Quantum circuits trained in a classical simulator often lose most of their accuracy when run on a real quantum processor. The best-known fix, noise-aware training, adds random errors to the simulator that match the error rates the hardware vendor last reported. Those reported rates are a snapshot, though, and real devices drift by the hour. A circuit is trained against the device as it was yesterday and then run on the device as it is when it clears the queue. We propose to forecast the device's noise before training. The forecast is a calibrated probability distribution over each qubit's error rates at the expected execution time. Training then injects noise sampled from that distribution rather than from the stale snapshot.
This borrows the central lesson of probabilistic forecasting in power grids. There, operators plan reserves against a distribution of possible futures rather than a single best guess, and they score forecasts with proper scoring rules, which reward a forecaster only for reporting its true beliefs. We apply that lesson to the gap between simulated and hardware performance, and we test it where it counts: accuracy measured on the hardware. The design also includes controls that tell us whether forecasting is actually needed. Simply randomizing noise over a broad range might work just as well, and we check for that directly.
Core Claim
The claim is conditional. For a fixed training budget, injecting noise sampled from a conditional forecast of device error rates at deployment time should beat two alternatives: static snapshot injection and unconditional noise randomization. The advantage should grow with the delay between the calibration snapshot and execution. It should vanish when that delay is near zero, and it should vanish on devices whose error rates barely move.
Formally, let $\eta$ be the vector of per-qubit and per-link noise parameters: relaxation time $T_1$, dephasing time $T_2$, readout error, and single- and two-qubit gate error. Let $H_t$ be the calibration history available at training time $t$, and $\Delta$ the expected delay to execution. We train circuit parameters $\theta$ by
$$\min_\theta \; \mathbb{E}_{\eta \sim p(\eta_{t+\Delta} \mid H_t)} \big[ L(\theta; \eta) \big],$$
where $L(\theta;\eta)$ is the task loss of the circuit simulated under noise $\eta$. The static baseline replaces $p$ with a point mass at the latest snapshot. The randomization baseline replaces $p$ with the unconditional historical distribution of $\eta$. The new ingredient is the conditioning on recent history and on $\Delta$.
Where This Sits
Three results motivate the proposal by exposing the staleness problem. The first is RoQNN / QuantumNAT (also QuantumNAT). It injects Pauli errors matched to backend calibration data, normalizes and quantizes measurement outputs, and reports up to 43% accuracy gains on real IBM devices. Its noise model, however, is a fixed snapshot. We keep its training machinery intact and change only where the injected noise comes from.
The second is Detecting and tracking drift in quantum information processors. It finds drift by applying a spectral transform to time-ordered measurement bitstrings, reports a 0.92 correlation between qubit detuning and lab temperature, and gets a 10x idle-gate error reduction from recalibrating every 45 minutes. It looks backward: it detects drift but does not predict it. We turn that kind of signal into forward-looking predictive distributions.
The third is When Does Adaptation Win?. It meta-learns pulse controls whose adaptation gain scales with device variance, and it explicitly assumes noise stays stationary within an episode. That constrains our design. Their result says robustness gains depend on the spread of device states, which is why we must separate the value of conditional forecasting from the value of merely having spread.
On the forecasting side, we reuse methods and do not invent them. Multi-APLF is an online multi-series state-space forecaster that updates recursively and discounts old data with a forgetting factor. We adapt it to jointly forecast correlated qubit parameters, since neighboring qubits and couplers plausibly drift together. The Arctic wind study supplies the physical-bounds idea. Clamping its prediction intervals at rated capacity cut the interval score from 2.53 to 1.16. We impose the analogous quantum bounds: $T_2 \le 2T_1$ and error rates between 0 and 1. The CAISO ramp study and scoring-rule virtual power plants motivate judging a forecast by what it does for the downstream decision and by the Continuous Ranked Probability Score (CRPS), not by point error alone.
DreamQAS is the nearest quantum work that uses predictive uncertainty. It uses ensemble disagreement over simulated circuit energies to guide architecture search, but only in noise-free simulation, and its authors call hardware noise an open question. We forecast the hardware noise itself.
Research Program
We first build a calibration-history dataset. To our knowledge the file lists no public longitudinal corpus of cloud-device calibration logs, so we record the vendor-published calibration properties several times a day over a few months for the IBM-Q devices we can access. That set overlaps the kind of fleet benchmarked in QASMBench. We supplement the logs with short, scheduled characterization circuits so that some ground truth is measured by us rather than reported. For sub-hour dynamics, we reuse the trapped-ion drift data from Proctor et al. as a second forecasting testbed, including lab temperature as an outside input where it is available.
We then fit constrained probabilistic forecasters at several delays $\Delta$, from hours up to about two days. These are a persistence model with empirical quantiles, the Multi-APLF-style joint state-space model, and a small recurrent quantile model. We score each on CRPS, pinball loss, and interval coverage.
