the three-dimensional space coordinates, It’s time; represents a fixed, non-training variable random Gaussian matrix, represents the Fourier eigenmap function, is the number of Fourier features; Represents a coordinate encoding vector. By concatenating the spatiotemporal coordinates of the query point , preventing the network from losing the low-frequency basis when learning high-frequency information.

[0043] 2. Local feature extraction module, responsible for generating local features related to the query point.

[0044] In this embodiment, a multi-layer perceptron is used to extract local features based on the coordinate encoding vector. ,include network layers, with parameters , The forward propagation process is expressed as:

[0045]

[0046] in, It is Hidden representation of the layer, is the weight matrix, is the bias vector; is the coordinate encoding vector, Represents the macro context vector. is an activation function, preferably a periodic activation function , because of its continuous and everywhere differentiable characteristics, it has unique advantages for the subsequent introduction of physical constraints that require the calculation of high-order derivatives and the characterization of natural oscillation signals of the field.

[0047] through Layer forward propagation obtains , is the dimension of local features; As a local feature, .

[0048] 3. Layered fusion module, including several global dependency aggregation modules.

[0049] 1) The local features of all query points in the system to be simulated are used as the initial local feature matrix of the system to be simulated.

[0050] In this embodiment, the initial local feature matrix is ​​expressed as , N is the total number of query points in the system to be simulated. The formula is:

[0051] 2) Perform hierarchical fusion through several global dependency aggregation modules to capture the long-range dependencies between different query points in the system. The initial local feature matrix is ​​used as the initial input of the hierarchical fusion process, denoted as .

[0052] pass A global dependency aggregation module to obtain deep fusion features:

[0053] To enhance information flow and gradient propagation, each global dependency aggregation module can contain residual connections. The update rule can be expressed as:

[0054] in, , Index of the global dependency aggregation module; The global information aggregation feature matrix output by the previous global dependency aggregation module; Indicates the A global dependency aggregation module, For the The global information aggregation feature matrix output by the global dependency aggregation module; pass A global dependency aggregation module, The output of a global dependency aggregation module is the deep fusion feature, denoted as .matrix Each line Represents the query point The high-level feature representation after fully integrating the global context information provides a rich and robust feature basis for subsequent decoding mapping.

[0055] The global dependency aggregation module includes: a multi-head self-attention sublayer and a position feedforward network sublayer, each of which uses residual connections and layer normalization.

[0056] Data processing of the global dependency aggregation module:

[0057]

[0058] in, To represent attention features, MHSA (Multi-Head Self-Attention) calculates the correlation weights between a batch of query point features, aggregating and upgrading the local features of each point into features rich in global context information. In this embodiment, the standard scaled dot-product attention is combined with a multi-head attention structure and connected with residual connections, layer normalization, and a feedforward network to build a deeper and more stable network, further enhancing expressive power.

[0059] The multi-head self-attention data processing process includes: Scaled Dot Product Attention Mechanism: Via Three Trainable Linear Projection Matrices , mapping the overall features into query (Query), key (Key) and value (Value) matrices, the formula is expressed as:

[0060]

[0061]

[0062] Calculate the attention weights and apply them to the values:

[0063] in, is a scaling factor to ensure gradient stability. To enhance expressiveness, this process is usually performed in parallel in multiple "heads", and the results are concatenated and linearly projected again.

[0064] Through residual connection and layer normalization, we obtain:

[0065] in, The attention feature.

[0066] In this embodiment, FFN is a position feedforward network, which includes two layers of MLP. Deep fusion features are obtained based on attention features through the position feedforward network. The formula is expressed as:

[0067] in, represents the deep fusion feature, i.e. .matrix Each line is the corresponding point feature representation rich in contextual information.

[0068] The parameters of the global dependency aggregation module include: .

