Hybrid quantum-classical physics-informed neural networks for solving nonlinear PDEs: when and where hybridization is effective?

Abstract: Physics-informed neural networks (PINNs) often struggle on nonlinear partial differential equations (PDEs) with sharp gradients, stiff dynamics, high-frequency content, or multiscale structure. Such limitations, rooted in spectral bias, ill-conditioned optimization, and unstable convergence, restrict PINN accuracy in regimes where advanced solvers are most needed. In this work, we develop a hybrid quantum-classical physics-informed neural network (HQPINN) that integrates a classical neural-network backbone with a parameterized quantum circuit (PQC) to enrich the solution representation. The framework is benchmarked against a classical PINN on three representative nonlinear PDEs: Burgers' equation, the Allen-Cahn equation, and the Korteweg-de Vries (KdV) equation. The framework is further examined through a systematic sensitivity analysis of qubit count, circuit depth, PQC placement, collocation density, and classical-network width. Across all benchmarks, HQPINNs exhibit smoother training dynamics, reduced loss oscillations, and improved final accuracy, with the largest gains occurring in stiff and multiscale regimes. Relative L2 error decreases by about fourfold for Burgers' equation and fivefold for the Allen-Cahn equation, while improvements for the KdV equation are more moderate. Overall, the results demonstrate that carefully co-designed hybrid quantum-classical architectures can mitigate key limitations of classical PINNs and provide practical design guidance for near-term quantum-enhanced PDE solvers.

• The paper introduces a hybrid architecture embedding a parameterized quantum circuit into a classical neural network to mitigate spectral bias in PINNs for nonlinear PDEs.
• This approach achieves up to 4-5 times lower relative L2 errors in Burgers' and Allen-Cahn equations, demonstrating improved training convergence and localized error suppression.
• Results indicate that optimal performance is attained when the quantum circuit is placed after the classical network, emphasizing the importance of careful circuit design and placement.

Hybrid Quantum-Classical Physics-Informed Neural Networks for Nonlinear PDEs: Analysis and Implications

Introduction

Physics-informed neural networks (PINNs) have established themselves as a compelling approach for solving deterministic and inverse problems related to PDEs by embedding physics-based residuals and constraints directly into the loss function. However, their application to nonlinear PDEs featuring sharp gradients, stiff dynamics, or multiscale structure is consistently limited by spectral bias, poor optimization conditioning, and slow or unstable convergence. These difficulties restrict the practical applicability and reliability of PINNs in physiologically and technologically significant settings.

Hybrid quantum-classical architectures, in particular the HQPINN proposed in this work, aim to mitigate these obstacles by combining a classical neural network backbone with a parameterized quantum circuit (PQC). The inclusion of a PQC is motivated by its theoretically superior expressive capacity in high-dimensional Hilbert spaces and the possibility of enriching the feature space in ways that classical neural nets cannot, thus potentially addressing the spectral bias and trainability bottlenecks of conventional PINNs.

HQPINN Architecture and Training

The introduced HQPINN comprises a deep classical feed-forward neural network (6 layers, 40 neurons per layer, tanh activations), which produces a bounded latent representation. This representation is angle-encoded onto an $n_q$-qubit register through single-qubit rotations. A variational quantum circuit, structured as $m$ layers of parameterized single-qubit rotations and ring-pattern CNOT entanglers, generates a quantum feature map. Measurement proceeds via Pauli-Z expectation values, which are passed to a final classical linear readout layer to yield the field solution prediction.

The entire network is end-to-end differentiable and trained with composite physics-informed loss (sum of meansquared PDE residual, initial, and boundary condition errors). Gradients traverse both classical and quantum subsystems unequivocally via automatic differentiation. Training leverages a two-stage optimization protocol: initial Adam phases for rapid landscape exploration, followed by L-BFGS-B for local refinement.

All benchmarks are conducted in classical quantum-circuit simulation, guaranteeing noiseless results and obviating device-related confounds.

Benchmarking and Sensitivity Analysis

Three canonical 1D nonlinear PDEs delineate the empirical regime: Burgers' equation (shock formation), Allen-Cahn equation (stiff metastable dynamics), and Korteweg-de Vries (KdV, dispersive nonlinear waves). Controlled evaluation entails identical architectures and training schedules, with the quantum component as the primary variable between HQPINN and PINN models.

