Title: Boosting the Efficiency of Physics-Informed Neural Networks in Full Waveform Inversion via a Hybrid Quantum-Classical Finite-Basis Framework

Abstract:

Full waveform inversion (FWI) is a technique used to reconstruct heterogeneous material properties from receiver data, yet it is notoriously computationally intensive. While physics-informed neural networks (PINNs) and their domain-decomposed counterparts (FBPINNs) provide a mesh-free solution, they often struggle with convergence when attempting to model complex velocity fields. To address these limitations, we introduce a hybrid quantum-classical FBPINN tailored for acoustic FWI. This approach integrates quantum computing with classical machine learning, structuring both the decomposed wavefield network and the global velocity network as classical-to-quantum pipelines that conclude with parameterized quantum circuits (PQCs). These PQCs are executed as differentiable JAX statevector simulators, which facilitate end-to-end automatic differentiation across the classical PINN, the quantum circuit, and the physics-informed loss function.

In testing on a geophysical anomaly benchmark, our quantum hybrid model achieved a lower L1 velocity error than the primary classical FBPINN baseline. Remarkably, this superior performance was attained in roughly 8 times fewer training iterations and with approximately 33% fewer trainable parameters. Furthermore, the hybrid model surpassed all 15 classical hyperparameter variants evaluated in the study. A secondary benchmark involving a checkerboard pattern validated the robustness and generality of the inversion pipeline, demonstrating that the quantum hybrid architecture can successfully recover structured spatial variations beyond the scope of the localized anomaly tests. This framework holds broad applicability for wave-based inverse problems across various fields, extending beyond geophysics to include medical ultrasound tomography and non-destructive evaluation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC