Abstract
Process variations and physical defects can degrade the performance of a circuit, or even drastically affect its operation. It is therefore essential to verify the performance of each circuit produced in order to ensure the quality of manufactured devices shipped to the customers. This is the role of the testing process. This process represents a significant part of the total cost of an IC, especially for analog and RF circuits, whose performance must be measured with sophisticated and expensive test equipment. In order to reduce testing costs, one attractive solution is to adopt an indirect test strategy, which consists in measuring parameters that require only low-cost test resources and correlating these measurements, called Indirect Measurements (IMs), with the device specifications. This correlation is generally established using machine-learning algorithms during an initial learning phase. Then, during the production testing phase, every new device is evaluated using only the low-cost indirect measurements. While the indirect test strategy seems attractive, its deployment in an industrial context is viable only if sufficient test quality can be achieved. In this thesis, we have developed a methodology that permits to assist and guide the test engineer in its practical choices for an efficient implementation. Different aspects have been explored, such as the use of different types of regression models, the definition of pertinent metrics to evaluate the test efficiency, or the proposition of an original adaptive test flow in order to trade-off test quality and test cost. We have also proposed an adaptation of the indirect test strategy allowing to perform on-line monitoring of the device performance within its application. All results presented in this thesis have been evaluated using industrial test data on various case studies, which fully support the developed innovations.