A Machine Learning Rooted Pre-characterization Method for Floating Random Walk Based Capacitance Extraction of Multi-dielectric VLSI Interconnects

Green's Function Table(GFT) and Weight Value Table (WVT) with more than two dielectric layers suffer from a significant growth in storage and memory. A machine learning based method to approximate three-dielectric GFTs using two-dielectric ones during run-time random walk procedure is presented in this article. We explore the numerical relationship between two- and three-dielectric GFTs, and accordingly apply deep neural network (DNN) model to compute three-dielectric GFTs out of the pre-characterized two-dielectric GFTs during run-time hence no pre-characterization of three-dielectric GFTs is required. We utilize automated machine learning (AutoML) tool to find the satisfied DNN configurations. Two test cases under an advanced process node are employed to validate our method, where the self capacitance error is less than 5% in over 93% probability compared with two-dielectric FRW process, and the memory usage is proved to be 390X times lower than the direct storage of three-dielectric GFTs.