随着集成电路工艺的进步，芯片特征尺寸不断缩小，芯片集成度不断增加，传统电子设计自动化（Electronic Design Automation，EDA）方法面临重大挑战。近年来，机器学习技术的发展为解决集成电路EDA难题提供了新的研究思路。本文针对集成电路EDA流程中最为复杂费时的物理设计阶段，提出机器学习方法解决设计瓶颈问题，并将物理设计技术应用于前沿微流控生物芯片设计中，提升了设计质量与效率。主要贡献包括： 首先，针对物理设计的点工具评估问题以及流程设计问题，本文研究了机器学习辅助的物理设计预测与自动设计方法。一方面，提出了一种基于卷积神经网络与迁移学习的设计规则违规（Design Rule Violation，DRV）数量预测模型，可以在布线阶段之前准确预测不同尺寸设计的DRV数量。与现有方法相比，DRV数量的预测误差平均降低72.2%。另一方面，提出了一种积木式物理设计框架，集成了多种开源物理设计点工具，支持灵活切换和应用不同的点工具以构建不同的设计流程，并且支持集成用于评估点工具性能的多种机器学习预测模型。进一步提出了一种机器学习辅助的设计流程自动构建算法，应用机器学习预测模型对不同阶段的点工具进行评估，从而自动为每个设计阶段选择点工具。其次，本文将机器学习辅助的预测技术应用于微流控混合器设计中的评估问题，首次提出了基于图神经网络的微流控混合器浓度预测模型。一方面，提出了基于图卷积神经网络的浓度预测方法，可以预测固定流速的微流控混合器的输出浓度。与现有方法相比，预测误差减少了88%。另一方面，提出了基于图注意力网络的浓度预测方法，可以预测任意流速的微流控混合器的输出浓度。与现有方法相比，预测误差减少了85%。与商业有限元分析软件相比，提出的基于图神经网络的微流控混合器预测方法可以实现50倍以上的加速。最后，本文将机器学习辅助的自动设计技术应用于微流控混合器设计问题，首次提出了基于人工神经网络与迁移学习的微流控混合器自动设计方法。所提出的人工神经网络模型，可以根据浓度需求自动修改微流控混合器的设计结果，从而降低微流控混合器输出浓度与浓度需求之间的误差。与现有方法相比，所提出的方法可以将浓度误差减少88%。所提出的迁移学习方法，可以显著减少模型训练所需的数据集规模，将数据库构建时间从四个月缩短到两天，时间开销减少98%。

With the continuous reduction of chip feature sizes and the continuous increase of chip integration, traditional Electronic Design Automation (EDA) methods are confronted with significant challenges. In recent years, the development of machine learning technology has provided new research ideas for addressing the EDA problems for integrated circuits. This thesis targets the physical design stage, which is the most complex and time-consuming stage in the EDA flow of integrated circuits, and proposes machine learning approaches to address the design bottleneck problems. Moreover, this thesis applies the physical design technologies in the design of cutting-edge microfluidic biochips. The design quality of results (QoR) and efficiency are significantly enhanced. The main research work includes:Firstly, aiming at the point tool evaluation problem and the physical design process design problem in physical design, this thesis studies machine learning-assisted physical design prediction and automatic design technologies. On one hand, a Design Rule Violation (DRV) number prediction model based on convolutional neural networks and transfer learning is proposed, which can accurately predict the DRV number of different size designs before the routing stage. Compared with existing methods, the prediction error of the DRV number is reduced by 72.2% on average. On the other hand, a building-block physical design framework is proposed, which integrates a variety of open-source physical design point tools and supports flexible switching and application of different point tools to build different design processes. The framework also supports integrating various machine learning prediction models for evaluating the performance of point tools. And a machine learning-assisted automatic construction algorithm is proposed, which applies machine learning prediction models to evaluate point tools at different stages so that the point tools can be automatically selected for each design stage according to the optimization objectives.Secondly, aiming at the evaluation problem in microfluidic mixer designs, applying machine learning-assisted prediction technologies, this thesis first proposes concentration prediction models for microfluidic mixers based on graph neural networks. On the one hand, a concentration prediction method based on graph convolutional neural networks is proposed, which can predict output concentrations of microfluidic mixers with fixed flow rates. Compared with existing methods, the prediction error is reduced by 88%. On the other hand, a concentration prediction method based on graph attention networks is further proposed, which can predict output concentrations of microfluidic mixers at any flow rate. Compared with existing methods, the prediction error is reduced by 85%. Compared with the commercial finite element analysis software, the proposed microfluidic mixer prediction method based on graph neural networks can achieve more than 50 times acceleration.Finally, aiming at the microfluidic mixer design problem and applying machine learning-assisted automatic design technologies, this thesis first proposes an automatic design method for microfluidic mixers based on artificial neural networks and transfer learning. The proposed artificial neural network model can automatically modify design results of microfluidic mixers according to concentration requirements, thus reducing the error between output concentrations of microfluidic mixers and concentration requirements. Compared with the existing methods, the proposed method can reduce concentration error by 88%. The proposed transfer learning method can greatly reduce the dataset size required for model training, shorten database construction time from four months to two days, and reduce the time cost by 98%.