Multi-stream physics hybrid networks for solving Navier-Stokes equations

Understanding and solving fluid dynamics equations efficiently remains a fundamental challenge in computational physics. Traditional numerical solvers and physics-informed neural networks struggle to capture the full range of frequency components in partial differential equation solutions, limiting their accuracy and efficiency. Here, we propose the Multi-stream Physics Hybrid Network, a novel neural architecture that integrates quantum and classical layers in parallel to improve the accuracy of solving fluid dynamics equations, namely ‘‘Kovasznay flow’’ problem. This approach decomposes the solution into separate frequency components, each predicted by independent Parallel Hybrid Networks, simplifying the training process and enhancing performance. We evaluated the proposed model against a comparable classical neural network, the Multi-stream Physics Classical Network, in both data-driven and physics-driven scenarios. Our results show that the Multi-stream Physics Hybrid Network achieves a reduction in root mean square error by 36% for velocity components and 41% for pressure prediction compared to the classical model, while using 24% fewer trainable parameters. These findings highlight the potential of hybrid quantum-classical architectures for advancing computational fluid dynamics.

💡 Research Summary

The paper introduces a novel neural architecture called the Multi‑stream Physics Hybrid Network (MPHN) that combines quantum and classical layers in parallel to solve the Navier‑Stokes equations more accurately than conventional approaches. The authors motivate the design by pointing out that traditional numerical solvers and pure physics‑informed neural networks (PINNs) often struggle to capture high‑frequency components of PDE solutions, leading to reduced accuracy and higher computational cost. MPHN addresses this by decomposing the solution into separate frequency components and assigning each component to an independent Parallel Hybrid Network (PHN).

Each PHN consists of two sub‑networks: a small, parameterized two‑qubit quantum circuit and a shallow classical multilayer perceptron (MLP) with ten hidden neurons. The quantum circuit is intended to model the periodic (high‑frequency) part of the solution, while the classical MLP captures the slowly varying or linear part. Their outputs are merged through an affine combination (w0 + w1·Qout + w2·Cout + w3·Qout·Cout), allowing the model to exploit the expressive power of quantum states together with the stability of classical training.

The authors evaluate MPHN on the classic Kovasznay flow problem, a 2‑D laminar flow behind a grid that possesses an exact analytical solution for velocity components (vx, vy) and pressure (p). This choice enables a precise error analysis. Two training regimes are considered: (1) a data‑driven setting where the exact solution is sampled on a uniform 30 × 40 grid (1,200 points) and the mean‑squared error (MSE) between predictions and ground truth is minimized; (2) a physics‑driven setting where the loss comprises the PDE residual (Navier‑Stokes) and boundary‑condition terms, following the standard PINN formulation.

Both MPHN and a baseline Multi‑stream Classical Network (MCN) are trained with the Adam optimizer (learning rate = 10⁻²) for 100 epochs, using identical hyper‑parameters for a fair comparison. MPHN contains 936 trainable parameters (quantum + classical), whereas MCN, built solely from classical layers, uses 1,239 parameters to match or exceed MPHN’s capacity.

Results show that MPHN achieves a substantial reduction in root‑mean‑square error (RMSE): approximately 36 % lower for the velocity components and 41 % lower for pressure, despite having 24 % fewer parameters. This demonstrates that the quantum sub‑network can capture high‑frequency features efficiently, reducing the need for a large number of classical weights. The authors also discuss practical considerations: the quantum circuit is deliberately shallow to avoid barren‑plateau gradients and to be compatible with current noisy intermediate‑scale quantum (NISQ) hardware, which cannot yet support deep, high‑precision circuits.

The paper’s contributions are threefold: (i) integrating a variational quantum circuit into a PINN framework, showing that quantum expressivity can improve PDE solving; (ii) proposing a multi‑stream decomposition that simplifies training by isolating different physical scales; (iii) providing empirical evidence on a benchmark problem with an exact solution, establishing both accuracy gains and parameter efficiency.

Limitations are acknowledged. The current quantum circuit is minimal (two qubits, shallow depth), which may restrict its ability to model more complex, highly nonlinear interactions. Training is performed on a classical simulator, so the impact of quantum noise and real‑hardware constraints remains untested. Future work is suggested to explore deeper variational circuits, error‑mitigation techniques, and deployment on actual quantum processors, potentially extending the approach to higher‑dimensional or turbulent flow problems where exact solutions are unavailable.

In summary, the Multi‑stream Physics Hybrid Network demonstrates that a carefully designed quantum‑classical hybrid architecture can enhance the performance of physics‑informed neural solvers for fluid dynamics, offering a promising direction for integrating quantum machine learning into scientific computing.