[0069] 4. Mapping module, which obtains structured potential coding based on deep fusion features. The data processing process includes: Deep fusion features Map and decompose into basic potential codes that represent different attributes respectively: project the general high-order feature vector into multiple semantic-specific subspaces, each subspace corresponds to a simulation element attribute of geometry, appearance, physics and behavior.

[0070] This is achieved through parallel processing of four independent mapping head networks. In this embodiment, the mapping head network is a small multi-layer perceptron and is denoted as:

[0071] The parameters are: .

[0072] for Each row of the matrix has eigenvectors , Used to index query points and perform the following parallel mapping operations: Generate geometric potential code:

[0073] Generate appearance potential code:

[0074] Generate physical potential code:

[0075] Generate behavioral potential code:

[0076] These four independent basic potential codes are spliced ​​along the dimension direction to form a structured potential code.

[0077] Mapping module generated Vectors are abstract potential codes. As an intermediate representation, they do not directly correspond to specific physical quantities or geometric forms. The subsequent decoder module is responsible for translating these abstract potential codes into ultimately usable, specific simulation parameters.

[0078] Structured latent encoding for a single query point , expressed as:

[0079] Therefore, the structured neural potential field model transforms the space-time coordinate points and global context , which is deterministically mapped into a structured latent code .

[0080] S2. Determine the training method of structured neural potential field model based on self-supervised learning.

[0081] 1. Through self-supervised learning, in the structured neural potential field model Structured latent encoding of output In this paper, a structured potential space with mandatory semantic decoupling is constructed. Without relying on external labels, an information-theoretic internal supervision mechanism is used to ensure that the basic potential encodings of the structured potential encodings are Establishing unique, separable causal relationships with specific attributes of system elements (geometry, appearance, physics, and behavior) lays the foundation for precise, independent attribute control and analysis in subsequent stages. Geometric attributes: Control the spatial presence of objects, SDF fields. Appearance attributes: Control the visual properties of objects, PBR fields. Physical attributes: Control the material type of objects, physics fields. Behavioral attributes: Control the actions and decision logic of intelligent agents.

[0082] In this embodiment, the following loss terms related to information theory are introduced: In this embodiment, the mutual information between the basic potential encoding of each attribute and the corresponding attribute is maximized through the Information Maximizing Generative Adversarial Network (InfoGAN), the structure of which is as follows: Figure 4 shown.

[0083] Forward propagation and loss calculation of information maximization generative adversarial network: 1) Sampling potential codes from prior distribution ,in is the structured potential code to be decoupled, It is a standard unstructured noise potential code. (as a generator ) Generate simulation element samples . Through a composite loss function To optimize:

[0084] in, is the standard adversarial loss; is the mutual information term, and the supervision signal of the mutual information term is composed of an auxiliary network By Reconstruction The error provided by ,forms a self-supervisory closed loop; is a preset weight coefficient greater than zero, used to adjust the strength of the mutual information regularization term.

[0085] Encoding potential For each sub-vector that needs to be decoupled, forward propagation and loss calculation are iteratively performed until all dimensions are effectively decoupled.

[0086] 2. Physical information and structured truth data are fine-tuned in a coordinated and supervised manner to achieve strict alignment with the objective laws of the physical world and the structured truth data in the engineering field.

[0087] Physics information and structured ground truth data are used to supervise fine-tuning: A differentiable adapter agent is used to transmit high-level, structured ground truth data supervision signals to the continuous output of the decoder, and automatic differentiation is used to convert physical laws into computable loss terms. The two work together to solve the problems of low physical credibility of the generated model and its inability to accept structured data supervision. Its overall structure is as follows: Figure 5 shown.

[0088] The model training is carried out with the optimization goal of minimizing the composite loss function.

[0089] Composite loss function:

[0090] in, is the physical information loss, is the supervision loss term for structured ground truth data; represents the physical information loss weight and Represents the weight of the structured truth data supervision loss term, which are all positive hyperparameters.

