A system simulation modeling method based on a neural potential field model combined with physical constraints

Abstract

The application provides a system simulation modeling method based on a neural potential field model combined with physical constraints, and belongs to the technical field of system simulation. A structured neural potential field model is constructed based on a deep neural network combined with a potential field method, including: mapping low-dimensional input coordinates 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 vector of the query point to generate local features of the query point, and further performing global information aggregation to obtain deep fusion features; mapping the deep fusion features to basic potential coding representing different attributes of a system to be simulated; splicing the basic potential coding of different attributes to obtain structured potential coding; receiving system data to be simulated by using the trained structured neural potential field model, and outputting structured potential coding of the system to be simulated to realize system simulation. A single and continuous parameterized function is used to represent elements in the system, thereby improving the fidelity of system simulation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer simulation, in particular to a system simulation modeling method based on a neural potential field model combined with physical constraints. BACKGROUND

[0002] In the production scheduling and equipment coordination of intelligent manufacturing workshops, it is often necessary to accurately simulate the strong coupling effects of the thermal effects (thermal) of the machining equipment under long-time high-load operation, the mechanical structure deformation and material fatigue (mechanics) caused by thermal stress, and the reverse influence (production process and logistics dynamics) of these deformations on the process flow, line rhythm and product precision.

[0003] Currently, system simulation is mainly realized by combining multiple professional simulation technologies, including: geometric and environmental modeling schemes, entity geometry and environmental background are usually constructed based on computer-aided design models, geographic information system data or remote sensing data. The physical processes involved in the physical process modeling scheme, such as mechanics, thermodynamics, fluid, electromagnetism, etc., are usually simulated by independent professional solvers. For example, finite element analysis software is used to simulate structural stress, computational fluid dynamics software is used to simulate air or liquid flow, and multi-body dynamics engine is used to simulate mechanical system motion. Independent simulation tools are connected through high-level architecture-based connection. However, in actual production, the simulators of different physical domains such as FEA and CFD usually use mathematical models and discretization methods with great differences in principle. The collaborative simulation technology only exchanges discrete data at the macro level, which can cause serious numerical errors, oscillations and delays at the interface, making it difficult to accurately simulate cross-domain strong coupling effects. While the pure data-driven machine learning model can learn and generate complex behavior patterns, its "black box" nature makes its output may violate the basic physical laws or domain constraint conditions, leading to unreliable simulation results, which cannot be used for analysis, planning and decision support scenarios with high reliability requirements, and there is a huge "reliability gap". In summary, the current system simulation method has the problem of insufficient cooperation between multiple models or insufficient physical constraints, resulting in insufficient simulation accuracy, poor adaptability and insufficient physical reliability of system simulation.

[0004] Therefore, a high-fidelity and highly adaptable system simulation modeling method is needed. SUMMARY

[0005] Therefore, the present application provides a system simulation modeling method based on a neural potential field model combined with physical constraints, which uses a single, continuous parameterized function to represent the geometry, appearance, physical and behavior attributes of each element in the system, replacing the traditional federated architecture composed of multiple discrete models, meeting the high-fidelity requirements of simulation.

[0006] To this end, the present application provides the following technical solutions:

[0007] A system simulation modeling method based on a neural potential field model combined with physical constraints, comprising:

[0008] A structured neural potential field model is constructed based on a deep neural network combined with a potential field method;

[0009] The structured neural potential field model comprises:

[0010] The low-dimensional input coordinates of the spatio-temporal coordinates of the query point are mapped to a high-dimensional feature space through a position encoding technique to obtain a coordinate encoding vector of the query point;

[0011] The coordinate encoding vector of the query point is locally feature-extracted to generate a local feature of the query point;

[0012] The local features of each query point are globally aggregated to obtain a deep fusion feature;

[0013] The deep fusion feature is mapped to a basic potential encoding representing different attributes of the system to be simulated;

[0014] The basic potential encodings of different attributes are spliced to obtain a structured potential encoding; the attributes include geometry, appearance, physics, and behavior;

[0015] The structured neural potential field model is trained; the trained structured neural potential field model is used to receive data of the system to be simulated and output a structured potential encoding of the system to be simulated;

[0016] The structured potential encoding of the system to be simulated is used to realize system simulation.

[0017] Further, training the structured neural potential field model comprises:

[0018] An information maximization generative adversarial network is used to perform self-supervised learning training on the structured neural potential field model;

[0019] The structured neural potential field model is trained based on physical laws as constraints combined with structured ground truth data.

[0020] Further, it also includes determining an element configuration scheme based on the structured potential encoding:

[0021] A reinforcement learning framework is constructed based on the structured potential encoding;

[0022] An optimized structured potential encoding is obtained through the reinforcement learning framework by taking an optimization goal as a reward function;

[0023] The optimized structured potential encoding is used as an element configuration scheme of the simulation system.

