Neutral-atom architecture model

Reconfigurable neutral-atom arrays are a different kind of quantum machine than the fixed-coupling devices most compilers target. There are no SWAP gates: connectivity is rewritten at runtime by physically moving atoms. That single fact reshapes the entire compilation problem — placement, routing, scheduling, and cost are all about movement rather than gate insertion. This page documents the abstract hardware model Quon’s quon_na backend approximates, the target JSON schema every pass is built against, and how the compiler schedules movement and entanglement on top of it.

This is a web-oriented condensation of the normative reference in the repository. Every modeled mechanism is attributed to a specific paper, and every numeric constant is either pinned to a cited source or explicitly labeled an illustrative placeholder. For the full field-by-field schema, the resource-report formats, and the complete citation list, read the source document in the repo.

What is being modeled

The backend targets an abstract reconfigurable neutral-atom array: atoms held in optical traps on a 2D plane, where some traps are static (a spatial light modulator, SLM) and some are mobile (a crossed 2D acousto-optic deflector, AOD). Entangling gates are mediated by the Rydberg blockade between atoms brought within interaction range; connectivity is reconfigured at runtime by physically moving atoms rather than by inserting SWAP gates. This is the family called DPQA (dynamically field-programmable qubit arrays) in [OLSQ-DPQA] and [Enola], and FPQA/RAA in [Atomique].

Two variants are modeled, in two compiler stages:

1. Flat reconfigurable array — a single plane, one SLM grid plus one or more AOD grids, a global Rydberg laser illuminating the whole plane. “Which pairs may interact” is a distance and scheduling notion, not a spatial one. The model of [OLSQ-DPQA], [Enola], and [Atomique].
2. Zoned architecture — physically separate regions with different capabilities: a storage zone (dense static traps, shielded, long coherence), an entanglement zone (paired traps under a zone-restricted Rydberg beam), and optionally a readout zone (mid-circuit measurement without disturbing other atoms). The model of [AbstractModel], [RAP], and the experimental architecture of [Bluvstein24].

These are two genuinely different hardware models from two different lines of literature. The backend builds both deliberately — the flat AOD movement planner first, the zoned joint placement-routing scheduler second — and results from one stage must never be quoted as evidence about the other.

What is explicitly not modeled

Drawing the boundary of the model is as important as drawing its contents. Several things a reader might expect are deliberately absent:

• No full QEC decoder. The QEC layer is a logical-op and resource-accounting abstraction: code blocks expand to atom counts and syndrome rounds are schedulable operations, but no syndrome decoding, error propagation, or logical-failure-rate simulation is performed.
• No vendor-specific hardware claims. The model is grounded entirely in published techniques and their stated parameters. The target descriptor is a generic reconfigurable neutral-atom machine; it does not describe any specific commercial device.
• No pulse-level physics. Gates are discrete scheduled actions with durations and fidelities; no Hamiltonian, pulse-shape, or blockade-strength simulation.
• No atom-loss or heating simulation. [Atomique] models movement-induced heating and probabilistic atom loss; this backend’s v0 cost model does not. Movement cost is time-based; heating-aware cost is a possible later extension.
• No continuous trajectories. Movement is modeled at the granularity of rearrangement steps between discrete placements, as in [RAP]. Collision-freedom within a step is guaranteed by the AOD ordering constraints, not by path simulation.

Fixed, reconfigurable, and zoned arrays

Three architecture families appear in the neutral-atom compilation literature. The distinction matters because compiler problems and cost models differ per family:

Family	Traps	Connectivity	Compiler problem
Fixed array	Static only	Fixed local neighborhood (within an interaction radius)	Qubit mapping + SWAP insertion, as on any fixed-coupling device. Baselines “FAA” in [Atomique].
Reconfigurable array (DPQA/FPQA)	Static SLM + mobile AOD	Any pair can be brought within range by movement	Joint placement, movement scheduling, gate layering under AOD constraints ([OLSQ-DPQA], [Enola], [Atomique]).
Zoned architecture	Static traps in operation-specific zones + AOD transport between zones	Entangling only inside the entanglement zone; storage shielded; readout isolated	Zone-aware placement and routing; shuttle scheduling storage↔entanglement↔readout ([AbstractModel], [RAP], [Bluvstein24]).

Quon’s existing gate-model backend covers the first family generically. The quon_na backend covers the second and third via TargetKind::NeutralAtomReconfigurable.

