Reinforcement Learning for Hardware-Aware Reconfigurable Atom Array Compilation

Circuit compilation for reconfigurable neutral-atom arrays (RAA) is challenging due to hardware constraints on the movement of acousto-optic deflectors (AODs) during qubit transport. Existing compilers typically separate qubit placement and routing into two stages, limiting their ability to discover globally optimal strategies that exploit RAA hardware structure, particularly in zoned architectures where atoms must be transported between storage and entangling regions for high-fidelity entangling operations. We introduce a reinforcement learning (RL)-based compiler that jointly optimizes placement and routing by learning AOD movement policies directly from hardware constraints. To address the exponential growth of possible movement combinations, we factorize the action space into sequential single-AOD decisions, while using attention-based policy networks to enable flexibility across different circuit sizes and types. We validate our approach by training on random circuits and evaluating performance on practical quantum algorithms, demonstrating that joint optimization via RL leads to efficient AOD movement strategies that improve the fidelity of the compiled circuit. Our results establish reinforcement learning as a scalable framework for jointly optimizing placement and routing in neutral atom quantum processors.