Researchers Propose Algorithm to Rebuild Fluid Flows from Limited Data

Reconstructing fields governed by nonlinear partial differential equations (PDEs) from sparse measurements presents key challenges owing to equation nonlinearity and restricted observation locations. Fluid velocity fields exemplify these situations. A new variational quantum algorithm is proposed for simultaneously reconstructing solutions throughout the complete spacetime domain. This differs from conventional techniques using temporal marching as it encodes the whole discrete spacetime solution into one variational quantum state, allowing combined optimisation of every time point. The related cost function incorporates both a discrepancy between sparse measurements and a physics-informed PDE violation term.

Variational Quantum Algorithm Enables High Accuracy Fluid Velocity Reconstruction

Root-mean-square errors fell to approximately 10− consistently when reconstructing fluid velocities. Previously, achieving globally consistent solutions from such sparse data was unattainable due to computational limitations with traditional time-marching methods. The new variational quantum algorithm reconstructs these velocity fields by representing all space and time as a single optimised quantum state via combined physics and measurement discrepancies.

Avoiding explicit temporal progression, a resource intensive process prone to error, the approach circumvented significant hurdles in modelling complex active behaviour governed by nonlinear partial differential equations like Burgers and Kuramoto, Sivashinsky equations. This method offers sharp efficiency gains when dealing with large datasets because it reduces circuit preparations needed for measurement residuals from a number proportional to the total time steps down to a constant value regardless of simulation length.

Two complementary spatial representations exist: real-space directly corresponds amplitudes to field values at grid points evaluating derivatives through finite difference operators sharing the spacetime encoding ansatz but employing different reconstruction suited functions; while Galerkin-reduced basis expands fields within truncated domain adapted bases encoding coefficients into states capturing dominant structures using fewer modes permitting analytical derivative evaluation avoiding errors.

Hybrid data-equation methods reconstruct partial differential equation solutions

Data and governing equations are simultaneously utilised to constrain solutions, as demonstrated by numerical simulations applied to the one-dimensional Burgers and Kuramoto, Sivashinsky equations. Variational quantum algorithms with a spacetime encoding scheme provide a compact framework for reconstructing nonlinear PDE dynamics according to these results. Reconstructing high-dimensional signals or fields from limited measurements is a pervasive challenge within scientific computing, image processing, and experimental sensing.

Often, accessible observations are fewer than the degrees of freedom needed to represent the quantity of interest; this renders inverse problems underdetermined unless structural assumptions are imposed. Compressed sensing addresses this issue showing recovery can remain accurate when unknowns have low complexity, typically sparsity or compressibility in an appropriate representation, alongside measurement processes preserving relevant information. For fields governed by partial differential equations (PDEs), physics-based structure frequently complements or replaces sparsity-based priors because governing equations impose strong constraints on reconstructions.

A common strategy involves posing sparse-sensor reconstruction as a physics-regularized inverse problem balancing agreement with measurements against consistency with underlying dynamics. Physics-informed neural networks (PINNs) exemplify this approach, representing unknown fields using deep neural networks and optimising a loss coupling sensor mismatch with residual penalties for the PDE; enabling reconstruction from limited data across various forward and inverse fluid mechanics problems. PIKANs employ similar objectives but replace fixed activation functions of multilayer perceptrons with learnable univariate functions along network edges offering an alternative representation for complex solutions.

Neural operator architectures like DeepONet learn mappings to input functions or parameters to complete solution fields. Its extension, physics-informed DeepONets, incorporates governing equation residuals during training allowing reconstructions across families of PDE instances rather than optimising separate networks per realisation. As demand grows for spatiotemporal resolution, for example in near-turbulent flows, resolving multiscale features often requires prohibitively large networks imposing substantial computational burden.

Quantum computing offers a route by representing and simulating solutions within exponentially large Hilbert spaces. Several quantum algorithms have been proposed for solving PDEs potentially outperforming classical counterparts. To accommodate current and near-term device constraints several variational quantum algorithms (VQAs) encode the discretised solution into a parameterised quantum state; these approaches assume sparse representation exists within accessible Hilbert space obtained via varying circuit parameters to minimise physics-motivated objectives.

