Beyond two transmons, and what to do before hardware¶
Who offers pulse-level access today¶
Pulse-level access to commercial quantum computers has narrowed: IBM removed pulse-level control from all its processors in February 2025, and a 2026 survey of thirteen vendor stacks describes public access to pulse and control-electronics interfaces as "bifurcated" [malarchick2026]. What remains through Amazon Braket (checked 24 September 2026):
| Device | Technology | Control exposed | Relevance |
|---|---|---|---|
| QuEra Aquila | Neutral atoms, Rydberg | Analog Hamiltonian simulation: global amplitude, phase and detuning of the drive over time, up to 256 atoms | The most open analog control available; natural home for RL-shaped global pulses |
| Rigetti Ankaa-3 | Superconducting, tunable couplers | Braket Pulse (OpenPulse frames, ports, arbitrary waveforms) | Same device family as our model. The documentation's example port uses 1 ns samples; at 1 ns the paper's RL pulse loses its gate (J_T 0.29, home page), so pulses must be trained under the device's sampling |
| IonQ, AQT, IQM | Trapped ions, superconducting | Gate-based | No pulse access through Braket |
Candidate platforms¶
| Platform | Why it fits | What it would take |
|---|---|---|
| Tunable-coupler transmons (v3) | Continuity with the paper; pulse access on Rigetti Ankaa-3; current records use tunable couplers [sung2021] | Realistic coupler model (flux map, no RWA, direct coupling), 1 ns sampling, named-gate reward |
| Rydberg atoms | Open pulse-level hardware (Aquila); two-qubit gates at 99.5 % designed by optimal control [evered2023], with time-optimal pulses known [jandura2022]; drift dimension grows with the atom count (per-atom position, laser intensity and detuning inhomogeneity), which is exactly the scaling H1 needs | A JAX Rydberg-blockade environment (global drive, N atoms, blockade), drift over atom positions and laser parameters |
| Trapped ions | Robust entangling gates by pulse shaping are mature [leung2018]; slow drift of motional modes is a natural drift variable | Mølmer-Sørensen model with motional modes; no open pulse access on public clouds, so hardware would need a lab partner |
Recommendation¶
Go bigger, in two tracks, and keep the transmon line:
- Rydberg atoms next. Hardware you can actually submit pulses to today, a control problem (global drive) where RL has a structural advantage (one pulse must serve many atoms with different local parameters), and a drift dimension that grows with system size, which is the direct test of the hypothesis. It also connects to the DeepScar idea of controlling many-body dynamics in Rydberg arrays.
- Transmons, hardware-shaped. The v3 model, constrained to what Rigetti Ankaa-3 exposes (sampling, bandwidth, amplitude), so that pulses are submittable when access comes.
- Ions later, with a lab collaborator.
What to do until there is hardware¶
- Simulate the hardware contract, not just the physics. Train and refine only pulses the target device would accept: its sampling interval, amplitude and bandwidth limits, pulse-length quantisation, and for Aquila the waveform constraints stated in its device properties. A pulse that cannot be submitted is not a result.
- Count shots, not simulator samples. Replace exact J_T by estimates from finite measurements with state-preparation and measurement errors, and redo the cost comparison in shots (test T4). This is where the hypothesis says RL should gain, and it costs nothing but compute.
- Drift as a time series. Replace static drift boxes with drift trajectories (random walks and jumps calibrated to the measured magnitudes on Hardware drift ranges) and measure fidelity-hours: how long each pipeline keeps a device above target per unit of calibration effort. This is the operational metric a hardware group cares about.
- Dress rehearsal on vendor simulators. Braket's local analog simulator runs the same program format as Aquila, so the full pipeline (policy → program → results → refinement) can be tested end to end before the first paid shot.
- Publish a benchmark. Package the JAX environments (transmon, and Rydberg next) with the drift models and baselines as an open benchmark for drifting-device quantum control. It makes the work reusable and is a concrete reason for groups with hardware to get in touch.
- Keep the first hardware experiment small. A few-atom Rydberg experiment on Aquila (a blockade gate or a global-pulse state preparation, refined across two sessions) is a feasible first test of "trained in simulation, refined on the device".