High-performance calculation method suitable for multi-scale simulation
Through technologies such as convolutional neural networks and deep deterministic strategy gradient algorithms, the problems of low computational efficiency and difficulty in selecting feature scales in multi-scale simulation are solved. The tree data structure and ICP algorithm are used to represent dynamic geometry, and the numerical value is evaluated using error integral method. Stability, realizes high-performance multi-scale simulation calculation, and improves simulation accuracy and efficiency.
Patent Information
- Application Number
- CN202411096024.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-09
- Publication Date
- 2024-12-03
AI Technical Summary
The traditional multi-scale simulation calculation methods have problems with low computational efficiency, difficulty in selecting time and space feature scales, difficulty in dynamic geometric representation and time scale feature analysis, time discretization error and numerical stability, resulting in inaccurate simulation and large resource consumption.
Convolutional neural network is used to judge the time and space feature scales, deep deterministic strategy gradient algorithm establishes a time step regulation system, tree-like data structure realizes the octree adaptive grid representation, and ICP algorithm captures geometric structure evolution, clustering and patterns Identify and analyze the characteristics of physical quantities, evaluate discretization error and numerical stability by the error integration method, and adjust the time step length through gradient iteration to optimize the calculation accuracy.
It realizes automatic extraction of time and spatial feature scales, optimizes time step selection, improves calculation accuracy and efficiency, and is suitable for multi-scale physical simulation and simulation fields.
Smart Images

Figure CN119066913A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of information technology, and in particular to a high-performance computing method suitable for multi-scale simulation. Background Art
[0002] Things change over time. The process of simulating changes in time and space is a process that requires high-performance computing, such as simulating the growth of plants, or mechanical devices that wear out over time or other changes that produce objects of multiple scales. With the continuous development of science and technology and the growing demand for applications, multi-scale simulation has become an important issue in many fields. However, there are currently some challenges and problems. First, traditional computing methods have the problem of low computational efficiency in multi-scale simulation. Since physical processes involve multiple scales, traditional methods usually require a large amount of calculations to simulate and predict the behavior of the system, resulting in long computing time and high consumption of computing resources. Second, the selection of temporal and spatial characteristic scales in multi-scale simulation is also a key issue. Physical processes at different scales have different characteristics and behaviors, so it is necessary to determine the appropriate temporal and spatial characteristic scales for simulation. However, how to accurately select these scales remains a challenging problem. In addition, dynamic geometric representation and time scale feature analysis are also difficult points in multi-scale simulation. The geometric shapes in multi-scale systems are usually dynamically changing, so appropriate representation methods are needed to capture their dynamics. In addition, the analysis of time scale characteristics is also crucial for accurately simulating and predicting physical processes, but how to effectively analyze and utilize time scale characteristics remains a challenging problem. Finally, traditional computational methods also have certain problems in time discretization errors and numerical stability of predictions. Due to time discretization and numerical stability issues, traditional methods may not be able to provide accurate simulation and prediction results. Therefore, how to evaluate time discretization errors and numerical stability of predictions, and take corresponding measures to improve the accuracy and efficiency of simulations, remains an issue that needs to be addressed. Summary of the invention
[0003] The present invention provides a high-performance computing method suitable for multi-scale simulation, which mainly includes:
[0004] Convolutional neural networks are used to determine the temporal and spatial characteristic scales in multi-scale physical processes. Based on the obtained temporal and spatial characteristic scales, a deep deterministic policy gradient algorithm is used to establish an automated time step control system. The octree adaptive grid implemented with a tree data structure is used to represent dynamic geometry. The ICP algorithm is used to capture the evolution of geometric structures and build a deformation prediction model. The time scale characteristics of different physical quantities are analyzed with the help of clustering and pattern recognition. A coupling step selection strategy is designed based on numerical stability indicators and computational costs. The error integral method is used to evaluate the time discretization error and predict numerical stability. Based on the error integral method and parameter identification method, online evaluation of components is achieved. The value of the next time step is adjusted through gradient iteration to optimize the calculation accuracy.
[0005] In one embodiment, determining the temporal and spatial characteristic scales in a multi-scale physical process by a convolutional neural network includes:
[0006] Acquire multi-scale data in the physical process, extract the spatial and temporal features in the data through convolutional neural networks, and obtain spatial and temporal feature maps; use wavelet transform to perform multi-scale analysis on the feature maps, judge the characteristic responses at different scales, and determine the corresponding spatial and temporal characteristic scales; based on the determined spatial and temporal characteristic scales, use the equivalent characteristic scale criterion to integrate the spatial and temporal dimensions to obtain the equivalent characteristic scale; use the sliding window method to capture the evolution information of the characteristic scale by sliding a fixed-size window on the time series data.
[0007] In one embodiment, the method uses a deep deterministic policy gradient algorithm to establish an automated time step control system based on the obtained temporal and spatial feature scales, including:
[0008] Obtain the temporal and spatial characteristic scales of the physical process, construct a data set, and use it as environmental data for the deep deterministic policy gradient; set the time step as the action of the time step control system, and use the error metric and stability indicator as rewards; use the deep deterministic policy gradient algorithm to construct a deep neural network policy model, receive the current state as input, obtain the probability distribution of each action, interact with the environment, realize the autonomous setting of the time step, and obtain the optimization strategy of the time step; obtain simulation data of different physical processes, retrain through transfer learning, perform parameter weighting adjustment, and optimize the weights of the deep neural network policy model; when a new physical situation occurs, obtain data under the new situation and adjust the policy network.
