Coherent Floquet quantum reservoirs for molecular property prediction

AI summary

Overview

Research area: Quantum machine learning / quantum reservoir computing, applied to molecular property prediction (cheminformatics and excited-state molecular dynamics).

Technical level: Advanced. The paper assumes familiarity with quantum reservoir computing, Floquet driving, discrete time crystals, density-matrix channels, and classical kernel/linear decoders.

Scope (one sentence): The paper builds a discrete-time-crystal-based quantum reservoir (DTC-QRC) with a controlled reset channel, feeds it molecular graph events and surface-hopping trajectory frames, and benchmarks it against matched echo-state networks for inhibitor-activity classification, blood–brain-barrier permeability classification, and ethene electronic-gap forecasting, including dephasing simulations and runs on the Quafu superconducting cloud platform.

What This Paper Is About

Molecular property prediction from either a static structure or a time-resolved trajectory requires a fixed-width representation that retains relations spread across many local observations. The authors ask whether a periodically driven quantum system near the discrete-time-crystal regime can serve as that representation, with a tunable reset that controls how much earlier inputs influence the final measurement. They test this on two very different molecular event streams: breadth-first-search (BFS) graph events for classification, and surface-hopping frames for forecasting future electronic energy gaps.

Key Contributions

1. A DTC-QRC architecture with controlled reset for molecular event streams. The reservoir applies a per-event input rotation, then three Floquet cycles of a near-π pulse followed by longitudinal disorder and nearest-neighbor Ising couplings, with a global reset channel applied either after every graph token or between physical trajectory frames. Unrolling the recurrence yields a geometric "reset-induced event-age kernel" with survival weight (1 − p)^ℓ, giving an explicit physical knob for how strongly older events contribute.

2. Two molecular interfaces sharing one reservoir rule. BFS-derived local chemical tokens (atom discovery, parent–child bonds, ring closures, traversal depth, padding) drive a 12-qubit reservoir; SHNITSEL A01 ethene surface-hopping frames (15 interatomic distances plus backward-difference radial velocities) drive a 15-qubit reservoir, with final-state Z + ZZ observables supplying fixed-width feature vectors while reservoir parameters stay fixed during decoder training.

3. Matched-baseline benchmarking against echo-state networks. At identical input lengths, output widths, data splits, seeds, and decoder grids, DTC-QRC is compared with ESNs on long-prefix graph classification and on four ethene gap targets, with paired bootstrap or crossed seed-by-trajectory confidence intervals reported.

4. Coherence and hardware evidence. Event-wise dephasing simulations quantify the dependence on coherent propagation, and experiments on the Quafu superconducting quantum cloud platform examine whether pair observables retain task information under device noise.

Main Findings

• Long-prefix graph classification favors DTC-QRC. With Q = 12, T = 30, and p = 0.10, BACE test ROC-AUC reaches 0.798 ± 0.029 versus 0.762 ± 0.034 for the matched 78-node ESN (difference 0.0366, 95% CI [0.0246, 0.0503]). At T = 24, BACE is 0.791 ± 0.022 versus 0.754 ± 0.029 (difference 0.0368, CI [0.0180, 0.0589]).

• BBBP shows the same pattern. At T = 24, BBBP AUC is 0.761 ± 0.036 versus 0.706 ± 0.040 (difference 0.0547, CI [0.0424, 0.0658]); at T = 30, 0.763 ± 0.037 versus 0.727 ± 0.033 (difference 0.0355, CI [0.0147, 0.0579]). All six paired seeds favor QRC at T = 24 and 30 in both datasets.

• The decoder matters as much as the reservoir. At T = 30, p = 0.10, BACE AUC rises from 0.656 with an affine logistic head to 0.798 with RBF-SVR, and BBBP follows the same pattern. Weak reset slightly lowers affine-decoder AUC relative to p = 0 while improving RBF-SVR, indicating that nonlinear decoding is needed to exploit the measured relations.

• Controlled reset improves all four gap forecasts over the matched ESN. With Q = 15, T = 20, final-state readout, and RBF-SVR, validation-selected reset gives: 1-fs gap 0.716 ± 0.034 eV versus 1.306 ± 0.057 eV (difference −0.591 eV, CI [−0.688, −0.495]); 5-fs gap 1.000 ± 0.047 versus 1.207 ± 0.045 (difference −0.207); 10-fs gap 1.440 ± 0.024 versus 1.529 ± 0.020 (difference −0.089); next-5-fs minimum 0.800 ± 0.047 versus 1.097 ± 0.040 (difference −0.298). Selected reset values differ by target: p = 0.90, 0.65, 0.80, and 0.85 respectively.

