DiffVortex: A quantum-inspired fast solver for incompressible flows

Yu Jun Loo; Silas Alben; Shravan Veerapaneni

Home/Archive/arXiv

arXiv · 2610.30012

DiffVortex: A quantum-inspired fast solver for incompressible flows

Yu Jun Loo·Silas Alben·Shravan Veerapaneni

View original source ↗Read source document ↗Export metadata ↓

Abstract

Matrix-product-state (MPS) and quantized tensor-train (QTT) representations offer a promising framework for computational fluid dynamics, with logarithmic run time and memory scaling in mesh size when bond dimensions remain bounded. Existing MPS-based solvers are Eulerian: they compress velocity fields and enforce immersed-body boundary conditions through static masks or body-fitted coordinates. Moving or flexible bodies complicate this approach: deformation requires updating MPS masks every time step, while filamentary vorticity in high-Reynolds-number flows may resist low-rank compression. We propose a complementary approach: compress the Biot--Savart kernel rather than the flow state, evolve vorticity using Lagrangian vortex particles, and represent immersed bodies by bound vortex sheets on Lagrangian surface points. Our main theoretical result is a lifting theorem: QTT rank $r$ of the kernel implies matrix-product-operator (MPO) rank at most $2^d r$ for the corresponding convolution operator in $d$ dimensions, independent of grid resolution. This bound is sharp for our interleaved ordering. With quantics tensor cross interpolation, the moment-to-local stage of a single-level fast multipole method becomes a precomputed tensor-network contraction with quasi-linear complexity, retaining Lagrangian flexibility for immersed bodies. Our PyTorch implementation (\texttt{DiffVortex}) is end-to-end differentiable, runs on GPUs, and handles $6.7\times10^{7}$ particles on a single device. We validate against reference vortex-method results for an impulsively started cylinder and exact and reference solutions for inviscid elliptical vortices. We also demonstrate two-dimensional flows past rigidly moving, multiple, and deforming bodies, including gradients propagated through the full time integration.

Explore related subjects

physics.comp-ph physics.flu-dyn

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Yu Jun Loo, Silas Alben, Shravan Veerapaneni. 2026-10-21. DiffVortex: A quantum-inspired fast solver for incompressible flows. https://arxiv.org/abs/2610.30012

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Matrix-free phase-field modeling of fracture in micromechanical testing simulations of inelastic materials

Resolving steep damage gradients across diffuse cracks in the phase-field modeling of fracture favors the use of high-order finite elements, for which matrix-free methods can provide superior performance and scalability. Here, we implement the Peric & Dettmer constitutive framework for visco-elastoplastic materials in an open source solid mechanics library supporting matrix-free operators for high-order finite elements with p-multigrid preconditioning on GPUs. We introduce a rheological fracture element assembled in series with the Peric & Dettmer block, keeping fracture and inelastic material properties mutually independent. We conduct numerical simulations of tensile and compressive quasistatic tests on synthetic particle-matrix microstructures using an El Capitan high-performance computing prototype. The results reproduce representative inelastic responses and fracture patterns, including X-shaped cracking observed in polymer-bonded explosives and cup-and-cone fracture in ductile metals.

physics.comp-ph↗

arXiv2026

Tractable Inference with Kolmogorov-Arnold Networks

Machine-learning surrogates in computational science often require not only accurate predictions but also expectations, variances, conserved integrals, and normalization constants. Conventional neural networks, including Kolmogorov--Arnold networks (KANs), generally evaluate these quantities through Monte Carlo sampling or multidimensional quadrature, making repeated inference costly, particularly in high dimensions. We introduce KANs with Tractable Operators for Uncertainty and Constraints (TOUCKANs), whose finite-degree polynomial outer functions yield a decomposable sum--product representation. Under factorizing input measures, expectations and higher moments then reduce algebraically to deterministic recursions over one-dimensional leaf integrals. Independent and shared inner-block variants form a parameter-efficiency hierarchy, and a universal-approximation proof for its most restricted member extends to the complete class. The same deterministic operators allow conserved integrals and normalization to be imposed throughout training. Across uniform and Gaussian benchmarks through 50 dimensions, deterministic mean and variance errors remained below the median fixed-budget Monte Carlo errors in every case. TOUCKANs also conserved Cahn--Hilliard mass to machine precision and imposed normalization and orthogonality through deterministic moments, enabling joint computation of the lowest four Schrödinger eigenpairs in two, three, and five dimensions. TOUCKANs therefore connect KAN approximation with tractable probabilistic circuits, making statistical operators and integral constraints directly available within an expressive neural surrogate.

physics.comp-ph↗

arXiv2026

Physics-informed sampler for solver-free data generation in operator learning

Neural operators offer fast surrogates for parametric PDEs but typically rely on costly datasets generated by repeated numerical solves. We propose a physics-informed sampler (PI-sampler) that generates PDE-consistent input--solution pairs without numerical PDE solves. The sampler draws candidate solution fields from neural-network priors and uses automatic differentiation to evaluate the governing differential, boundary, and initial-condition operators, yielding consistent training tuples by construction. Unlike solution-first methods based on analytical bases or precomputed solutions, it requires neither problem-specific basis functions nor seed simulations. The resulting PDE inputs follow an implicit pushforward distribution, which can be characterized for linear operators in the Gaussian-process limit of Bayesian neural networks but is generally non-Gaussian for nonlinear operators. We integrate the PI-sampler with a function encoder and Transformer-based neural operator and test it on reaction--diffusion equations, Poisson equations on parameterized geometries, time-dependent Sine--Gordon dynamics, a five-dimensional PDE, and flexible-cylinder vortex-induced vibration. The models generalize to independently prescribed forcing distributions, zero forcing, mixed boundary conditions, and unseen geometries. Experiments with DeepONet confirm that the generated data also support other operator architectures. In problem-specific benchmarks, sampling is approximately 51 times faster on CPU and 850 times faster on GPU than numerical solvers. The PI-sampler thus provides an efficient approach to data generation and augmentation that complements high-fidelity simulation.

physics.comp-ph↗