Accelerating physics-informed neural networks for full waveform inversion using a hybrid quantum-classical finite-basis architecture

This paper introduces a hybrid quantum-classical finite-basis physics-informed neural network (FBPINN) that utilizes parameterized quantum circuits to significantly accelerate full waveform inversion, achieving lower velocity errors with fewer trainable parameters and training iterations compared to classical baselines.

The Gist

Imagine you are trying to figure out what's inside a giant, opaque block of rock just by listening to the echoes of a sound you tap on one side. This is the core challenge of Full Waveform Inversion (FWI). It's like trying to guess the shape of a hidden object inside a box by only hearing the sound waves bounce around inside it.

Usually, solving this puzzle is incredibly slow and requires massive supercomputers. It's like trying to solve a giant jigsaw puzzle by moving one tiny piece at a time, checking if it fits, and then moving it back if it doesn't.

The New Approach: A "Quantum-Hybrid" Team

The authors of this paper propose a new way to solve this puzzle using a team-up between classical computers (the ones we use every day) and quantum computers (a futuristic technology that uses the weird rules of quantum physics).

Think of their solution as a relay race:

1. The Classical Runner: First, a standard computer network takes the raw data (the coordinates of the rock) and simplifies it into a "secret code" (a low-dimensional latent space).
2. The Quantum Runner: This secret code is then handed to a "quantum circuit." Imagine this circuit as a special, highly efficient machine that can mix and twist the information in ways a normal computer struggles to do quickly. It processes the data and spits out a result.
3. The Finish Line: The result is passed back to the classical computer, which translates it into a final map of the rock's speed.

Why is this special?

The researchers tested this "Quantum-Hybrid" team against a team of only "Classical Runners" (standard AI) on two specific test cases:

1. The "Hidden Anomaly" Test:
They tried to find a specific, slow-moving patch of rock hidden in a faster background.

• The Result: The Quantum-Hybrid team found the hidden patch 8 times faster (in terms of training steps) than the best classical team.
• The Efficiency: Even though the Quantum-Hybrid team had fewer players (about 33% fewer adjustable settings or "parameters"), they still did a better job. It's like a small, elite special forces unit solving a problem that usually requires a whole army of regular soldiers.

2. The "Checkerboard" Test:
They tried to reconstruct a complex pattern of alternating fast and slow rock speeds (like a checkerboard).

• The Result: The Quantum-Hybrid team successfully mapped this complex pattern without needing any extra tuning, proving their method works for different types of shapes, not just the first one.

How did they do it? (The Secret Sauce)

The paper suggests three reasons why the quantum part helped:

• Efficient Mixing: The quantum circuit mixes information in a way that uses fewer "knobs" to turn but creates more complex patterns.
• Built-in Rhythm: The way the quantum machine reads the data naturally creates a "rhythmic" or wave-like structure. This helps it understand the wiggly, fast-moving waves of sound much better than standard AI, which often gets stuck trying to learn simple, slow patterns first.
• Smart Boundaries: The system is built with hard rules that prevent it from guessing impossible speeds, keeping the solution realistic.

Important Reality Checks

The authors are very careful to state what this is not:

• It's not magic yet: They didn't use a real, physical quantum computer. They used a simulator (a program that pretends to be a quantum computer) running on a normal computer.
• It's not a "Quantum Advantage" yet: Because they used a simulator, they aren't claiming that quantum computers are currently faster than supercomputers in real life. They are showing that the mathematical structure of the quantum approach is very efficient.
• It's still a work in progress: The test was done on a simple 2D map with just one sound source. Real-world oil exploration or medical imaging is much more complex (3D, many sound sources).

The Bottom Line

This paper shows that by borrowing a specific mathematical trick from quantum computing and plugging it into a standard AI, we can solve complex wave-puzzles much faster and with fewer resources. While it's currently a simulation, it suggests that when real quantum computers are ready, they could be the secret weapon for making these complex imaging tasks much more efficient.

Technical Summary

Technical Summary: Accelerating Physics-Informed Neural Networks for Full Waveform Inversion Using a Hybrid Quantum-Classical Finite-Basis Architecture

Problem Statement
Full Waveform Inversion (FWI) is a critical technique for reconstructing heterogeneous material properties from waveform data, with applications ranging from subsurface velocity imaging to medical ultrasound tomography. However, FWI remains computationally demanding, requiring iterative solutions to the forward wave equation and gradient computations via adjoint-state methods. While Physics-Informed Neural Networks (PINNs) offer a mesh-free alternative by minimizing partial differential equation (PDE) residuals, standard PINNs struggle with high-frequency wave propagation and multi-scale physics due to "spectral bias"—a tendency to learn low-frequency features before high-frequency ones. Although Finite-Basis PINNs (FBPINNs) have addressed this via domain decomposition, their application to waveform inversion remains unexplored. Furthermore, while quantum computing has shown promise in scientific machine learning, its integration into the specific, highly nonlinear inverse problem of FWI—where wavefields and velocity models are coupled—has not been thoroughly investigated.

