Quantum Circuit and Tensor Network Implementation of the 2D Acoustic Wave Equation

AI summary

Overview

Research area: Quantum computing and tensor-network methods applied to seismic wave simulation (computational geophysics). Technical level: Advanced — the paper assumes familiarity with finite-difference wave propagation, quantum circuits, and matrix product states/tensor trains. Scope (one sentence): The paper derives a quantum-circuit formulation of the explicit 2D acoustic wave equation, maps it onto an equivalent Matrix Product State solver, and demonstrates the MPS version on a 256x256 subset of the Marmousi velocity model.

What This Paper Is About

Accurate simulation of seismic wave propagation is the computational bottleneck behind imaging and inversion techniques such as Reverse Time Migration (RTM) and Full Waveform Inversion (FWI), and high-resolution surveys push classical high-performance computing to its architectural limits. The authors ask whether the standard explicit finite-difference time-domain (FDTD) update can be re-expressed as a quantum circuit, and whether that same structure can be executed deterministically on classical hardware as a tensor network. Their goal is to show that the tensor-network version can actually generate valid 2D seismic wavefields, not just exist on paper.

Key Contributions

1. A qubit encoding of the 2D acoustic wavefield on a 2^n x 2^n grid using 2n qubits, with the squared velocity field represented as a diagonal operator implemented via QROM architectures or a Walsh series approximation.
2. A circuit construction of the discrete Laplacian using incrementer/decrementer (ripple-carry adder) circuits for the forward and backward shift operators, avoiding quantum Fourier Transform-based translation and its phase errors and routing overhead.
3. A Linear Combination of Unitaries (LCU) block-encoding of the non-unitary transition operator that advances the wavefield using the current and previous time steps, with two working registers and controlled-SWAP operations for the subtraction term.
4. An explicit Matrix Product Operator construction of the same shift operators (bond dimension exactly 2) and the resulting MPS time-stepping scheme, implemented and compared against an FDTD solver on the Marmousi model.

Main Findings

• MPS reproduces FDTD wavefields qualitatively: Figure 2 shows the FDTD wavefield (top) and the MPS wavefield (bottom) after 300 time steps from a center source; the authors report "close similarity" between the two.
• Differences remain: The authors state that some differences are visible on closer inspection and attribute them to the parameterization of the MPS solver.
• All three formulations share one structure: Table 1 compares the FDTD, quantum, and MPS update equations and shows that in every case the next wavefield is a linear combination of the current and previous wavefields at local and neighboring grid points; only the representation of the grid and the field values differs.
• Compression is mandatory: Because adding two MPSs or applying an MPO to an MPS increases the bond dimension, a QR orthogonalization sweep followed by an SVD sweep with truncation to a maximum bond dimension is required after each operator application.
• Low entanglement is the enabling assumption: The paper argues that propagation through smoothly varying velocity models preserves a low-entanglement signature, making the 2^(2n)-dimensional state vector compressible into a low-rank tensor network.
• No quantitative benchmarks reported: The paper does not report numerical error metrics, timing, memory usage, speedup, or a bond-dimension convergence study; the comparison with FDTD is visual.

Methodology in Plain English

The authors start from the textbook second-order finite-difference update for the acoustic wave equation: the next wavefield value at a grid point depends on the two previous time steps at that point and on the four neighboring points at the current step. They then rewrite this rule in the language of qubits, where each grid point is a computational basis state of 2n qubits and the whole wavefield is stored as an unnormalized quantum state vector.

Moving the wavefield one grid cell sideways is done with quantum incrementer and decrementer circuits — essentially ripple-carry adders built from multi-controlled Pauli-X (Toffoli) gates — rather than with Fourier transforms. Adding the four shifted copies and subtracting four times the original gives the discrete Laplacian as a sum of unitary operators. Since the actual time step involves numerical coefficients (the mesh ratio squared times the squared velocity), the resulting operator is non-unitary, so the authors use a standard Linear Combination of Unitaries protocol to implement it with a small ancilla-assisted circuit, keeping two registers to represent consecutive time steps.

Because fault-tolerant quantum hardware for this is not available, they translate the identical logic into tensor networks: the grid is treated as a chain of 2n two-level systems, the wavefield becomes a Matrix Product State with bond dimension χ, and the shift operators become Matrix Product Operators built from 2x2 blocks containing the identity and the raising/lowering operators, with a virtual "carry bit" index flowing from least to most significant bit. After every operator application the state is compressed by discarding singular values below a cutoff and capping the bond dimension. Finally, they run this classical MPS solver and an FDTD solver on the same 256x256 Marmousi subset — spatial sampling Δx = Δz = 10 m, Δt = 1 ms, impulsive source of strength 1 at (1280 m, 1280 m), χ_max = 128, SVD cutoff reported as 10^-7 — and propagate for 300 steps.

Why This Matters

The work offers a bridge between two normally separate research programs: near-term quantum algorithm design for PDEs and practical classical tensor-network simulation. It shows that the same algebraic structure can be run today on classical hardware while quantum hardware matures, and it gives geophysicists a concrete path for compressing wavefield state rather than storing full grids.

Real-world applications, grounded in the paper's stated context:

• Seismic imaging via Reverse Time Migration, which repeatedly propagates wavefields.
• Full Waveform Inversion for building velocity models of the subsurface.
• Reflection seismology in complex media, including high-contrast structures such as salt bodies.
• Any large-scale multidimensional grid simulation where spatial correlations stay low-rank and memory, not compute, is the binding constraint.

Industry relevance: seismic acquisition and processing groups, HPC vendors serving energy exploration, and quantum-hardware/algorithm teams looking for realistic PDE workloads. The paper frames wavefield propagation as a candidate application where exponential memory scaling of the quantum representation would matter — while being explicit that hardware at that scale does not yet exist.

Future Directions

• Quantifying the MPS solver against FDTD: the authors state they are conducting further analysis to evaluate the discrepancies and the solver's potential capabilities, but no error metrics appear in this paper.
• Extending to 3D seismic wavefields, which the authors say they are confident the MPS solvers can generate.
• Systematic study of the parameterization — bond dimension, SVD cutoff, and compression frequency — since the observed differences are attributed to these choices.
• Establishing cost and scaling comparisons between the MPS approach and conventional FDTD on classical hardware, and eventually executing the LCU circuit on fault-tolerant quantum hardware.

Target Audience

Quantum computing researchers interested in PDE simulation and tensor-network emulation; computational geophysicists and seismic processor developers curious about low-rank and tensor-train alternatives to FDTD; and graduate students or advanced practitioners in either field who want a worked, end-to-end derivation connecting circuit diagrams, MPO tensors, and a standard finite-difference stencil. Readers without background in quantum circuits or tensor networks will find the mathematics demanding.

Authors’ abstract

We present a cohesive framework for simulating seismic wave propagation utilizing quantum computing paradigms and their classical tensor network equivalents. We detail a quantum circuit-based formulation for the explicit finite-difference time-domain (FDTD) solution of the two-dimensional acoustic wave equation and map this quantum architecture onto a tensor train representation, namely for Matrix Product State (MPS). The MPS solver enables deterministic simulation of large-scale wavefield dynamics on classical high-performance computing systems. We demonstrate the MPS representation by computing 2D seismic wavefields on the Marmousi model. Our results indicate that the MPS representation is a viable direction for computing and scaling wavefield propagation.