[0009] In one embodiment, the octree adaptive grid representation of dynamic geometry implemented using a tree data structure includes:
[0010] The initial three-dimensional geometric model of the object is obtained, and the three-dimensional geometric model is segmented by an octree. The mesh topology relationship represented by the tree structure is obtained through recursive segmentation. The geometric features of each unit in the mesh are determined, and according to the feature changes and the preset accuracy requirements, a local refinement / coarsening method is used to insert a refined mesh in the outer area of the surface and generate a coarsening mesh inside the area to achieve adaptive adjustment of the mesh. The topological evolution information of the geometric structure is extracted, the topological relationship of the octree mesh is updated, and the octree nodes are inserted or deleted to achieve the update of the dynamic mesh topology. The mesh accuracy is determined again for the updated mesh to check whether it meets the requirements for the representation of dynamic geometry. If not, the mesh distribution is adjusted and optimized through the local refinement / coarsening method until the requirements for the representation of dynamic geometry are met.
[0011] In one embodiment, capturing the evolution of the geometric structure by the ICP algorithm and constructing a deformation prediction model includes:
[0012] The point cloud data of the structure at different times are obtained, and the point cloud data at each time is registered to achieve the alignment of the data in the same coordinate system; the motion and displacement information of the feature points on the surface of the structure are determined by point cloud registration, and the rigid body motion and non-rigid body deformation are decomposed by matrix variation method; based on the feature points, a deformation prediction model is constructed using a neural network algorithm, and the coordinates and time information of the feature points are used as input. The deformation prediction model is trained to predict the displacement of the feature points in real time to obtain the deformation trend of the structure; according to the predicted displacement of the feature points, a deformation dynamics equation is constructed, and the structure is divided into a finite number of units, each unit contains a number of nodes, and is discretized into algebraic equations on a finite number of nodes. The discretized equations are solved by the finite element method to simulate the deformation process of the structure; the deformation termination and convergence conditions are set, and the deformation process ends when the deformation is less than the set threshold.
[0013] In one embodiment, analyzing the time scale characteristics of different physical quantities by means of clustering and pattern recognition includes:
[0014] Acquire multi-source heterogeneous sequence data containing different physical quantities, normalize the sequence data to eliminate dimensional differences; use the normalized multi-source heterogeneous sequence data as input, build a clustering model based on cosine similarity, identify the category patterns in the data, and obtain pattern sequences representing different physical processes; based on the pattern sequence, use the wavelet transform multi-scale analysis method to decompose the energy distribution of the sequence at different time scales, determine the main response scale of each category of physical quantity sequence, and determine the corresponding time characteristic scale; use the EM clustering algorithm to calculate the probability that each energy distribution obtained by the wavelet transform multi-scale analysis belongs to each category, update the parameters of each category, and distinguish the physical process patterns at different time scales.
[0015] In one embodiment, the integrated numerical stability index and computational cost are used to design a coupled step size selection strategy, including:
[0016] Taking the time step, the numerical solution of the gas phase and solid phase solver, as well as the grid size and grid node layout as input, a discrete error estimation model based on matrix permutation theory is constructed to obtain the dynamic growth rate of the numerical solution in the alternating iteration process of the gas phase and solid phase solver under different time steps, and obtain the numerical stability index; determine the number of memory accesses and the amount of floating-point calculations, and estimate the computational cost under different step sizes; combine the numerical stability index and the computational cost, and weight them according to the determined weights to obtain the compromise optimal step size; obtain the analytical form of the coupling equation, and take the sensitivity of the determined equation to the coupling step size as the target layer, the different equations as the preparation layer, and the specific equations as the solution layer to construct a judgment matrix for pairwise comparison of factors at different levels; calculate the weight vector of each level by performing eigenvalue decomposition on the judgment matrix to represent the relative importance of the corresponding factors; determine the sensitivity of different equations to the coupling step size based on the calculated weight vector; determine the coupling sequence according to the order of equation sensitivity, and narrow the search space for step size selection; if a critical point where the coupling gain changes sharply occurs, use the minimum generation condition to search for potential singular points, avoid the boundary step size where the numerical solution collapses, and obtain a globally optimized coupling step size combination scheme.
[0017] In one embodiment, the method of evaluating the time discretization error and predicting the numerical stability by using an error integration method includes:
[0018] Establish an analytical solution for the continuous system as an exact solution; construct an equivalent time-discrete equation group, and use different time steps Δt for space-time discretization; solve the discrete equation group separately to obtain the corresponding numerical solution, compare it with the analytical solution, use the error integration method to compare the numerical solution with the analytical solution, and quantitatively analyze the discretization error of each physical quantity; according to the error estimation result, if the error increases with time, use the matrix gain analysis theory to calculate the spectral radius of the numerical solution matrix of the two step schemes to judge the numerical stability; determine the spectral radius of the gain matrix, and when the spectral radius is greater than the preset spectral radius threshold, it is judged to be unstable; if there is an unstable state, adjust the time step Δt, solve and evaluate again until a stable solution is obtained; use the polynomial fitting method to establish a relationship model between the discretization error and Δt according to the stable Δt range, and realize the prediction of numerical stability under arbitrary Δt.
