Search

CN-122021308-A - Chip array field intensity distribution prediction method and system based on spatial information

CN122021308ACN 122021308 ACN122021308 ACN 122021308ACN-122021308-A

Abstract

The invention discloses a chip array field intensity distribution prediction method and system based on spatial information, and relates to the technical field of chip array electromagnetic field prediction. The method comprises the steps of constructing a multi-channel input tensor, carrying out pixel-level fusion on a spatial information pixel matrix representing geometric layout and a frequency matrix representing working state, enabling a network to learn local physical dependency, adopting an asymmetric downsampling codec architecture E-300Net, and carrying out convolution and deconvolution operation with a mixed configuration stride of 2 and a stride of 3 in a downsampling and upsampling path so as to ensure that feature images which are input in any size can be accurately aligned and spatial details with high fidelity are reserved. The invention obviously improves the simulation efficiency, shortens the simulation time to the second level, simultaneously realizes the electric field strength prediction with high precision and high generalization capability, and meets the engineering requirements of chip electromagnetic compatibility and signal integrity analysis.

Inventors

  • Fan Dailai
  • FENG NAIXING

Assignees

  • 安徽大学

Dates

Publication Date
20260512
Application Date
20260129

Claims (9)

  1. 1. The chip array field intensity distribution prediction method based on the spatial information is characterized by comprising the following specific steps of: obtaining an original data set by using COMSOL simulation software, wherein the original data set comprises a pixel matrix, a frequency matrix and an electric field intensity matrix; The original data set is divided into a training set, a verification set and a test set according to a preset proportion after data preprocessing; performing pixel-level fusion on the pixel matrix and the frequency matrix to generate a multi-channel input tensor as input data of subsequent training, wherein the electric field intensity matrix is used as target data of training; Constructing a convolution encoder-decoder architecture of an asymmetric downsampling structure, and mixing convolution and deconvolution operations with steps of 2 and 3 in a downsampling and upsampling path; The convolutional encoder-decoder architecture is trained by the training set, and input data is processed using the trained convolutional encoder-decoder architecture to generate a predicted electric field strength distribution.
  2. 2. The method for predicting field intensity distribution of a chip array based on spatial information as set forth in claim 1, wherein said pixel matrix is used for characterizing geometric layout and material distribution of the chip array, and said frequency matrix has a dimension identical to a dimension of said pixel matrix for specifying an operating frequency of a corresponding position at a pixel level.
  3. 3. The method for predicting field intensity distribution of chip array based on spatial information as set forth in claim 1, wherein said original dataset is stored as a CSV file, and said pixel matrix, said frequency matrix, and said electric field intensity matrix stored in a CSV file format are converted into NPY file format, respectively.
  4. 4. The method of claim 1, wherein the preprocessing of the raw data set includes normalizing, in which global normalization parameters are calculated based only on the data distribution of the training set, and the normalization parameters are determined using a percentile clamping strategy.
  5. 5. The method for predicting field strength distribution of chip array based on spatial information as set forth in claim 1, wherein a smoothed L1 loss function is used as a loss function and an Adam optimizer is used for parameter optimization in training the convolutional encoder-decoder architecture.
  6. 6. The method for predicting field intensity distribution of chip array based on spatial information as set forth in claim 1, wherein the performance of the current model on a key performance index is evaluated after each round of training or verification is completed, wherein the operation of storing the model is triggered if and only if the performance index of the current round is better than the historical optimal value recorded by all previous rounds, the parameter state of the current model is stored as the optimal model in a lasting manner, and the storing is not performed if the performance of the current round does not exceed the historical optimal value.
  7. 7. The method of claim 1, wherein the pixel matrix and the frequency matrix are stacked along a channel dimension to generate the multi-channel input tensor.
  8. 8. A chip array field intensity distribution prediction system based on space information is characterized by comprising a data set acquisition module, a data set dividing module, an input data construction module, a deep learning model construction module and a training module, wherein, The data set acquisition module is used for acquiring an original data set by utilizing COMSOL simulation software, wherein the original data set comprises a pixel matrix, a frequency matrix and an electric field intensity matrix; The data set dividing module is used for dividing the original data set into a training set, a verification set and a test set according to a preset proportion after data preprocessing; the input data construction module is used for carrying out pixel level fusion on the pixel matrix and the frequency matrix to generate a multi-channel input tensor which is used as input data of subsequent training, and the electric field intensity matrix is used as target data of training; the deep learning model building module is used for building a convolution encoder-decoder framework of an asymmetric downsampling structure, and in a downsampling and upsampling path, convolution and deconvolution operations with steps of 2 and 3 are mixed and configured; the training module is configured to train the convolutional encoder-decoder architecture through the training set, and process input data using the trained convolutional encoder-decoder architecture to generate a predicted electric field intensity distribution.
  9. 9. The spatial information based chip array field strength distribution prediction system of claim 8, further comprising a visualization module for visually displaying training results, validation results, and test results.