[0091] 1) Physical information loss: physical laws are imposed as hard constraints on the model generation process to achieve physical credibility. The specific implementation process is as follows: a. Determine the physical laws as predefined partial differential equation operators.

[0092] b. Use automatic differentiation to construct the residual function: In each iteration of training, a batch of configuration points are randomly sampled from the simulation domain .

[0093] Using the built-in computer automatic differentiation function of the deep learning framework, it is possible to calculate the structured neural potential field model and decoder Composite functions About its input Any higher-order derivative of .

[0094] c. Perform residual calculation: Substitute the derivative value obtained by automatic differentiation into the preset PDE operator , get the physical residual at each sampling point This residual value This numerically represents the degree to which the data field generated by the current model deviates from the physical laws. If the model completely complies with the physical laws, the residual should be zero.

[0095] d. Generate optimized signal: the physical residual of all sampling points The mean square norm of the aggregation is used as the physical information loss , the formula is:

[0096] Physical information loss value As a clear and quantifiable optimization signal, its relationship with the model parameters is calculated through the back propagation algorithm The gradient is used to update and adjust the model parameters .

[0097] The abstract physical laws are transformed into an executable control and correction mechanism that directly acts on the model parameters, thereby forcing the model to converge to a solution that strictly adheres to the physical laws during the learning process, ensuring the inherent physical consistency of its output results.

[0098] 2) Structured ground truth data supervision loss , used to ensure that the model output can reproduce real, structured engineering data: a. Introducing the differentiable adapter agent: Introducing a pre-defined, mathematically differentiable adapter agent module The function of this module is to convert the decoder The output point-like and continuous values ​​are aggregated and transformed into a structured true value data Corresponding prediction data in data structure .

[0099]

[0100] This module is part of the training method, which is designed with differentiability as the first principle.

[0101] b. Loss calculation: Based on forecast data With true data The loss is calculated by the difference between:

[0102] because is differentiable, and the gradient can be obtained from Back propagation to .

[0103] S3. Based on the model performance requirements, a reinforcement learning framework is constructed based on structured potential coding to optimize the model and determine the factor configuration plan.

[0104] In this embodiment, the reinforcement learning framework: the structured neural potential field model It is considered as a parameterized, generalized strategy that can generate all elements of the system. By maximizing the reward signal directly related to the final performance indicators of the system, it can not only optimize the behavior of the intelligent agent, but also optimize the representation of all non-intelligent elements in the system, including geometry, appearance, physics, and behavior, giving it powerful reverse design and autonomous optimization capabilities during the model construction phase. Mature policy optimization algorithms in the field of reinforcement learning, such as policy gradient and proximal policy optimization, can be used to calculate losses and update the structured neural potential field model. The parameters of the Figure 6 shown.

[0105] 1) Definition of reinforcement learning framework: Define the complete smart factory simulation system as the environment. Structured neural potential field model Considered a strategy , which is in state The potential coding generated by That is action The environmental status here is a time period that contains the current moment , the agent's own coordinates , and the macro context of the moment The structured representation of , thus with The input matches.

[0106] 2) Determine the reward function: A scalar reward function directly linked to the system's final task performance indicator Its supervisory signal comes from the performance evaluation results after the simulation system runs.

[0107] 3) Strategy Optimization Finding the optimal parameters To maximize the expected cumulative reward

[0108] .

[0109] Calculate the loss through algorithms such as policy gradient , and using its gradient Structured neural potential field model The parameters are iteratively updated until the model performance converges.

[0110] Example 2 Based on the method of Example 1, the process of a six-axis industrial robot and its end gripper gripping a target car door and completing high-speed handling in a smart factory system is simulated, and the design of the gripper is optimized.

[0111] First, the CAD model and material properties of the six-axis industrial robot and its end gripper are determined, and the geometry, mass distribution, and material properties of the car door are determined as the environment.

[0112] 1. Capture the three-dimensional coordinates of the gripper: The initial space-time coordinates are obtained by random sampling or grid sampling.

