Fixed-Reservoir vs Variational Quantum Architectures for Chaotic Dynamics: Benchmarking QRC and QPINN on the Lorenz System

Abstract: Deploying quantum machine learning on NISQ devices requires architectures where training overhead does not negate computational advantages. We systematically compare two quantum approaches for chaotic time-series prediction on the Lorenz system: a variational Quantum Physics-Informed Neural Network (QPINN) and a Quantum Reservoir Computing (QRC) framework utilizing a fixed transverse-field Ising Hamiltonian. Under matched resources ($4$--$5$ qubits, $2$--$3$ layers), QRC achieves an $81\%$ lower mean-squared error (test MSE $3.2 \pm 0.6$ vs. $47.9 \pm 36.6$ for QPINN) while training $\sim 52,000\times$ faster ($0.2$\,s vs. $\sim 2.4$\,h per seed). Drawing on the classical delay-embedding principle, we formalize a temporal windowing technique within the QRC pipeline that improves attractor reconstruction by providing bounded, structured input history. Analysis reveals that QPINN instability stems from capacity limitations and competing loss terms rather than barren plateaus; gradient norms remained large ($10^3$--$10^4$), ruling out exponential suppression at this scale. These failure modes are absent by construction in the non-variational QRC approach. We validate robustness across three canonical systems (Lorenz, Rössler, and Lorenz-96), where QRC consistently achieves low test MSE ($3.1 \pm 0.6$, $1.8 \pm 0.1$, and $12.4 \pm 0.6$, respectively) with sub-second training. Our findings suggest the fixed-reservoir architecture is a primary driver of QRC's advantage at these scales, warranting further investigation at larger qubit counts and on hardware where quantum-specific advantages are expected to emerge.

• The paper demonstrates that fixed-reservoir quantum architectures (QRC) achieve an 81% lower train MSE and 93% lower test MSE compared to variational quantum circuits (QPINN) on chaotic systems.
• The paper details how QRC leverages closed-form ridge regression readouts and temporal windowing to rapidly and accurately model the Lorenz attractor.
• The paper contrasts these quantum methods with classical ESNs, highlighting QRC's scalability and hardware amenability despite current validation on classical simulations.

Head-to-Head Evaluation of Fixed-Reservoir and Variational Quantum Architectures for Chaotic Dynamics

Overview

This paper delivers a systematic, resource-matched comparison between two leading quantum machine learning paradigms for modeling chaotic dynamical systems: Quantum Reservoir Computing (QRC) and Quantum Physics-Informed Neural Networks (QPINN). The experimental focus is on the Lorenz system, with extension to Rössler and Lorenz-96 systems. The empirical evidence establishes a decisive practical advantage for fixed-reservoir quantum architectures over end-to-end variational quantum circuits on NISQ-era devices, both in prediction performance and computational efficiency.

Quantum Architectures: Implementations and Principles

Quantum Physics-Informed Neural Network (QPINN)

QPINN leverages parameterized quantum circuits (PQCs) to fit the mapping $t \mapsto (x(t), y(t), z(t))$, encoding the ODE dynamics of the Lorenz attractor directly into a variational loss. The architecture exploits rotation-based time encoding and a layered PQC ansatz, with observables mapped to physical variable ranges. Training minimizes a physics-informed loss—a compound of dynamical residuals and boundary terms—via Adam optimizer.

However, for the resource regime explored (4 qubits, 3 layers, 45 parameters), QPINN exhibits persistent oscillatory loss, high variance across seeds, and capacity-limited convergence far from the optimum. Gradient vanishing (barren plateaus) was not observed; gradient norms remained large throughout optimization.

Figure 2: QPINN training converges slowly and oscillates, confirming capacity, not barren-plateau, driven failure.

Quantum Reservoir Computing (QRC)

QRC utilizes a fixed random quantum circuit initialized with a transverse-field Ising model topology. At each step, the Lorenz state is encoded via angle rotations, and the resulting quantum state is measured across all qubits, yielding a $2^n$-dimensional feature vector. Measurement features are collected into temporal windows and subsequently aggregated by a classical ridge regression readout layer—requiring no quantum circuit training.

This design eliminates gradient-based quantum optimization, accelerates model construction via closed-form solution, and incorporates trajectory context through temporal windowing.

