Search

CN-121525525-B - Electromagnetic simulation iso-surface extraction method based on spatial rearrangement and deep learning prediction

CN121525525BCN 121525525 BCN121525525 BCN 121525525BCN-121525525-B

Abstract

The invention belongs to the field of computational electromagnetics, and provides an electromagnetic simulation equivalent surface extraction method based on spatial rearrangement and deep learning prediction, which is used for solving the problems of computational resource waste, access bottleneck of unstructured grid data, insufficient instantaneity and the like in the prior art; firstly, carrying out memory rearrangement on unstructured grid data output by electromagnetic simulation by using Morton codes, forcibly continuing geometrically adjacent field points in a physical memory, converting random access into stream access, greatly improving the bandwidth utilization rate, then constructing a lightweight fully-connected neural network, inputting field scalar values and target equivalence of the vertexes of grid units, outputting the probability that the grid units pass through an isosurface, carrying out quick batch reasoning on the rearranged grid data by using a trained network model, eliminating most of background airspace with extremely low calculation cost, generating an active unit candidate list, and finally, executing accurate geometric interpolation on the active units to generate high-quality isosurface grids.

Inventors

  • ZHANG LI
  • XU LI
  • WANG HAO
  • HU YULU
  • LIU BINGQI
  • LI BIN

Assignees

  • 电子科技大学

Dates

Publication Date
20260512
Application Date
20260114

Claims (6)

  1. 1. The electromagnetic simulation iso-surface extraction method based on spatial rearrangement and deep learning prediction is characterized by comprising the following steps of: Step 1, three-dimensional unstructured grid data generated by electromagnetic simulation and corresponding electromagnetic field scalar values are obtained, and Morton codes are adopted to conduct memory rearrangement on the grid data, so that space rearrangement is completed; step 2, constructing a fully-connected neural network and completing training to obtain a prediction model, wherein the input of the prediction model is that the field scalar values of all vertexes of a grid unit are equivalent to corresponding targets, and the prediction model outputs the active probability that the grid unit contains an equivalent surface; The fully-connected neural network is formed by connecting an input layer, two hidden layers and an output layer, wherein the input layer receives input data , Representing the first of the grid cells The field scalar values for the individual vertices, As the total number of vertices of the grid cell, The neuron number of the two hidden layers is 128 and 64 in sequence, and the activation function is a linear rectification function; step 3, inputting the grid data subjected to spatial rearrangement into a prediction model in batches for probability prediction, setting a probability threshold value, and reserving all grid cells with probability prediction values larger than the probability threshold value to form an active cell candidate list; And 4, executing geometric interpolation on the grid cells in the active cell candidate list to generate the isosurface grid.
  2. 2. The electromagnetic simulation iso-surface extraction method based on spatial rearrangement and deep learning prediction according to claim 1, wherein the specific process of step 1 is as follows: step 1.1, calculating the barycenter coordinates of each grid cell in three-dimensional unstructured grid data generated by electromagnetic simulation; Step 1.2, quantizing the triaxial coordinates of the centroid coordinates of each grid unit into integers, and generating Morton codes by adopting a bit crossing algorithm; step 1.3, performing global ascending order on the grid cells and the corresponding electromagnetic field data by taking the Morton code as an index.
  3. 3. The electromagnetic simulation iso-surface extraction method based on spatial rearrangement and deep learning prediction according to claim 2, wherein in step 1.2, the quantization process adopts an integer centroid quantization method.
  4. 4. The method for extracting an electromagnetic simulation iso-surface based on spatial rearrangement and deep learning prediction according to claim 1, wherein in step 2, the field scalar value includes an electric field modulus value or a magnetic field modulus value.
  5. 5. The electromagnetic simulation iso-surface extraction method based on spatial rearrangement and deep learning prediction according to claim 1, wherein in step 3, the threshold value is in a value range of 0.01.
  6. 6. The electromagnetic simulation iso-surface extraction method based on spatial rebinning and deep learning prediction according to claim 1, wherein in step 4, a mobile tetrahedron algorithm is used for geometric interpolation.

