摘要
针对封装基板的翘曲预测问题,提出一种基于循环神经网络(RNN)与双向长短期记忆(Bi-LSTM)网络相结合的机器学习方法,构建封装基板翘曲预测模型。该模型可预测非对称基板翘曲分布,并有效提高预测效率与准确性。为获取模型训练所需数据集,开发了随机游走自动布线算法,生成不同特征的基板布线结构,并利用铜迹线强化有限元分析(FEA)方法获取翘曲分布数据。研究结果表明,Bi-LSTM网络模型在80个训练周期内误差收敛至0.05 mm^(2)以下,结构相似性衡量指标(SSIM)均大于0.7;在非训练集铜布线验证样本上表现出良好的泛化能力,并且预测时间仅需数秒,预测速度显著快于FEA,为基板设计提供了快速、准确的翘曲预测新途径,有助于提高优化迭代效率。
To address the warpage prediction problem of packaging substrates,a machine learning method combining recurrent neural networks(RNN)and bidirectional longshort-termmemory(Bi-LSTM)networks was proposed to construct a packaging substrate warpage prediction model.The model can predict the distribution of non-symmetrical substrate warping and effectively improve prediction effi-ciency and accuracy.To obtain the dataset required for model training,a random walk-based automatic routing algorithm was developed to generate substrate routing structures with different features.Copper trace reinforcement was applied using the finite element analysis(FEA)method to acquire warpage dis-tribution data.The research results show that the Bi-LSTM network model converges to an error below 0.05 mm^(2)within 80 training cycles,with the structural similarity index measurement(SSIM)all greater than O0.7.It exhibits good generalization ability on non-training set copper routing validation samples.The prediction time is only a few seconds,and the prediction speed is significantly faster than that of FEA.This provides a fast and accurate new approach for warpage prediction in substrate design,which can help to improve the efficiency of optimization iterations.
作者
王昊舟
王珺
Wang Haozhou;Wang Jun(College of Smart Materials and Future Energy,Fudan University,Shanghai 200438,China)
出处
《半导体技术》
2025年第10期1057-1066,共10页
Semiconductor Technology
作者简介
通信作者:王珺,(1972-),男,云南曲靖人,博士,教授,硕士生导师,主要从事集成电路封装可靠性研究;王昊舟(2000-),男,陕西西安人,硕士研究生,从要从事封装仿真技术研究。