Figure 3: Schematic of the 5-qubit, 2-layer fixed QRC circuit and temporal windowing for feature extraction.

Empirical Results

Lorenz System: Direct Comparison

With matched quantum resources, QRC yields an average train MSE of $17.1 \pm 3.7$ and test MSE $3.2 \pm 0.6$, outperforming QPINN by 81% lower train MSE and 93% lower test MSE (QPINN: $91.3 \pm 21.9$ train, $47.9 \pm 36.6$ test, across 5 seeds). QRC's training time per seed is ∼0.2 seconds, compared to QPINN's ∼2.4 hours—a 52,000× speedup.

Figure 4: QRC surpasses QPINN in both accuracy (train MSE) and computational efficiency across matched seeds.

Qualitative trajectory predictions confirm QRC's ability to track the Lorenz attractor far into the extrapolation regime.

Figure 1: QRC replicates true Lorenz dynamics (reference and prediction), demonstrating stability and accuracy even for chaotic flow.

Temporal Windowing

Ablation over window sizes establishes windowing as critical to performance. Increasing temporal context from $w=1$ (no window) to $w=5$ (full window) reduces train MSE by 54%, from 48.0 to 22.1.

Figure 5: Longer temporal windowing significantly augments QRC prediction accuracy.

Cross-System Generalization

QRC was validated on Rössler and Lorenz-96, maintaining sub-second training and consistent performance: test MSE $1.8 \pm 0.1$ (Rössler), $12.4 \pm 0.6$ (Lorenz-96). Increased system dimensionality (Lorenz-96) reveals QRC's limitations for larger state spaces when employing only 5 qubits.

Contextualizing with the Classical Baseline

A classical Echo State Network (ESN) baseline with 500 neurons surpasses both quantum approaches (train MSE $2^n$0, test MSE $2^n$1). The classical random reservoir remains larger in effective dimensionality ($2^n$2 for 5-qubit QRC vs 500 dim ESN), but QRC demonstrates its key advantage—tractable quantum-native feature extraction—will likely manifest as qubit counts scale beyond classical simulability.

Analysis: Fixed Reservoirs Dominate in Practicality

The study shows that QRC's performance edge is not quantum-specific at the small qubit scales, but rather derives from the reservoir computing paradigm—fixed, high-dimensional nonlinear transformations, and closed-form classical readout training. Variational approaches, as exemplified by QPINN, are fundamentally encumbered by expressivity limitations and the computational load of quantum-gradient estimation, particularly in noisy regimes and for chaotic, high-complexity signals.

Task asymmetry exists: QRC leverages windowed historical state input for prediction, while QPINN attempts parametric function synthesis from time alone. Thus, QRC's true architectural advantage may be somewhat overstated in this context.

Theoretical and Practical Implications

QRC provides a scalable and hardware-amenable path for quantum representations of nontrivial dynamical systems, as it dispenses with the difficulties of quantum gradient training and is robust to random initialization. With hardware progress, QRC's accessible Hilbert space will grow, and the crossover point where quantum reservoirs can systematically outperform classical ESNs is an urgent target for empirical verification.

QPINN's lack of barren plateau issues at this scale confirms that optimization failure is dominated by parameter inefficiency, not quantum-intrinsic trainability barriers. Practical improvements will likely require richer ansätze and data-supported loss terms, or hybridization with windowed inputs (a windowed variational quantum circuit, as noted).

Limitations and Future Work

All results were obtained via classical quantum simulation. Hardware-native benchmarking, particularly for the QRC regime, must be performed to establish real-world readiness and noise tolerance. Exploration at higher qubit counts and with more elaborate temporal or system encodings will clarify where QRC's theoretical scaling yields substantive real-world advantage. Full ablation studies controlling for information symmetry (windowed QPINN) are necessary for definitive architectural ranking.

Conclusion

Quantum Reservoir Computing, as implemented with fixed, randomly-initialized circuits and classical readout, substantially outperforms resource-matched variational quantum circuits on canonical chaotic time-series tasks at the NISQ scale—both in predictive accuracy and compute efficiency. The key drivers are the closed-form readout optimization and the ability to incorporate temporal context. As quantum hardware evolves and system scales increase, QRC will become essential for leveraging quantum-enhanced modeling of nonlinear dynamical phenomena. The released codebase ensures reproducibility and provides a strong foundation for future empirical work on quantum machine learning for complex systems.