Description

Electromagnetic simulation iso-surface extraction method based on spatial rearrangement and deep learning prediction Technical Field The invention belongs to the field of computational electromagnetics, relates to a method for rapidly extracting an isosurface (Isosurface) of large-scale unstructured grid data in electromagnetic field simulation (Electromagnetic Simulation) post-processing, and particularly provides an electromagnetic simulation isosurface extraction method based on spatial rearrangement and deep learning prediction. Background In computational electromagnetics (Computational Electromagnetics, CEM), finite element methods (FINITE ELEMENT methods, FEM) or finite volume methods (Finite Volume Method, FVM) are often used to solve maxwell's equations to produce massive amounts of three-dimensional unstructured grid data, and engineers need to analyze antenna radiation patterns, radar cross-sections (Radar Cross Section, RCS) or electromagnetic compatibility (Electromagnetic Compatibility, EMC) problems by extracting isosurfaces (e.g., electric field strength isosurfaces, magnetic flux density isosurf). However, the existing iso-surface extraction method still faces many challenges when processing electromagnetic simulation large-scale data, firstly, calculation waste caused by sparse features is generated, electromagnetic field data often has extremely strong spatial locality, for example, a high field intensity region is only concentrated at a feed source or a gap, a traditional moving cube/tetrahedron algorithm (Marching Cubes/TETRAHEDRA) adopts a full-scale traversal strategy, more than 90% of calculation time is wasted in a background region with a gentle field intensity value or a near zero in data of thousands of grid units, secondly, the electromagnetic simulation often adopts a tetrahedron adaptive grid to attach to complex geometric boundaries, the storage sequence of the grid in a memory is usually random, so that the Cache (Cache) hit rate of a central processing unit/graphics processor (CPU/GPU) is extremely low when field value interpolation is performed, the interactive frame rate of post-processing software is seriously restricted, thirdly, the real-time performance is insufficient, the construction time of an existing acceleration structure (such as an octree) is long, and the real-time analysis of parameters (PARAMETRIC SWEEP) or transient fields (TRANSIENT FIELD) in electromagnetic simulation cannot be met. Aiming at the problems, the invention provides an electromagnetic simulation iso-surface extraction method based on spatial rearrangement and deep learning prediction. Disclosure of Invention The invention aims to provide an isosurface extraction method combining space filling curve rearrangement (Spatial Reordering) and deep neural network Prediction (Deep Neural Network Prediction, DNN Prediction) to solve the problems of calculation resource waste, access bottleneck of unstructured grids, insufficient instantaneity and the like in the prior art. In order to achieve the above purpose, the invention adopts the following technical scheme: The electromagnetic simulation iso-surface extraction method based on spatial rearrangement and deep learning prediction comprises the following steps: Step 1, three-dimensional unstructured grid data generated by electromagnetic simulation and corresponding electromagnetic field scalar values are obtained, and Morton codes are adopted to conduct memory rearrangement on the grid data, so that space rearrangement is completed; step 2, constructing a fully-connected neural network and completing training to obtain a prediction model, wherein the input of the prediction model is that the field scalar values of all vertexes of a grid unit are equivalent to corresponding targets, and the prediction model outputs the active probability that the grid unit contains an equivalent surface; step 3, inputting the grid data subjected to spatial rearrangement into a prediction model in batches for probability prediction, setting a probability threshold value, and reserving all grid cells with probability prediction values larger than the probability threshold value to form an active cell candidate list; And 4, executing geometric interpolation on the grid cells in the active cell candidate list to generate the isosurface grid. Further, the specific process of the step 1 is as follows: step 1.1, calculating the barycenter coordinates of each grid cell in three-dimensional unstructured grid data generated by electromagnetic simulation; Step 1.2, quantizing the triaxial coordinates of the centroid coordinates of each grid unit into integers, and generating Morton codes by adopting a bit crossing algorithm; step 1.3, performing global ascending order on the grid cells and the corresponding electromagnetic field data by taking the Morton code as an index. Further, in step 1.2, the quantization process uses an integer centroid quantization method. Further, in