Reproduce “Hybrid Quantum-Classical PINNs for Scientific Computing: A Multi-GPU Open-Source Framework” on OpenResearch

About this paper

Ziv Chen, Hemanth Chandravamsi, Shimon Pisnoy, Aaron Goldgewert, Gal Shaviner, Boris Shragner, Steven H. Frankel

QPINNACLE is an open-source framework from Technion that integrates advanced training strategies, multi-GPU acceleration, and hybrid quantum-classical architectures for Physics-Informed Neural Networks. The framework demonstrated that QPINNs can achieve parameter efficiency and accuracy improvements when augmented with energy regularization and revealed near-linear multi-GPU scaling for classical PINNs.

Key insights

• Combining specific classical PINN enhancements (e.g., Random Fourier Features, loss balancing, strict periodic boundary conditions) is crucial for achieving high accuracy and stable convergence across diverse PDE types, often yielding superior results compared to baseline configurations.
• Hybrid quantum-classical PINNs can achieve parameter efficiency and accuracy gains, but stable training requires specific regularization techniques like an energy conservation penalty, which is detrimental to classical PINNs, highlighting fundamental differences in their optimization landscapes.
• While Distributed Data Parallelism effectively scales PINN training by distributing memory and accelerating runtime on multiple GPUs, its efficiency is subject to communication overhead, leading to diminishing returns beyond a certain number of devices.

Summary by alphaXiv.