Finally, we plug the best-scoring forecaster into QuantumNAT-style training on its published tasks. We train at time $t$ and execute on hardware at $t+\Delta$, holding circuit architecture, optimizer, and number of training steps fixed across arms. As a secondary use of the same forecasts, we test a recalibration trigger: re-characterize the device only when the forecast upper quantile of a critical error rate crosses a threshold. We compare this trigger against fixed-interval recalibration at equal characterization time.
Baselines
The naive reference point is noise-free simulated training executed directly on hardware. It establishes the gap, which RoQNN reports at over 60% on MNIST-4 on IBMQ-Yorktown.
The strongest existing method is QuantumNAT itself. It combines static snapshot noise injection with its normalization and quantization steps, and it reports over 94% 2-class and 80% 4-class accuracy on real quantum computers and up to 43% improvement. It is the strongest competitor because it already closes most of the gap using the same calibration files we start from. We run it from its published pipeline, using the snapshot available at training time, on the same tasks, devices, execution times, shot counts, and step budget as our method, with the same hyperparameter search budget. Beating it means higher mean hardware accuracy across paired runs at matched $\Delta$, with the gain statistically distinguishable from device-to-device and day-to-day variation, and ideally growing with $\Delta$.
A second strong baseline is snapshot-refreshed QuantumNAT. It retrains or fine-tunes on the freshest snapshot immediately before execution, and it is charged the extra compute and time that retraining costs.
The ablation that isolates our key ingredient is unconditional randomization: sample $\eta$ from the pooled historical distribution for that device, which gives the same spread without conditioning. A second ablation strips the physical bounds from the forecaster. A third replaces the joint forecaster with independent per-qubit forecasts. For the forecasting stage alone, the reference is persistence, meaning tomorrow equals today with empirical error quantiles. We found no reported result on Lacuna for probabilistic forecasting of quantum calibration drift, so all forecasting numbers will be new.
Why not just…?
Why not just inject broader random noise instead of forecasting it? This is the real threat. The adaptation-gain scaling in When Does Adaptation Win? suggests that spread alone buys robustness. If drift on cloud devices is mostly unpredictable jitter around a stable mean, unconditional randomization will match forecast sampling. The unconditional-randomization ablation settles this directly. If it ties, we will report that conditional forecasting is unnecessary for training and redirect the forecasts to the recalibration-trigger use.
Why not just retrain on the newest calibration snapshot right before execution? When the queue delay is short and retraining is cheap, this may win, and our claim predicts no advantage at $\Delta \approx 0$. The vendor snapshot is itself a stale point estimate, and queue delays are not under the user's control. We measure both arms across $\Delta$, with retraining cost included, to find where each dominates.
Why not just recalibrate on a fixed timer? Proctor et al. show that periodic recalibration helps. When Does Adaptation Win? notes that recalibration can consume hours of daily uptime, and much of that time is spent on devices that have not moved. The trigger experiment compares a forecast-quantile trigger against fixed intervals at equal characterization time. The measures are the resulting gate error and the fraction of drift excursions each schedule catches.
Why not just use an off-the-shelf deep forecaster? We include one. In this regime, however, a data-light state-space model or plain persistence may simply win. We select the forecaster by CRPS on held-out time periods, not by preference.
Decisive Checks
Three results decide the proposal. The first is whether forecast-sampled training beats unconditional randomization on real-hardware accuracy for the QuantumNAT tasks (MNIST-2, MNIST-4, Fashion-MNIST, Vowel-4). The second is whether its margin over static QuantumNAT grows with $\Delta$ and shrinks on the devices our forecasts identify as stable; that dose-response pattern is the signature that forecasting, rather than noise magnitude, is doing the work. The third is whether the constrained forecasters beat persistence on CRPS and coverage at the delays that matter.
If the first check fails while the third succeeds, forecasting quantum drift is still useful for scheduling recalibration but not for training. If the third fails, device drift at these timescales is effectively unforecastable, and broad robust training is the correct answer.
References
- QuantumNAT: Quantum Noise-Aware Training with Noise Injection, Quantization and Normalization
- Detecting and tracking drift in quantum information processors
- When Does Adaptation Win? Scaling Laws for Meta-Learning in Quantum Control
- Adaptive Multi-task Learning for Probabilistic Load Forecasting
- Probabilistic forecasts of wind power generation in regions with complex topography using deep learning methods: An Arctic case
- Using probabilistic solar power forecasts to inform flexible ramp product procurement for the California ISO
- Cooperative Virtual Power Plant Formation Using Scoring Rules
- DreamQAS: Learning a Decision-Useful World Model for VQE-Efficient Quantum Architecture Search
- QASMBench: A Low-Level Quantum Benchmark Suite for NISQ Evaluation and Simulation