|
Practical machine learning (ML) techniques are being deployed, at an accelerated pace, in an expanding set of application domains. The newsworthy examples in games strategy, image recognition, automated translation and autonomous driving are only the tip of the iceberg of a massive revolution in industrial manufacturing. Semiconductor IC design and manufacturing are also starting to see a number of ML applications, albeit of limited scope. In this work we present both a novel computational tool for ML of physical design layout styles (constructs, patterns) and also a general technical framework for the implementation of ML solutions across the design to mask to silicon chain.
ML applications derive their mathematical foundations from Computational Learning Theory, which establishes “learning” as a computational process. The quantitative characterization of the learnability space and its associated Vapnik-Chervonenkis, (VC) dimension, is therefore a pre-requisite of any meaningful ML application. Various types of geometric constructs in the 3D Euclidian space of physical design layouts provide an ideal learnability domain. Specifically, the entire set of generalized Design Rules, silicon retargeting, Optical proximity Correction (OPC), post-OPC verification (ORC) and mask manufacturing constraints (MRC) can be demonstrated to be in a learnable set. This means that, given a suitable feature extraction model, classifiers systems can be built to perform data analytics and optimization, with a quantifiably higher performance (in terms of speed and scale) than any of the currently used engineering-based heuristics. Additionally layout styles, particularly the ones generated by Place-and-Route (P&R) tools and even manual layout for custom blocks are amenable to automated learning (and subsequent parametric-space optimization).
Two complete examples in the Design to Mask flow, for advanced technology node applications will be used to validate the ML methodological framework and to illustrate extendibility to other areas, such as process and fab flow optimization.
References
[1] “A Theory of the Learnable”, L. Valiant – Communications of the ACM, 27 (1984)
[2] “On the Learnability of Boolean Formulae”, M. Kearns et al. - ACM Symposium on Theory of computing (1987)
[3] “Classifying learnable geometric concepts with the Vapnik-Chervonenkis dimension”, Blumer, et al. - ACM Symposium on Theory of computing (1986)
[4] Optimization of complex high-dimensional layout configurations for IC physical designs using graph search, data analytics, and machine learning, V. Dai, et al. - Proc. SPIE 10148, DPTCO XI, (April 2)
|
