A computational method for optimizing shapes in changing fluid flows produced nearly the same final designs as a full-time adjoint calculation while using substantially less reported computing capacity, according to an arXiv preprint. In the comparison, objective values differed by approximately 0.003%, while average runtime per iteration was reported as approximately 59.3% lower and required storage approximately 99.96% lower.
The work concerns topology optimization, the numerical adjustment of a design’s configuration to pursue a chosen flow objective. Its central idea is to identify a representative time window in which the flow has developed, then use that window for the objective and sensitivity calculations that guide each design update.
A narrower view of an unsteady flow
The framework checks consecutive time windows rather than relying only on a time interval chosen in advance. It monitors the average drag coefficient and compares that quantity across consecutive windows to identify the representative period.
Once a window is identified, it is held fixed while the objective and discrete-adjoint analysis are carried out within that design iteration. After the design changes, the representative window is identified again. The calculations use a regularized lattice Boltzmann large-eddy simulation method, together with a partial bounce-back model for fluid-solid boundaries.
That distinction mattered in one comparison. An early prescribed window produced a lower objective but an unreasonable configuration, whereas the adaptive criterion and a later prescribed window produced closely resembling streamlined, teardrop-shaped configurations. When both were reevaluated under a common criterion, their objective values differed by approximately 0.1%.
Tests moved from a benchmark to more demanding shapes
The numerical evaluation covered backward-facing-step solver validation, circular-cylinder topology optimization, wake-flow recovery and U-bend optimization. The backward-facing-step test examined whether the flow solver reproduced global reattachment characteristics and local velocity distributions; the reported comparison found qualitative agreement with the benchmark results.
The researchers also checked the method’s shape sensitivities against finite-difference calculations. Here, sensitivity means an estimate of how the objective changes when the design is perturbed. Across the tested steady and unsteady regimes, the adjoint sensitivity distributions had relative errors below 1% compared with the finite-difference results.
In the wake-recovery cases, the optimized structures at Re 60 and Re 72 were recovered to approximately circular-cylinder shapes. The wake-recovery objective was reduced by more than 99% in the reported cases.
The U-bend cases showed pressure-drop objective reductions of 88.16% at Re 400 and 93.96% at Re 750. The reported flow changes included suppression of large downstream recirculation and the formation of more localized recirculation structures.
Similar outputs, lower computational burden
The closest test of the proposed approach compared it with full-time adjoint propagation. After common reevaluation, the final designs were nearly identical and their objective values differed by approximately 0.003%. The same comparison reported approximately 59.3% less average runtime per iteration and approximately 99.96% less required storage for the proposed framework.
These figures come from a reported computational-cost comparison, without an uncertainty interval or replicate-variability analysis. The results therefore support the method as a computational strategy within the cases tested, rather than as a guarantee that every lower objective corresponds to a physically representative flow window.
The evidence remains computational
The study’s evidence comes from numerical simulations and optimization runs in the listed flow cases. It does not report a physical prototype or experimental validation of the optimized topologies.
The tested case suite does not establish how robust the representative-window criterion would be in three-dimensional turbulent flows, coupled thermal-flow optimization or other untested unsteady regimes. Those extensions remain open questions rather than demonstrated results.
The document is an arXiv preprint, not a report of a journal-published study in the supplied metadata. The authors state that supporting code and data are available upon reasonable request.
The authors report support from three Chinese national or municipal funding programs. The findings offer a computational result about the tested flow-optimization problems, with further validation needed before the approach is extended to physical systems or broader flow regimes.
Paper data and sources
Original title: Adaptive Time Windows for Discrete Adjoint Topology Optimization of Unsteady Flows
Authors: Zongyuan Liu, Kentaro Yaji, Musaddiq Al Ali, Shengfeng Zhu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text