[0019] In one embodiment, the online evaluation of components based on the error integration method and the parameter identification method includes:
[0020] The spatiotemporal multi-resolution data generated by the simulation calculation is obtained, and different physical quantity fields in the spatiotemporal multi-resolution data are normalized to eliminate the influence of dimensional errors. The exact solution or experimental data is extracted as a benchmark, and an error integration model is constructed. The complex step size algorithm is used to divide the integration interval into several sub-intervals, and different step sizes are used in each sub-interval. The numerical solution is compared with the benchmark to quantitatively analyze the numerical errors of each field variable. According to the error analysis results, it is judged whether the simulation accuracy meets the requirements. If not, the feature tracking algorithm based on the optical flow method is used to identify the parameters with reduced calculation accuracy, including grid density and time step. The identified parameters with reduced calculation accuracy are used as input to construct a parameter correction regression model, train the parameter sensitivity coefficients of the parameter correction regression model, obtain the corrected parameters, and update the parameters online to control the accuracy within the allowable range. When the parameter correction regression model or the grid is dynamically adaptively adjusted, the online evaluation component is rebuilt to achieve closed-loop control of the simulation accuracy.
[0021] In one embodiment, the step of adjusting the next time step value by gradient iteration to optimize the calculation accuracy includes:
[0022] In the decoupling model of the current time step, a test source term is introduced to perturb each coupling variable; the response fluctuation of the coupling field is Fourier transformed, the spectral response of different components is analyzed, and the main time scale of the coupling field is determined; based on the spectral analysis, if there is a fast-scale component that does not match the current time step well, the information of the gradient field at this scale is extracted as feedback for the next time step adjustment; in the current time step, the gradient field amplitude of each coupling field is measured and compared with the preset spectral response range; if the gradient field exceeds the range, the gradient descent method is used to iteratively optimize and adjust the next time step so that the gradient field response conforms to the expected decoupling model; the test source excitation, gradient detection and time step adjustment process are repeated to achieve adaptive optimization control of the time domain resolution.
[0023] The technical solution provided by the embodiment of the present invention may have the following beneficial effects:
[0024] The present invention discloses a high-performance computing method suitable for multi-scale simulation. The method can automatically extract time and space characteristic scales, and automatically select a suitable time step according to numerical stability and computational cost, thereby optimizing computational accuracy and efficiency. The method has high accuracy and practicality and can be widely used in the field of physical simulation and simulation. BRIEF DESCRIPTION OF THE DRAWINGS
[0025] Figure 1 The present invention is a flowchart of a high-performance computing method suitable for multi-scale simulation.
[0026] Figure 2A schematic diagram of a high performance computing method suitable for multi-scale simulation according to the present invention. DETAILED DESCRIPTION
[0027] In order to enable those skilled in the art to better understand the technical solutions in this specification, the technical solutions in the embodiments of this specification will be clearly and completely described below in conjunction with the drawings in the embodiments of this specification. Obviously, the described embodiments are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this specification.
[0028] A high performance computing method suitable for multi-scale simulation in this embodiment may specifically include:
[0029] S101. Determine the temporal and spatial characteristic scales in multi-scale physical processes through convolutional neural networks.
[0030] Acquire multi-scale data in the physical process, extract the spatial and temporal features in the data through convolutional neural networks, and obtain spatial and temporal feature maps. Use wavelet transform to perform multi-scale analysis on the feature map, judge the characteristic response at different scales, and determine the corresponding spatial and temporal characteristic scales. According to the determined spatial and temporal characteristic scales, use the equivalent characteristic scale criterion to integrate the spatial and temporal dimensions to obtain the equivalent characteristic scale. Use the sliding window method to capture the evolution information of the characteristic scale by sliding a fixed-size window on the time series data.
[0031] For example, there is a data set of a physical process, which contains multiple observation points of temperature changes over time. Convolutional neural networks can be used to extract spatial and temporal features from the data. By performing convolution operations on the data, a set of feature maps can be obtained, each of which corresponds to a spatial and temporal feature. Next, these feature maps can be analyzed at multiple scales using wavelet transform. Wavelet transform can decompose the signal into frequency components of different scales and obtain characteristic responses at each scale. By analyzing the characteristic responses at different scales, the corresponding spatial and temporal characteristic scales can be determined. Based on the determined spatial and temporal characteristic scales, the equivalent characteristic scale criterion can be used to comprehensively consider the spatial and temporal dimensions and obtain equivalent characteristic scales. This can help better understand the multi-scale behavior of physical processes. Finally, the sliding window method can be used to capture the evolution information of the characteristic scale. By sliding a fixed-size window on the time series data, the change of the characteristic scale over time can be observed. The mean or variance of the characteristic scale within each window and the trend of the characteristic scale over time can be calculated.
[0032] S102. Based on the obtained temporal and spatial feature scales, a deep deterministic policy gradient algorithm is used to establish an automated time step control system.
[0033] Obtain the time and space characteristic scales in the physical process, construct a data set, and use it as the environmental data for the deep deterministic policy gradient. Set the time step as the action of the time step control system, and the error metric and stability indicator as the reward. Use the deep deterministic policy gradient algorithm to build a deep neural network policy model, receive the current state as input, obtain the probability distribution of each action, interact with the environment, realize the autonomous setting of the time step, and obtain the optimization strategy of the time step. Obtain simulation data of different physical processes, retrain through transfer learning, adjust the weighted parameters, and optimize the weights of the deep neural network policy model. When a new physical situation occurs, obtain data under the new situation and adjust the policy network.
