Skip to main content
QUICK REVIEW

[论文解读] Data-Driven Modeling of Geometry-Adaptive Steady Heat Transfer based on Convolutional Neural Networks: Heat Conduction

Jiang-Zhou Peng, Xianglei Liu|arXiv (Cornell University)|Oct 8, 2020
Advanced Numerical Analysis Techniques参考文献 64被引用 5
一句话总结

本文提出了一种基于卷积神经网络(CNN)的数据驱动模型,用于对具有任意几何形状的二维物体实现快速、精确的稳态热传导预测。通过使用符号距离函数(SDF)编码复杂形状,并在多样化几何构型上进行训练,该模型在保持对未见过几何形状高精度的同时,推理速度比传统数值方法快三到四个数量级。

ABSTRACT

Numerical simulation of steady-state heat conduction is common for thermal engineering. The simulation process usually involves mathematical formulation, numerical discretization and iteration of discretized ordinary or partial differential equations depending on complexity of problems. In current work, we develop a data-driven model for extremely fast prediction of steady-state heat conduction of a hot object with arbitrary geometry in a two-dimensional space. Mathematically, the steady-state heat conduction can be described by the Laplace's equation, where a heat (spatial) diffusion term dominates the governing equation. As the intensity of the heat diffusion only depends on the gradient of the temperature field, the temperature distribution of the steady-state heat conduction displays strong features of locality. Therefore, in current approach the data-driven model is developed using convolution neural networks (CNNs), which is good at capturing local features (sub-invariant) thus can be treated as numerical discretization in some sense. Furthermore, in our model, a signed distance function (SDF) is proposed to represent the geometry of the problem, which contains more information compared to a binary image. For the training datasets, the hot objects are consisting of five simple geometries: triangles, quadrilaterals, pentagons, hexagons and dodecagons. All the geometries are different in size, shape, orientation and location. After training, the data-driven network model is able to accurately predict steady-state heat conduction of hot objects with complex geometries which has never been seen by the network model; and the prediction speed is three to four orders faster than numerical simulation. According to the outstanding performance of the network model, it is hoped that this approach can serve as a valuable tool for applications of engineering optimization and design in future.

研究动机与目标

  • 开发一种数据驱动模型,实现对任意二维几何形状的超快速稳态热传导预测。
  • 克服传统数值求解器在不规则域中求解拉普拉斯方程时的计算成本和复杂性。
  • 利用CNN的局部特征提取能力,近似热传导方程的数值离散化。
  • 引入符号距离函数(SDF)作为几何表示方法,相较于二值掩码更有效地编码形状细节。
  • 训练单一模型,使其能够泛化至未见过的几何形状,包括训练集中未出现的复杂和新颖形状。

提出的方法

  • 该模型使用卷积神经网络(CNN)学习从几何输入到稳态温度场的映射关系。
  • 通过符号距离函数(SDF)编码几何信息,提供连续且可微的形状信息,超越简单的二值掩码。
  • 训练数据集包含五种基础形状——三角形、四边形、五边形、六边形和十二边形——在尺寸、方向和位置上均有变化的二维热传导问题。
  • 通过最小化预测解与真实解之间的均方误差,训练CNN以预测稳态温度分布。
  • 通过端到端训练,模型隐式学习拉普拉斯方程的数值离散化,利用温度梯度的局部性特征。
  • 通过数据增强和CNN对局部模式识别的归纳偏差,实现对未见几何形状的泛化能力。

实验结果

研究问题

  • RQ1在多样化二维几何形状上训练的深度学习模型,能否泛化至预测完全未见过的形状的稳态热传导?
  • RQ2与二值图像表示相比,使用符号距离函数(SDF)在几何保真度和模型性能方面有何差异?
  • RQ3CNN在不显式进行数值离散化的情况下,能在多大程度上近似拉普拉斯方程的解?
  • RQ4与传统的有限元法或有限差分求解器相比,该数据驱动模型的加速比是多少?
  • RQ5当预测复杂且不可训练的几何形状的温度场时,模型的精度如何?

主要发现

  • 所提出的CNN模型预测速度比传统数值模拟快三到四个数量级。
  • 该模型能准确泛化至未见过的几何形状,包括训练集中未出现的复杂形状。
  • 使用符号距离函数(SDF)作为输入,相比二值掩码,能提供更优的几何表示并提升模型性能。
  • 该模型在各种形状、尺寸、方向和位置下均保持高精度,展现出强大的泛化能力。
  • 网络通过捕捉局部温度梯度,隐式学习了热传导的物理规律,有效模拟了数值离散化过程。
  • 由于其高速度和高精度,该方法可实现工程应用中热设计的快速探索与优化。

更好的研究,从现在开始

从阅读论文到最终审阅,大幅缩短您的研究时间。

无需绑定信用卡

本解读由 AI 生成,并经人工编辑审核。