SA-DQAS: Self-attention Enhanced Differentiable Quantum Architecture Search

TL;DR

SA-DQAS introduces a self-attention mechanism into differentiable quantum architecture search, improving quantum circuit design for variational algorithms and demonstrating better performance, stability, and hardware compatibility.

Contribution

This paper presents the first integration of self-attention with DQAS, enhancing architecture learning by capturing inter-placeholder dependencies.

Findings

SA-DQAS outperforms prior QAS methods in multiple tasks.

Circuits trained on small graphs generalize to larger instances on hardware.

Enhanced stability and noise resilience in quantum circuits.

Abstract

We introduce SA-DQAS, a novel framework that enhances Differentiable Quantum Architecture Search (DQAS) by integrating a self-attention mechanism, enabling more effective quantum circuit design for variational quantum algorithms. Unlike DQAS, which treats placeholders independently, SA-DQAS captures inter-placeholder dependencies to improve architecture learning. We evaluate SA-DQAS across multiple tasks, including MaxCut, Job-Shop Scheduling Problem (JSSP), quantum chemistry simulation, and error mitigation. Experimental results show that SA-DQAS outperforms baselines and prior QAS methods in most cases, producing architectures with better stability, convergence, and noise resilience. To assess scalability and hardware readiness, we further test SA-DQAS-generated circuits on IBM's quantum device using the MaxCut problem. Circuits trained on small graphs are stacked to solve larger instances without retraining, demonstrating generalization to real hardware and larger problem sizes. Our results suggest that SA-DQAS not only improves circuit quality during training but also enables practical deployment on near-term quantum devices. This research represents the first successful integration of self-attention mechanism with DQAS.