[0034] For example, to build a deep deterministic policy gradient algorithm to optimize a physical process of reverse swinging. In this process, the temporal and spatial feature scales can be represented as time steps and swing angles, respectively. First, collect simulation data of reverse swinging as a training dataset. Each data sample includes the current swing angle and speed as state input, and the time step as action. Also define an error metric, the deviation of the swing angle, and a stability indicator, such as the oscillation amplitude of the swing. These indicators will be used as reward signals to guide the optimization of the policy network. Next, build a deep neural network policy model that will receive the current state as input and output a probability distribution of each possible action. Use the deep deterministic policy gradient algorithm to interact with the environment, and update the weights of the policy network by selecting an action and observing the feedback from the environment. By continuously iterating this process, the policy network can be optimized to enable it to select the appropriate time step based on the current state. In the initial training stage, the reverse swing data collected previously can be used for training. Then, transfer learning can be performed to retrain with data of similar physical processes but with different parameters, and the policy model can be optimized by adjusting the parameter weights of the neural network. When faced with a new physical situation, you can collect data from the new situation and use this data to adjust the policy network to adapt to the new physical characteristics. If you optimize the physics of an inverted pendulum, you can collect simulation data of the inverted pendulum under different parameter settings and adjust the policy network based on this data.
[0035] S103, using an octree adaptive grid implemented with a tree data structure to represent dynamic geometry.
[0036] Obtain the initial 3D geometric model of the object, perform octree segmentation on the 3D geometric model, and obtain the mesh topology relationship represented by the tree structure through recursive segmentation. Determine the geometric features of each unit in the mesh, and according to the feature changes and the preset accuracy requirements, use the local refinement / coarsening method to insert refined meshes in the outer area of the surface and generate coarsening meshes inside the area to achieve adaptive adjustment of the mesh. Extract the topological evolution information of the geometric structure, update the topological relationship of the octree mesh, insert or delete octree nodes, and update the dynamic mesh topology. Determine the mesh accuracy again for the updated mesh to check whether it meets the requirements for the representation of dynamic geometry. If not, continue to adjust and optimize the mesh distribution through the local refinement / coarsening method until the requirements for the representation of dynamic geometry are met.
[0037] For example, when representing the complex flow geometry of a wing-shaped body, first import a three-dimensional airfoil geometry model containing 10,000 triangular facets. Then, the octree recursive segmentation algorithm is used to divide the model space into grid units of different sizes, and the internal geometric features of each unit are relatively uniform. Next, the volume change rate of each grid unit is calculated. According to the preset 1% volume change rate accuracy requirement, a local refinement / coarsening method is used to insert smaller grids in areas with large flow gradients, and expand the grid size in areas with relatively gentle internal flows. After 5 adaptive adjustment iterations, the volume change rate of each unit is less than 1%. If the wing encounters a large torsional deformation during movement due to gusts, the deformed geometric model is first quickly reconstructed based on the octree space segmentation relationship. Then, the topological feature change information of the wing airfoil profile before and after deformation is extracted, including newly appearing and disappearing feature points and feature lines. The topological information is used to update the node connection relationship of the octree grid model, and the tree nodes are inserted or merged to achieve grid topology adjustment to ensure that the grid topology is consistent with the geometric feature changes. The adjusted adaptive grid is subjected to geometric feature extraction and accuracy evaluation again. If more than 15% of the units still fail to meet the 1% volume change rate requirement, the grid optimization is continued to make the grid distribution fully adapt to the dynamic deformation process of the airfoil surface.
[0038] S104. Capture the evolution of geometric structure through ICP algorithm and build a deformation prediction model.
[0039] Obtain the point cloud data of the structure at different times, align the point cloud data at each time, and realize the alignment of the data in the same coordinate system. The motion and displacement information of the feature points on the surface of the structure is determined by point cloud registration, and the matrix variation method is used to decompose the rigid body motion and non-rigid body deformation. Based on the feature points, a deformation prediction model is constructed using a neural network algorithm. The coordinates and time information of the feature points are used as input. The deformation prediction model is trained to predict the displacement of the feature points in real time to obtain the deformation trend of the structure. According to the predicted displacement of the feature points, the deformation dynamics equation is constructed, and the structure is divided into a finite number of units. Each unit contains several nodes, which are discretized into algebraic equations on a finite number of nodes. The discretized equations are solved by the finite element method to simulate the deformation process of the structure. Set the deformation termination and convergence conditions. When the deformation is less than the set threshold, the deformation process ends.
[0040] For example, there is a building's structural data, and three point cloud data are measured at different times, namely, the point cloud data at time t1, t2, and t3. The three point cloud data are registered to align them in the same coordinate system. First, the ICP algorithm can be used to register the point cloud data at time t1 and t2. In this way, a transformation matrix of rigid body motion can be obtained, indicating that the point cloud data at time t1 is aligned with the point cloud data at time t2 after being transformed by the transformation matrix. Next, the ICP algorithm can be used to register the point cloud data at time t1 and t3. Similarly, a transformation matrix can be obtained, indicating that the point cloud data at time t1 is aligned with the point cloud data at time t3 after being transformed by the transformation matrix. By comparing these two transformation matrices, the motion and displacement information of the feature points on the surface of the structure can be obtained. The coordinates of a feature point at time t1 are P1, the coordinates at time t2 are P2, and the coordinates at time t3 are P3. The motion and displacement information of the feature point can be obtained by calculating the displacement vector of P2 relative to P1 and the displacement vector of P3 relative to P1. Based on these feature points, a deformation prediction model can be constructed. The model takes the coordinates and time information of the feature points as input to predict the displacement of the feature points. A neural network can be used to train this model so that it can accurately predict the displacement of the feature points. With the displacement prediction of the feature points, the deformation dynamics equation can be constructed. The structure is divided into several units, each of which contains multiple nodes. The deformation process of the structure can be discretized into algebraic equations on a finite number of nodes. Then, the finite element method can be used to solve the discretized set of equations to simulate the deformation process of the structure. By iteratively solving these equations, the deformation of the structure at different times can be obtained. During the simulation process, deformation termination and convergence conditions can be set. When the deformation of the structure is less than a set threshold, the deformation process can be considered to be over.