[0024] Further, the self-supervised learning training of the structured neural potential field model by the information maximization generative adversarial network comprises:

[0025] The information maximization generative adversarial network maximizes the mutual information between each basic potential code and the corresponding attribute; the loss function of the information maximization generative adversarial network is:

[0026]

[0027] Wherein, is a standard adversarial loss; is a mutual information term, is a structured potential code to be decoupled; the simulation element sample is generated by a decoder , is a preset weight coefficient greater than zero, used to adjust the strength of the mutual information regularization term.

[0028] Further, the structured neural potential field model is trained by combining the structured ground truth data with the physical law as a constraint, comprising:

[0029] The physical law is taken as a hard constraint by a physical information loss, and a composite loss function is constructed by combining a structured ground truth data supervised loss term;

[0030] The structured neural potential field model is trained with the optimization objective of minimizing the composite loss function;

[0031] The composite loss function is:

[0032]

[0033] Wherein, is a physical information loss, is a structured ground truth data supervised loss term; represents the weight of the structured ground truth data supervised loss term and represents the weight of the physical information loss, both of which are positive value parameters.

[0034] Further, the reinforcement learning framework comprises:

[0035] The structured neural potential field model is taken as a strategy;

[0036] The potential code generated by the structured neural potential field model is taken as an action;

[0037] The structured representation of the current time, coordinates and macro context is taken as an environment state.

[0038] Further, the low-dimensional input coordinates of the spatio-temporal coordinates of the query point are mapped to a high-dimensional feature space by the position encoding technology to obtain a coordinate encoding vector of the query point, comprising:

[0039]

[0040]

[0041] wherein, and are applied element by element; is a spatio-temporal coordinate, is a three-dimensional space coordinate, is time, is a random Gaussian matrix, is the number of Fourier features.

[0042] Further, the coordinate encoding vector of the query point is locally feature extracted to generate a local feature of the query point, comprising:

[0043] coordinate encoding vector and macro context vector , the local feature is extracted by a multi-layer perception machine;

[0044] The forward propagation process of each layer network of the multi-layer perception machine is represented as:

[0045]

[0046] wherein, is the hidden representation of the layer, is the hidden representation of the layer; is a weight matrix, is a bias vector, is the total number of layers of the multi-layer perception machine; is an activation function; the output is a local feature , is the dimension of the local feature.

[0047] Further, the local features of each query point are globally aggregated to obtain a deep fusion feature, comprising:

[0048] The local features of all query points of the to-be-simulated system are taken as an initial local feature matrix of the to-be-simulated system;

[0049] The initial local feature matrix is hierarchically fused by a plurality of global dependence aggregation modules to obtain a deep fusion feature, comprising:

[0050]

[0051] in, , where m is the index of the global dependency aggregation module; This is the global information aggregation feature matrix output by the previous global dependency aggregation module; Indicates the first A global dependency aggregation module, For the first The global information aggregation feature matrix output by each global dependency aggregation module;

[0052] The global information aggregation feature matrix output by the last global dependency aggregation module is used as the deep fusion feature.

[0053] Furthermore, the data processing procedure of the global dependency aggregation module includes:

[0054]

[0055]

[0056] in, Indicating attentional characteristics, MHSA This represents a multi-head attention mechanism, and FFN represents a feedforward network. This indicates the normalization layer.

[0057] Advantages and positive effects of the present invention:

[0058] This method uses a structured neural potential field model for system simulation: it uses a single and continuous parameterized function to characterize the geometric, physical, and behavioral attributes of each element in the system and outputs a semantically decoupled structured potential code, thus solving the problem of insufficient system simulation fidelity caused by combining multiple discrete models in existing system simulations.

[0059] 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 fitting ability of continuous fields; it captures non-local interactions between points in the field through a self-attention mechanism; the hierarchical structure and independent mapping heads ensure the decoupling of the model's expressive power and the structured output; and it integrates physical laws, operating procedures and safety regulations into the structured neural potential field model through a physical information loss term to ensure the physical consistency of the system simulation.

[0060] 2) This method uses a reinforcement learning framework: structured potential encoding based on system simulation obtains the element configuration scheme of the preset target through a preset reward function, such as generating element configuration schemes of structure, materials, behavior and other elements that meet specific engineering goals, realizing the functional extension from "forward simulation analysis" to "reverse optimization design". Attached Figure Description

[0061] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0062] Figure 1 The flowchart shows a system simulation modeling method based on a neural potential field model combined with physical constraints.

[0063] Figure 2 This is a structural diagram of a structured neural potential field model.