The movement model

This backend does not model free-grid Manhattan movement. Atoms in an AOD are trapped at the intersections of a set of AOD rows and columns; the only movement controls are the Y coordinate of each row and the X coordinate of each column ([OLSQ-DPQA]: “we cannot move AOD traps individually”). A movement plan that moves one atom independently of its row and column mates is not realizable on this hardware and is rejected by the movement-legality verifier. The enforced constraints, each pinned to a source:

• M1 — Coupled motion. All atoms in an activated AOD row (column) move together when that row (column) moves.
• M2 — Order preservation (no crossing). Within one AOD, a row cannot move past another row, nor a column past another column — crossing rows would violate minimum separation and cause heating/atom loss.
• M3 — No merging. Two rows (columns) of the same AOD cannot occupy the same coordinate.
• M4 — Static traps are static. SLM-trapped atoms do not move; a position change requires a trap transfer into an AOD, a costed, fidelity-bearing action.
• M5 — Occupancy. One trap holds at most one atom; two atoms may never occupy the same site at the same time.

M2 and M3 bind within one AOD only; rows/columns of different AODs may cross. In the zoned model the same constraints reappear as the three rearrangement constraints of [RAP]: non-crossing, preservation (atoms starting in the same AOD row stay in it), and ghost spots (activating rows and columns affects all their grid intersections, so loading a subset needs offset moves — this is why zone transfer is not free even when distances are short).

Movement timing. One rearrangement step covering maximum distance $d$ takes

t_move(d) = sqrt(d / a),  a = 2750 m/s²   (e.g. d = 110 µm → t = 200 µs)

cited to [RAP] and [Enola]; [OLSQ-DPQA] states the same law as $t = T_0\sqrt{D/D_0}$ with $T_0 = 200,\mu s$, $D_0 = 110,\mu m$ (“to maintain constant heating”). Each trap transfer adds 15 µs. A known literature divergence: [Atomique] instead charges a fixed 300 µs per movement stage regardless of distance. The backend implements the √-law because it is what the reproduced paper’s metric uses; the divergence is documented so nobody “fixes” one to match the other.

The interaction model

For the flat array, the Rydberg laser is global, which makes interaction a scheduling problem more than a geometry one:

• R1 — Range. Two atoms can perform an entangling gate iff their distance is $\le r_b$ (the Rydberg blockade radius) while the Rydberg laser is on.
• R2 — Compulsion. The flat-array laser illuminates the whole plane: every pair within $r_b$ when the laser fires undergoes a gate, wanted or not. Parking two non-partner atoms within $r_b$ at a Rydberg stage is illegal.
• R3 — Isolation. Non-interacting atoms must be separated by $> 2.5,r_b$; among any three atoms at most one pairwise distance may be below this when the laser is on. This is the source of the min_rydberg_spacing_um field.
• R4 — Stage-based error. Because the laser is global, every illuminated idle atom accrues error each Rydberg stage. Minimizing the number of Rydberg stages — not the gate count — is the correct flat-array objective.

In the zoned model, R1–R3 hold inside the entanglement zone: traps there come in pairs placed so that pair members interact and distinct pairs do not. The beam covers only that zone, so storage-zone atoms are shielded and R4’s idle penalty applies only to atoms left isolated in the entanglement zone.

Scheduling

Compilation proceeds in three moves: extract an interaction graph from the circuit, color its entangling layers, then plan movement and compact the schedule.

Edge coloring. Each 2-qubit gate layer is a graph whose edges are the gates and whose vertices are the atoms. Scheduling a layer means partitioning its edges into Rydberg stages where no two edges sharing a vertex fire together — an edge coloring. The backend uses the [Enola] Misra–Gries bound, which guarantees at most $S_{opt} + 1$ stages, so the number of Rydberg stages stays within one of optimal.

ASAP scheduling and compaction. The flat movement planner greedily builds maximal sets of AOD-compatible parallel moves (maximal independent sets in a move-conflict graph, after [Enola] and [Atomique]). An exclusive-cycle ASAP serializer lays independent layers out sequentially for a merge-free baseline, then a greedy pass recovers legal entangle-only parallelism. This compaction pass is engineering glue — not a paper reproduction, and not Enola-optimal ASAP; [Enola] is cited only for the critical-path lower bound.