[0041] S105. Analyze the time scale characteristics of different physical quantities with the help of clustering and pattern recognition.
[0042] Obtain multi-source heterogeneous sequence data containing different physical quantities, normalize the sequence data, and eliminate dimensional differences. Take the normalized multi-source heterogeneous sequence data as input, build a clustering model based on cosine similarity, identify the category patterns in the data, and obtain pattern sequences representing different physical processes. According to the pattern sequence, the wavelet transform multi-scale analysis method is used to decompose the energy distribution of the sequence at different time scales, determine the main response scale of each category of physical quantity sequence, and determine the corresponding time characteristic scale. Using the EM clustering algorithm, calculate the probability that each energy distribution obtained by the wavelet transform multi-scale analysis belongs to each category, update the parameters of each category, and distinguish the physical process patterns at different time scales.
[0043] For example, when analyzing the velocity field and pressure field during the evolution of the vortex street, the multi-field Euler integral node data including the velocity components in the three directions of U, V, and W and the static pressure are first obtained. In order to eliminate the dimension difference, the data of each field is dimensionless by using the standardization method. Then, a clustering model based on vector cosine similarity is constructed to identify the two main modal structures representing the vortex core and the shear layer. The peak center of the vortex core mode is mainly distributed in the low-speed area, while the shear layer mode is concentrated in the high-strain area. Then, the two modal time series data are wavelet transformed respectively, and the Db4 wavelet basis is used to realize the 4-layer decomposition, and the energy distribution of each component is calculated. The analysis results show that the vortex core mode includes two main time scale components of 0.01s and 0.08s, while the main energy of the shear layer mode is concentrated in the 0.02s scale. In view of the coexistence of multiple characteristic time scales of different physical fields, the EM algorithm is then used to implement soft clustering. While retaining the maximum likelihood estimate, the association rules are integrated to distinguish the large eddy structure evolution corresponding to the low-frequency scale of the vortex core and the micro-vortex birth and death process corresponding to the high-frequency scale of the shear layer.
[0044] S106. Design a coupling step selection strategy based on comprehensive numerical stability indicators and computational costs.
[0045] The time step, the numerical solution of the gas phase and solid phase solver, as well as the grid size and grid node layout are used as input to construct a discrete error estimation model based on matrix permutation theory. The dynamic growth rate of the numerical solution in the alternating iteration process of the gas phase and solid phase solver under different time steps is obtained, and the numerical stability index is obtained. The number of memory accesses and floating-point calculations are determined to estimate the computational cost under different step lengths. The numerical stability index and the computational cost are combined, and the optimal step length is obtained by weighting according to the determined weight. The analytical form of the coupling equation is obtained, and the sensitivity of the determined equation to the coupling step length is taken as the target layer, the different equations as the preparation layer, and the specific equations as the solution layer. A judgment matrix for comparing factors at different levels is constructed. By performing eigenvalue decomposition on the judgment matrix, the weight vector of each level is calculated to represent the relative importance of the corresponding factor. According to the calculated weight vector, the sensitivity of different equations to the coupling step length is determined. The coupling sequence is determined according to the order of equation sensitivity to narrow the search space for step length selection. If a critical point where the coupling gain changes sharply occurs, the minimum generation condition is used to search for potential singular points, avoid the boundary step where the numerical solution collapses, and obtain a globally optimized coupling step combination scheme.
[0046] For example, when solving the numerical stability and calculation time of the gas phase and solid phase in the coupled solution, a discrete error estimation model based on matrix permutation theory is first constructed to calculate the dynamic growth rate of the numerical solution in the alternating iteration process of the gas phase and solid phase solvers at different time steps to obtain the numerical stability index. Then, the memory read and write volume and floating point calculation volume of the GPU are counted to estimate the calculation cost under different step lengths. Next, a numerical stability target with a priority weight ratio of 0.6 is set, and the weighted optimization method is used to obtain the optimal step length of the compromise criterion value of 10 microseconds. In addition, according to the analytical form of the equation, the temperature field in the solid phase equation is more sensitive to the change of the coupling step length. The solid phase step length is first fixed, and then the gas phase step length is determined by traversal to narrow the search range. During the traversal process, the coupling gain index is monitored. When the coupling gain of the temperature field and the velocity field changes drastically, it is considered to be close to the singular area, and the critical step length is recorded and the restricted area is demarcated. Finally, within the reasonable step size range, a coupling step size scheme with a numerical stability index greater than 0.85 and the shortest calculation time is selected. At the same time, an online monitoring program is set up, and the adjustment of the coupling step size is fed back by the spatiotemporal distribution of the temperature field and velocity field residuals to achieve coordinated optimization of computational stability and efficiency.
[0047] S107. Use the error integration method to evaluate the time discretization error and predict numerical stability.
