Large-scale-aware data augmentation for reduced-order models of high-dimensional flows

APL Machine Learning, Год журнала: 2025, Номер 3(1)

Опубликована: Янв. 13, 2025

Convolutional autoencoders have proven to be an adequate tool perform reduced-order modeling for high-dimensional nonlinear dynamical systems. Their goal is reduce dimensionality strongly while preserving the most characteristic features of system. Here, we show that these models rely sensitively on completeness provided data. This particularly challenging fully turbulent flows with their coherent structures ranging from large-scale superstructures dissipative eddies over orders magnitude in time and space. As a result, unrealistically large number data snapshots would required properly cover all essential dynamics, whereas small length scales require only respective flow, especially long lasting are difficult characterize either numerically or experimentally. We demonstrate three types missing representation leads failures training process. suggest method mitigate this shortcoming. includes transformation samples new structures, which enhance Furthermore, skip augmentations more detrimental model performance. evaluate our datasets, two numerical simulations Rayleigh–Bénard convection one laboratory experiment flow past array cylinders. can substantially improve utility In way, avoid intensive grid search through possible augmentation combinations without further knowledge about underlying

Язык: Английский

---

On the spatial prediction of the turbulent flow behind an array of cylinders via echo state networks

Engineering Applications of Artificial Intelligence, Год журнала: 2025, Номер 144, С. 110079 - 110079

Опубликована: Янв. 23, 2025

Язык: Английский

---

Impact of tracer particles on the electrolytic growth of hydrogen bubbles

Physics of Fluids, Год журнала: 2024, Номер 36(1)

Опубликована: Янв. 1, 2024

The thermocapillary effect at gas bubbles growing micro-electrodes seems well understood. However, the interfacial flow measured in upper bubble part decays faster than found first simulations by Massing et al. [“Thermocapillary convection during hydrogen evolution microelectrodes,” Electrochim. Acta 297, 929 (2019)]. Recently, Meulenbroek attributed origin of difference to influence surfactants being present electrolyte [“Competing Marangoni effects from a stagnant cap on interface attached microelectrode,” 385, 138298 (2021)]. Surprisingly, presence tracer particles added for measuring its was not yet considered. Our recent experiments reveal that varying small amount influences shape, dynamics, and also nearby. We therefore model describe particle attraction dynamics interface, which allows us quantify impact. Corresponding are validated against measurements different bulk concentrations show good agreement tangential velocity profile caused thermo- solutocapillary effects. Depending concentration, parts become stagnant. results allow deeper insight into complex phenomena electrolytic further put attention careful application particle-based measurement techniques gas–liquid systems.

Язык: Английский

---

On the prediction of the turbulent flow behind cylinder arrays via Echo State Networks

Machine Learning Science and Technology, Год журнала: 2024, Номер 5(3), С. 035005 - 035005

Опубликована: Июнь 4, 2024

Abstract This study aims at the prediction of turbulent flow behind cylinder arrays by application Echo State Networks (ESN). Three different arrangements seven cylinders are chosen for current study. These represent regimes: single bluff body flow, transient and co-shedding flow. allows investigation flows that fundamentally originate from wake yet exhibit highly diverse dynamics. The data is reduced Proper Orthogonal Decomposition (POD) which optimal in terms kinetic energy. Time Coefficients POD Modes (TCPM) predicted ESN. network architecture optimized with respect to its three main hyperparameters, Input Scaling (INS), Spectral Radius (SR), Leaking Rate (LR), order produce best predictions Weighted Prediction Score (WPS), a metric leveling statistic deterministic prediction. In general, ESN capable imitating complex dynamics even longer periods several vortex shedding cycles. Furthermore, mutual interdependencies TCPM well preserved. However, hyperparameters depend strongly on characteristics. Generally, as become faster more intermittent, larger LR INS values result better predictions, whereas less clear trends SR observable.

