Towards Scaling Quantum Fine-Tuning of Foundational Time Series Models for Classification

AI summary

Overview

Research area: Quantum machine learning applied to time-series foundation models — specifically, replacing the classical fine-tuning head of a frozen forecasting model with a parameterized quantum circuit, evaluated on a power-grid event-classification benchmark.

Technical level: Intermediate to Advanced. The paper assumes familiarity with parameterized quantum circuits, data re-uploading, post-selection, and hybrid quantum–classical training loops, although the core narrative (add qubits that carry new information, not just more parameters) is accessible to a motivated non-specialist.

Scope (one sentence): The paper tests whether a quantum head on frozen Chronos-T5 embeddings can beat classical baselines on the five-class PSML-5 power-grid benchmark, diagnoses why a fixed-width circuit saturates, and introduces a "wing" module that scales qubit count by adding input bandwidth rather than circuit depth.

What This Paper Is About

Large time-series foundation models such as Chronos produce rich embeddings, but it is unclear whether quantum circuits can exploit those embeddings and how such hybrid heads should scale. The authors freeze a Chronos-T5 backbone, attach either a quantum circuit or a classical multilayer perceptron as a classification head, and test both on PSML-5, a five-class power-system disturbance dataset. When the quantum head's accuracy stops improving despite richer inputs, they build a new circuit topology to widen the head's data intake instead of enlarging the circuit.

Key Contributions

1. Quantum fine-tuning extended to time-series foundation models. The authors convert a forecasting model (Chronos-T5) into a multi-class event classifier on PSML and report that this setup surpasses the best specialized classical baseline originally published for that dataset (MLSTM-FCN at 74.2% ± 2.9% balanced accuracy).
2. A scaling study of a fixed-width quantum head. They sweep the sketch-fingerprint budget fed to a 12-qubit core and show it saturates, while on bit-identical inputs the quantum head plateaus roughly 1.7–2.0 percentage points above a larger classical MLP head.
3. The "wing" architecture. A self-contained three-qubit module with its own sPQC encoder that couples into the core circuit only through a few fixed, one-way CNOT gates at staggered points, never coupling to other wings, and leaving core depth constant — so qubit count grows intake capacity without adding depth or fully connected width.
4. Attribution ablations. Two controlled ablations at the 19-qubit operating point separate added circuit capacity from added payload information, showing that capacity growth without new information gains nothing and that mismatched wing payloads perform worse than having no wing at all.

Main Findings

• Type-group pooling gives the largest single jump. Grouping the 91 per-channel Chronos embeddings by PMU measurement type (voltage, active power, reactive power) before mean/std summarization lifts both heads to roughly 82% balanced accuracy (81.93% quantum, 82.10% classical), up from 74.43% for a plain classical head (E0) that already matches the published 74.20% baseline.
• The quantum head leads on identical inputs once finer features are added. From a fingerprint budget of two onward, the quantum head outperforms a larger classical MLP head on bit-identical cached inputs by 1.7–2.0 percentage points of balanced accuracy. The quantum head has 3,240 trainable parameters versus 1.2–3.0 million for the MLP.
• Fixed-width intake saturates. The quantum head's trajectory is 84.21% → 84.29% → 84.39% → 84.37% at budgets 8, 16, 96, and 160. Budgets 96 and 160 switch from the base sketch to the base+raw family, so the 16 → 96 point is a family switch rather than "more of the same"; the near-flat result (84.29% vs. 84.39%) indicates the fixed 12-qubit width, not feature supply, is the binding constraint.
• Wings increase accuracy monotonically. Under a preregistered four-seed (42–45) protocol on a fixed 12-qubit core with fixed features: 83.63% (13 qubits, zero wings) < 84.62% (16 qubits, one wing) < 85.23% (19 qubits, two wings). The 12-qubit reference sits below all rungs at 83.39% (and at 83.53% for the n=4 mean at learning rate 0.010).
• A selection-free check preserves the ordering. Sharing one learning rate (0.008) across all four configurations with no per-rung tuning: 12-qubit 83.01% < 13q 83.63% < 16q 84.08% < 19q 84.87%. The authors note this rate is near-optimal for the smallest rung and off-peak for larger ones, biasing against the ordering it nonetheless preserves.
• More circuit capacity without new information does not help. Un-sharing the data-reuploading block raises rotation-parameter count from 150 to 204 while inputs stay identical; balanced accuracy drops from 85.23% to 83.85% (Δ = −1.38 points), with a deficit of −1.4 to −1.9 points at every candidate learning rate.
• Mismatched wing payloads are actively harmful. Shuffling wing input channels across samples (identical dimensionality, distribution and parameter count, but wrong sample pairing) drops balanced accuracy to 81.59% — a fold-paired deficit of 4.20 ± 3.92 points over the 15 shared folds, with 13 of 15 folds negative, one tie, and one positive by 0.47 points (paired t = −4.15, n = 15, p ≈ 0.001, reported descriptively; seed-level differences are −5.22, −4.56, −2.82 points). This is 2.04 points below the wingless 13-qubit model (83.63%).
• Initialization does not rescue a grown model. The enlarged ablation circuit starts as an exact functional copy of its smaller baseline yet still ends below it; every grown model in the ladder is trained from scratch.
• The benchmark is hard for the right reasons. PSML instances show high intra-class variance (six generator-trip examples vary widely in baseline, onset, and amplitude on channel #66), yet a 50-component PCA gives inter-to-intra-class distance ratios of 1.171 (standard scaling) and 1.131 (robust scaling), so a class-level signal remains recoverable.