[0048] Establish the analytical solution of the continuous system as the exact solution. Construct an equivalent time-discrete equation group and use different time steps Δt for space-time discretization. Solve the discrete equation group separately to obtain the corresponding numerical solution, compare it with the analytical solution, use the error integration method to compare the numerical solution with the analytical solution, and quantitatively analyze the discretization error of each physical quantity. According to the error estimation results, if the error increases with time, the matrix gain analysis theory is used to calculate the spectral radius of the numerical solution matrix of the two step schemes to judge the numerical stability. Determine the spectral radius of the gain matrix, and when the spectral radius is greater than the preset spectral radius threshold, it is judged to be unstable. If there is an unstable state, adjust the time step Δt, solve and evaluate again until a stable solution is obtained. Using the polynomial fitting method, a relationship model between the discretization error and Δt is established according to the stable Δt range to achieve the prediction of numerical stability under any Δt.
[0049] For example, when evaluating the time discretization error of the material diffusion equation, we first assume that the continuous system is the classic Fick second-order diffusion equation, and the analytical solution is obtained by the Boltzmann transformation method. Then, two time steps of 0.01s and 0.02s are used to establish the time discretization format of Crank-Nicolson finite difference form, and the spatial discretization adopts the second-order central difference format. The two discretization schemes are solved and compared with the analytical solution respectively. The L2norm error of the concentration field is quantitatively analyzed by the complex step method. The error curves show a trend of increasing with time. Then, based on matrix theory, the spectral radius of the numerical solution matrix of the two step schemes is calculated. The spectral radius of the 0.01s step is 0.98, and the spectral radius of the 0.02s step is greater than 1. It is predicted that the latter will cause numerical instability. To verify the prediction, the step size of 0.015s is used to re-solve and analyze. Under this step size, the error curve continues to increase, and the matrix spectral radius also exceeds 1, which verifies the prediction result. Finally, several stable step sizes between 0.01 and 0.014 s were integrated, and a cubic polynomial fitting was used to establish a relationship model between step size and error. This enabled the prediction of discrete error and stability under any step size, providing a basis for step size selection.
[0050] S108. Based on the error integration method and parameter identification method, online evaluation of the components is realized.
[0051] Obtain the spatiotemporal multi-resolution data generated by the simulation calculation, normalize the different physical quantity fields in the spatiotemporal multi-resolution data, and eliminate the influence of dimensional errors. Extract the exact solution or experimental data as a benchmark, build an error integration model, use the complex step size algorithm to divide the integration interval into several sub-intervals, and use different step sizes in each sub-interval, compare the numerical solution with the benchmark, and quantitatively analyze the numerical errors of each field variable. According to the error analysis results, judge whether the simulation accuracy meets the requirements. If not, use the feature tracking algorithm based on the optical flow method to identify the parameters with reduced calculation accuracy, including grid density and time step. Use the identified parameters with reduced calculation accuracy as input to build a parameter correction regression model, train the parameter sensitivity coefficients of the parameter correction regression model, obtain the corrected parameters, and update the parameters online to control the accuracy within the allowable range. When the parameter correction regression model or the grid is dynamically adaptively adjusted, rebuild the online evaluation component to achieve closed-loop control of simulation accuracy.
[0052] For example, when evaluating the accuracy of numerical simulation of heat transfer of aerogels, the temperature field data set calculated under different grid densities is first obtained, and the experimental measured temperature data under corresponding conditions are collected at the same time. In order to eliminate the dimension difference, the 0-1 normalization method is used to preprocess the numerical and experimental data. Then, the experimental data is selected as the benchmark, the L2 error integral model of the temperature field is constructed, and the temperature field error under each grid density is calculated using the third-order complex step algorithm. The analysis results show that the error of the 1 million unit grid is the largest, and the error of the 500,000 unit grid is smaller but still exceeds the accuracy requirement. Therefore, the feature tracking algorithm based on the optical flow method is used to analyze the gel volume change rate in the calculation process as the key parameter affecting the accuracy. Then, a parameter correction regression model is constructed to determine the volume change rate threshold as 0.5%, and the grid adaptation and reconstruction process are adjusted accordingly to control the volume change under the threshold requirement. After the model and grid are adjusted, the reconstruction error verification module is re-verified. After three iterative optimizations, the temperature field errors of each grid scheme meet the accuracy requirements, verifying the effectiveness of the evaluation component. When the gel breaks or reassembles, the error checking and parameter adjustment process is repeated to ensure the accuracy and stability of the numerical calculation.
[0053] S109. Adjust the next time step value through gradient iteration to optimize calculation accuracy.
[0054] In the decoupling model of the current time step, a test source term is introduced to perturb each coupling variable. The response fluctuation of the coupling field is Fourier transformed, the spectral response of different components is analyzed, and the main time scale of the coupling field is determined. According to the spectrum analysis, if there is a fast scale component that does not match the current time step well, the information of the gradient field at this scale is extracted as feedback for the next time step adjustment. In the current time step, the gradient field amplitude of each coupling field is measured and matched and compared with the preset spectral response range. If the gradient field exceeds the range, the gradient descent method is used to iteratively optimize and adjust the next time step so that the gradient field response conforms to the expected decoupling model. Repeat the test source excitation, gradient detection and time step adjustment process to achieve adaptive optimization control of time domain resolution.
