SCP-FTLE and SCP-FSLE: High-performance finite-time and finite-size Lyapunov-exponent computations for CPUs and GPUs
Abstract
Finite-time and finite-size Lyapunov exponents (FTLE and FSLE) characterise material stretching and transport barriers over prescribed time intervals and separation scales, but their evaluation from three-dimensional velocity sequences requires large computational costs. We introduce SCP-FTLE and SCP-FSLE, two modules of StreamCenter+ for CPU- and GPU-accelerated Lagrangian analysis. The modules support structured and native unstructured meshes, selectable integration and interpolation schemes, and graphical and command-line workflows. Rolling snapshot reuse, memory-bounded FTLE batching, device-resident trajectory state and active-set FSLE advancement reduce data movement and redundant computation, achieving the state-of-the-art (SOTA) computational performance among peer packages. For a billion-seed benchmark, strong-scaling efficiencies on 32 GPUs reach 86.3% for FTLE and 95.2% for FSLE. FTLE is verified with published steady and unsteady Arnold-Beltrami-Childress (ABC) flows and stationary concentrated vortex model (SCVM), while FSLE is verified in linear stretching and rotational shear flows. As examples, applications reveal the spanwise reorganisation of stretching layers in a shear-thinning lid-driven cavity, and resolve three-dimensional stretching structures associated with contrasting tracer dispersion around an urban building canopy at NUS campus. These results demonstrate a scalable workflow in large-scale Lagrangian diagnostics and analysis of flow data.
Related
- Improved Finite-Time Lyapunov Exponent framework for Comprehensive Synchronization Analysis in Multilayer Networks
- Butterfly Effects in Turbulence: A Unified View through the Finite Size Lyapunov Exponent
- Black holes with quantum corrections in 3d: The case of Page curve in Lindblad, greybody factor, and Lyapunov exponent
- Matter Induced Bounds: Photon Ring, Lyapunov Exponent, Induced Temperature
- Full Lyapunov Spectra of Random Neural Networks, Part I: Finite-Time and Asymptotic Theory
BibTeX
@misc{wei2026scpftle,
title = {{SCP-FTLE and SCP-FSLE: High-performance finite-time and finite-size Lyapunov-exponent computations for CPUs and GPUs}},
author = {Huanxia Wei and Mengqi Zhang},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.22609},
eprint = {2610.22609},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.22609},
}
RIS
TY - UNPB
TI - SCP-FTLE and SCP-FSLE: High-performance finite-time and finite-size Lyapunov-exponent computations for CPUs and GPUs
AU - Huanxia Wei
AU - Mengqi Zhang
PY - 2026
DA - 2026/10/15
PB - arXiv
DO - 10.48550/arxiv.2610.22609
UR - https://arxiv.org/abs/2610.22609
AB - Finite-time and finite-size Lyapunov exponents (FTLE and FSLE) characterise material stretching and transport barriers over prescribed time intervals and separation scales, but their evaluation from three-dimensional velocity sequences requires large computational costs. We introduce SCP-FTLE and SCP-FSLE, two modules of StreamCenter+ for CPU- and GPU-accelerated Lagrangian analysis. The modules support structured and native unstructured meshes, selectable integration and interpolation schemes, and graphical and command-line workflows. Rolling snapshot reuse, memory-bounded FTLE batching, device-resident trajectory state and active-set FSLE advancement reduce data movement and redundant computation, achieving the state-of-the-art (SOTA) computational performance among peer packages. For a billion-seed benchmark, strong-scaling efficiencies on 32 GPUs reach 86.3% for FTLE and 95.2% for FSLE. FTLE is verified with published steady and unsteady Arnold-Beltrami-Childress (ABC) flows and stationary concentrated vortex model (SCVM), while FSLE is verified in linear stretching and rotational shear flows. As examples, applications reveal the spanwise reorganisation of stretching layers in a shear-thinning lid-driven cavity, and resolve three-dimensional stretching structures associated with contrasting tracer dispersion around an urban building canopy at NUS campus. These results demonstrate a scalable workflow in large-scale Lagrangian diagnostics and analysis of flow data.
ER -