[0064] Figure 3 A schematic diagram showing the integration of the decoder and external application adapter;

[0065] Figure 4 This is a schematic diagram of a structured potential space based on self-supervised decoupling;

[0066] Figure 5 A schematic diagram illustrating the collaborative supervision of physical information and structured data;

[0067] Figure 6 A schematic diagram of the online reinforcement learning optimization process driven by model performance requirements. Detailed Implementation

[0068] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0069] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0070] This invention provides a system simulation modeling method based on a neural potential field model combined with physical constraints. The method constructs a structured neural potential field model based on a deep neural network and potential field approach, including: mapping low-dimensional spatiotemporal input coordinates to a high-dimensional feature space using position encoding technology to obtain coordinate encoding vectors; extracting local features based on the coordinate encoding vectors to generate local features of query points; aggregating global information from the local features of each query point to obtain deep fusion features; mapping the deep fusion features to structured potential codes representing different attributes; concatenating the structured potential codes of different attributes to obtain structured potential codes; and using the trained structured neural potential field model to receive data from the system to be simulated and output the structured potential codes of the system to be simulated to achieve system simulation.

[0071] Example 1

[0072] Combination Figure 1 A system simulation modeling method based on a neural potential field model combined with physical constraints includes:

[0073] S1. Construct a structured neural potential field model, receive the data to be simulated, and output the structured potential code for a single query point.

[0074] 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. This is to achieve the mapping of spatiotemporal and contextual coordinates to a structured potential code.

[0075] Structured neural potential model:

[0076]

[0077] in, This represents all trainable parameters of the structured neural potential model; These are the three-dimensional spatial coordinates of the query point; It is the timestamp of the query point; It is a macro-context vector. It is a preset positive integer; It is used to characterize the global state that affects the entire simulation system, that is, a fixed-length vector data structure composed of multiple named context variables. It is the structured potential code of the output, where d is a preset integer. It is composed of multiple basic potential codes that represent different attributes and are semantically decoupled.

[0078] S12. Determine the structure of the structured neural potential model:

[0079] To achieve the goal of uniformly representing all elements of the system as a single, continuous function, It is a layered converged network architecture, with the structure as follows:Figure 2 As shown. Overall trainable parameters It is the collection of parameters of all its submodules.

[0080] The sub-modules of the structured neural potential model include:

[0081] The module consists of an input encoding module, a local feature extraction module, a hierarchical fusion module, and a mapping module.

[0082] 1. The input encoding module maps low-dimensional input coordinates such as spatiotemporal coordinates to a high-dimensional feature space to enhance the network's ability to capture high-frequency detail information, thereby achieving accurate representation of the continuous field.

[0083] In this embodiment, the input encoding module is implemented based on positional encoding technology:

[0084] Using Fourier feature mapping based on periodic functions, the spatiotemporal coordinates of the query point are... Mapped to a higher-dimensional vector To preserve the original information, the encoding vector... spatiotemporal coordinates of the query point Its high-dimensional mapping It was pieced together.

[0085] The formula is expressed as:

[0086]

[0087]

[0088] in, and Apply per element; Represents the spatiotemporal coordinates of the query point. It is a three-dimensional spatial coordinate. It is time; Let represent a fixed, non-trained, variable random Gaussian matrix. Represents the Fourier feature mapping function. It is the number of Fourier features; This represents a coordinate encoding vector. It is constructed by concatenating the spatiotemporal coordinates of the query point. This prevents the network from losing low-frequency fundamentals while learning high-frequency information.

[0089] 2. Local feature extraction module, responsible for generating local features related to the query point.

[0090] In this embodiment, a multilayer perceptron is used to extract local features based on coordinate encoding vectors. Multilayer perceptron ,include There are 1 network layer with parameters as follows: , The forward propagation process is represented as:

[0091]

[0092]

[0093] in, It is the first Hidden representation of layers, It is a weight matrix. It is the bias vector; Coordinate encoding vector, Represents the macro-context vector. It is an activation function, preferably a periodic activation function. Due to its continuous and differentiable properties, it has unique advantages for introducing physical constraints that require the calculation of higher-order derivatives and for characterizing the natural oscillation signals of the field.

[0094] through Layer forward propagation , It is the dimension of local features; As a local feature, it is denoted as .

[0095] 3. Layered fusion module, including several global dependency aggregation modules.

[0096] 1) Use the local features of all query points in the system to be simulated as the initial local feature matrix of the system to be simulated.

[0097] In this embodiment, the initial local feature matrix is ​​represented as follows: N represents the total number of query points in the system to be simulated. The formula is expressed as:

[0098]

[0099] 2) Layered fusion is performed through several global dependency aggregation modules to capture long-range dependencies between different query points in the system. The initial local feature matrix is ​​used as the initial input for the layered fusion process, denoted as... .

[0100] pass A global dependency aggregation module is used to obtain deep fusion features:

[0101]

[0102] To enhance information flow and gradient propagation, each global dependency aggregation module can contain residual connections, and the update rule in this case can be expressed as:

[0103]

[0104] in, , This is an index for globally dependent aggregation modules; This is the global information aggregation feature matrix output by the previous global dependency aggregation module; Indicates the first A global dependency aggregation module, For the first The global information aggregation feature matrix output by each global dependency aggregation module;

[0105] pass The first global dependency aggregation module, the... The output of each global dependency aggregation module For deep fusion features, denoted as .matrix each line Represents the query point The high-order feature representation, which fully integrates global contextual information, provides a rich and robust feature foundation for subsequent decoding and mapping.