[0055] For example, when coupling the aerogel phase change and heat transfer process, the current time step is 0.1 microseconds. First, a Gaussian pulse heat source term with a peak amplitude of 10K is added to the calculated temperature field at 100 microseconds to obtain the response fluctuation sequence of each field variable under excitation. Then, the response sequence is subjected to a fast Fourier transform, and it is determined that there is a high-frequency scale component near 0.2 microseconds, which cannot be effectively resolved by the current step. Therefore, the gradient distribution of the temperature field and phase field at 98 microseconds is extracted, and it is found that the gradient value of the local area corresponding to the pulse excitation point is large, and the peak value reaches the maximum allowable value of the temperature gradient of 30K / mm. Next, the preset temperature gradient spectrum range is set to 0-20K / mm, and it is determined that the current gradient exceeds the expected range. Therefore, a gradient error curve is constructed, and the gradient descent method is used to iterate and adjust the next time step to 0.08 microseconds. After updating the time step, the response sequence is parsed, and the temperature gradient field at 98 microseconds is checked. After optimization, the gradient peak is reduced to 15K / mm, which meets the expected range requirements. Through iterative energy saving of gradient spectrum matching, adaptive optimization of the time domain resolution of the phase change and heat transfer coupling process is achieved, thereby improving the accuracy and efficiency of numerical calculations.
[0056] The preferred embodiments of the present invention disclosed above are only used to help explain the present invention. The preferred embodiments do not describe all the details in detail, nor do they limit the invention to only specific implementation methods. Obviously, many modifications and changes can be made according to the content of this specification. This specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the present invention, so that those skilled in the art can understand and use the present invention well. The present invention is limited only by the claims and their full scope and equivalents.
Claims
1. A high performance computing method suitable for multi-scale simulation, characterized in that: The method comprises: Convolutional neural networks are used to determine the temporal and spatial characteristic scales in multi-scale physical processes. Based on the obtained temporal and spatial characteristic scales, a deep deterministic policy gradient algorithm is used to establish an automated time step control system. The octree adaptive grid implemented with a tree data structure is used to represent dynamic geometry. The ICP algorithm is used to capture the evolution of geometric structures and build a deformation prediction model. The time scale characteristics of different physical quantities are analyzed with the help of clustering and pattern recognition. A coupling step selection strategy is designed based on numerical stability indicators and computational costs. The error integral method is used to evaluate the time discretization error and predict numerical stability. Based on the error integral method and parameter identification method, online evaluation of components is achieved. The value of the next time step is adjusted through gradient iteration to optimize the calculation accuracy.
2. The method according to claim 1, wherein: The method of determining the time and space characteristic scales existing in the multi-scale physical process by using a convolutional neural network includes: Acquire multi-scale data in the physical process, extract the spatial and temporal features in the data through convolutional neural networks, and obtain spatial and temporal feature maps; use wavelet transform to perform multi-scale analysis on the feature maps, determine the characteristic responses at different scales, and determine the corresponding spatial and temporal feature scales; According to the determined spatial and temporal characteristic scales, the equivalent characteristic scale criterion is adopted to integrate the spatial and temporal dimensions to obtain the equivalent characteristic scale; The sliding window method is used to capture the evolution information of the characteristic scale by sliding a fixed-size window on the time series data.
3. The method according to claim 1, wherein: According to the obtained time and space feature scales, a deep deterministic policy gradient algorithm is used to establish an automated time step control system, including: Obtain the temporal and spatial characteristic scales of physical processes and construct a data set as the environment data for deep deterministic policy gradients; Set the time step as the action of the time step control system, and the error metric and stability index as the reward; use the deep deterministic policy gradient algorithm to build a deep neural network policy model, receive the current state as input, obtain the probability distribution of each action, interact with the environment, realize the autonomous setting of the time step, and obtain the optimization strategy of the time step; Acquire simulation data of different physical processes, retrain through transfer learning, adjust parameter weights, and optimize the weights of the deep neural network strategy model; when new physical situations arise, obtain data under the new situations and adjust the strategy network.
4. The method according to claim 1, wherein: The octree adaptive grid representation of dynamic geometry implemented by the tree data structure includes: The initial three-dimensional geometric model of the object is obtained, and the three-dimensional geometric model is segmented by an octree. The mesh topology relationship represented by the tree structure is obtained through recursive segmentation. The geometric features of each unit in the mesh are determined, and according to the feature changes and the preset accuracy requirements, a local refinement / coarsening method is used to insert a refined mesh in the outer area of the surface and generate a coarsening mesh inside the area to achieve adaptive adjustment of the mesh. The topological evolution information of the geometric structure is extracted, the topological relationship of the octree mesh is updated, and the octree nodes are inserted or deleted to achieve the update of the dynamic mesh topology. The mesh accuracy is determined again for the updated mesh to check whether it meets the requirements for the representation of dynamic geometry. If not, the mesh distribution is adjusted and optimized through the local refinement / coarsening method until the requirements for the representation of dynamic geometry are met.
5. The method according to claim 1, wherein: The ICP algorithm is used to capture the evolution of the geometric structure and construct a deformation prediction model, including: Obtain the point cloud data of the structure at different times, align the point cloud data at each time, and realize the alignment of the data in the same coordinate system; The motion and displacement information of the feature points on the surface of the structure is determined through point cloud registration, and the matrix variation method is used to decompose the rigid body motion and non-rigid body deformation; based on the feature points, a deformation prediction model is constructed using a neural network algorithm, and the coordinates and time information of the feature points are used as input. The deformation prediction model is trained to predict the displacement of the feature points in real time to obtain the deformation trend of the structure; based on the predicted displacement of the feature points, a deformation dynamics equation is constructed, and the structure is divided into a finite number of units, each unit contains a number of nodes, and discretized into algebraic equations on a finite number of nodes. The discretized equations are solved by the finite element method to simulate the deformation process of the structure; Set the deformation termination and convergence conditions. When the deformation amount is less than the set threshold, the deformation process ends.
