A highly accurate strategy for data-driven turbulence modeling

Computational and Applied Mathematics, Год журнала: 2024, Номер 43(1)

Опубликована: Янв. 19, 2024

Язык: Английский

---

Constrained re-calibration of two-equation Reynolds-averaged Navier–Stokes models

Theoretical and Applied Mechanics Letters, Год журнала: 2024, Номер 14(2), С. 100503 - 100503

Опубликована: Фев. 7, 2024

Machine-learned augmentations to turbulence models can be advantageous for flows within the training dataset but often cause harm outside. This lack of generalizability arises because constants (as well as functions) in a Reynolds-averaged Navier–Stokes (RANS) model are coupled, and un-constrained re-calibration these (and disrupt calibrations baseline model, preservation which is critical model's generalizability. To safeguard behaviors beyond dataset, machine learning must constrained such that basic like law wall kept intact. letter aims identify constraints two-equation RANS so future work performed without violating constraints. We demonstrate identified not limiting. Furthermore, they help preserve model.

Язык: Английский

---

Shape optimization and hydrodynamic simulation of a Magnus anti-rolling device based on fully parametric modeling

Physics of Fluids, Год журнала: 2023, Номер 35(5)

Опубликована: Май 1, 2023

Ship anti-rolling devices are an essential component of modern vessels. The core the Magnus effect-based ship device is a rotating cylinder, hereinafter referred to as cylinders. In this paper, fully parametric three-dimensional modeling cylinders was performed, and design space dimension reduced using Sobol optimization method while still providing accurate reliable results. generates quasi-random sequences that more uniformly spaced in search can efficiently cover entire solution space. shape study cylinder carried out conjunction with computational fluid dynamics find geometry excellent hydrodynamic performance. Critical parameters include diameters ends length cylinder. flow field characteristics before after were compared. results show there be multiple local optimal values for lift drag within increase decrease drag. effect primarily influences position vortex-shedding separation point at surface deflects wake one side. For optimized distribution pressure velocity significantly different. This research forms basis improving practical application devices.

Язык: Английский

---

Data-driven turbulence modeling for fluid flow and heat transfer in peripheral subchannels of a rod bundle

Physics of Fluids, Год журнала: 2024, Номер 36(2)

Опубликована: Фев. 1, 2024

This study presents a comparison of the performance machine learning (ML) techniques, specifically multi-dimensional gene expression programming (MGEP), tensor basis neural network (TBNN), and also proposes novel universally interpretable architecture to model turbulent scalar flux (UIML-s) enhance turbulence models for fluid flows at different Prandtl numbers in channels with complex shapes walls channel cross section. In particular, peripheral subchannels rod bundles are primary interest. However, accuracy mean velocity distributions predicted by commonly used still poses challenge compared data extracted from high-fidelity eddy-resolving numerical simulations, particularly engineering applications involving geometry flows. present study, utilizing an explicit algebraic nonlinear Reynolds-stress term obtained through both evolutionary MGEP optimization TBNN, secondary flow structure has been adequately cross-wise square duct rectangular three longitudinal rods. is observed concurrent runs performed direct simulation (DNS) but completely absent results produced baseline Reynolds-averaged Navier–Stokes (RANS) closure, which employs linear eddy viscosity Reynolds stress tensor. Comparison TBNN shown their nearly equal flow; however, works better more Furthermore, based on field RANS-MGEP model, ML modification gradient diffusion hypothesis, integrated into aforementioned RANS-ML called as UIML-s, significantly improves bumps serving prototype subchannel bundle. The normalized root squared error decreases 13.5% 7.6%, bringing closer DNS data, near-wall region. Another approach, MGEP-s, yields acceptable results, identical those UIML-s. These findings highlight potential using data-driven calibration closures predictability RANS simulations flows, heat, mass transfer geometry.

Язык: Английский

---

A data‐driven turbulence modeling for the Reynolds stress tensor transport equation

International Journal for Numerical Methods in Fluids, Год журнала: 2024, Номер 96(7), С. 1194 - 1214

Опубликована: Март 26, 2024

Abstract The long lasting demand for better turbulence models and the still prohibitively computational cost of high‐fidelity fluid dynamics simulations, like direct numerical simulations large eddy have led to a rising interest in coupling available datasets popular, yet limited, Reynolds averaged Navier–Stokes through machine learning (ML) techniques. Many recent advances used stress tensor or, less frequently, force vector as target these corrections. In present work, we considered an unexplored strategy, namely use modeled terms transport equation ML predictions, employing neural network approach. After that, solve coupled set governing equations obtain mean velocity field. We apply this strategy flow square duct. obtained results consistently recover secondary flow, which is not baseline that model. were compared with other approaches literature, showing path can be useful seek more universal turbulence.

