Abstract
The rapid growth of worldwide semiconductors demand leads to numerous innovations and improvements in cost, speed, and power consumption.The constant shrinking of the transistors feature size allows the development of new applications and capabilities but reveals new types of manufacturing defects.Efficient test and precise diagnosis are crucial to guarantee the creation of quality products and to improve the production yield.To keep test and diagnosis in phase with the new types of defects, new methodologies and faults models have been invented and deployed.The Cell-Aware (CA) fault models abstract the subtle defects found inside the standard cells used to design digital ICs, at the transistor level.Cell-aware test uses these fault models with ATPG to create test patterns explicitly targeting cells internal defects.Cell-aware diagnosis tools use the CA data to identify the location and type of intra-cell defects, providing valuable insights in the context of large and complex standard cells.While becoming an industry standard, the CA methodology has a large and costly development overhead, involving numerous electrical simulations to characterize standard cells.Due to the high number of standard cells to be characterized, and, for each cell, the high numbers of potential intra-cell defects and cell-level patterns to consider, the CA characterization phase represents a heavy usage and cost of simulator licenses and computational power.This thesis presents an innovative flow using Machine-Learning (ML) to reduce the CA test method runtime and ease its adoption for industrial usage.Experiments using different technology nodes demonstrated an over 99% runtime reduction for 80% of combinational cells.To ensure the generation of a quality CA model for all cells, while decreasing the CA characterization time, a hybrid flow is proposed, mixing ML-based CA models prediction method with the conventional method using electrical simulations.This hybrid-flow includes a decision algorithm, which leverage ML techniques to decide whether the CA characterization of a new standard cell should be ML-based or simulation-based, thus allowing to decrease the CA characterization runtime while maintaining high quality CA models for all cells.Experimental results demonstrate the high performance of the new decision algorithm and the quality of the obtained CA models.The coverage of real cell-internal defects of ATPG patterns using ML-predicted CA data proves that our predicted CA data can accurately replace those obtained by running extensive analog simulations, thus proving the effectiveness and pertinence of the proposed methodology.