Algorithms for Verification of Analog and Mixed-Signal Integrated Circuits

Abstract

Over the past few decades, the tremendous growth in the complexity of analog and mixed-signal (AMS) systems has posed great challenges to AMS verification, resulting in a rapidly growing verification gap. Existing formal methods provide appealing completeness and reliability, yet they suffer from their limited efficiency and scalability. Data oriented machine learning based methods offer efficient and scalable solutions but do not guarantee completeness or full coverage. Additionally, the trend towards shorter time to market for AMS chips urges the development of efficient verification algorithms to accelerate with the joint design and testing phases.

This dissertation envisions a hierarchical and hybrid AMS verification framework by consolidating assorted algorithms to embrace efficiency, scalability and completeness in a statistical sense. Leveraging diverse advantages from various verification techniques, this dissertation develops algorithms in different categories.

In the context of formal methods, this dissertation proposes a generic and comprehensive model abstraction paradigm to model AMS content with a unifying analog representation. Moreover, an algorithm is proposed to parallelize reachability analysis by decomposing AMS systems into subsystems with lower complexity, and dividing the circuit's reachable state space exploration, which is formulated as a satisfiability problem, into subproblems with a reduced number of constraints. The proposed modeling method and the hierarchical parallelization enhance the efficiency and scalability of reachability analysis for AMS verification.

On the subject of learning based method, the dissertation proposes to convert the verification problem into a binary classification problem solved using support vector machine (SVM) based learning algorithms. To reduce the need of simulations for training sample collection, an active learning strategy based on probabilistic version space reduction is proposed to perform adaptive sampling. An expansion of the active learning strategy for the purpose of conservative prediction is leveraged to minimize the occurrence of false negatives.

Moreover, another learning based method is proposed to characterize AMS systems with a sparse Bayesian learning regression model. An implicit feature weighting mechanism based on the kernel method is embedded in the Bayesian learning model for concurrent quantification of influence of circuit parameters on the targeted specification, which can be efficiently solved in an iterative method similar to the expectation maximization (EM) algorithm. Besides, the achieved sparse parameter weighting offers favorable assistance to design analysis and test optimization.