[0113] Three-dimensional space coordinates : Any point in the gripper design space , timestamp : At any moment in the handling process simulation. Macro context : is a fixed vector representing the ID of the gripper design task. In more complex scenarios, It can also include the encoded factory environment temperature and humidity, grid voltage fluctuation level, production line cycle time, and process parameter ID of the current production batch.

[0114] 2. Output the potential code of the clamp: ; in: is the vector representation that determines the three-dimensional geometry of the gripper. A vector representation used to determine the appearance of each part of the gripper. Encodes the potential of physical properties. This is a vector representation of the robot's motion strategy when handling a door, including acceleration and velocity curves. The dimensions of each code are preset positive integers.

[0115] 3. Self-supervised decoupling training: 1) Random The vector is input into a pre-trained geometry decoder, preferably a DeepSDF decoder, to generate a large number of 3D model samples with different geometries.

[0116] 2) Training using the InfoGAN framework , so that it can generate geometric features consistent with the distribution of the above 3D model samples, while forcing Maximize the mutual information with the generative geometry.

[0117] 3) Yes A similar approach is used to enforce association with specific properties of the material distribution (e.g., density, elastic modulus).

[0118] After training is completed, Can execute the following instructions: give a specific vector, can stably generate the corresponding gripper geometry field; given a specific Vectors can generate corresponding material property fields, which successfully achieves semantic decoupling.

[0119] 4. Physical information and data collaborative supervision training: 1) Determine the physical laws of the gripper: Continuum mechanics equations: .in, is the body force, is the acceleration, is stress, is the density.

[0120] 2) Structured truth data supervision standard: Calculate the stress field truth data through traditional FEA software , as the model calibration value.

[0121] 3) Integrate the physical information of the gripper and the structured ground truth data supervision standard to construct the loss function:

[0122] in, Loss of physical information: random sampling within the gripper design domain points , calculated by automatic differentiation technique The generated stress and displacement fields are substituted into the above elastic mechanics equations to obtain the mean square value of the residual. The results generated by the driving model must comply with the laws of physics. :Stress field predicted by the model and FEA true value data The mean square error between .

[0123] Once trained, the model not only generates geometry and materials, but also generates fields that are physically self-consistent and accurate.

[0124] 5. Determine optimization goals: Lightest weight ;The most stable clamping , i.e. the maximum displacement and vibration amplitude of the door during transportation); lowest cost ; Strongest structure , that is, under the action of the inertial force generated by high-speed movement, the maximum stress of the clamping structure is far lower than the yield limit of the material.

[0125] 6. Achieving optimization goals through reinforcement learning: 1) Environment: A differentiable simulator constructs a reinforcement learning environment. The simulator receives the parameters generated by the model. , and can calculate the final performance indicators of the elements to be optimized through simulation.

[0126] 2) Strategy: Structured Neural Potential Field Model As a strategy.

[0127] 3) Action: At each optimization iteration, Generate a set of potential codes: .

[0128] 4) Reward function:

[0129] and Obtained by volume integration of the geometry and material fields generated by the model. and (Stress) is calculated using a differentiable physics solver with built-in elasticity equations. is the weight coefficient used to balance different optimization objectives.

[0130] 5) Optimization process: using policy gradient algorithm.

[0131] Model Generate a configuration of the elements of the gripper, that is, the action The simulation environment executes the element configuration plan, calculates various performance indicators, and Given a scalar reward value, calculate the policy gradient based on the reward value and backpropagate to update the model Parameters Through repeated iterations, the model will autonomously explore the design space and continuously generate factor configurations with higher reward values, making the gripper lighter, more stable, cheaper, and stronger.

[0132] This method can construct a parameter-fixed, highly optimized structured neural potential field model It can not only represent complex systems with high fidelity and self-consistency, but also contains the optimal element configuration plan for specific task performance.

