A Review of Physics-Informed Machine Learning in Fluid Mechanics

Energies, Год журнала: 2023, Номер 16(5), С. 2343 - 2343

Опубликована: Фев. 28, 2023

Physics-informed machine-learning (PIML) enables the integration of domain knowledge with machine learning (ML) algorithms, which results in higher data efficiency and more stable predictions. This provides opportunities for augmenting—and even replacing—high-fidelity numerical simulations complex turbulent flows, are often expensive due to requirement high temporal spatial resolution. In this review, we (i) provide an introduction historical perspective ML methods, particular neural networks (NN), (ii) examine existing PIML applications fluid mechanics problems, especially Reynolds number (iii) demonstrate utility techniques through a case study, (iv) discuss challenges developing mechanics.

Язык: Английский

---

Physics-informed graph convolutional neural network for modeling fluid flow and heat convection

Physics of Fluids, Год журнала: 2023, Номер 35(8)

Опубликована: Авг. 1, 2023

This paper introduces a novel surrogate model for two-dimensional adaptive steady-state thermal convection fields based on deep learning technology. The proposed aims to overcome limitations in traditional frameworks caused by network types, such as the requirement extensive training data, accuracy loss due pixelated preprocessing of original and inability predict information near boundaries with precision. We propose new framework that consists primarily physical-informed neural (PINN) graph convolutional (GCN). GCN serves prediction module predicts computational domain considering mutual influence between unstructured nodes their neighbors. On other hand, PINN acts physical constraint embedding control equation into function network, ensuring inference results comply constraints equation. advantages this lie two aspects. First, computation mechanism is more line actual evolution temperature fields. Second, enhances cognitive ability toward field information. It accurately describes changes gradient at boundary position reduces model's demand data. To validate model, we gradually analyzed geometric adaptability predictive from single cylinder case double case. also investigated impact number sampling points compared those purely data-driven model. show exhibits good stability. With only 20 mean error predicting velocity less than 1% 0.6% cylinder, 2% case, while 9.4% 6.4% These findings demonstrate effectiveness physics-informed allowing accurate fluid flow heat using

Язык: Английский

---

Prediction and optimization of airfoil aerodynamic performance using deep neural network coupled Bayesian method

Physics of Fluids, Год журнала: 2022, Номер 34(11)

Опубликована: Ноя. 1, 2022

In this paper, we proposed an innovative Bayesian optimization (BO) coupled with deep learning for rapid airfoil shape to maximize aerodynamic performance of airfoils. The coefficient prediction model (ACPM) consists a convolutional path and fully connected path, which enables the reconstruction end-to-end mapping between Hicks–Henne (H–H) parameterized geometry coefficients airfoil. computational fluid dynamics (CFD) is first validated data in literature, numerically simulated lift drag were set as ground truth guide training validate network based ACPM. average accuracy predictions are both about 99%, determination R2 more than 0.9970 0.9539, respectively. Coupled ACPM, instead conventional expensive CFD simulator, method improved ratio by 43%, where optimized parameters coincide well results CFD. Furthermore, whole time less 2 min, two orders faster traditional BO-CFD framework. obtained demonstrate great potential BO-ACPM framework fast accurate design.

Язык: Английский

---

Rapid spatio-temporal flood modelling via hydraulics-based graph neural networks

Hydrology and earth system sciences, Год журнала: 2023, Номер 27(23), С. 4227 - 4246

Опубликована: Ноя. 30, 2023

Abstract. Numerical modelling is a reliable tool for flood simulations, but accurate solutions are computationally expensive. In recent years, researchers have explored data-driven methodologies based on neural networks to overcome this limitation. However, most models only used specific case study and disregard the dynamic evolution of wave. This limits their generalizability topographies that model was not trained in time-dependent applications. paper, we introduce shallow water equation–graph network (SWE–GNN), hydraulics-inspired surrogate GNNs can be rapid spatio-temporal modelling. The exploits analogy between finite-volume methods solve SWEs GNNs. For computational mesh, create graph by considering cells as nodes adjacent being connected edges. inputs determined topographical properties domain initial hydraulic conditions. GNN then determines how fluxes exchanged via learned local function. We time-step constraints stacking multiple layers, which expand considered space instead increasing time resolution. also propose multi-step-ahead loss function along with curriculum learning strategy improve stability performance. validate approach using dataset two-dimensional dike breach simulations randomly generated digital elevation high-fidelity numerical solver. SWE–GNN predicts unseen mean average errors 0.04 m depths 0.004 m2 s−1 unit discharges. Moreover, it generalizes well locations, bigger domains, longer periods compared those training set, outperforming other deep-learning models. On top this, has speed-up up 2 orders magnitude faster than Our framework opens doors new replace solvers time-sensitive applications spatially dependent uncertainties.

