Mean-Flow Adjoint Sensitivity Analysis of Unsteady Flow Around Porous Cylinders Using a Homogenized Lattice Boltzmann Method
Shota Ito, Johannes L. Grafen, Fedor Bukreev, Adrian Kummerländer, Mathias J. Krause · 2026
Read Mean-Flow Adjoint Sensitivity Analysis of Unsteady Flow Around Porous Cylinders Using a Homogenized Lattice Boltzmann Method by Shota Ito, Johannes L. Grafen, Fedor Bukreev, Adrian Kummerländer, Mathias J. Krause (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-31707.txt.
Introduction
Ito and four colleagues present a mean-flow adjoint sensitivity method for unsteady flow around porous cylinders. Adjoint methods give design gradients cheaply, but unsteady flows force costly storage of the full time history, and chaotic flow can break the gradients. The team uses a homogenized lattice Boltzmann method, which models solid bodies as local porous material through Brinkman penalization.
They test adjoint gradients for drag and energy dissipation from steady laminar flow through unsteady flow to turbulence. For the turbulent case at Reynolds number 3900, automatic differentiation builds the adjoint kernels, including a large-eddy turbulence model. That setup lets them compare directly against the frozen turbulence assumption, a common shortcut.
Catalog record
| CREATOR | Shota Ito, Johannes L. Grafen, Fedor Bukreev, Adrian Kummerländer, Mathias J. Krause |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 60 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.31707
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Mean-Flow Adjoint Sensitivity Analysis of Unsteady Flow Around Porous Cylinders Using a Homogenized Lattice Boltzmann Method Shota Ito, Johannes L. Grafen, Fedor Bukreev, Adrian Kummerländer, Mathias J. Krause Abstract Adjoint-based sensitivity analysis is an indispensable tool for large-scale fluid-dynamic design and distributed control problems, yet its application to unsteady and turbulent flows is frequently hindered by the prohibitive memory footprint of transient checkpointing and the divergence of gradients in chaotic regimes. To address these computational bottlenecks, this paper presents a mean-flow adjoint sensitivity analysis framework for unsteady flows around porous cylinders using the homogenized lattice Boltzmann method (HLBM). Within this framework, solid structures are efficiently modeled as local porous media utilizing a Brinkman penalization approach. We systematically investigate HLBM-based adjoint gradients for drag and energy dissipation objective functionals, transitioning from steady laminar to unsteady, and finally to turbulent flow regimes. For the turbulent case at Re = 3900, a proof-of-concept is conducted where the framework relies on automatic differentiation to automatically generate adjoint kernels containing subgrid-scale (SGS) turbulence models for large eddy simulations (LES), circumventing manual derivation and allowing for a direct comparison against the frozen turbulence assumption (FTA). 1 Introduction The optimization of fluid dynamic design and the optimal control of flow problems pose key disciplines in aerodynamic engineering [ 8 , 18 , 46 ] . Therein, it is an established procedure to couple methods of mathematical optimization and computational fluid dynamics (CFD) simulations to perform design improvements in an automated manner, while enforcing physicality of the optimized solution by numerically approximating the governing partial differential equations of the underlying flow phenomenon [ 10 , 22 ] . Therein, gradient-based optimization (GBO) is of particular interest as the iterative execution of CFD simulations for industrial-scale problems dominates the computational cost, where GBO, utilizing local gradient information, in general outperforms other approaches in terms of convergence rates [ 54 , 14 ] . Adjoint-based sensitivity analysis has established itself as an indispensable tool for solving large-scale gradient-based fluid-dynamic design and flow control problems [ 7 , 16 ] . Because the computational cost of evaluating an objective gradient via the adjoint method remains independent of the number of design variables, it is suited for distributed control problems such as fluid topology optimization [ 5 , 38 , 52 , 29 ] or shape optimization [ 33 , 51 ] . Within this context, the lattice Boltzmann method (LBM) represents a highly attractive discretization scheme due to its great suitability for parallel computing [ 21 , 26 , 25 ] . Beyond offering remarkable computational efficiency within iterative optimization loops, LBM is highly compatible with the adjoint analysis, as the corresponding adjoint problem can be again discretized by an LBM, providing the same performance benefits [ 15 ] . To incorporate the