Shivanshu Siyanwal

Conference paper

TensNet: Memory-Efficient Tensor Network Methods for 2D Incompressible Flow Simulation

S. Siyanwal, N. Khan, S. S. Pandey, K. Y. Islur, K. Dave, S. Chhipa

ICSSCC 2026 · Accepted, in processing

Overview

TensNet is a library implementing tensor network methods, using Matrix Product States (MPS) and Matrix Product Operators (MPO), for simulating the 2D incompressible Navier–Stokes equations. It is designed to reduce memory use compared with standard grid-based solvers, keeping accuracy through adaptive bond-dimension truncation; it uses the DMRG algorithm to solve the implicit pressure-correction equations and cross-approximation to handle complex immersed boundaries.

Problem / motivation

Computational fluid dynamics simulations demand enormous memory for storing high-resolution velocity fields, which is a barrier to practical simulations on resource-constrained systems.

Methods

TensNet represents flow fields as Matrix Product States and operators as Matrix Product Operators, compressing them by SVD truncation with adaptive bond dimensions. The two-site DMRG algorithm solves the implicit Poisson equations in the pressure-correction scheme, and cross-approximation represents immersed boundaries and obstacle geometries as MPS. Benchmarks: a cylinder in cross-flow, a square obstacle and a 3×2 array of cylinders, compared against a dense finite-difference solution on the same grid.

Results / outcomes

Reported in the accepted manuscript (ICSSCC 2026, in processing).

Setup: Intel i7 workstation with 16 GB RAM and no GPU; 256 × 64 grid, Δt = 10⁻³, ν = 0.01; accuracy measured against a dense finite-difference solution on the same grid.

At D_max = 30 the L2 error is below 10⁻² and memory falls from about 8 GB to 5.8 MB, a reduction of roughly 1.4 × 10³. The L2 velocity error falls roughly exponentially with D_max at Re = 15, 30, 100 and 500; higher Re needs a moderately larger D_max.

Scalability: on a 1024 × 256 grid with D_max = 50, TensNet uses about 35 MB; the manuscript estimates about 2 TB for a dense finite-difference solver.

Memory vs accuracy, cylinder in cross-flow (t = 1.0, 100 time steps)
D_maxMemoryL2 errorTime
102.3 MB0.08745 s
204.1 MB0.03268 s
305.8 MB0.00892 s
508.2 MB0.002145 s
Dense reference8,000 MB—3,600 s
Cylinder flow statistics, TensNet at D_max = 30 vs literature
ReCd (TensNet)Cd (reference)St (TensNet)St (reference)
152.412.35–2.40––
301.681.62–1.680.1980.190–0.200
1001.321.25–1.350.1660.160–0.170
5001.141.06–1.200.2110.205–0.220
Computational cost, 256 × 64 grid, 100 steps
SolverMemoryRuntime
TensNet (D_max = 30)5.8 MB92 s
Dense finite-difference8,000 MB3,600 s
GPU-optimised finite-difference6,000 MB450 s

My role

TensNet was originally developed by N. Khan and S. S. Pandey. I carried out the testing, code troubleshooting, benchmarking and visualisation, and prepared the manuscript.

ProblemApproachResultsStatus
Schematic: structure of the work (from the write-up's sections)

Project write-up with figures →