A systematic sensitivity analysis interrogates five critical factors:

• Number of qubits in the PQC ($n_q=3..7$)
• Number of variational layers ($m=3..7$)
• PQC placement (early, mid, or post-classical network)
• Number of collocation (training) points
• Width of the classical backbone

Main Results

Accuracy and Convergence

HQPINNs demonstrate more stable training (smoother loss and error trajectories, reduced oscillations) and substantially reduced final errors in stiff and multiscale PDE contexts. Relative L2 error is reduced by a factor of four for Burgers' equation and five for Allen-Cahn versus classical PINNs. Gains for the smoother KdV equation are present but considerably less pronounced. Notably, for Burgers' and Allen-Cahn, the HQPINN displays sharper adaptation during second-stage (L-BFGS-B) optimization, a regime where classical PINNs stagnate or oscillate.

Local Error Suppression

Analysis of spatiotemporal error distributions indicates that the HQPINN better confines errors to regions of sharp gradients (Burgers') or interfaces (Allen-Cahn), substantially weakening localized error bands compared to PINNs. This demonstrates a pronounced benefit in capturing steep or stiff features.

Sensitivity to Quantum Circuit Design

Performance is non-monotonic in both qubit count and circuit depth. No universal optimal configuration exists; rather, each PDE admits a distinct sweet spot in the qubit-layer parameter space. Excess capacity degrades or plateaus accuracy due to optimization hardness, consistent with observations of barren plateaus and vanishing gradients in quantum circuits of increasing size.

PQC Placement

Empirically, the quantum circuit achieves best efficacy when placed after the final classical hidden layer, acting on a structured latent space. Input-stage or mid-network placements are suboptimal, impairing convergence and final accuracy. This is consistent with recent quantum ML literature emphasizing the synergy of classical preprocessing and quantum transformation for regression.

Data and Network Width Analysis

HQPINN exhibits superior robustness to sparse training point regimes for Burgers' and Allen-Cahn equations, whereas for KdV, it becomes advantageous only past a certain collocation density threshold. With increasing classical network width, HQPINN accuracy improves more consistently and substantially, particularly for problems with stiffness and sharp features—suggesting the hybrid model's effectiveness is partly contingent on rich classical representations prior to quantum transformation.

Mechanistic Interpretation

The observed HQPINN advantages arise from the synergistic interaction between the classical backbone (which constructs a problem-adapted latent space) and the quantum circuit (which injects additional nonlinear, oscillatory, and high-frequency content). The PQC acts as a feature enrichment module that, due to the embedding and variational structure, can realize trigonometric function classes inaccessible to standard classical networks. Simultaneously, the hybrid structure manifests smoother and more favorable optimization landscapes, as reflected in improved L-BFGS-B convergence.

This interaction specifically addresses spectral bias and balances gradient contributions in the multi-term PINN loss, leading to effectiveness in regimes of rapid spatial variation and stiff temporal evolution.

Limitations

• All results are limited to noiseless quantum simulation. Realistic hardware noise (decoherence, sampling, crosstalk) may negate or diminish observed accuracy and convergence gains. The findings are algorithmic, not evidence of near-term hardware feasibility.
• Conclusions are architecture-specific; broader ansatzes, encoding schemes, or measurement strategies are not explored.
• The reported optima depend on the training budget and optimizer protocol—different schedules or hyperparameters may shift observed trends.
• The collocation/sampling strategy is fixed; adaptive or problem-specific sampling could alter comparative performance.
• Benchmarks are limited to 1D PDEs; extension to multidimensional, multiphysics, or engineering-scale systems—where quantum models may face increased challenge or display new advantages—is deferred to future work.
• No runtime or wall-clock analyses are provided; the computational cost of classical quantum-circuit simulation is high.

Implications and Future Directions

The results directly inform practical design of hybrid quantum-classical solvers for nonlinear PDEs: quantum layers should be judiciously architected (not simply deepened), placed after substantial classical processing, and balanced with adequate network width and data sampling. There is potential for HQPINNs to become essential in computational physics domains where stiffness, shocks, or underresolved solution features degrade classical PINN performance.

However, translation to noisy intermediate-scale quantum (NISQ) hardware will require robust noise-aware design, ansatz search, and realistic benchmarking. The co-design of classical pre-processing and quantum circuits, along with rigorous training and architecture search, will be essential for practical impact. Advances in quantum circuit differentiation, expressivity/overparametrization theory, and error mitigation will further shape future developments.

Conclusion

The study establishes that HQPINNs, when appropriately architected, overcome core limitations of classical PINNs—most notably, spectral bias and convergence instability—in challenging nonlinear PDE domains. The quantum-classical hybrid advantage is sharply problem-dependent, not universal. The empirical and architectural insights presented herein serve as a roadmap for future quantum-enhanced scientific machine learning, while underscoring the imperative for noise-robust algorithmic innovation and comprehensive hardware validation in advancing quantum PDE solvers (2606.04679).