Methodology
The authors propose a hybrid quantum-classical FBPINN framework for acoustic FWI. The architecture replaces the standard neural network heads of both the decomposed wavefield network ($\phi$) and the global velocity network ($\alpha$) with classical-to-quantum pipelines.

• Architecture: The input coordinates are processed by classical pre-networks (fully-connected layers with tanh activations) to map into low-dimensional latent spaces. These latent vectors are then encoded into parameterized quantum circuits (PQCs) via angle encoding ($R_y(\beta_i z_i)$). The PQCs consist of variational layers containing single-qubit rotations ($R_x, R_y, R_z$) and entangling CNOT gates, followed by measurement in the Pauli-Z basis. The bounded quantum outputs are rescaled by classical output layers to physical units.
• Simulation Environment: To isolate architectural benefits from hardware noise, the PQCs are implemented as differentiable JAX statevector simulators. This enables exact, end-to-end automatic differentiation through the classical PINN, the quantum circuit, and the physics-informed loss, avoiding the computational overhead of parameter-shift rules required on physical hardware.
• Training: The framework utilizes a composite loss function incorporating PDE residuals (source-free acoustic wave equation), initial conditions (wavefield snapshots), data misfit (receiver observations), and free-surface boundary constraints. The model is trained using the Adam optimizer.

Key Contributions

1. First Application of Hybrid Quantum-Classical FBPINNs to FWI: The work extends the domain-decomposition advantages of FBPINNs to waveform inversion, coupling a decomposed wavefield network with a global velocity network via quantum circuits.
2. End-to-End Differentiable Pipeline: By implementing PQCs as JAX-native statevector simulators, the authors enable efficient gradient computation without the parameter-shift rule, facilitating the optimization of the coupled forward-inverse problem.
3. Systematic Benchmarking: The study evaluates the hybrid architecture against a purely classical FBPINN baseline and 15 distinct classical hyperparameter variants (including over-parameterized networks and alternative subdomain partitions) to ensure the results are not artifacts of sub-optimal classical tuning.

Results
The authors evaluated the framework on two geophysical benchmarks:

• Anomaly Benchmark: In a 2D acoustic domain with a localized low-velocity anomaly, the hybrid model achieved a lower $L_1$ velocity error (2.4 $\times$ 10$^{-2}$) in approximately 65,000 training iterations. In contrast, the classical baseline required 500,000 iterations to reach a slightly higher error (2.9 $\times$ 10$^{-2}$). This represents an approximately 8-fold reduction in optimization iterations to reach comparable accuracy, despite the hybrid model using 33% fewer trainable parameters (101,812 vs. 151,836). None of the 15 classical hyperparameter variants matched this performance within the tested window.
• Checkerboard Benchmark: Using a higher-frequency source (60 Hz) to resolve a 3 $\times$ 3 checkerboard velocity pattern, the hybrid model successfully recovered the structured spatial variations using the exact same network configuration as the anomaly benchmark (with only increased collocation point density). This confirms the generality of the inversion pipeline beyond localized anomalies.

Significance and Claims
The paper claims that the hybrid quantum-classical architecture offers a significant acceleration in convergence speed and parameter efficiency for acoustic FWI compared to classical FBPINN baselines. The authors attribute this performance to three properties of the classical-to-PQC pipeline:

1. Parameter Efficiency: The PQC acts as a nonlinear map where parameters are shared across feature mixing and output generation.
2. Finite-Frequency Fourier Structure: Angle encoding endows the output with a trainable finite-frequency Fourier structure, potentially mitigating spectral bias more effectively than fixed Fourier-feature layers or standard tanh networks.
3. Structured Correlations: Entangling CNOT layers generate correlations among latent variables through circuit dynamics rather than stacked affine-plus-nonlinearity layers.

Limitations and Modesty
The authors explicitly state that these results do not constitute a "quantum computational advantage" in the strict sense, as the circuits are classically simulated and the function class is classically representable. The observed gains are in optimization iteration counts, not wall-clock time. The study is limited to:

• Two-dimensional acoustic wave propagation (not 3D elastic).
• A single-source, one-sided receiver configuration (a severely ill-posed problem).
• Classical statevector simulation (avoiding NISQ noise and barren plateaus).
• A comparison against tanh-based classical architectures; the authors note that a matched classical control with explicit Fourier-feature or sinusoidal embeddings is a necessary future step to isolate whether the advantage stems from the quantum parameterization or the finite-frequency inductive bias.

The framework is presented as broadly applicable to wave-based inverse problems beyond geophysics, including medical ultrasound tomography and non-destructive evaluation, provided the challenges of scaling to 3D and real hardware are addressed.