Description

Chip array field intensity distribution prediction method and system based on spatial information Technical Field The invention relates to the technical field of chip array electromagnetic field prediction, in particular to a chip array field intensity distribution prediction method and system based on spatial information. Background With the continued advancement of semiconductor technology, chips, and particularly chip arrays and systems on a chip, are evolving toward high integration, high operating frequencies, and high power densities. With this, the electromagnetic environment inside and between chip arrays is becoming increasingly complex, and problems such as electromagnetic interference, signal integrity, and power integrity have become key challenges that limit chip performance. The electric field distribution inside the chip array is accurately predicted, and is important for early performance evaluation, electromagnetic compatibility analysis and later optimization of the design. Traditional numerical simulation methods, such as Finite Element Method (FEM) or finite difference time domain method (FDTD), are currently accepted "gold standards". However, its high accuracy relies on fine meshing and complex iterative calculations, resulting in extremely time consuming calculation processes. The simulation bottleneck is seriously delayed from the requirement of rapid iteration of modern complex chip array design, so that the development of an alternative technology capable of rapidly and accurately predicting chip array field intensity distribution has great research significance and application value for improving chip design efficiency and shortening research and development period. To address this challenge, researchers have begun exploring the use of deep learning methods, particularly Convolutional Neural Networks (CNNs), to build proxy models of physical fields to replace or accelerate traditional physics simulations. However, the direct application of such known architectures to physical field prediction presents fundamental challenges. First, the physical field prediction task is highly dependent on accurate characterization of spatial information and physical parameters. Physically, the spatial information is a boundary condition that determines the field distribution, and the physical parameter is the excitation source. Conventional convolutional networks lack a mechanism to explicitly distinguish and process these two heterogeneous information. In the depth feature extraction process, the high-fidelity spatial information is irreversibly coupled with the physical parameter information. In this mode, the effectiveness of the jump connection mechanism, which is often used to preserve shallow details, is also fundamentally limited. While this mechanism is intended to preserve shallow information for subsequent reconstruction, the features they preserve and convey are themselves already data-mixed products, rather than pure, high-fidelity spatial geometry information. When a decoder attempts to reconstruct the field strength, it cannot effectively decouple the pure spatial boundary constraints from such hybrid features, resulting in a reconstructed field strength distribution that necessarily loses accuracy. Therefore, when high-fidelity physical field prediction is required, the lack of spatial awareness exposes the fundamental limitation of the known architecture, which results in low spatial fidelity, blurred edges and peak position deviation of the prediction result, and is difficult to meet the strict chip design requirements. Therefore, it is a urgent problem to be solved for those skilled in the art how to design a novel convolution codec intelligent algorithm specifically optimized for physical field prediction, so as to implement fast and high-precision prediction of field intensity distribution. Disclosure of Invention The invention aims to provide a chip array field intensity distribution prediction method and system based on spatial information, which are used for solving the problems in the background technology, and can rapidly and highly accurately predict the electric field intensity distribution of a chip array according to the structural layout and the working frequency of the chip array and deduce the global electric field response under a complex electromagnetic environment. In order to achieve the above purpose, the invention provides the following scheme, on the one hand, a chip array field intensity distribution prediction method based on spatial information is provided, and the specific steps comprise: obtaining an original data set by using COMSOL simulation software, wherein the original data set comprises a pixel matrix, a frequency matrix and an electric field intensity matrix; The original data set is divided into a training set, a verification set and a test set according to a preset proportion after data preprocessing; performing pixel-level fusion on the pixel