講演情報
[10p-F212-13]Reinforcement Learning Enhanced Algorithm for Fast Atom-Array Generation
〇(B)Inhoe Koo1, Luis Fernandez1, Wenchao Xu1 (1.ETH Zurich)
キーワード：
Quantum Computing、Neutral Atom、Reinforcement Learning
Scalable neutral atom quantum processors require rapid rearrangement of probabilistically loaded optical tweezer arrays into defect-free target regions. Conventional rule-based methods may not fully exploit the global occupation structure and can generate unnecessary or long-distance atom movements, which leads to atom loss and low fidelity. Here, we demonstrate that reinforcement learning (RL) can substantially reduce the transport cost of neutral atom rearrangement.
We train an agent on occupation maps using structured row- and column-packing actions. At the largest tested system size, containing N=1444 target atoms, the learned policy reduces the estimated distance-dependent movement cost by 68.3% relative to the Tetris baseline. Power-law fitting gives DRL ∼ N0.92, compared with DTetris ∼ N1.05. The learned policy also favors shorter movements and suppresses the long-distance tail, explaining the reduction in transport cost.
To connect this RL-enhanced rearrangement algorithem with real experimental implementation, which involves the use of acousto-optic deflector (AOD) for steering tweezer beams, we measure the optical response delay across the 60-100 MHz RF range used in actual rearrangement. The measured delay is 8.26±0.39μs, indicating weak frequency dependence across the RF tones. This shows that RL can reduce transport costs without being significantly affected by frequency variations in AOD response delay, providing a practical strategy for AOD-based neutral atom rearrangement.
We train an agent on occupation maps using structured row- and column-packing actions. At the largest tested system size, containing N=1444 target atoms, the learned policy reduces the estimated distance-dependent movement cost by 68.3% relative to the Tetris baseline. Power-law fitting gives DRL ∼ N0.92, compared with DTetris ∼ N1.05. The learned policy also favors shorter movements and suppresses the long-distance tail, explaining the reduction in transport cost.
To connect this RL-enhanced rearrangement algorithem with real experimental implementation, which involves the use of acousto-optic deflector (AOD) for steering tweezer beams, we measure the optical response delay across the 60-100 MHz RF range used in actual rearrangement. The measured delay is 8.26±0.39μs, indicating weak frequency dependence across the RF tones. This shows that RL can reduce transport costs without being significantly affected by frequency variations in AOD response delay, providing a practical strategy for AOD-based neutral atom rearrangement.