How AI Accelerates Atom‑Array Assembly for Future Quantum Computers

Background

Neutral‑atom quantum computers use individual atoms trapped in a two‑dimensional optical‑tweezer lattice as qubits. Arranging thousands of atoms into a desired pattern requires generating holographic light fields that move the atoms, a process that becomes computationally demanding as the array grows.

AI‑guided assembly

Lu Chaoyang and Pan Jianwei (University of Science and Technology of China) integrated a deep‑learning model into the control system of a rubidium‑atom tweezer platform. The model was trained on a large dataset of holographic phase patterns and the resulting atom trajectories, learning to predict the optimal phase mask for any target configuration.

Training procedure

Generate a random target lattice (e.g., 2 × 1024 sites).

Compute a set of candidate holographic phase patterns using a standard Gerchberg–Saxton algorithm.

Record the actual atom movements produced by each pattern.

Use the (pattern, outcome) pairs to train a convolutional neural network that maps target configurations to the phase mask that minimizes rearrangement time.

Experimental results

With the AI‑predicted phase masks the team assembled a 2 × 1024 array containing 2 024 rubidium atoms in 60 ms. The same size array previously required ~1 s for ~800 atoms using conventional optimization. The speedup factor exceeds 15× for comparable atom numbers.

To demonstrate the rapid reconfiguration, the system was programmed to spell the outline of Schrödinger’s cat; the atoms emitted fluorescence when the laser pulse illuminated the pattern, producing a visible image.

Scalability

The neural network inference time scales sub‑linearly with the number of target sites, allowing the method to handle tens of thousands of atoms without a noticeable increase in latency. The authors anticipate extensions to arrays of 10 k–100 k atoms, although reaching the million‑qubit regime will still require further hardware advances.

References

Original research article: Physical Review Letters, 8 August 2025. URL: https://journals.aps.org/prl/abstract/10.1103/2ym8-vs82

Related news coverage: https://www.nature.com/articles/d41586-025-02577-9

Code example

来源：ScienceAI
本文
约1300字
，建议阅读
5
分钟
这
是一种快速组装原子网格的最佳方法，该原子网格未来很有可能成为量子计算机的
「
大脑
」
。