6. The method according to claim 1, wherein: The time scale characteristics of different physical quantities are analyzed by clustering and pattern recognition, including: Acquire multi-source heterogeneous sequence data containing different physical quantities, normalize the sequence data, and eliminate dimensional differences; use the normalized multi-source heterogeneous sequence data as input, build a clustering model based on cosine similarity, identify the category patterns in the data, and obtain pattern sequences representing different physical processes; based on the pattern sequence, use the wavelet transform multi-scale analysis method to decompose the energy distribution of the sequence at different time scales, determine the main response scale of each category of physical quantity sequence, and determine the corresponding time characteristic scale; The EM clustering algorithm is used to calculate the probability that each energy distribution obtained by wavelet transform multi-scale analysis belongs to each category, update the parameters of each category, and distinguish the physical process patterns at different time scales.
7. The method according to claim 1, wherein: The comprehensive numerical stability index and computational cost are used to design a coupled step size selection strategy, including: Taking the time step, the numerical solutions of the gas phase and solid phase solvers, as well as the mesh size and mesh node layout as input, a discrete error estimation model based on matrix permutation theory is constructed to obtain the dynamic growth rate of the numerical solutions in the alternating iteration process of the gas phase and solid phase solvers at different time steps, and obtain the numerical stability index; Determine the number of memory accesses and floating-point calculations, and estimate the computational cost under different step lengths; combine numerical stability indicators and computational costs, and perform weighted calculations according to the determined weights to obtain the optimal step length of compromise; obtain the analytical form of the coupling equation, determine the sensitivity of the equation to the coupling step length as the target layer, different equations as the preparation layer, and the specific equation as the solution layer, and construct a judgment matrix that compares factors at different levels in pairs; calculate the weight vector of each level by performing eigenvalue decomposition on the judgment matrix to indicate the relative importance of the corresponding factors; determine the sensitivity of different equations to the coupling step length based on the calculated weight vector; determine the coupling sequence based on the order of equation sensitivity to narrow the search space for step length selection; If a critical point where the coupling gain changes sharply occurs, the minimum generation condition is used to search for potential singular points, avoid the boundary step where the numerical solution collapses, and obtain a globally optimized coupling step combination scheme.
8. The method according to claim 1, wherein: The method of using an error integration method to evaluate the time discretization error and predict numerical stability includes: Establish the analytical solution of the continuous system as the exact solution; construct an equivalent time-discrete equation group, and use different time steps Δt for space-time discretization; solve the discrete equation group separately, obtain the corresponding numerical solution, compare it with the analytical solution, use the error integration method to compare the numerical solution with the analytical solution, and quantitatively analyze the discretization error of various physical quantities; according to the error estimation results, if the error increases with time, use the matrix gain analysis theory to calculate the spectral radius of the numerical solution matrix of the two step schemes to judge the numerical stability; Determine the spectral radius of the gain matrix. When the spectral radius is greater than the preset spectral radius threshold, it is judged to be unstable. If an unstable state exists, adjust the time step Δt, solve and evaluate again until a stable solution is obtained. Use the polynomial fitting method to establish a relationship model between the discretization error and Δt according to the stable Δt range, and realize the prediction of numerical stability under arbitrary Δt.
9. The method according to claim 1, wherein: The online evaluation of components based on the error integration method and the parameter identification method includes: Obtain the spatiotemporal multi-resolution data generated by simulation calculations, normalize the different physical quantity fields in the spatiotemporal multi-resolution data, and eliminate the influence of dimensional errors; Extract the exact solution or experimental data as the benchmark, build the error integration model, use the complex step algorithm to divide the integration interval into several sub-intervals, and use different step sizes in each sub-interval, compare the numerical solution with the benchmark, and quantitatively analyze the numerical errors of each field variable; Based on the error analysis results, determine whether the simulation accuracy meets the requirements; If not, use the feature tracking algorithm based on the optical flow method to identify the parameters where the calculation accuracy decreases, including grid density and time step; The identified parameters with reduced calculation accuracy are used as input to build a parameter correction regression model, and the parameter sensitivity coefficients of the parameter correction regression model are trained to obtain the corrected parameters. The parameters are then updated online to keep the accuracy within the allowable range. When the parameter correction regression model or the grid is dynamically adaptively adjusted, the online evaluation component is rebuilt to achieve closed-loop control of the simulation accuracy.
10. The method according to claim 1, wherein: The step of adjusting the next time step value by gradient iteration to optimize the calculation accuracy includes: In the decoupling model of the current time step, a test source term is introduced to perturb each coupling variable; the response fluctuation of the coupling field is Fourier transformed, the spectrum response of different components is analyzed, and the main time scale of the coupling field is determined; According to the spectrum analysis, if there is a fast-scale component that does not match the current time step well, the information of the gradient field at this scale is extracted as feedback for the next time step adjustment; in the current time step, the gradient field amplitude of each coupling field is measured and matched and compared with the preset spectrum response range; if the gradient field exceeds the range, the gradient descent method is used to iteratively optimize and adjust the next time step so that the gradient field response meets the expected decoupling model; Repeat the test source excitation, gradient detection and time step adjustment process to achieve adaptive optimization control of time domain resolution.
Citation Information
Cited By
Wind wave and flow coupling environment modeling method of FPSO system based on modelica
CN119337782A
Multi-scale cloud system dynamic evolution simulation modeling method and system based on digital twinning
CN119885657A
Internal temperature simulation system of gravimeter system
CN119918363A
Step length data coupling simulation method for considering multi-time scale equipment
CN120745435A
A step data coupling simulation method for accounting for multi-time scale devices
CN120745435B
