Quantum-enhanced convergence of data-driven physics-informed neural networks for Earth system modeling
Nils
Klement
German Aerospace Center, Institute of Atmospheric Physics
Talk
Partial differential equations (PDEs) describe key processes in Earth system models, including atmospheric dynamics. While physics-informed neural networks (PINNs) provide a flexible and data-efficient framework for solving PDEs, their slow convergence and challenging optimization limit their practical use in climate-scale applications.
In this work, we investigate whether hybrid quantum–classical neural networks can accelerate PINN-based PDE solvers. We develop architectures that integrate parametrized quantum circuits with classical neural networks, train them using statevector simulations and systematically evaluate their performance across nonlinear PDEs and boundary conditions representative of complex physical systems. Consistent with prior findings, both classical and hybrid models achieve comparable solution accuracy given sufficient training. However, we show that quantum-enhanced models converge to accurate solutions in substantially fewer training epochs, particularly for more complex problem settings.
These results suggest that quantum-enhanced PINNs may help alleviate a key computational bottleneck in Earth system modeling. By reducing training costs while maintaining accuracy, hybrid quantum approaches offer a promising pathway toward accelerating the development of climate models, especially in regimes characterized by high complexity and dimensionality.
In this work, we investigate whether hybrid quantum–classical neural networks can accelerate PINN-based PDE solvers. We develop architectures that integrate parametrized quantum circuits with classical neural networks, train them using statevector simulations and systematically evaluate their performance across nonlinear PDEs and boundary conditions representative of complex physical systems. Consistent with prior findings, both classical and hybrid models achieve comparable solution accuracy given sufficient training. However, we show that quantum-enhanced models converge to accurate solutions in substantially fewer training epochs, particularly for more complex problem settings.
These results suggest that quantum-enhanced PINNs may help alleviate a key computational bottleneck in Earth system modeling. By reducing training costs while maintaining accuracy, hybrid quantum approaches offer a promising pathway toward accelerating the development of climate models, especially in regimes characterized by high complexity and dimensionality.