Publications

Time integration as filtering: a space-time discretization-aware LES formulation

Syver Døving Agdestein
arXiv preprint

Why do LES closures ignore the time step? Forward Euler is itself a top-hat filter in time, so the fully discrete scheme is an exactly filtered equation with one extra temporal residual that grows to nearly 60% of the closure target at high CFL.

@misc{agdesteinTimeIntegrationFiltering2026,
  title = {Time Integration as Filtering: A Space-Time Discretization-Aware {{LES}} Formulation},
  author = {Agdestein, Syver D{\o}ving},
  year = 2026,
  month = jun,
  number = {arXiv:2606.17759},
  eprint = {2606.17759},
  publisher = {arXiv},
  archiveprefix = {arXiv},
  doi = {10.48550/arXiv.2606.17759},
}

Data-driven discrete closure models for large-eddy simulation of incompressible turbulence

Syver Døving Agdestein
PhD thesis

What if the grid, filter, and fluxes are defined before any turbulence modeling? Neural closures then become stable with cheap a-priori training, the unresolved stress gets an exact discrete expression, and a differentiable GPU solver ties it together.

@phdthesis{agdesteinDatadrivenDiscreteClosure2026,
  type = {Phd {{Thesis}} 1 ({{Research TU}}/e / {{Graduation TU}}/e)},
  title = {Data-Driven Discrete Closure Models for Large-Eddy Simulation of Incompressible Turbulence},
  author = {Agdestein, Syver D{\o}ving},
  year = 2026,
  month = may,
  address = {Eindhoven},
  isbn = {978-90-386-6703-4},
  school = {Eindhoven University of Technology},
}

A differentiable software suite for accelerated simulation of turbulent flows

Syver Døving Agdestein and Benjamin Sanderse
arXiv preprint

How do you train a neural closure model inside a running turbulence solver? IncompressibleNavierStokes.jl is a matrix-free, GPU-capable Julia solver with adjoints for every operator, running DNS up to 840³ on a single H100.

@misc{agdesteinDifferentiableSoftwareSuite2026,
  title = {A Differentiable Software Suite for Accelerated Simulation of Turbulent Flows},
  author = {Agdestein, Syver D{\o}ving and Sanderse, Benjamin},
  year = 2026,
  month = apr,
  number = {arXiv:2604.18536},
  eprint = {2604.18536},
  primaryclass = {math},
  publisher = {arXiv},
  archiveprefix = {arXiv},
  doi = {10.48550/arXiv.2604.18536},
}

Exact Expressions for the Unresolved Stress in a Finite-Volume Based Large-Eddy Simulation

Syver Døving Agdestein, Roel Verstappen, and Benjamin Sanderse
Journal of Computational Physics 556

What does a finite-volume LES actually need to close? An exact, non-symmetric, non-local residual stress that reproduces filtered DNS to machine precision, where the classical stress leaves errors of 16 to 21%.

@article{agdesteinExactExpressionsUnresolved2026,
  title = {Exact Expressions for the Unresolved Stress in a Finite-Volume Based Large-Eddy Simulation},
  author = {Agdestein, Syver D{\o}ving and Verstappen, Roel and Sanderse, Benjamin},
  year = 2026,
  month = jul,
  journal = {Journal of Computational Physics},
  volume = {556},
  pages = {114810},
  doi = {10.1016/j.jcp.2026.114810},
}

A New Data-Driven Energy-Stable Evolve-Filter-Relax Model for Turbulent Flow Simulation

Anna Ivagnes, Toby van Gastelen, Syver Døving Agdestein, Benjamin Sanderse, Giovanni Stabile, and Gianluigi Rozza
Computer Methods in Applied Mechanics and Engineering 450

Can the evolve-filter-relax method be made both accurate and provably stable? Learn the filter from DNS data with one least-squares problem per Fourier mode, and pick the relax parameter online so kinetic energy can never grow.

