基于多模态卷积神经网络的翼型管翅片散热器形状与布局参数优化

Shape and Layout Parameter Optimization of an Airfoil Tube-Fin Radiator Based on a Multimodal Convolutional Neural Network

摘要: 针对翼型管翅片散热器多参数设计中CFD逐点寻优成本高、纯几何代理模型难以表征局部流场演化的问题，提出一种融合流场图像与几何参数的多模态CNN形状与布局参数优化方法。以NACA0012对称翼型管阵列为基准构型，构建包含弯度、最大弯度位置、厚度、纵向间距、横向间距和交错位移的6维参数化设计空间，并基于COMSOL获得不同结构下的速度场、温度场、压力场及Nu、阻力系数f等性能指标。在此基础上，建立引入CBAM注意力机制的多模态双流CNN代理模型，并将其作为差分进化算法的快速适应度评估器。测试结果表明，模型对Nu和f的预测决定系数分别达到0.9756和0.9633，较纯几何MLP表现出更高预测精度。经CFD复核，最优结构的Nu、总换热率$Q_{total}$ 和综合评价因子$\eta$ 分别提升6.03%、0.25%和3.93%，翼型管平均温升降低约5.45%，同时阻力系数f增加6.18%，说明该结构在散热增强与流阻代价之间取得了一定折中。该方法可通过“代理预测–智能筛选–CFD复核”流程降低候选结构筛选成本，为紧凑式散热器快速优化提供参考。

Abstract: To reduce the high computational cost of point-by-point CFD optimization and overcome the limited ability of geometry-only surrogate models to describe local flow-field evolution, a multimodal CNN-based shape and layout parameter optimization method integrating flow-field images with geometric parameters is proposed for an airfoil tube-fin radiator. Taking the NACA 0012 symmetric airfoil-tube array as the baseline configuration, a six-dimensional parametric design space is constructed, including camber, maximum camber position, thickness, longitudinal spacing, transverse spacing, and staggering offset. The velocity, temperature, and pressure fields, as well as the Nusselt number and resistance coefficient of different configurations, are obtained using COMSOL. On this basis, a multimodal dual-stream CNN surrogate model with a CBAM attention mechanism is developed and used as a fast fitness evaluator in the differential evolution algorithm. The test results show that the coefficients of determination for Nu and f reach 0.9756 and 0.9633, respectively, indicating higher prediction accuracy than the geometry-only MLP model. According to independent CFD verification, the optimal structure increases Nu, the unit-depth equivalent heat transfer rate $Q_{total}$ , and the comprehensive evaluation factor $\eta$ by 6.03%, 0.25%, and 3.93%, respectively, while reducing the average temperature rise of the airfoil tube by approximately 5.45%; meanwhile, the resistance coefficient f increases by 6.18%. The results indicate that the optimized structure achieves a reasonable compromise between heat dissipation enhancement and flow resistance penalty. The proposed “surrogate prediction - intelligent screening - CFD verification” procedure can reduce the screening cost of candidate structures and provide a reference for the rapid optimization of compact radiators.