摘要
探讨利用机器学习方法预测PINMAP封装的辐射。首先,为解决全波仿真方法在计算封装电磁辐射时的时效性问题,构建的卷积神经网络(CNN)模型在5 GHz以上频率实现了小于2%的平均相对误差,且预测时间达到毫秒级,显著提高了封装远场辐射的计算效率。其次,为了解决封装设计中的辐射抑制问题,引入了逆向结构设计网络以达到辐射预期。该网络融合逆向Transformer与正向CNN的级联架构,优化了网络的收敛性和预测准确性。最后,开发了一种接地焊球位置优化方案来优化封装结构并抑制辐射,通过微调接地焊球的位置,设计了多个新的封装模型,并使用已训练的CNN模型进行辐射预测,从而选出辐射抑制效果最佳的优化模型。并通过实例分析验证了这些方法的有效性,这两种辐射抑制技术在缓解越来越小型化的电子设备的电磁兼容问题上具有广泛的应用前景。
This paper delves into the application of machine learning methods to predict the radiation of PINMAP packages.Firstly,in response to the timeliness issue of calculating package electromagnetic radiation using fullwave simulation methods,the convolutional neural network(CNN)model established in this paper can achieve an average relative error of less than 2%after 5 GHz.The model has a prediction time of milliseconds for radiation,which can effectively improve the calculation time of package far-field radiation.Secondly,in response to the radiation suppression issue inherent in package design,this paper proposes a reverse structural design network to meet radiation expectations.The inverse network adopts a cascaded architecture of inverse Transformer and forward CNN to make it easier to converge and predict more accurately.Finally,in response to the radiation suppression problem after package design,this paper proposes a position optimization scheme for the ground balls to achieve structural optimization of package to suppress radiation.This method designs several new package models by adjusting the positon of ground balls in a small range,and combines them with the trained CNN model for radiation prediction.After calculation and comparison,it is found that the new model with the best radiation reduction effect is the model optimized by the structure.The effectiveness of the above method was verified through typical cases.Therefore,the two radiation suppression methods proposed in this paper have great potential in mitigating the electromagnetic compatibility(EMC)issues of increasingly miniaturized electronic devices.
作者
李燕
杨啸天
Li Yan;Yang Xiaotian
基金
国家自然科学基金面上项目(62071424)
浙江省自然科学基金(LD21F010002)。
关键词
电磁干扰
机器学习
辐射预测
辐射抑制
逆向设计
electromagnetic interference
machine learning
radiation prediction
radiation suppression
inverse design