[0133] Example 3 This implementation demonstrates the feasibility of applying this method to obtain potential coding in existing characterization software. The structure is as follows Figure 3 As shown: After the model training is completed, its parameters With decoder set Together, they form a unified, continuous digital foundation. As the information source for the simulated world, it provides external services through a standardized query API. Therefore, the technical boundaries of this approach are clearly defined as providing a continuous field containing all the system's information, which can be queried on demand.

[0134] The process of converting this continuous information base into the discrete, explicit assets required by specific downstream applications, such as rendering engines and physics solvers, is accomplished by an external "adapter," a non-core component of this invention. The adapter is a "driver" for a specific application, responsible for calling the query API of this invention and performing the necessary engineering transformations.

[0135] 1. Query API paradigm This is done through a standardized query interface. The core paradigm of this interface is to allow external applications to input specific query parameters such as spatiotemporal coordinates and contextual information, and request specific types of output information such as SDF values, PBR parameters, physical properties, and behavioral intentions at that location.

[0136] As an implementation example of this paradigm, its interface function can be abstracted as:

[0137] in, Specify the type of query information. Contains the parameters required for the query. An example is as follows: Query the SDF value: , returns a floating point number .

[0138] Query PBR parameters: , returns a parameter vector .

[0139] Query physical properties: , returns a parameter vector .

[0140] Query behavioral intent:

[0141] Returns an intermediate semantic representation In this query, the API first uses Corresponding space-time coordinates and macro context Call get , and then and the incoming micro-context Send it to the behavior decoder , ultimately returning highly contextualized behavioral intentions .

[0142] Downstream Application Adapter Integration Example The following examples illustrate several typical operating modes for external application adapters. These adapters are built by downstream application developers based on the standardized query API specifications provided by this invention and their own specific application requirements. They are designed to convert the unified continuous field provided by this invention into discrete assets or instructions usable by specific downstream engines. These examples are for illustrative purposes only and do not constitute limitations on this invention.

[0143] Rendering Adapter In virtual reality (VR) training or digital twin visualization applications in smart factories, the rendering adapter is responsible for converting the continuous geometry and material fields provided by the present invention into the discretized assets (such as triangular meshes and PBR textures) required by specific 3D rendering engines (such as Unreal Engine 5 and Unity).

[0144] The adapter is an independent software module, and its typical workflow is as follows: (1) Geometric discretization: The adapter internally implements a well-known surface extraction algorithm, such as "Marching Cubes." To execute this algorithm, the adapter needs to obtain the signed distance function (SDF) values ​​of a large number of points within a spatial region. Therefore, the adapter cyclically calls the query interface provided by the present invention within a preset spatial grid: .

[0145] (2) 3D asset generation: The “marching cubes” algorithm uses a series of SDF scalar values ​​returned by the previous query step to calculate the vertex and face information of the triangular mesh describing the surface of the object, and can save it into a standard 3D file format (such as .fbx) that the rendering engine can recognize.

[0146] (3) Material texture generation: The adapter calculates UV coordinates for the generated 3D mesh according to standard computer graphics procedures. The adapter then traverses the UV space and calls the query interface of the present invention again for each texel to be calculated:

[0147] The spatial point coordinates Mapped from the current texel's UV coordinates and grid position, the adapter "bakes" the returned PBR parameter vector into one or more 2D texture maps.

[0148] In this flow, the present invention is strictly limited to responding to atomic queries for SDF values ​​and PBR parameters. All subsequent processing, including the specific implementation of the "marching cubes" algorithm, construction of the 3D mesh data structure, UV unwrapping, texture baking, and file input / output operations, is performed independently by the rendering adapter according to the specific requirements of its target engine and is outside the core technical scope of this invention.

[0149] Physical adapter When performing high-fidelity physics simulations, the physics adapter acts as an "on-demand virtual material database" to provide material properties at any point in space to commercial physics solvers (such as Ansys and Abaqus) in real time during the calculation process.