Standard VQA methods rely on sequential time-marching preparing or updating states at successive steps. Recently, spacetime-encoded VQAs avoid explicit time marching which is resource intensive and error-prone. Existing algorithms find solutions from known initial conditions where velocity fields are available across the entire discrete spatial lattice at some starting point. However, sparse sensor reconstruction presents a distinct challenge as boundary conditions take form of field values only at certain locations throughout the window precluding explicit time marching requiring globally consistent spacetime solutions satisfying measurements alongside nonlinear dynamics from undersampled data.

The proposed algorithm with spacetime encoding reconstructs velocity fields using sparse time resolved measurements representing all space and time degrees of freedom in one variational quantum state optimised via cost function combining governing equation residual with sensor mismatch enabling incorporation of fixed location sensors over full observation without constructing separate circuits per timestep. Fluid velocity fields serve as a representative example.

A variational quantum algorithm is proposed that reconstructs solutions over the entire spacetime domain simultaneously rather than step-by-step in time; it encodes the full discrete spacetime solution within a single quantum state optimising all-time points together. The cost function combines discrepancies between sparse measurements with physics-informed PDE violations allowing both data and governing equation to constrain the solution concurrently, demonstrated through numerical simulations using one-dimensional Burgers and Kuramoto, Sivashinsky equations. Two complementary spatial representations of velocity fields exist.

Quantum-state amplitudes directly correspond to field values at discrete grid points evaluating spatial derivatives via finite-difference operators in the real-space approach. This shares the spacetime encoding and ansatz from previous work but uses a different cost function suited to reconstruction without an initial condition. Expanding the field in a truncated basis adapted to domain conditions encodes time-dependent expansion coefficients into the quantum state within the Galerkin-reduced-basis approach; it can capture dominant structures using few modes achieving spectral convergence for smooth solutions while analytically evaluating derivatives within that space providing stability.

For fixed operator and discretization order, circuit count needed for cost evaluation is independent of qubit number excluding state preparation by the variational ansatz and measurement data encoding. The real-space method’s circuit count also depends on finite difference stencil order used approximating spatial derivatives whereas Galerkin formulation evaluates these directly avoiding errors and overhead.

Validation through numerical reconstructions of one-dimensional Burgers and Kuramoto, Sivashinsky equations employs a problem agnostic hardware efficient ansatz; improvements may be possible incorporating dynamics into architecture adaptively. Reconstructed velocity fields achieve root-mean-square errors around 10−2 using measurements from no more than 25% of grid points demonstrating recovery globally consistent spacetime fields.

Quantum algorithms reconstruct full histories of fluid motion from minimal sensing

Reconstructing fluid behaviour from limited data remains a core challenge across disciplines like weather forecasting and ocean modelling. Accurately predicting these complex systems demands insight into their underlying dynamics despite incomplete observations. This approach simultaneously uses both observed measurements and the underlying physics governing these systems via an optimised cost function, enabling reconstruction even when observations are limited. By jointly optimising all points in space and time, this method circumvents computational bottlenecks associated with traditional numerical simulations of nonlinear partial differential equations like Burgers and Kuramoto-Sivashinsky equations. It’s important to acknowledge that quantum computing remains a developing field with questions around scalability and practical implementation; however, this work offers a valuable proof of principle nonetheless.

The research demonstrated a variational quantum algorithm successfully reconstructed complete spacetime solutions for fluid dynamics using sparse measurement data. This means the full history of flow was rebuilt despite having information from fewer than 25% of grid points, achieving root-mean-square errors around 10−2.

By combining observational data with the known physics governing these systems, specifically the one-dimensional Burgers and Kuramoto, Sivashinsky equations, the method provides globally consistent reconstructions without relying on time-step calculations typical in numerical simulations. The authors suggest future improvements may be possible by adapting the computational architecture to incorporate system dynamics more effectively.

More like this

Quantum AlgorithmsResearchers Achieve Np-Hardness for Approximating Ising Model Properties

Quantum AlgorithmsResearchers Simulate Molecular Dynamics Including Nuclear Effects

Quantum AlgorithmsResearchers Limit Quantum Noise Loss to Three Percentage Points

Quantum AlgorithmsResearchers Unlock Accurate Spectra Via Optimised Density Functionals