1 May 2020 in Physics of Fluids
Improved delayed detached eddy simulation is performed to investigate aero-optical distortions induced by Mach 0.5 flow over a cylindrical turret with a flat window at a Reynolds number of Re = 5.9 × 105 based on the turret radius. Optical wave-front distortions and associated far-field intensities at elevation angles of 90°, 100°, and 120° are calculated using the geometric ray-tracing method and Fourier optics theory, respectively. The time-averaged properties of the flow fields and the optical distortions are compared with the experimental results, and reasonable agreements are obtained. It is shown that large-scale coherent structures significantly increase the optical distortions and, consequently, degrade the optical performance in terms of the far-field intensity level, while the optical effects of the attached turbulent boundary layers and separation bubbles are trivial and negligible. Very little difference is observed in both instantaneous and statistical optical results calculated with and without defocus and astigmatism components, indicating the adequacy of the removing piston and tilt components for aiding in the design of adaptive optical systems. Both co-flow and blowing jet fluid control methods are introduced with a steady mass flow to alleviate wave-front distortions, and preliminary simulations demonstrate the practicability of these fluid control methods in suppressing the optical distortions. It is shown that both the active control methods are competent to reduce the overall wave-front root-mean-square of the optical path difference.
10 November 2026 in International Journal of Heat and Fluid Flow
23 February 2026 in SSRN Electronic Journal
This paper presents a bi-fidelity framework to accelerate CFD-driven evolutionary algorithms for turbulence closure discovery in RANS models. The framework leverages the statistical correlation between inexpensive a priori evaluations of symbolic model terms (our low fidelity in this paper) and costly a posteriori performance obtained from CFD-in-the-loop simulations (our high fidelity). A kernel density estimation-based probabilistic mapping between the two fidelity levels is learned and refined during evolution, enabling selective high-fidelity evaluation of the most promising candidate models. Symbolic representations of turbulence model-form error terms, including the Reynolds-stress anisotropy correction and the turbulent kinetic energy production deficit, are constructed using a tensor-based gene expression programming formulation, yielding compact and interpretable closures that can be directly embedded into CFD solvers. The methodology is demonstrated on rectangular duct flows as well as flows over periodic hills and a curved backward-facing step. Across all considered configurations, the bi-fidelity strategy reduces the number of high-fidelity CFD evaluations by approximately 60–70% without compromising the predictive accuracy of the resulting models compared to conventional CFD-driven approaches
5 June 2026 in SSRN Electronic Journal
This paper presents an iteration-based multi-fidelity framework for reducing the computational cost of CFD-in-the-loop turbulence model discovery. The framework is built upon tensor-based gene expression programming (TB-GEP), which generates symbolic turbulence models that can be embedded directly into CFD solvers and evaluated a posteriori during training. Instead of advancing every newly generated candidate model to full convergence in each evolutionary generation, the framework first evaluates all candidates at a partially converged CFD state, treating it as an intermediate-fidelity level for low-cost fitness estimation. The resulting estimates are then used to preselect a small subset of promising candidates for fully converged high-fidelity CFD evaluations. The proposed framework is applied to the discovery of Reynolds-stress anisotropy corrections for turbulent rectangular duct flows, a challenging benchmark in which the baseline 𝑘–𝜔 SST model fails to predict Prandtl’s secondary flow of the second kind. Compared with conventional single-fidelity CFD-in-the-loop training, the proposed approach reduces the effective computational cost by up to 92% while maintaining comparable predictive accuracy of the discovered models. The results demonstrate that partially converged CFD states can provide reliable fitness guidance during model evolution, offering a practical route toward more efficient CFD-in-the-loop turbulence model discovery.