Due to inherent complex behaviors and stringent requirements in analog and mixed-signal (AMS) systems, verification and testing become key bottlenecks in the product development cycle. Rare failure detection in a high-dimensional parameter space using minimal expensive simulation/measurement data is a major challenge.
For rare failure detection in the verification flow, this dissertation proposes to put machine learning models, that mimic the circuit behavior, under verification, which greatly relaxes the simulation/measurement requirements and improves the verification efficiency.
We first present a hybrid formal/machine-learning verification technique (HFMV) to combine the best of the two worlds. HFMV adds formalism on the top of a probabilistic learning model while providing a sense of coverage for extremely rare failure detection. On the other hand, we also study Bayesian optimization (BO) based approaches to the challenging problem of verifying AMS circuits with stringent low failure requirements. We simultaneously leverage multiple optimized acquisition functions to explore varying degrees of balancing between exploitation and exploration. Furthermore, this dissertation proposes a BO framework under high dimensional space to further improve the verification efficiency. Two techniques are explored here: 1) random embedding to linearly embed input into a low dimensional space and 2) sensor fusion networks to identify important nonlinear features transformed by reversible neural networks. The proposed approaches are very effective in finding very rare failures in high dimensional space which existing statistical techniques can miss.
On the subject of AMS testing, this dissertation proposes to utilize self-supervised learning methods to detect extremely rare customer failure. First, we study a transformation-based self-labeling technique to reliably screen out rare customer return defects. The normality score to an unseen input data is the goodness of the multi-class classification model trained by self-labeled data via a set of reversible transformations. Furthermore, this dissertation suggests a contrastive learning framework for semi-supervised learning and prediction of wafer map patterns. Contrastive learning is applied for the unsupervised encoder representation learning supported by augmented data generated by different transformations (views) of wafer maps. Experimental results demonstrate that the self-supervised learning framework greatly improves test accuracy compared to traditional supervised methods.