The zoned scheduler reproduces the joint placement-routing formulation of [RAP]: placements are chosen per 2Q layer by an A* search whose node cost is a routing cost — implied movements grouped greedily into AOD-compatible parallel groups, $\text{cost}(p) = \sum_G \sqrt{d_{max}(G)}$ over groups $G$, extended with reuse and one-layer look-ahead. Placement quality is measured in rearrangement steps and durations, not raw travel distance — that is what “routing-aware” means versus the sequential distance-minimizing placement of the ZAC baseline it improves on. The regression anchor pins to [RAP] Table I (e.g. the 42-qubit ising benchmark: 22 rearrangement steps routing-agnostic vs 9 routing-aware).

The target JSON schema

The NeutralAtomTarget payload is loaded with quonc --target <path>. Field names are normative; the checked-in sample is targets/neutral_atom/generic_rna_v0.json. Units: lengths in µm, times in µs, fidelities as probabilities in $[0, 1]$.

The top level ties the pieces together:

Field	Type	Meaning
id	string	Target identifier
kind	"neutral_atom_reconfigurable"	TargetKind discriminant
grid	object	Bounding box: width_um, height_um
zones	array of Zone	Zone list — occupancy, capacity, operation legality
movement	object	AOD movement model
interaction	object	Rydberg parameters
native_gates	array of string	Gate names executable natively, e.g. ["cz", "rz_local", "ry_global", "measure_z"]
timing	object	Operation durations
fidelity	object	Operation fidelities + coherence time
error_model	object, optional	Explicit physical error probabilities for QEC (never derived as 1 − fidelity)
cost_model	object, optional	Four weights. Omitted fields use the placeholder defaults in the source document §9

A Zone declares a region’s capability. A flat-array target is expressed as a single entanglement zone covering the whole grid, so the zone constraints degenerate to the flat model; a zoned target declares at least one storage and one entanglement zone.

Field	Type	Meaning
zone_id	integer	Unique zone identifier
kind	"storage" | "entanglement" | "readout"	Zone capability
rows, cols	integer	Static-trap grid extent (for entanglement: trap pairs)
origin_um	[number, number]	Lower-left corner of the trap grid
site_pitch_um	[number, number]	Trap spacing in x and y
pair_gap_um	number, entanglement only	Distance between the two traps of a pair

The Movement object selects the AOD row/column-coupled semantics (the only legal value; "free_manhattan" is deliberately not a variant) and caps the simultaneously usable rows/columns and independent AOD units, which is where M2 and M3 bind.

Field	Type	Meaning
model	"aod_row_column_coupled"	Enables M1–M5 in the movement-legality verifier
aod_rows, aod_cols	integer	Max simultaneously usable AOD rows/columns
num_aods	integer	Number of independent AOD units
min_row_col_separation_um	number	Minimum spacing of same-AOD rows/columns
speed_model	{ "kind": "sqrt", "acceleration_m_s2": 2750 }	Move duration $t = \sqrt{d/a}$
trap_transfer_us	number	Duration of one pick-up or drop-off

The Interaction object pins the Rydberg geometry that gates R1–R3:

Field	Type	Meaning
rydberg_range_um	number	$r_b$: max distance for an entangling pair
min_rydberg_spacing_um	number	Isolation distance for non-partners, $= 2.5,r_b$ per R3
max_parallel_entangling_pairs	integer	Cap on simultaneous 2Q gates per Rydberg stage

Timing and fidelity are flat objects of operation durations and per-action fidelities — cz_us, single_qubit_us, measurement_us, reset_us; cz, single_qubit, atom_transfer, coherence_time_us. The optional error_model is a sibling of fidelity, not a replacement: it carries explicit physical error probabilities (per Rydberg stage, per measurement round, per rearrangement step, per transfer, per µs of idle) for QEC resource accounting. Per ADR-0017, you must never convert rates as 1 − fidelity.*.

Cost model

The v0 cost_model is one dot product. The zoned placer scores with it, and the verified resource report recomputes the same total. Field names, units, placeholder defaults, schema rules, and the time / error-budget CLI modes are in the source document §9.

total = rydberg_stage_weight · rydberg_stages
      + movement_time_weight · movement_time_us
      + trap_transfer_weight · trap_transfers
      + idle_time_weight · idle_time_us

rydberg_stages and trap_transfers are counts. movement_time_us and idle_time_us are microseconds already stamped on the verified schedule. The weights are dimensionless placeholders, not published measurements.

QEC overlay

The QEC abstraction layer expands each logical qubit’s code block into atoms by an exact per-CodeFamily formula. $N$ below counts atoms per logical qubit, including syndrome/check ancillas unless stated otherwise.

