Elivagar: Resource-efficient Quantum Circuit Architecture Search Guided by Quantum Noise

Near-term quantum machine learning tasks are limited by two key factors: device-level noise and learning ability of variational quantum circuits. These observations have inspired attempts to search for performant variational circuits that are robust to hardware noise. However, current Quantum-circuit Architecture Search (QAS) methods adopt classically-inspired designs that (1) do not account for high training cost on quantum systems (2) are incapable of finding effective data embedding. These two limitations lead to severe performance bottleneck of current approaches.

In this work, we present Elivagar, a resource-efficient, noise-guided Quqntu framework. The design of Elivagar is based on the key observation that a large portion of candidate circuits in the search space can be rejected cheaply by leveraging device noise information. By introducing a device topology-aware, zero-cost-to-compute metric, and 2 easy-to-compute noise-related metrics, Elivagar decouples the high-cost initial training process from circuit search. Due to its resource-efficiency, Elivagar can further co-search for data embedding in combination with trainable variational circuits, significantly improving the searched quantum model's learning ability and circuit sizes.