PhD Defense: "Learning Approaches to Analog and Mixed Signal Verification and Analysis"

Samantha Alt

December 18th (Thursday), 4:00pm
Harold Frank Hall, Room 4164 (ECE Conference Room)

This work focus on addressing various problems related to the design automation of analog and mixed signal circuits. Analog circuits are typically highly specialized and fined tuned to fit the desired specifications for any given system reducing the reusability of circuits from design to design. This hinders the advancement of automating various aspects of analog design, test, and layout. At the core of many automation techniques, simulations or data collection are required. Unfortunately, for some complex analog circuits, a single simulation may take many days. This prohibits performing any type of behavior characterization or verification of the circuit. This leads us to the first fundamental problem with the automation of analog devices. How can we reduce the simulation cost while maintaining the robustness of transistor level simulations? As analog circuits can vary vastly from one design to the next and are hardly ever comprised of standard library based building blocks, the second fundamental question is how to create automated processes that are general enough to be applied to all or most circuit types? Finally, what circuit characteristics can we utilize to enhance the automation procedures?

Statistical learning techniques for improving the circuit verification efficiency is studied. To enable the application, circuit simulation is modeled as an event propagation process through a system consisting of primitive elements. Then, the efficiency improvements are achieved with two approaches. By assuming that the output space can be represented by a set of selected events, unsupervised learning is applied to search the input events that correspond to the selected output events. Only the selected input events are simulated, resulting in saving of the simulation time. During the simulation, low-complexity primitive elements with low information content are modeled by supervised learning models. Event propagation through these primitive elements is achieved by model prediction rather than by actual simulation, resulting in further saving of the simulation time. This chapter explains the statistical learning concepts and the techniques to implement the two approaches and demonstrates their effectiveness with experimental results in the context of voltage domain analysis of several analog circuit designs. The work is extended to provide critical node and environmental analysis of large analog and mixed signal systems.

About Samantha Alt:

photo of samantha alt Samantha Alt received her B.S. and M.S. degrees in electrical and computer engineering from the University of California Santa Barbara, in 2007 and 2011, respectively. She is currently employed at Intel in the Rotational Engineering Program. During her study at University of California Santa Barbara, she hold internships at Mentor Graphics and Intel, where she worked on transition fault testing, and virtual design under test generation, RTL test generation using SMT solvers, test point insertion, and mixed signal simulation analysis. Her research interests include machine learning applications in design and verification of mixed-signal circuits, diagnosis, and test.

Hosted by: Malgorzata Marek-Sadowska and Li-C. Wang