• A non-quantum baseline can match part of the gap-forecasting gain. An Extra Trees regressor on the same inputs gives a 10-fs test MAE of 1.404 eV, compared with 1.440 eV for QRC, so the advantage over the ESN does not automatically extend to all classical models.

• Complete reset isolates the value of reservoir memory. Setting p = 1 gives MAEs of 0.722 ± 0.027 eV (1 fs), 1.144 ± 0.018 eV (5 fs), 1.461 ± 0.044 eV (10 fs), and 0.852 ± 0.026 eV (next-5-fs minimum). The difference from the validation-selected p is +0.0065, +0.1442, +0.0217, and +0.0527 eV respectively, so retaining history helps most at the intermediate 5-fs horizon.

• Reset preferences differ between the two applications. Weak reset improves several longer graph prefixes, with the clearest gains on longer BBBP prefixes, while stronger reset emphasizes recent motion in the trajectory forecasts. Strong reset degrades graph classification, consistent with loss of useful structural context.

• A common low-error Floquet-drive window exists. Averaging percentage excess over each configuration's scanned minimum and averaging within-configuration drive ranks over the 20 target–reset configurations with p > 0 favor g = 0.88 and g = 0.86 respectively, identifying g ≃ 0.86–0.88 as a shared low-error window near the DTC reference g_c ≃ 0.84. At g = 0.84, 17 of the 20 dissipative RBF-SVR/final configurations lie within 5% of their own scanned minima.

• The drive matters even without reservoir memory. At p = 1, the 1-fs target with RBF-SVR and final readout falls from 1.403 eV at g = 0 to 0.722 eV at g = 0.84, showing that the drive governs the quality of the input-to-observable map itself.

• Dephasing degrades both applications. For BACE at λ = 0.5 and k = 8 Pauli trajectories, the closed-reservoir AUC falls from 0.783 to 0.504; weak reset reduces but does not eliminate the loss. Gap forecasts show target-balanced MAE ratios above unity for both reset settings and both trajectory counts (k = 8 and k = 32), referenced to coherent evolution at the same p.

• Hardware behavior on Quafu. The four-point drive scan uses a balanced 400-molecule BACE subset with Q = 6, T = 6, L = 3, W = 0.6, p = 0, 15 connected pair coordinates C_ij^ZZ, ten nested scaffold/decoder analyses, and three Shenglian acquisition batches. The authors report better retention of task information in second-order correlations in the DTC regime, with the transition edge giving the highest mean hardware classification performance under full Z + ZZ readout among the sampled drives; the available text truncates before the numerical hardware AUC values, which are not reported in the excerpt.

Methodology in Plain English

Each molecule is turned into a sequence of small "events." For graphs, the molecule is parsed from a canonical SMILES string, traversed breadth-first, and each step is encoded into triples of numbers describing atom type and degree, bond context, ring closure, and traversal depth. Whole-molecule descriptor slots are fixed at zero, so any predictive information must come from accumulating local events. For trajectories, each frame contributes 15 interatomic distances and their radial velocities, standardized, clipped to [−5, 5], and scaled by 0.5 before injection.

The quantum reservoir works like this: every qubit starts in the all-zero state. An input event rotates each qubit by an amount set by its three coordinates, then the system undergoes three Floquet cycles consisting of a near-π X pulse followed by disorder and nearest-neighbor Z–Z couplings. Because these operations do not commute, the response to a new event depends on what came before. A global reset channel, applied with probability p, mixes the state back toward the all-zero state, either after every graph token or between trajectory frames.

At the end of the sequence, the researchers measure all single-qubit Z expectations and all pairwise Z–Z correlations, giving 78 numbers for 12 qubits and 120 for 15 qubits; a temporal-multiplexing variant (TM3) concatenates readouts at the endpoint and after one and two extra Floquet cycles, giving 360 numbers for 15 qubits. These features go into classical decoders — RBF support-vector regression, ridge regression, or a logistic head — trained while the quantum reservoir stays fixed.