solid body in the fluid topology optimization, a Brinkman penalization approach is often applied, where the pressure drop is modeled by a local porous medium imitating the immersed solid structure [ 16 , 52 , 19 ] . By controlling the permeability of the porous medium, the topology can be optimized regarding different objective functionals such as drag or pressure drop. Krause et al. [ 20 ] proposed the homogenized lattice Boltzmann method (HLBM), extended by Kummerländer et al. [ 25 ] , which efficiently employs the Brinkman penalization to model solid structures in the fluid. In prior works, the HLBM method has been successfully applied for topology optimization problems, such as in [ 16 , 19 ] . The application of adjoint sensitivity analysis to transient flow control, especially in LBM, remains very limited in the existing literature. In the realm of unsteady flows, standard transient adjoint simulations require checkpointing or storing the entire forward state history to be traversed in reverse time, creating an often prohibitive memory bottleneck [ 35 , 32 ] . Prior work reports efficient checkpointing approaches such as the local-in-time method , in which checkpointing is combined with approximations of unsteady adjoint solutions to reduce the memory footprint in the adjoint simulation [ 7 , 53 ] . However, for large-scale simulations, especially for three-dimensional problems, the storage of the primal state solution becomes quickly computationally unfeasible [ 50 ] . Especially for turbulent flows, computational resources become prohibitively large when resolving transient structures. Furthermore, the chaotic flow dynamics lead to diverging solution trajectories and thus to failure of conventional gradient computation, requiring more sophisticated and numerically expensive approaches, such as the least squares shadowing method [ 32 , 4 ] . In classical approaches, the adjoint analysis is applied to mean flow solutions of turbulent flows by combining it with Reynolds averaged Navier-Stokes (RANS) simulations to circumvent the diverging gradients [ 31 , 37 , 42 ] . Additionally, by driving the reversed-time adjoint equations using the time-averaged primal flow field rather than instantaneous transient snapshots, Dwight et al. demonstrated that computationally tractable and physically meaningful surface sensitivities could be recovered [ 11 ] . Meliga et al. investigated the adjoint 2D squared cylinder flow using a finite element method for Reynolds numbers up to 22000 . A RANS model is applied in the primal simulation, where the adjoint problem is constructed on the averaged solutions to handle the chaotic flow behavior in the turbulent case [ 32 ] . In LBM, Cheylan et al. proposed the mean-flow adjoint-based sensitivity analysis for aerodynamic shape optimization to address drag minimization problems for a vehicle [ 8 ] . Therein, the primal state solution is averaged and used to evaluate the Jacobian expressions in the adjoint simulation, in which the controlled body shape is incorporated into the flow field via adjoint curvature-capturing boundary conditions. To the best of the authors’ knowledge, the investigation of mean-flow-based adjoint sensitivities for LBM simulations is not yet applied in the context of topology optimization, e.g., for a Brinkman penalization model such as HLBM. Therefore, the present work systematically investigates HLBM-based adjoint gradients for a porous cylinder with drag-type functionals, from steady laminar to unsteady and finally turbulent flow regimes. Both works [ 8 ] and [ 32 ] , neglect the subgrid-scale (SGS) turbulence model in the adjoint sensitivity analysis by employing the frozen turbulence assumption (FTA) due to the challenging manual derivation of the non-linear adjoint large eddy simulation (LES) based on the Smagorinsky model, despite the loss in gradient accuracy [ 8 , 42 ] . So far, the comparison of computed gradients from fully differentiated turbulence models and those obtained by the FTA has been reported only for RANS models [ 31 , 42 ] . This manuscript demonstrates preliminary comparison results of adjoint Smagorinsky-LES models and the FTA for the turbulent flow past a 3D cylinder. The framework developed by Ito et al. is used, where the adjoint kernels, containing SGS turbulence models, are automatically generated from their primal implementations using automatic differentiation (AD), omitting the need for manual derivation [ 15 ] . The research questions of the present work are formulated as follows: 1. Can HLBM adjoints compute accurate drag-type sensitivities for a porous cylinder flow? 2. How do mean-flow adjoint gradients behave for unsteady vortex shedding using the HLBM setting? 3. Can the same framework be extended to 3D and