Methodology in Plain English

The team took a pretrained Chronos-T5 base model (200M parameters, d_model = 768) and never updated its weights. Each PSML-5 sample is a window of shape (91, 960) — 91 PMU-type sensor channels over 960 millisecond steps. Each channel is passed independently through the frozen encoder, the per-timestep hidden states are mean-pooled over the window, and the resulting 768-dimensional per-channel embeddings are concatenated into one raw feature vector.

From that shared representation, two heads are built in parallel: a classical MLP (1.2–3.0 million trainable parameters) and a quantum circuit (3,240 trainable parameters). The quantum head compresses the feature vector with a simulated-PQC encoder — which consistently beat a classical dense encoder — into a 12-qubit core running data re-uploading with R = 2 repetitions. Readout is local Z measurement on five core wires feeding a linear head into five classes. A dedicated post-selection qubit is coupled late through controlled-R_Y rotations initialized near zero (θ₀ = 0.05); only the |0⟩ outcome is kept, and the loss penalizes acceptance rates below a floor of 0.25.

To test saturation, the authors grew the classical feature vector fed to the fixed 12-qubit core with "sketch fingerprints" — compact deterministic per-block summaries, including the base sketch from pooled blocks and the raw-channel sketch from per-channel streams. When accuracy plateaued, they added wings: three-qubit registers, each with its own sPQC encoder and its own per-channel feature stream, coupled into the core by a few fixed CNOTs at staggered points. Core depth and the 96-fingerprint core feature set stay bit-identical across all rungs, so 0, 1, and 2 wings give 13, 16, and 19 qubits including the post-selection qubit.

Hyperparameters were tuned with Optuna's TPE sampler using five-fold cross-validation on the 439-sample training split (549 samples total, deterministic 80/20 split by sorted row index, yielding 110 test samples). The scaling ladder uses a bracket search on seed 42 to find peak learning rate, then evaluates the peak and its two neighbors across seeds 42–45. All circuits were simulated noiselessly with state-vector simulation.

Why This Matters

Impact on research. The paper reframes what "scaling" means for quantum fine-tuning: added qubits help when they carry added inputs, not merely more trainable parameters. The ablation showing that an enlarged, better-initialized circuit can end up below its smaller baseline is a concrete counterexample to the assumption that more capacity plus a good warm start is automatically better. The wing design also offers a shallow, sparse route to wider quantum intake that avoids the two-qubit-gate overhead and trainability pressure of deepening or fully connecting the register.

Real-world applications (as motivated by the paper's domain):

• Power-grid disturbance classification, distinguishing generator trips, branch trips, bus trips, branch faults, and bus faults from synchrophasor measurements.
• Real-time isolation of physical failures in transmission and distribution networks, where automated control systems must classify events quickly.
• Preventing cascading blackouts and maintaining grid stability as renewable energy integration increases.
• Statistical event triage on millisecond-resolution PMU data, distinguishing intentional breaker openings from active short circuits.

Industry relevance. The work originates from IonQ, QuantumBasel, and the University of Basel, and positions near-term quantum hardware as a fine-tuning head on existing AI infrastructure rather than a replacement for it. Because foundation-model weights stay frozen, the quantum resources required scale with the size of the head, not the backbone — an attractive property for organizations that already hold pretrained models and want a small, task-specific quantum component.

Future Directions

• Execution on quantum hardware. Every circuit here was simulated noiselessly; the authors call for running the trained heads on a QPU and characterizing shot-noise sensitivity at inference.
• Extending the ladder beyond two wings and replicating the attribution ablations at other scales, since the current ablations were run only at the 19-qubit operating point.
• A gradient-based trainability analysis of wing-extended circuits. The current trainability argument is made at the design level (constant per-module depth, restricted connectivity) and supported only by the ladder training successfully at every rung.
• Systematic backbone and initialization studies. The authors did not evaluate Chronos-2, Lag-Llama, or TimeGPT, and they flag initialization for grown models as an unanticipated and unresolved question; additional datasets are also needed, since all numbers come from a single benchmark.

Target Audience

Quantum machine-learning researchers working on hybrid quantum–classical architectures and NISQ-era circuit design; practitioners applying foundation models to time-series classification who want to know whether a quantum head adds anything over a classical one; and power-systems and energy-sector data scientists interested in event classification from synchrophasor data. Readers focused on proof-of-concept hardware demonstrations should note that all results here are noiseless state-vector simulations and internal five-fold cross-validation on the training split.

Authors’ abstract

Time-series foundation models produce rich embeddings, but whether quantum models can exploit them, and how far hybrid classical-quantum architectures scale, remains unclear. We address this by fine-tuning Chronos for power-grid event classification (PSML-5) with a quantum head on the model's embeddings. Grouping embeddings by physical sensor type before summarization already surpasses the best published baseline built for this benchmark, and with finer-grained features the quantum head outperforms a larger classical multilayer perceptron on identical inputs by 1.7-2.0 percentage points of balanced accuracy. Yet the gains saturate: past a point, feeding more information to the same fixed-width register yields no improvement. We show the bottleneck is neither the supply of information nor circuit expressiveness, but the bandwidth of the data intake. To overcome this limitation, we introduce the wing module, a self-contained few-qubit circuit that feeds additional information into the core circuit through a sparse, one-way coupling. Under a preregistered four-seed protocol, we attach wings to a fixed 12-qubit core with fixed features. Balanced accuracy increases with each added wing, from 83.6% with no wings (13 qubits, including a post-selection qubit) to 85.2% with two (19 qubits). Ablations establish that a circuit enlarged without new information gains nothing, while a wing fed information from the wrong sample harms accuracy. These results reframe scaling for quantum fine-tuning: added qubits help when they carry added inputs, not merely more parameters. Wings offer a modular and stable route to widening that bandwidth.