A direct comparison of two tools for moving neutral atoms in reconfigurable atom arrays:
- AOD (Acousto-Optic Deflector) - the traditional tool that moves atoms in a grid geometry configuration.
- RIPA (Re-imaging Phased Array) - a fast SLM proposed in arXiv:2601.08906 that can move atoms in arbitrary patterns.
Routing identical atoms, for example, in the case of assembling a defect-free array, is a good example to compare the two tools.
The RIPA completes the assembly within 497 us, while both aod schedulers compiles to ~2ms operation times.
Click to see details
The RIPA scheduler (left) used here is a simple heuristic scheduler RIPAPebbleAdvScheduler. We choose two AOD schedulers to compare with, one is the Tetris (middle) suitable for stochastic reservoir loading with postselection; and the other one is a Sqrt-time (right) that fully utilizes the parallelism of the AOD to solve an arbitrary binary-grid reconfiguration problem.
When routing atoms encoded with quantum information, they are no longer treated as identical, and the routing is a labeled, pair-wise rearrangement problem.
Here we demonstrate an inversion of a 1x6 atom array.
- RIPA (left) performs the inversion in a 1-step fashion
- AOD (middle, right) needs multiple steps.
Click to see notes and 2D examples
Note 1: Moreover, RIPA does not require a hand-off between AOD and SLM traps which could potentially introduce atom loss.
Note 2: For the AOD log(N) steps arbitrary re-arrangement algorithm, as demonstrated in the middle panel. See the original paper constant_overhead_paper and the code algorithm here.
Note 3: With a 6x6 atom array, where AOD arranger (middle, right) gains more parallelism, now the total time cost to rearrange is limited by the maximum throughput. We can see the RIPA scheduler (left) still takes less time due to its flexible routing capability.
The RIPA shows its power in more complex, 2d routing problems. For example, a logical hadamard gate for a surface code includes a transversal hadamard for each atoms, followed by a patch rotation that exchanges the rough and smooth boundaries.
The patch rotation can be implemented with a set of horizontally aligned 2D-ADO together with an extra set of 2D-AOD (Zihan Chen, et al.), diagonally aligned.
With RIPA-native movements, we can implement the patch rotation in much shorter time.
Modern magic state factory benefits from Fold-transversal surface code cultivation, which needs complicated atom movements.
What is so different about a RIPA ?
- It can move atoms independently in an arbitrary pattern (continuously along the cartesian grid).
- It can be scheduled asynchronously.
For detailed comparison, see aod_vs_ripa.md
Click to see details
- Continuous frequency sweeps move spots smoothly along either a row or a column.
- Using frequency- or polarization- multiplexing to replicate the primary channel (row channel, continuously moving along x) and rotate it by 90 deg, so that the second channel (col channel) continuously moves along y.
- The EOM RF drive switches each channel on/off independently, and they can be simultaneously on, they can have independent timing for multiple frequency tones (so does multiple tweezer traps).
- To move an atom A → B: pick up atom from A, transport continuously along row/column. At the intersection of row/column (i.e., at integer grid point (i,j)), hand-off from one channel to the other. After several row(column)-continuous transport, the atom reaches B.
RIPA scheduling is a very hard problem, mostly in its vast parameter space and its asynchronous nature. This repo implements some heuristic algorithms, as well as a searching algorithm, see scheduler/readme.md.
The RIPA scheduler, in its geometric nature, is similar to the pebble motion problem, see Pebble Problem. We also adapted some path-planning algorithms like Continuous-CBS.
Different schedulers will yield different total arrangement times. See inversion 6x6 benchmark
src/: source code, see src/readme.md.src/scheduler/: contains the scheduling algorithms, see scheduler/readme.md.example/: example scripts to demonstrate the scheduling and to do benchmarking.