Reinforcement Learning for Hardware-Aware Reconfigurable Atom Array Compilation
ABSTRACT
Quantum circuit compilation for reconfigurable neutral atom arrays (RAA) maps an abstract circuit onto a sequence of hardware-executable operations: atom placement, acousto-optic deflector (AOD) movements between storage and entanglement zones, and gate scheduling under tight constraints on what can be parallelized. Conventional compilers produce the entire compilation sequence in a single pass: exact solvers achieve high fidelity but scale poorly with qubit number, while heuristic compilers remain fast but leave a substantial fidelity gap. We introduce the Neutral Atom Compiler Agent (NACA), a reinforcement-learning compiler that instead generates the compilation sequence step by step, choosing each action conditioned on the current hardware state. NACA uses an LLM backbone to capture long-range temporal correlations across compilation steps, and is pre-trained via supervised imitation of a heuristic compiler before reinforcement-learning fine-tuning on circuit fidelity. We validate the approach on small random circuits, where NACA consistently surpasses the heuristic baseline, and further evaluate it on practical quantum algorithms with larger and deeper circuits to demonstrate that learned, state-conditioned compilation closes a meaningful portion of the gap to optimal while remaining tractable at scale.