Comparisons use echo-state networks with the same token arrays, the same number of recurrent coordinates (78 for graphs, 120 for trajectories), spectral radius 0.90, input scale 0.75, and leak rate 0.65, plus the same splits, seeds, and decoder grids. Statistical uncertainty is estimated by paired bootstrap (20,000 draws, seed 20260905) for graphs and by crossed seed-by-trajectory resampling for trajectories. Dephasing is simulated by inserting a per-qubit bit-flip-style channel with strength λ after every event, averaging over 8 or 32 Pauli trajectories.

Why This Matters

The work links two things that are usually studied separately: the physics of periodically driven, strongly interacting spin systems near a discrete-time-crystal transition, and the practical machine-learning problem of representing molecular structure and motion. It shows that a single fixed quantum dynamical rule plus one tunable reset parameter can serve both a static screening task and a time-resolved forecasting task, and it provides controlled comparisons against classical reservoirs at matched size rather than against unmatched baselines.

Real-world applications:

• Drug discovery screening: inhibitor-activity prediction (BACE) and blood–brain-barrier permeability prediction (BBBP) are standard steps in prioritizing compounds.
• Photochemistry and materials design: forecasting electronic energy gaps several femtoseconds ahead along excited-state trajectories supports non-adiabatic dynamics studies.
• Molecular simulation acceleration: using a small measured feature vector instead of storing long histories could reduce the classical post-processing burden in trajectory analysis.
• Quantum hardware benchmarking: the reset and drive scans offer a concrete protocol for tuning a driven quantum device toward a useful operating regime.

Industry relevance: pharmaceutical and biotech pipelines that rely on molecular classification, and computational chemistry groups working with excited-state dynamics, are the most direct beneficiaries. Hardware providers also gain a benchmark that explicitly separates coherent information mixing from noise resilience, which is a practical criterion for reservoir selection and drive tuning on near-term superconducting devices.

Future Directions

• Extending beyond the truncated hardware study. The Quafu results in the available text stop before the numerical device AUCs; a fuller hardware comparison across more drive points, larger Q, and the four-target regression tasks is the obvious next step.
• Understanding when classical models suffice. Extra Trees reaches 1.404 eV on the 10-fs gap versus 1.440 eV for QRC, so a systematic comparison against stronger classical sequence models and message-passing networks is needed to delimit the quantum advantage.
• Making reset selection cheaper. Reset strengths were chosen per target by validation MAE over 27 candidate values; whether a single p can serve multiple targets, or be predicted from trajectory statistics, remains open.
• Noise mitigation and error modeling. Dephasing substantially degrades both tasks, and the paper's interpretation that dephasing suppresses off-diagonal elements that later convert into populations suggests testing whether coherence-preserving encodings or mitigation strategies recover performance on hardware.

Target Audience

Quantum machine learning researchers working on reservoir computing and near-term quantum advantage; physicists interested in discrete time crystals and Floquet-engineered dynamics as computational resources; computational chemists and cheminformaticians interested in event-stream representations of molecules and trajectories; and quantum hardware teams looking for application-level benchmarks that separate coherent dynamics from noise effects.

Authors’ abstract

Quantum reservoir computing (QRC) uses quantum dynamics to represent input histories for prediction through a trained classical readout. Discrete time crystals (DTCs) exhibit robust subharmonic responses under periodic driving, and previous work has used their dynamics to construct DTC-QRC. Here we construct a DTC-based reservoir architecture to predict molecular properties from structural and dynamical observations. Coherent Floquet evolution processes local molecular graph events and surface-hopping frames, while controlled reset regulates the contribution of earlier inputs. Measurements at the end of each input sequence yield a feature vector of fixed dimension. Trained classical decoders use this vector for inhibitor-activity and blood--brain-barrier permeability classification and electronic-gap forecasting, while the reservoir parameters remain fixed during training. With matched input lengths and output widths, DTC-QRC outperforms echo-state networks on long-prefix graph classification and the studied ethene gap forecasting tasks. Dephasing lowers performance in both applications, consistent with a role for coherent propagation. Experiments on the Quafu superconducting quantum cloud platform show that pair observables retain task information under device noise. The architecture provides a common framework for molecular screening and time-resolved property prediction using quantum reservoir computing.