[0106] The global dependency aggregation module includes a multi-head self-attention sub-layer and a location feedforward network sub-layer, with each sub-layer employing residual connections and layer normalization.

[0107] Data processing steps that depend on the global aggregation module:

[0108]

[0109]

[0110] in, Representing attention features, MHSA (Multi-Head Self-Attention) aggregates and upgrades the local features of each point into features rich in global context information by calculating the weights of the relationships between features of a batch of query points. In this embodiment, a multi-head attention structure is combined with standard scaled dot product attention, and connected with residual connections, layer normalization, and feedforward networks to build a deeper and more stable network, further enhancing its expressive power.

[0111] The multi-head self-attention data processing procedure includes:

[0112] Scaling dot product attention mechanism: using three trainable linear projection matrices The overall features are mapped to a matrix of queries, keys, and values, expressed by the following formula:

[0113]

[0114]

[0115]

[0116] Calculate attention weights and apply them to the values:

[0117]

[0118] in, This is a scaling factor to ensure gradient stability. To enhance expressiveness, this process is typically performed in parallel across multiple "heads," and the results are concatenated and then linearly projected again.

[0119] Through residual connections and layer normalization, we obtain:

[0120]

[0121] in, This is an attention feature.

[0122] In this embodiment, FFN is a location-based feedforward network containing two MLP layers. It obtains deep fusion features based on attention features through the location-based feedforward network, as expressed by the formula:

[0123]

[0124] in, This indicates deep fusion features, i.e. .matrix Each line Corresponding points A feature representation rich in contextual information.

[0125] The parameters of the global dependency aggregation module include: .

[0126] 4. Mapping module: Based on deep fusion features, structured potential codes are obtained. The data processing includes:

[0127] Deep fusion features Mapping and decomposing into basic potential codes that represent different attributes: Projecting a general high-order feature vector into multiple semantically specific subspaces, each subspace corresponding to a simulation element attribute in terms of geometry, appearance, physics, and behavior.

[0128] This is achieved through parallel processing using four independent mapping head networks. In this embodiment, the mapping head networks are small multilayer perceptrons, denoted as follows:

[0129]

[0130] The parameters are as follows: .

[0131] for eigenvectors of each row in the matrix , Used for index query points, the following parallel mapping operations are performed:

[0132] Generative geometric potential coding:

[0133] Generate appearance potential coding:

[0134] Generate physical potential coding:

[0135] Generate behavioral potential encoding:

[0136] These four independent basic potential codes are spliced ​​together along the dimensional direction to form a structured potential code.

[0137] The mapping module generates Vectors are abstract potential codes that serve as an intermediate representation and do not directly correspond to specific physical quantities or geometric shapes. Subsequent decoder modules are responsible for translating these abstract potential codes into final, usable, and concrete simulation parameters.

[0138] Structured potential coding for a single query point , is represented as:

[0139]

[0140] Therefore, the structured neural potential field model will use spatiotemporal coordinates. and global context Deterministically mapped to structured potential encoding .

[0141] S2. A method for determining the training of a structured neural potential field model based on self-supervised learning.

[0142] 1. Through self-supervised learning, in the structured neural potential field model Output structured potential coding In this study, a structured potential space with forced semantic decoupling is constructed. Without relying on external labels, an internal supervision mechanism based on information theory ensures the proper encoding of each fundamental potential. Establishing unique and separable causal relationships between system elements and specific attributes (geometric, appearance, physical, and behavioral) lays the foundation for precise and independent attribute control and analysis in subsequent stages. Geometric attributes: control the spatial existence of objects, SDF field. Appearance attributes: control the visual attributes of objects, PBR field. Physical attributes: control the material type of objects, physical field. Behavioral attributes: control the actions and decision-making logic of the agent.

[0143] In this embodiment, this is achieved by introducing an information-theoretic loss term:

[0144] In this embodiment, the Information Maximizing Generative Adversarial Network (InfoGAN) is used to maximize the mutual information between the basic potential encoding of each attribute and the corresponding attribute, as shown in the structure below. Figure 4 As shown.

[0145] Forward propagation and loss calculation in information-maximizing generative adversarial networks:

[0146] 1) Sampling potential codes from prior distributions ,in It is the structured potential encoding to be decoupled. It is a standard unstructured noise latent energy code. (Through the decoder...) (as a generator) Generate simulation element samples Through a composite loss function Optimize:

[0147]

[0148] in, Standard adversarial loss; For mutual information terms, the monitoring signal for mutual information terms is provided by an auxiliary network. By from Reconstruction The error provided forms a self-supervised closed loop; This is a preset weight coefficient greater than zero, used to adjust the strength of the mutual information regularization term.

[0149] Encoding potential For each subvector that needs to be decoupled, forward propagation and loss calculation are performed iteratively until all dimensions are effectively decoupled.