Язык: Английский

---

Multiscale graph neural network autoencoders for interpretable scientific machine learning

Journal of Computational Physics, Год журнала: 2023, Номер 495, С. 112537 - 112537

Опубликована: Окт. 11, 2023

Язык: Английский

---

Fast performance prediction and field reconstruction of gas turbine using supervised graph learning approaches

Aerospace Science and Technology, Год журнала: 2023, Номер 140, С. 108425 - 108425

Опубликована: Июнь 5, 2023

Язык: Английский

---

Simulating multiphase flow in fractured media with graph neural networks

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

Опубликована: Фев. 1, 2024

Numerical modeling of flow dynamics with multiple fluid phases in subsurface fractured porous media is great significance to numerous geoscience applications. Discrete fracture-matrix (DFM) approaches become popular for simulating reservoirs the last decade. Data-driven surrogate models can provide computationally efficient alternatives high-fidelity numerical simulators. Although convolutional neural networks (CNNs) are effective at approximating space-time solutions multiphase flowing processes, it remains difficult CNNs operate upon DFMs unstructured meshes. To tackle this challenge, we leverage graph (GNNs) an embedded DFM model. The results two-dimensional cases complex fracture systems show that learned surrogates precisely capture effect variations connectivity and forecast dynamic pressure saturation high accuracy. Furthermore, our GNN-based exhibit promising generalizability different geometries numbers fractures not encountered from training dataset.

Язык: Английский

---

Rapid and sparse reconstruction of high-speed steady-state and transient compressible flow fields using physics-informed graph neural networks

Physics of Fluids, Год журнала: 2024, Номер 36(4)

Опубликована: Апрель 1, 2024

Explosion flow fields are characterized by shock waves with varying intensity and position (i.e., explosive loads), which the primary causes of structural damage. Accurate rapid prediction loads is crucial for blast-resistant design daily security management. While existing empirical models numerical simulation methods can capture propagation characteristics waves, high-precision requires a massive computational workload, insufficient to meet fast demands various scenarios. To address this contradiction, study constructed sparse reconstruction model two-dimensional explosion based on machine learning algorithms. The utilizes observational data establish mapping relationship distribution entire field. built physics-informed graph neural network (PIGN). employed associate node features, while physical utilized control convergence, aiming enhance performance. Using dataset, PIGN was tested. Performance generalization capabilities were assessed comparing its results simulation. This evaluation analyzed relative error statistical reconstructed indicate that effectively reconstruct fields, an average in field below 4%. Furthermore, when number probe points reaches 10, close 6%. not only provides highly reliable overpressure pressure-time variations but also, well-trained model, accomplishes within 1 ms. It offers novel approach achieving reasonable or compressible fields.

Язык: Английский

---

Rapid Spatio-Temporal Flood Modelling via Hydraulics-Based Graph Neural Networks

Опубликована: Март 22, 2023

Abstract. Numerical modelling is a reliable tool for flood simulations, but accurate solutions are computationally expensive. In the recent years, researchers have explored data-driven methodologies based on neural networks to overcome this limitation. However, most models used only specific case study and disregard dynamic evolution of wave. This limits their generalizability topographies that model was not trained in time-dependent applications. paper, we introduce SWE-GNN, hydraulics-inspired surrogate Graph Neural Networks (GNN) can be rapid spatio-temporal modelling. The exploits analogy between finite volume methods, solve shallow water equations (SWE), GNNs. For computational mesh, create graph by considering finite-volume cells as nodes adjacent connected edges. inputs determined topographical properties domain initial hydraulic conditions. GNN then determines how fluxes exchanged via learned local function. We time-step constraints stacking multiple layers, which expand considered space instead increasing time resolution. also propose multi-step-ahead loss function along with curriculum learning strategy improve stability performance. validate approach using dataset two-dimensional dike breach simulations randomly-generated digital elevation models, generated highfidelity numerical solver. SWE-GNN predicts unseen mean average error 0.04 m depths 0.004 m2/s unit discharges. Moreover, it generalizes well locations, bigger domains, over longer periods time, outperforming other deep models. On top this, has speedup up two orders magnitude faster than Our framework opens doors new replacing solvers time-sensitive applications spatially-dependant uncertainties.

Язык: Английский

---

Perceiving flow fields and aerodynamic characteristics of turbomachinery via sparse detection data: a graph data mining model

Energy, Год журнала: 2025, Номер unknown, С. 136081 - 136081

Опубликована: Апрель 1, 2025

Язык: Английский

---