CodeFamily	Formula (atoms per logical qubit)	Parameters	Primary citation
SurfaceCodeLike	$2d^2 - 1$ (incl. checks)	distance $d$	[Bravyi24] §1; [BMD07]
RepetitionCodeToy	$2d - 1$ (incl. measure atoms)	distance $d$	[Kelly15] Fig. 1b
HighRateQldpcLike	$\lceil 1/r \rceil$ ($r$ = net rate, incl. checks)	rate $r$	[Bravyi24] §1, Table 1
AbstractBlockCode	$\lceil n/k \rceil$ (user convention)	$n, k$	[Gottesman97] §2.3

A few subtleties are worth recording. The rotated surface code’s $2d^2 - 1$ is exact ([Bravyi24]); the common shorthand $N \approx 2d^2$ is this rounded. For qLDPC, physical-per-logical overhead is the inverse of the rate, $n/k = 1/r$, not $k/r$ — an easy bug, called out in the issue. Constant-rate qLDPC families exist asymptotically, with concrete low-overhead points on reconfigurable arrays, which is the reason this family is worth modeling on this backend at all.

The hybrid QEC path. Code blocks are scheduling units: whole blocks move between zones and logical 2Q gates are physical-parallel transversal interleavings — the operational picture motivating this layer, demonstrated in [Bluvstein24]. The hybrid QEC schedule path (ADR-0016) adds memory_rounds — syndrome cycles that are schedulable operations — to the resource report. The atom-indexed interaction graph (ADR-0029) lets the scheduler treat each physical atom as a first-class node, so transversal logical gates over code blocks decompose into atom-level entangling layers that the same movement and edge-coloring machinery handles. Sizing helpers live in quon_qec, with quon_na consuming them for schedules.

Two artifacts come out of QEC evaluation and must never be fused (ADR-0020): a compiler analytic ResourceReport (schedule metrics, QEC metadata, an error_budget of rate × count, and an Enola-Eq.-(1) fidelity estimate) and a sampled Sinter CSV of logical failure rates from Monte Carlo. Neither is a threshold claim.

References

• [OLSQ-DPQA] D. B. Tan, D. Bluvstein, M. D. Lukin, J. Cong, “Compiling Quantum Circuits for Dynamically Field-Programmable Neutral Atoms Array Processors”, Quantum 8, 1281 (2024). arXiv:2306.03487.
• [Enola] D. B. Tan, W.-H. Lin, J. Cong, “Compilation for Dynamically Field-Programmable Qubit Arrays with Efficient and Provably Near-Optimal Scheduling”, ASPDAC 2025. arXiv:2405.15095.
• [Atomique] H. Wang et al., “Atomique: A Quantum Compiler for Reconfigurable Neutral Atom Arrays”, ISCA 2024. arXiv:2311.15123.
• [RAP] Y. Stade, W.-H. Lin, J. Cong, R. Wille, “Routing-Aware Placement for Zoned Neutral Atom-based Quantum Computing”, ICCAD 2025. arXiv:2505.22715. (The paper reproduced by the zoned scheduler.)
• [AbstractModel] Y. Stade, L. Schmid, L. Burgholzer, R. Wille, “An Abstract Model and Efficient Routing for Logical Entangling Gates on Zoned Neutral Atom Architectures”, IEEE QCE 2024. arXiv:2405.08068.
• [Bluvstein24] D. Bluvstein et al., “Logical quantum processor based on reconfigurable atom arrays”, Nature 626, 58 (2024). arXiv:2312.03982.
• [Bravyi24] S. Bravyi et al., “High-threshold and low-overhead fault-tolerant quantum memory”, Nature 627, 778 (2024). arXiv:2308.07915.
• [BMD07] H. Bombin, M. A. Martin-Delgado, “Optimal resources for topological two-dimensional stabilizer codes”, Phys. Rev. A 76, 012305 (2007). arXiv:quant-ph/0703272.
• [Kelly15] J. Kelly et al., “State preservation by repetitive error detection in a superconducting quantum circuit”, Nature 519, 66 (2015). arXiv:1411.7403.
• [Gottesman97] D. Gottesman, “Stabilizer Codes and Quantum Error Correction”, PhD thesis, Caltech (1997). arXiv:quant-ph/9705052.

For the normative full document — the complete mechanism-to-source attribution table, the resource-report field reference, the numeric-constant provenance table, and every error_budget multiplier — see architecture_model.md in the repository.

Continue to the compiler pipeline reference (contract — Reference) for how the neutral-atom stages fit into the overall quonc flow, or step back to the Language guide: QEC blocks (concept — Language guide) for the source-level concept a QEC block lowers into this hardware model.