[0150] 2. Physical information and structured truth data are collaboratively monitored and fine-tuned to achieve strict alignment with the objective laws of the physical world and structured truth data in the engineering field.

[0151] Co-supervised fine-tuning using physical information and structured truth data: A differentiable adapter is used to pass high-level, structured truth data supervision signals to the continuous output of the decoder, and automatic differentiation is used to transform physical laws into computable loss terms. The two work synergistically to address the problems of low physical reliability and inability to accept structured data supervision in generative models. Its overall structure is as follows: Figure 5 As shown.

[0152] The model is trained with the goal of minimizing the composite loss function.

[0153] Composite loss function:

[0154]

[0155] in, For physical information loss, For structured truth data, supervise the loss term; Represents the weights of physical information loss. The weights represent the supervisory loss terms for structured truth data, and are all positive hyperparameters.

[0156] 1) Physical information loss: Physical laws are applied as a hard constraint to the model generation process to achieve physical credibility. The specific implementation process is as follows:

[0157] a. Determine the physical laws as pre-defined operators for partial differential equations.

[0158] b. Constructing the residual function using automatic differentiation: In each iteration of training, a batch of placement points are randomly sampled from the simulation domain. .

[0159] By utilizing the built-in automatic differentiation capabilities of deep learning frameworks, it is possible to compute the structured neural potential model. and decoder Composite function Regarding its input Arbitrary higher-order derivatives of .

[0160] c. Perform residual calculation: Substitute the derivative value obtained through automatic differentiation calculation into the preset PDE operator. The physical residual at each sampling point is obtained. This residual value Numerically, this residual characterizes the degree to which the data field generated by the current model deviates from the physical law. If the model perfectly conforms to the physical law, this residual should be zero.

[0161] d. Generate optimized signal: physical residuals of all sampling points The mean square norm is aggregated as the physical information loss. The formula is expressed as:

[0162]

[0163] Physical information loss value As a clear and quantifiable optimization signal, its relationship with the model parameters is calculated using the backpropagation algorithm. The gradient is used to update and adjust the model parameters. .

[0164] 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 follows physical laws during the learning process, ensuring the inherent physical consistency of its output results.

[0165] 2) Structured truth data supervision loss term This is used to ensure that the model output can reproduce real, structured engineering data:

[0166] a. Introduce a differentiable adapter proxy:

[0167] Introduce a pre-defined, mathematically differentiable adapter proxy module. This module's function is to add a decoder. The output pointwise, continuous values ​​are aggregated and transformed into structured truth data. Corresponding prediction data in terms of data structure .

[0168]

[0169] This module is part of the training method, and its design prioritizes differentiability.

[0170] b. Loss Calculation:

[0171] Based on forecast data With true data Calculate the loss based on the difference between them:

[0172]

[0173] because It is differentiable, and the gradient can be derived from... Unimpeded backpropagation to .

[0174] S3. Based on the model performance requirements, construct a reinforcement learning framework using structured potential coding to optimize the model and determine the element configuration scheme.

[0175] In this embodiment, the reinforcement learning framework uses a structured neural potential field model. It is viewed as a parameterized, generalized policy capable of generating all elements of the system. By maximizing the reward signal directly related to the final performance index of the system, it can not only optimize the behavior of the agent, but also optimize the representation of all non-agent elements in the system, including geometry, appearance, physics, and behavior, thus endowing it with powerful inverse design and autonomous optimization capabilities during the model building stage. Mature policy optimization algorithms in the field of reinforcement learning, such as policy gradient and proximal policy optimization, can be used to calculate the loss and update the structured neural potential field model. The parameters, and its overall structure are as follows: Figure 6 As shown.

[0176] 1) Definition of reinforcement learning framework:

[0177] The complete intelligent factory simulation system is defined as the environment. Structured neural potential field model. regarded as strategy Its state Potential coding generated below That is, action The environmental conditions here It is a time that includes the current moment. The agent's own coordinates and the macro context of that moment. The structured representation of, i.e. Thus with The input matches.

[0178] 2) Determine the reward function:

[0179] Scalar reward function directly linked to the final performance indicators of the system's task The monitoring signal originates from the performance evaluation results after the simulation system has run.

[0180] 3) Strategy optimization

[0181] Finding the optimal parameters To maximize expected cumulative reward

[0182] .

[0183] Loss is calculated using algorithms such as policy gradient. and utilize its gradient Structured neural potential model The parameters are iteratively updated until the model performance converges.

[0184] Example 2

[0185] Based on the method in Example 1, the process of a six-axis industrial robot and its end effector gripper gripping a target car door and completing high-speed transport in a smart factory system is simulated, and the gripper design is optimized.

[0186] First, determine the CAD model and material properties of the six-axis industrial robot and its end effector, and then determine the geometry, mass distribution, and material properties of the door as the environment.

[0187] 1. Capture the three-dimensional spatial coordinates of the gripper:

[0188] Initial spatiotemporal coordinates are obtained through random sampling or grid sampling.