@article{ivagnesNewDatadrivenEnergystable2026,
  title = {A New Data-Driven Energy-Stable Evolve-Filter-Relax Model for Turbulent Flow Simulation},
  author = {Ivagnes, Anna and Van Gastelen, Toby and Agdestein, Syver D{\o}ving and Sanderse, Benjamin and Stabile, Giovanni and Rozza, Gianluigi},
  year = 2026,
  month = mar,
  journal = {Computer Methods in Applied Mechanics and Engineering},
  volume = {450},
  pages = {118654},
  doi = {10.1016/j.cma.2025.118654},
}

Discretize First, Filter next: Learning Divergence-Consistent Closure Models for Large-Eddy Simulation

Syver Døving Agdestein and Benjamin Sanderse
Journal of Computational Physics 522

Why are neural LES closures unstable? Because they are trained for filtered continuous equations but used in a discrete solver. Discretizing first and using a divergence-consistent face-averaging filter makes a CNN closure stable with a-priori training alone.

@article{agdesteinDiscretizeFirstFilter2025,
  title = {Discretize First, Filter next: {{Learning}} Divergence-Consistent Closure Models for Large-Eddy Simulation},
  author = {Agdestein, Syver D{\o}ving and Sanderse, Benjamin},
  year = 2025,
  month = feb,
  journal = {Journal of Computational Physics},
  volume = {522},
  pages = {113577},
  doi = {10.1016/j.jcp.2024.113577},
}

Discretize First, Filter next – a New Closure Model Approach

Syver Døving Agdestein and Benjamin Sanderse
8th European Congress on Computational Methods in Applied Sciences and Engineering (ECCOMAS)

What is the exact closure of a discretely filtered equation? On 1D linear convection with a non-uniform filter it is a single matrix, which lets three learning strategies be compared without any neural-network confounders.

@inproceedings{agdesteinDiscretizeFirstFilter2022,
  title = {Discretize First, Filter next -- a New Closure Model Approach},
  booktitle = {8th {{European Congress}} on {{Computational Methods}} in {{Applied Sciences}} and {{Engineering}}},
  author = {Agdestein, Syver D{\o}ving and Sanderse, Benjamin},
  year = 2022,
  publisher = {CIMNE},
  doi = {10.23967/eccomas.2022.094},
}

Practical Computation of the Diffusion MRI Signal Based on Laplace Eigenfunctions: Permeable Interfaces

Syver Døving Agdestein, Try Nguyen Tran, and Jing-Rebecca Li
NMR in Biomedicine 35.3

How do you simulate the diffusion MRI signal through permeable cell membranes? Extend the Laplace-eigenfunction matrix formalism with doubled interface nodes, then evaluate any gradient sequence one to three orders of magnitude faster than solving the Bloch-Torrey equation directly.

@article{agdesteinPracticalComputationDiffusion2022,
  title = {Practical Computation of the Diffusion {{MRI}} Signal Based on {{Laplace}} Eigenfunctions: Permeable Interfaces},
  author = {Agdestein, Syver D{\o}ving and Tran, Try Nguyen and Li, Jing-Rebecca},
  year = 2022,
  month = mar,
  journal = {NMR in Biomedicine},
  volume = {35},
  number = {3},
  pages = {e4646},
  doi = {10.1002/nbm.4646},
}

Développement de mélanges d'experts par apprentissage machine pour la conception avion

Development of mixtures of experts using machine learning for aircraft design

Syver Døving Agdestein
Master’s internship report · INSA Toulouse / IRT Saint Exupéry · August 2020

How can aircraft-design surrogates adapt across physical regimes? A modular machine-learning layer for GEMS combines clustering, classification, and local Kriging models into a mixture of experts validated on Burgers shocks and Airbus pressure fields, alongside an analytical benchmark for multidisciplinary optimization under uncertainty.

Report in French · not publicly available

Artery.FE: An Implementation of the 1D Blood Flow Equations in FEniCS

Syver Døving Agdestein, Kristian Valen-Sendstad, and Alexandra Diem
Journal of Open Source Software 3.32

How does a pulse travel through a network of arteries? Artery.FE solves the 1D blood flow equations in FEniCS from a single configuration file, with Windkessel outlets and bifurcation coupling, as the first ready-to-use FEniCS implementation.

@article{agdesteinArteryFEImplementation1D2018,
  title = {Artery.{{FE}}: {{An}} Implementation of the {{1D}} Blood Flow Equations in {{FEniCS}}},
  author = {Agdestein, Syver and {Valen-Sendstad}, Kristian and Diem, Alexandra},
  year = 2018,
  month = dec,
  journal = {Journal of Open Source Software},
  volume = {3},
  number = {32},
  pages = {1107},
  doi = {10.21105/joss.01107},
}