Identifying common root causes of systematic defects in a short time is crucial for yield improvement. Diagnosis driven yield analysis (DDYA) such as Root cause deconvolution (RCD) is a method to estimate root cause distribution by applying statistical analysis on volume diagnosis. By fixing identified common root causes, yield can be improved.
With advanced technologies, smaller feature size and more complex fabrication processes for manufacturing VLSI semiconductor devices lead to more complicated failure mechanisms. Lack of domain knowledge of such failure mechanisms makes identifying the emerging root causes more and more difficult. These root causes include but are not limited to layout pattern (certain prone to fail layout shapes) and cell internal root causes. RCD has proved to have certain degree of success in previous work, however, these root cause are not included and pose a challenge for RCD. Furthermore, complex volume diagnosis brings difficulty in investigation on RCD. To overcome the above challenges to RCD, improvement based on better understanding of the method is desired.
The first part of this dissertation proposes a card game model to create controllable diagnosis data which can be used to evaluate the effectiveness of DDYA techniques. Generally, each DDYA technique could have its own potential issues, which need to be evaluated for future improvement. However, due to limitation of real diagnosis data, it is difficult to, 1. Obtain diagnosis data with sufficient diversity and 2. Isolate certain issues and evaluate them separately. With card game model given correct statistical model parameters, impact of different diagnosis scenarios on RCD are evaluated. Overfitting problem from limited sample size is alleviated by the proposed cross validation method.
In the second part of this dissertation, an enhanced RCD flow based on pre-extract layout patterns is proposed to identify layout pattern root causes. Prone to fail layout patterns are crucial factors for yield loss, but they normally have enormous number of types which impact the effectiveness of RCD. Controlled experiment shows effectiveness of enhanced RCD on both layout pattern root causes and interconnect root causes after extending to layout pattern root causes. Test case from silicon data also validates the proposed flow.
The last part of this dissertation addresses RCD extension to cell internal root causes. Due to limitation of domain knowledge in both diagnosis process and defect behavior, parameters of RCD model are not perfectly accurate. As RCD moves to identify cell internal root causes, such limitation become an unescapable challenge for RCD. Due to inherent characteristics of cell internal root cause, RCD including cell internal root cause faces more difficulty due to less accurate model parameters. Rather than enhancing domain knowledge, supervised learning for more accurate parameters based on training data are proposed to improve accuracy of RCD. Both controlled experiments and real silicon data shows that with parameters learned from supervised learning, accuracy of RCD with cell internal root cause are greatly improved.