[0189] Three-dimensional spatial coordinates Any point within the gripper design space timestamp At any point during the simulation of the material handling process. (Macro-context) : It is a fixed vector representing the ID of the gripper design task. In more complex scenarios, It can also include coded factory environment temperature and humidity, power grid voltage fluctuation level, production line cycle time, and process parameter ID of the current production batch.

[0190] 2. Output the potential code of the gripper:

[0191] ;

[0192] in: Vector representation for determining the three-dimensional geometry of the gripper. Vector representations used to determine the appearance of each part of the gripper. Encodes the potential of physical attributes. The vector representation used to determine the robot's motion strategy when handling a car door includes acceleration and velocity curves. The dimensions of each encoding are preset positive integers.

[0193] 3. Self-supervised decoupling training:

[0194] 1) Randomly select The vector input is a pre-trained geometry decoder, preferably a DeepSDF decoder, which generates a large number of 3D model samples with different geometries.

[0195] 2) Training using the InfoGAN framework This enables it to generate geometric features consistent with the distribution of the 3D model samples mentioned above, while forcing Maximize mutual information with the generative geometry.

[0196] 3) To A similar approach is used to force a correlation between it and specific properties of material distribution (such as density and elastic modulus).

[0197] After training, It can execute the following instructions: give a specific A vector can stably generate the corresponding gripper geometry; given a specific... Vectors can generate corresponding material property fields, thus successfully achieving semantic decoupling.

[0198] 4. Collaborative supervised training using physical information and data:

[0199] 1) Determine the physical laws governing the grippers:

[0200] Equations of Continuum Mechanics: .in, For volume forces, For acceleration, For stress, For density.

[0201] 2) Structured True Data Supervision Standard: Calculating stress field true data using traditional FEA software. , as the model calibration value.

[0202] 3) Integrate the physical information of the grippers and structured truth data to construct a loss function:

[0203]

[0204] in, For physical information loss: random sampling within the gripper design domain Points Calculated using automatic differentiation techniques The mean square value of the residuals after substituting the generated stress and displacement fields into the above elasticity equations. The results generated by the driving model must obey physical laws. Structured ground truth data supervision loss. Stress field predicted by the model and FEA true data The mean square error between them.

[0205] After training, the model can not only generate geometry and materials, but also generate fields that are physically consistent and accurate.

[0206] 5. Define the optimization objective:

[0207] Lightest weight The most stable clamping (i.e., the maximum displacement and vibration amplitude of the vehicle door during transportation); lowest cost. The strongest structure That is, under the action of inertial force generated by high-speed motion, the maximum stress of the gripper structure is far lower than the material yield limit.

[0208] 6. Achieve optimization goals through reinforcement learning:

[0209] 1) Environment: A differentiable simulator constructs a reinforcement learning environment. This simulator receives parameters generated by the model. It can also calculate the final performance indicators of the elements to be optimized through simulation.

[0210] 2) Strategy: Structured Neural Potential Field Model As a strategy.

[0211] 3) Actions: In each optimization iteration step, Generate a set of potential codes:

[0212] .

[0213] 4) Reward function:

[0214]

[0215] and It is obtained by volume integration of the geometric and material fields generated by the model. and The stress is calculated using a differentiable physics solver that incorporates the equations of elasticity. These are weighting coefficients used to balance different optimization objectives.

[0216] 5) Optimization process: The policy gradient algorithm is used.

[0217] Model Generate a configuration scheme for the gripper's elements, i.e., actions. The simulation environment executes the configuration scheme of the elements, calculates various performance indicators, and applies the reward function. Given a scalar reward value, calculate the policy gradient based on the reward value and update the model through backpropagation. parameters Through repeated iterations, the model autonomously explores the design space, continuously generating element configuration schemes with higher reward values, enabling the grippers to achieve lighter, more stable, cheaper, and stronger configurations.

[0218] This method can construct a highly optimized structured neural potential field model with fixed parameters. It can not only represent complex systems with high fidelity and self-consistency, but also contains the optimal configuration scheme of elements for specific task performance.

[0219] Example 3

[0220] This implementation demonstrates the feasibility of applying the potential coding method to existing characterization software, with the structure as follows: Figure 3 As shown:

[0221] Once the model is trained, its parameters With decoder set Together, they form a unified and continuous digital foundation. As the information source of the simulated world, it provides services externally through a standardized query application interface. Therefore, the technical boundary of this method is clearly defined as providing a continuous field containing all the information of the system, which can be queried on demand.

[0222] The process of transforming the continuous information of this base into discrete, explicit assets required by specific downstream applications such as rendering engines and physics solvers is accomplished by an external "adapter," which is not part of the core 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.

[0223] 1. Query API paradigm

[0224] The call is implemented through a standardized query interface. The core paradigm of this interface is to allow external applications to take specific query parameters such as spatiotemporal coordinates and contextual information as input and request specific types of output information such as SDF value, PBR parameter, physical attributes, and behavioral intent at that location.