[0150] The adapter is usually implemented as a user-defined function (UDF) or plug-in supported by the solver. Its working mode is as follows: (1) Solver request: When an external physics solver performs a simulation (such as finite element analysis), it needs to obtain the material properties of a certain integration point (for example, Young's modulus and thermal conductivity that vary with temperature). At this time, the solver will call the pre-implanted UDF (i.e., physics adapter).

[0151] (2) API call: The UDF receives the coordinate points and current physical state (such as temperature) from the solver as the macro context. After encapsulating this information, it immediately calls the query interface of the present invention:

[0152]

[0153] (3) Data transfer: UDF receives the parameter vector returned by the present invention, which contains the precise physical properties of the point. , and passes it back to the solver seamlessly and in real time. The solver then uses these property values ​​to continue its subsequent equation-solving calculations.

[0154] Behavior Adapter The behavior adapter converts the high-level semantic intent output by the present invention into a fully contextualized , translated into specific, low-level agent decision and action instructions that can be executed by a robot controller (such as ROS) or a mission planning system.

[0155] The adapter can integrate mature motion planning algorithms (such as RRT*), task scheduling logic, or behavior trees. Its decision-making process is as follows: (1) Collecting micro-context: When a decision is required, the adapter first collects and encodes the entity’s current micro-interaction context from the simulation environment. , for example, readings from a virtual force sensor, the position of an obstacle detected by a lidar, or specific communication commands received.

[0156] (2) Querying high-level intent: The adapter calls the query interface of the present invention:

[0157]

[0158] .

[0159] It combines the entity ID, macro context and the micro-context just collected The API of this invention will return a highly contextualized semantic intent that has integrated macro-situations and micro-stimuli. .

[0160] (3) Downstream planning and execution: The adapter takes this explicit intention As the core input or goal of its internal planning algorithm. For example, if for , the adapter's motion controller will start a compliant control program; if for The adapter's RRT* path planner calculates a collision-free path to the charging station. Finally, the adapter generates specific low-level motion instructions (such as joint angles and velocities) and delivers them to the simulation engine for execution.

[0161] Through a unified structured neural potential field model, the present invention makes the interaction of multiple physical domains an intrinsic and continuous calculation process of the model, eliminates interface errors, realizes the inherent unification of multi-domain elements, and greatly improves the upper limit of simulation fidelity.

[0162] This method introduces physical information loss, ensuring that the model output strictly adheres to physical laws and system rules. This provides a strong physical credibility guarantee for AI-driven generative simulation models, significantly improves the verifiability of AI models, and enables their results to be used in critical engineering design, evaluation and demonstration, and command decision-making.

[0163] This approach provides a high-level "reactive intention" that responds to both the macro and micro environments. All the specific planning, calculation, and execution logic required to achieve this intent, such as path search, collision detection, or action sequence decomposition, are handled by the external behavior adapter and its integrated well-known algorithms, which ensures the versatility of this approach and the flexibility of the external execution engine.

[0164] To address the problem that single-domain model construction relies heavily on expert experience and manual parameter configuration, this method can automatically learn and generate entity behavior models with complex adaptability and emergence from data and rules. It can quickly build large-scale, highly adversarial simulation scenarios involving thousands of autonomous interacting entities, support automated modeling and highly parallel simulation of complex behaviors, and significantly improve the efficiency of model construction and simulation deduction of large-scale systems.

[0165] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A system simulation modeling method based on a neural potential field model combined with physical constraints, characterized by: include: Construct a structured neural potential field model based on deep neural network combined with potential field method; The structured neural potential field model includes: The low-dimensional input coordinates of the space-time coordinates of the query point are mapped to the high-dimensional feature space through the position encoding technology to obtain the coordinate encoding vector of the query point; Performing local feature extraction on the coordinate encoding vector of the query point to generate a local feature of the query point; Perform global information aggregation on the local features of each query point to obtain deep fusion features; Mapping the deep fusion features into basic potential codes that characterize different properties of the system to be simulated; splicing basic potential codes of different attributes to obtain structured potential codes; the attributes include: geometry, appearance, physics and behavior; Training the structured neural potential field model; using the trained structured neural potential field model to receive data of the system to be simulated, and outputting a structured potential code of the system to be simulated; The system simulation is realized by utilizing the structured potential coding of the system to be simulated.

