A new numerical method for simulating gas flows showed lower computing demands in a series of tests while remaining close to a reference calculation on the main flow variables. The preprint reports speedups of roughly six to seven and substantial reductions in GPU memory in several benchmarks, although the authors say implementation overhead limits how fully those gains translate into runtime savings.
The method, called MLVS-DUGKS, is designed to reduce redundancy in the velocity space used by a discrete unified gas-kinetic scheme, or DUGKS. It identifies each cell's local velocity-space needs from its distribution function, groups similar needs, dynamically constructs representative velocity spaces and assigns one representative space to each cell.
Accuracy held close to the reference calculation
The central test was whether that compression could preserve numerical accuracy and conservation. In a Sod shock-tube calculation, differences from the uniform-velocity-space reference stayed below 0.12% for density, 0.67% for velocity and 0.14% for temperature. Heat-flux differences were larger, ranging from 1.44% to 1.86%.
The method was also tested in a lid-driven cavity, a Mach-5 flow past a circular cylinder and a transient expansion between two connected cavities. In the cavity case, differences in the basic flow variables stayed below 1.03%, while heat-flux differences ranged from 1.84% to 2.00%. The reported speedup was 6.36 to 6.51 times, with GPU memory reduced by a factor of 5.25.
The cylinder calculations showed principal-field differences below 1.23%. At the stagnation point, where the flow meets the cylinder, the heat-flux difference stayed below 0.43%, including a reported difference of 0.077% at Kn = 0.01. The runs were reported as 7.01 to 7.51 times faster, with GPU-memory reduction factors of 6.15 to 6.74.
In the two-cavity expansion, the adaptive calculation used an average of 988.5 velocity points per physical cell, compared with 10,201 for the reference approach. That was a reported velocity-space compression factor of 10.32, alongside a speedup factor of 6.67 and a GPU-memory reduction factor of 6.88.
Conservation was a central test
The savings depend on moving information between velocity spaces without losing the quantities the solver must conserve. The method uses a linear moment-constrained correction, meaning it adjusts the discrete distribution to enforce compatibility and preserve prescribed moments during collision and mapping between velocity spaces.
In a compatibility-recovery test, the four-moment correction produced a maximum relative moment error of 1.11 × 10−16. The corresponding error was 2.63 × 10−3 without correction and 2.78 × 10−16 with a Newton iterative correction. The linear correction averaged 33.15 milliseconds per run, compared with 109.51 milliseconds for the iterative method, or about 30% of that cost.
A separate repeated-projection test examined whether higher-order information survived repeated transfers. The nine-moment correction preserved the basic conserved quantities and restored the constrained higher-order moments to their initial targets across all tested perturbation amplitudes and both perturbation types. Other treatments preserved only part of those requirements.
The conservation tests showed how much the reported errors differed between corrected and uncorrected calculations. In a uniform-flow test at Kn = 10−3, the reported mass error changed from 1.94 × 10−1 to 2.32 × 10−8, while total-energy error changed from 3.32 × 10−1 to 2.09 × 10−7 with the four-moment correction. In a sinusoidal-wave test, the corrected results had global conservation errors comparable to the reference solver across the three tested Knudsen numbers.
The gains are not yet a universal promise
The reported speedups should not be read as a fixed property of the algorithm. The authors say the single-GPU implementation was not extensively optimized, and that adaptation, transfers between velocity levels, mixed-level processing, synchronization and irregular memory access add overhead. As a result, reducing the number of velocity points does not translate directly into proportional reductions in wall-clock time.
The study also fitted a threshold-dependent runtime model using five cases from one benchmark. It reported a fit statistic of 0.925 and a maximum relative fitting error of 2.0%. The fit describes that tested case rather than establishing a general rule for every geometry or flow regime.
The main assessment covered deterministic numerical tests including the Sod shock tube, cavity flow, cylinder flow and two-cavity expansion. The reported evidence therefore concerns accuracy, conservation and efficiency within those benchmarks and this implementation. Whether the same balance holds when the implementation and hardware change remains open.
The manuscript is a preprint identified as arXiv:2609.35258v1 in physics.flu-dyn and dated 28 September 2026. It reports support from the National Natural Science Foundation of China, the Hunan Province Natural Science Foundation and the Hunan Provincial Innovation Foundation for Postgraduate.
Paper data and sources
Original title: A Conservative Multi-Level Adaptive Velocity-Space Method for the Discrete Unified Gas-Kinetic Scheme
Authors: Weijie Ren, Hang Yu, Zhengyu Tian et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-09-28
DOI: Not available
Original paper · Full text