[0225] As an example of an implementation of this paradigm, its interface functions can be abstracted as follows:

[0226]

[0227] in, Specify the type of information to query. Includes the parameters required for the query. An example is shown below:

[0228] Query SDF value:

[0229] Returns a floating-point number .

[0230] Query PBR parameters:

[0231] Returns a parameter vector .

[0232] Query physical properties:

[0233] Returns a parameter vector .

[0234] Intent of query behavior:

[0235]

[0236] Return an intermediate semantic representation In this query, the API internally first uses... Corresponding spatiotemporal coordinates and macro-context Call get Then With the incoming micro context Send them together to the behavior decoder Ultimately, it returns to highly contextualized behavioral intentions. .

[0237] Downstream application adapter integration example

[0238] The following describes several typical operating modes of external application adapters. These adapters are built by downstream application developers based on the standardized query API specifications provided by this invention and the specific needs of their applications, aiming to transform the unified continuous field provided by this invention into discrete assets or instructions usable by a specific downstream engine. These examples are for illustrative purposes only and do not constitute limiting terms of this invention.

[0239] Render adapter

[0240] 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 field provided by this invention into discretized assets (such as triangular meshes and PBR textures) required by a specific 3D rendering engine (such as Unreal Engine 5, Unity).

[0241] This adapter is a standalone software module, and its typical workflow is as follows:

[0242] (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 iteratively calls the query interface provided by this invention within a preset spatial grid:

[0243] .

[0244] (2) Generation of 3D assets: The "moving cube" algorithm uses a series of SDF scalar values ​​returned by the previous query to calculate the vertex and face information of the triangular mesh describing the surface of the object, and can save it as a standard 3D file format (such as .fbx) that the rendering engine can recognize.

[0245] (3) Material Texture Generation: The adapter calculates the UV coordinates for the generated 3D mesh according to the standard computer graphics process. Then, the adapter traverses the UV space, calling the query interface of this invention again for each texel to be calculated.

[0246]

[0247] The coordinates of the spatial point It is obtained by mapping the UV coordinates and mesh position of the current texel. The adapter will "bake" the returned PBR parameter vector into one or more 2D texture maps.

[0248] In this process, the responsibility of this invention is strictly limited to responding to atomic query requests for SDF values ​​and PBR parameters. All subsequent processing, including the specific implementation of the "moving cube" algorithm, the 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 not within the core technical scope of this invention.

[0249] physical adapter

[0250] In high-fidelity physics simulations, the physics adapter acts as an "on-demand virtual material database," providing commercial physics solvers (such as Ansys and Abaqus) with material properties at any point in space in real time during the calculation process.

[0251] This adapter is typically implemented as a user-defined function (UDF) or plugin supported by the solver. Its operating mode is as follows:

[0252] (1) Solver request: When an external physics solver is performing a simulation calculation (such as finite element analysis), it needs to obtain the material properties (e.g., Young's modulus, thermal conductivity, etc. as a function of temperature) at a certain integration point in its solution process. At this time, the solver will call the pre-embedded UDF (i.e., physics adapter).

[0253] (2) API call: The UDF receives the coordinates and current physical state (such as temperature, as a macroscopic context) from the solver. (A part of it). It encapsulates this information and immediately invokes the query interface of this invention:

[0254]

[0255]

[0256] (3) Data transmission: The UDF receives the parameter vector returned by the present invention, which contains the precise physical properties of the point. This information is then seamlessly and in real-time passed back to the solver. The solver then uses these attribute values ​​to continue its subsequent equation solving calculations.

[0257] Behavior adapter

[0258] The behavior adapter will output the fully contextualized high-level semantic intent of this invention. This translates to specific, underlying agent decisions and action instructions that can be executed by a robot controller (such as ROS) or task planning system.

[0259] This adapter can integrate mature motion planning algorithms (such as RRT*), task scheduling logic, or behavior trees. Its decision-making process is as follows:

[0260] (1) Acquiring micro-context: When a decision is required, the adapter first acquires and encodes the current micro-interaction context of the entity from the simulation environment. For example, readings from a virtual force sensor, the location of an obstacle detected by a lidar, or specific communication commands received.

[0261] (2) Query advanced intent: The adapter calls the query interface of this invention:

[0262]

[0263]

[0264] .

[0265] It includes the entity's ID and macro-context. And the micro-context just collected These are passed in as parameters. The API of this invention will return a highly contextualized semantic intent that integrates macro-level situational awareness and micro-level stimuli. .

[0266] (3) Downstream planning and execution: The adapter will convey this clear intention As the core input or objective of its internal planning algorithm. For example, if for The adapter's motion controller will then initiate 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 commands (such as joint angles and velocities) and delivers them to the simulation engine for execution.