2. The method according to claim 1, characterized in that Training the structured neural potential field model, comprising: The structured neural potential field model is trained by self-supervised learning using an information maximization generative adversarial network. The structured neural potential field model is trained using physical laws as constraints and structured true value data.

3. The method according to claim 1, characterized in that It also includes determining the element configuration scheme based on structured potential coding: Constructing a reinforcement learning framework based on structured potential coding; using the optimization objective as a reward function, and obtaining an optimized structured potential coding through the reinforcement learning framework; The optimized structured potential coding is used as the element configuration scheme of the simulation system.

4. The method according to claim 2, characterized in that The self-supervised learning training of the structured neural potential field model using an information maximization generative adversarial network includes: An information maximization generative adversarial network is used to maximize the mutual information between each basic potential code and the corresponding attribute. The loss function of the information maximization generative adversarial network is: in, is the standard adversarial loss; is the mutual information term, It is the structured potential code to be decoupled; the decoder generates simulation element samples , is a preset weight coefficient greater than zero, used to adjust the strength of the mutual information regularization term.

5. The method according to claim 2, characterized in that The training of the structured neural potential field model based on physical laws and structured true value data includes: The physical law is used as a hard constraint through physical information loss, and a composite loss function is constructed in combination with the structured true value data supervision loss term; the structured neural potential field model is trained with minimizing the composite loss function as the optimization goal; the composite loss function: in, is the physical information loss, is the supervision loss term for structured ground truth data; Represents the weights and sums of the structured truth data supervision loss terms Represents the physical information loss weight, which are all positive hyperparameters.

6. The method according to claim 3, characterized in that The reinforcement learning framework includes: The structured neural potential field model is used as the strategy; the potential encoding generated by the structured neural potential field model is used as the action; and the structured representation of the current moment, coordinates, and macro context is used as the environmental state.

7. The method according to claim 1, characterized in that The method of mapping the low-dimensional input coordinates of the space-time coordinates of the query point to the high-dimensional feature space by using the position encoding technology to obtain the coordinate encoding vector of the query point includes: in, and Apply by element; is the space-time coordinate, are the three-dimensional space coordinates, It's time, is a random Gaussian matrix, is the number of Fourier features.

8. The method according to claim 1, characterized in that Performing local feature extraction on the coordinate encoding vector of the query point to generate local features of the query point includes: Coordinate Encoded Vector and macro context vector , extract local features through multi-layer perceptron; The forward propagation process of each layer of the multilayer perceptron is expressed as: in, It is Hidden representation of the layer, It is Hidden representation of the layer; is the weight matrix, is the bias vector, is the total number of layers of the multilayer perceptron; Is the activation function; the output is the local feature , is the dimension of local features.

9. The method according to claim 1, characterized in that The global information aggregation of the local features of each query point to obtain deep fusion features includes: The local features of all query points of the system to be simulated are used as the initial local feature matrix of the system to be simulated; the initial local feature matrix is ​​hierarchically fused through several global dependency aggregation modules to obtain deep fusion features, including: in, , m is the index of the global dependency aggregation module; The global information aggregation feature matrix output by the previous global dependency aggregation module; Indicates the A global dependency aggregation module, For the The global information aggregation feature matrix output by the first global dependency aggregation module is used; the global information aggregation feature matrix output by the last global dependency aggregation module is used as the deep fusion feature.

10. The method according to claim 9, characterized in that The data processing process of the global dependency aggregation module includes: in, Represents the attention feature, MHSA represents the multi-head attention mechanism, FFN represents the feedforward network, represents a normalization layer.

Images