Evaluarea variației procesului de fabricație în analiza proiectării circuitelor bazată pe date, utilizând machine learning / Data-driven fabrication process variation assessment in circuit design analysis using machine learning

As the integrated circuits (IC) are growing exponentially in complexity, ensuring a high production yield has become a major challenge for the semiconductor industry. One of the most challenging problems in the IC development is handling the inherent fabrication process variation. Controlling this variation is vital in ensuring a high production yield, by obtaining the optimal circuit design that is least sensitive to process variation. In this context, this thesis focuses on developing strategies, methodologies and tools that assess the process variation impact on circuit design analysis, by linking the reality of an existing fabrication to a new IC design, translated as an accurate estimation of the relationship between IC performances and the process variation, early in the design and production process. First, the thesis introduces a comprehensive and automated methodology for modeling the functional and statistical dependencies between the circuit performances and the technology process variation, in pre-silicon stage. It enables an instant and accurate snapshot on the circuit performances behavior when technological changes occur, without the need for additional simulations or post-silicon measurements. Further, for an early diagnostic of circuit responses that are highly sensitive to process variation, two fast and cost-effective methodologies have been proposed for global and local sensitivity analysis. Next, as an insufficient coverage in pre-silicon often leads to redesign, an alternative approach to deal with the diagnostic of inaccuracies in simulation device model is introduced. Finally, the focus is shifted towards the yield prediction field, by introducing two computationally-low methodologies based on the conditional modelling of the IC performances distributions with the technology process variation distributions, for normal and non-normal cases.