[0267] This invention, through a unified structured neural potential field model, makes the interaction of multiple physical domains an intrinsic and continuous computational process within the model, eliminating interface errors, achieving the intrinsic unity of multi-domain elements, and greatly improving the upper limit of simulation fidelity.

[0268] This method introduces physical information loss to ensure that the model output strictly adheres to physical laws and system rules. This provides a strong guarantee of physical credibility for AI-driven generative simulation models, significantly improves the verifiability of AI models, and enables their results to be used for critical engineering design, evaluation and demonstration, and command and control decisions.

[0269] This method provides a high-level "reactive intent" that has responded to both macroscopic and microscopic environments. All the specific planning, computation, and execution logic performed to achieve this intent, such as path search, collision detection, or action sequence decomposition, is handled by the external behavior adapter and its integrated known algorithms. This ensures the versatility of the approach and the flexibility of the external execution engine.

[0270] To address the problem that single-domain model construction heavily relies on expert experience and manual parameter configuration, this method can automatically learn and generate entity behavior models with complex adaptability and emergent behavior from data and rules. It can quickly construct large-scale, highly adversarial simulation scenarios containing tens of thousands of autonomously interacting entities, support automated modeling and highly parallel simulation of complex behaviors, and significantly improve the efficiency of model construction and simulation inference of large-scale systems.

[0271] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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 in that, include: A structured neural potential field model is constructed based on deep neural networks combined with potential field methods; The structured neural potential field model is trained; the trained structured neural potential field model is used to receive the CAD model and material property data of the six-axis industrial robot and its end effector gripper, and outputs the structured potential code of the gripper. The structured potential coding of the gripper is used to realize the system simulation of the gripper; A reinforcement learning framework is built based on structured potential coding; Using the optimization objective as the reward function, the optimized structured potential encoding is obtained through the reinforcement learning framework. The optimized structured potential coding serves as the element configuration scheme for the gripper; The structured neural potential model includes: The low-dimensional input coordinates of the spatiotemporal coordinates of the query point are mapped to a high-dimensional feature space by using position encoding technology to obtain the coordinate encoding vector of the query point. Local feature extraction is performed on the coordinate encoding vector of the query point to generate local features of the query point; Deep fusion features are obtained by aggregating local features from each query point with global information. The deep fusion features are mapped to basic potential codes that characterize different attributes of the grippers; By splicing together the basic potential codes of different attributes, a structured potential code is obtained, including: a vector representation that determines the three-dimensional geometry of the gripper, a vector that determines the appearance of each part of the gripper, a potential code that represents physical attributes, and a vector representation that determines the motion strategy of the robot when moving the door. Training the structured neural potential model includes: The structured neural potential model is trained using a generative adversarial network that maximizes information. The structured neural potential model is trained using physical laws as constraints and structured ground truth data. Physical laws are used as hard constraints through physical information loss, and a composite loss function is constructed by combining a structured truth data supervision loss term. The structured neural potential model is trained with minimizing the composite loss function as the optimization objective. The composite loss function is as follows: in, For physical information loss, For structured truth data, supervise the loss term; The sum of the weights of the supervised loss terms for structured truth data The weights representing the physical information loss are all positive hyperparameters.

2. The method according to claim 1, characterized in that, The step of using an information-maximizing generative adversarial network to perform self-supervised learning training on the structured neural potential model includes: An information-maximizing generative adversarial network (GAN) is employed to maximize the mutual information between each basic potential encoding and its corresponding attribute; the loss function of the information-maximizing GAN is: in, Standard adversarial loss; For mutual information items, It is a structured potential encoding to be decoupled; simulation element samples are generated through a decoder. , This is a preset weight coefficient greater than zero, used to adjust the strength of the mutual information regularization term.

3. The method according to claim 1, 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 time, coordinates and macroscopic context is used as the environmental state.

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

5. The method according to claim 1, characterized in that, Local feature extraction is performed on the coordinate encoding vector of the query point to generate local features of the query point, including: coordinate encoding vector and macro-context vector Local features are extracted using a multilayer perceptron; The forward propagation process of each layer of the multilayer perceptron is represented as follows: in, It is the first Hidden representation of layers, It is the first Hidden representation of layers; This is the weight matrix. For bias vectors, This represents the total number of layers in the multilayer perceptron. It is an activation function; the output is local features. , It is the dimension of local features.

6. The method according to claim 1, characterized in that, The process of aggregating local features at each query point to obtain deep fusion features includes: 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; the initial local feature matrix is ​​then fused hierarchically through several global dependency aggregation modules to obtain deep fused features, including: in, , where m is the index of the global dependency aggregation module; This is the global information aggregation feature matrix output by the previous global dependency aggregation module; Indicates the first A global dependency aggregation module, For the first The global information aggregation feature matrix output by each 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.

7. The method according to claim 6, characterized in that, The data processing procedure of the global dependency aggregation module includes: in, Indicating attentional characteristics, MHSA This represents a multi-head attention mechanism, and FFN represents a feedforward network. This indicates the normalization layer.

Images