Язык: Английский

---

Validating the design optimisation of ultrasonic flow meters using computational fluid dynamics and surrogate modelling

International Journal of Heat and Fluid Flow, Год журнала: 2023, Номер 100, С. 109112 - 109112

Опубликована: Янв. 23, 2023

Domestic ultrasonic flow meters with an intrusive two-stand configuration present a complex behaviour due to their unique geometry, which offers interesting case evaluate optimisation methods in wall-bounded turbulent flows. In this study, the design and analysis of computer models by computational fluid dynamics is used predict perform robust meter. The accomplished surrogate modelling based on Kriging, Latin hypercube sampling, Bayesian strategies ensure high-quality space-filled response surface. A novel function quantify meter measurement uncertainty defined evaluated together pressure drop order define multi-objective problem. Pareto front shown compared numerically experimentally against laser Doppler velocimetry experiments, displaying performance gains geometrical changes 3D space. From various improved designs sampled experimentally, 4.9% reduction 37.4% have been analysed baseline case. applied methodology provides efficient framework changes, improving internal-flow problems similar features.

Язык: Английский

---

Revisiting Tensor Basis Neural Network for Reynolds stress modeling: Application to plane channel and square duct flows

Computers & Fluids, Год журнала: 2024, Номер 275, С. 106246 - 106246

Опубликована: Март 16, 2024

Язык: Английский

---

Data-driven Reynolds stress models based on the frozen treatment of Reynolds stress tensor and Reynolds force vector

Physics of Fluids, Год журнала: 2023, Номер 35(7)

Опубликована: Июль 1, 2023

For developing a reliable data-driven Reynold stress tensor (RST) model, successful reconstruction of the mean velocity field based on high-fidelity information (i.e., direct numerical simulations or large-eddy simulations) is crucial and challenging, considering ill-conditioning problem Reynolds-averaged Navier–Stokes (RANS) equations. It shown that frozen treatment Reynolds force vector (RFV) reduced even for cases with very high number; therefore, it has better potential to be used in development RANS models. In this study, we compare algebraic RST correction models are trained both RFV aforementioned potential. We derive vector-based framework similar tensor-based RST. Regarding complexity models, sparse regression set candidate functions multi-layer perceptron network. The training process applied data three cases, including square-duct secondary flow, roughness-induced periodic hills flow. results showed using discrepancy values, instead generally does not improve despite fact propagation shows lower errors all cases. complexity, improves prediction flows, but performance case hills.

Язык: Английский

---

Progressive augmentation of Reynolds stress tensor models for secondary flow prediction by computational fluid dynamics driven surrogate optimisation

International Journal of Heat and Fluid Flow, Год журнала: 2023, Номер 104, С. 109242 - 109242

Опубликована: Ноя. 7, 2023

Generalisability and the consistency of a posteriori results are most critical points view regarding data-driven turbulence models. This study presents progressive improvement models using simulation-driven Bayesian optimisation with Kriging surrogates where is achieved by multi-objective approach based on duct flow quantities. We aim for augmentation secondary-flow prediction capability in linear eddy-viscosity model k−ω SST without violating its original performance canonical cases e.g. channel flow. Progressively data-augmented explicit algebraic Reynolds stress (PDA-EARSMs) obtained enabling secondary flows that standard fails to predict. The new tested guaranteeing they preserve successful model. Subsequently, numerical verification performed various test cases. Regarding generalisability models, unseen demonstrate significant streamwise velocity. These highlight potential enhance fluid simulation while preserving robustness stability solver.

Язык: Английский

---

Mitigating ill-conditioning of Reynolds-averaged Navier–Stokes equations for experimental data-driven turbulence closures

Physics of Fluids, Год журнала: 2025, Номер 37(4)

Опубликована: Апрель 1, 2025

Experiments are a fundamental source of high-fidelity data in turbulence, providing reliable targets for modeling. As Direct Numerical Simulations (DNS), they can capture flow dynamics without an underlying model but less restrictive than DNS regarding geometry and Reynolds number limits. In this work, we conducted experimental campaign to generate quality on the turbulent quantities square duct flows. The recent literature has revealed ill-conditioned nature Reynolds-averaged Navier–Stokes (RANS) equations. These studies suggest that other target besides stress tensor (RST) compose closures lead more accurate solutions. We investigated statistical derived from these experiments feed RANS equations, testing various methods comparing different closure strategies. anticipated, using measured RST directly equations led significant error propagation predicted average velocity fields due equations' nature. Conversely, approaches successfully yielded propagated mean fields. Notably, implicit treatment linear components force vector provided precise results. By extending range tested numbers beyond those achievable by DNS, study underscores critical need careful selection data-driven turbulence highlights importance exploring new closing

Язык: Английский

---