Язык: Английский

---

A comprehensive study of low-Reynold flow through two tandem cylinders with various configurations using the Lattice Boltzmann method

Marine Geophysical Research, Год журнала: 2025, Номер 46(1)

Опубликована: Март 1, 2025

Язык: Английский

---

Spatial prediction of the turbulent unsteady von Kármán vortex street using echo state networks

Physics of Fluids, Год журнала: 2023, Номер 35(11)

Опубликована: Ноя. 1, 2023

The spatial prediction of the turbulent flow unsteady von Kármán vortex street behind a cylinder at Re = 1000 is studied. For this, an echo state network (ESN) with 6000 neurons was trained on raw, low-spatial resolution data from particle image velocimetry. During prediction, ESN provided one half domain fluid flow. task to infer missing other half. Four different decompositions termed forward, backward, forward–backward, and vertical were examined show whether there exists favorable region for which performs best. Also, it checked direction has influence network's performance. In order measure quality predictions, we choose velocity (VVPD). Furthermore, ESN's two main hyperparameters, leaking rate (LR) spectral radius (SR), optimized according VVPD values corresponding output. Moreover, each hyperparameter combination run 24 random reservoir realizations. Our results that are highest LR ≈ 0.6, quite independent SR all four approaches. maximum ≈0.83 achieved predictions while forward case VVPDmax=0.74 achieved. We found predicted fields predominantly align their respective ground truth. best overall accordance backward forward–backward scenarios. summary, conclude stable reconstructed over long period time, along simplicity machine learning algorithm (ESN), relied coarse experimental only, demonstrates viability as suitable method application in turbulence.

Язык: Английский

---

Numerical investigation of flow across three co-rotating cylinders in side-by-side arrangement

Physics of Fluids, Год журнала: 2023, Номер 35(11)

Опубликована: Ноя. 1, 2023

Flow across three side-by-side co-rotating cylinders is investigated at a Reynolds number of 100 and non-dimensional rotation rates varied from 0 to 8, for spacing ratios L/D=1.5, 2, 4 through two-dimensional numerical simulations, where D L are cylinder diameter the center-to-center between cylinders, respectively. For L/D=1.5 wakes classified into regime FF (flip-flopping) smaller SB (single-body) higher rates. Each can be further divided sub-regimes based on wake patterns. Regime flow switches two patterns intermittently. The L/D=1.5: vortex shedding (SB-VS), steady state (SB-SS), secondary instability (SB-SI) same as those single rotating gap too weak have effect global wake. A new sub-regime single-body quasi-steady (SB-QS) found L/D=2, shear layers in interact weakly with each other but do not form strong vortices. L/D=4, regimes found: 3V-to-3S (transition wake), suppressed consecutively one by increase rate, TB (two-body) behave body. L/D=4 has sub-regimes: (TB-SS) (TB-SI). effects force coefficients quantified. all 4, standard derivation drag lift significantly greater than that when occurs.

Язык: Английский

---

Flow stability and permeability in a nonrandom porous medium analog

Physics of Fluids, Год журнала: 2024, Номер 36(10)

Опубликована: Окт. 1, 2024

The estimation of the permeability porous media to fluids is fundamental importance in fields as diverse oil and gas industry, agriculture, hydrology, medicine. Despite more than 150 years since publication Darcy's linear law for flow media, several questions remain regarding range validity this law, constancy coefficient, how define transition from Darcy other regimes. This study a numerical investigation stability nonrandom quasi-tridimensional medium analog. effect increasing pressure gradient on velocity field Darcy–Forchheimer coefficients investigated three different obstacles radius. nonlinear behavior associated with formation jets outlet development instabilities. Different representations Reynolds number proved adequate detect deviation law. instantaneous calculated at each was sensitive velocity, agreement previous studies stating that cannot be conceptualized constant real flows.

Язык: Английский

---

On the Spatial Prediction of the Turbulent Flow Behind an Array of Cylinders Via Echo State Networks

Опубликована: Янв. 1, 2024

Язык: Английский

---

Numerical study on the flow of Bingham plastic fluids around an array of cylinders

Materials Today Proceedings, Год журнала: 2023, Номер 111, С. 78 - 85

Опубликована: Ноя. 23, 2023

Язык: Английский

---