Flow3DNet: A deep learning framework for efficient simulation of three-dimensional wing flow fields

Aerospace Science and Technology, Journal Year: 2025, Volume and Issue: unknown, P. 109991 - 109991

Published: Jan. 1, 2025

Language: Английский

---

Fast flow field prediction of three-dimensional hypersonic vehicles using an improved Gaussian process regression algorithm

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(1)

Published: Jan. 1, 2024

Conducting large-scale numerical computations to obtain flow field during the hypersonic vehicle engineering design phase can be excessively costly. Although deep learning algorithms enable rapid prediction with high-precision, they require a significant investment in training samples, contradicting motivation of reducing cost acquiring field. The combination feature extraction and regression also achieve high-precision fields, which is more suitable tackle three-dimensional small dataset. In this study, we propose reduced-order model (ROM) for utilizing proper orthogonal decomposition extract representative features Gaussian process improved automatic kernel construction (AKC-GPR) perform nonlinear mapping physical prediction. selection variables based on sensitivity analysis modal assurance criterion. underlying relationship unveiled between inflow conditions. ROM exhibits high predictive accuracy, mean absolute percentage error (MAPE) total less than 3.5%, when varying altitudes Mach numbers. During angle attack variations, only effectively reconstructs distribution by interpolation MAPE 7.02%. excellent small-sample fitting capability our AKC-GPR algorithm demonstrated comparing original AKC-GPRs maximum reduction 35.28%. These promising findings suggest that proposed serve as an effective approach accurate predicting, enabling its application analysis.

Language: Английский

---

Deep learning for airfoil aerodynamic-electromagnetic coupling optimization with random forest

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(1)

Published: Jan. 1, 2024

Reducing the design variable space is crucial in multi-objective airfoil profile optimization to improve efficiency and reduce computational costs. Based on random forest deep neural networks (DNNs), this work performs range reduction ten variables obtained through a fourth-order class shape transformation parameterization method for subsonic profiles. Three aerodynamic performance objectives (lift coefficient, drag lift-to-drag ratio) are evaluated using Reynolds-averaged Navier–Stokes equations, two radar stealth (horizontal vertical polarization cross sections) assessed of moments. By combining DNN architecture with an improved regression prediction capability, predictive models trained mapping objectives. The errors below 3% 1% particle swarm algorithm provides optimized profiles three scenarios. First higher lift coefficient lower section. Second Third coefficient. Increasing curvature reducing maximum thickness improves by 386 counts reduces 17 counts. curving leading edge, section transverse electric magnetic polarizations decreased 2.78 2.09 dBsm, respectively.

Language: Английский

---

Shock wave prediction in transonic flow fields using domain-informed probabilistic deep learning

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(1)

Published: Jan. 1, 2024

Transonic flow fields are marked by shock waves of varying strength and location crucial for the aerodynamic design optimization high-speed transport aircraft. While deep learning methods offer potential predicting these fields, their deterministic outputs often lack predictive uncertainty. Moreover, accuracy, especially near critical regions, needs better quantification. In this paper, we introduce a domain-informed probabilistic (DIP) framework tailored transonic with called DIP-ShockNet. This methodology utilizes Monte Carlo dropout to estimate uncertainty enhances flow-field predictions wall region employing inverse distance function-based input representation field. The obtained results benchmarked against signed function geometric mask representations. proposed further improves prediction accuracy in wave areas using loss function. To quantify our predictions, developed metrics assess errors location, achieving 6.4% 1%, respectively. Assessing generalizability method, tested it on different training sample sizes compared proper orthogonal decomposition (POD)-based reduced-order model (ROM). Our indicate that DIP-ShockNet outperforms POD-ROM 60% complete

Language: Английский

---

A comprehensive deep learning geometric shape optimization framework with field prediction surrogate and reinforcement learning

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(4)

Published: April 1, 2024

The optimization of aerodynamic components' geometric shapes demands a novel technical approach for adaptive and efficient exploration decision-making within the design space. In this study, we introduce an innovative shape framework that leverages deep reinforcement learning with neural network surrogate models. field prediction surrogate, realized by two distinct U-net architectures, can efficiently generate holistic solutions based on transformed mesh coordinates. Subsequently, inference engine dynamically calculates key metric flow fields, serving as objective function subsequent geometry-aware Deep Q (DQN)-based optimization. framework's efficacy is validated using rocket nozzle illustrative example. During validation, under both friction frictionless conditions, l1 errors entire vision transformer (ViT) convolutional (CNN) architectures are less than 0.4%. proposed ViT consistently outperforms CNN, superiority particularly evident in complex areas, outlet sections, vacuum thrust prediction. Following training, DQN model employed to explore variable B-spline defining profile successfully optimized final expanding segment improved thrust. Under it closely approaches theoretical optimum. practical condition considering friction, gains 2.96% improvement. results demonstrate framework, especially when coupled ViT, exhibits enhanced accuracy adaptability tasks.

Language: Английский

---

A boundary-assimilation Fourier neural operator for predicting initial fields of flow around structures

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(2)

Published: Feb. 1, 2025

Repeatedly solving flow around structures with varying parameters using computational fluid dynamics (CFD) is often essential for structural design. This study proposes a boundary-assimilation Fourier neural operator (BAFNO) method to address the challenges of manually setting initial conditions CFD. The focus BAFNO on generalization ability predict fields without relying observational data. addresses boundary constraint requirements existing physics-informed models in parametric geometries. Inspired by ghost node method, domain are assimilated into loss function instead adding penalty terms. Meanwhile, structure boundaries damping source term level set function. can flexibly handle geometries different shapes and quantities. Subsequently, series numerical experiments flow-around conducted confirm performance BAFNO. results indicate that has strong capability, + CFD obtain dynamic stable faster than direct

Language: Английский

---

Wake field prediction and optimization of submarine with fillet based on data-driven deep learning model

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(3)

Published: March 1, 2025

The fillet on the submarine is a rounded structure designed based body-stern appendages, which effectively weakens horseshoe vortex at junction between rudder and hull, thereby improving propeller's inflow quality. To investigate impact of stern shape wake flow, this research develops data-driven steady field prediction model for submarines U-Net architecture. By comparing computational fluid dynamics (CFD) simulation results with model, it demonstrated that efficiency flow significantly improved, accuracy can be maintained simultaneously. Furthermore, effects fillets different radii are analyzed optimal parameters identified. Compared to original optimized design reduces velocity non-uniformity by 20%.

Language: Английский

---

Nonlinear reduced-order modeling of compressible flow fields using deep learning and manifold learning

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(3)

Published: March 1, 2025

This paper presents a nonlinear reduced-order modeling (ROM) framework that leverages deep learning and manifold to predict compressible flow fields with complex features, including shock waves. The proposed DeepManifold (DM)-ROM methodology is computationally efficient, avoids pixelation or interpolation of field data, adaptable various grids geometries. consists four main steps: First, convolutional neural network-based parameterization network extracts shape modes directly from aerodynamic Next, applied reduce the dimensionality high-fidelity output fields. A multilayer perceptron-based regression then trained map input modes. Finally, back-mapping process reconstructs full predicted low-dimensional DM-ROM rigorously tested on transonic RAE2822 airfoil test case, which includes waves varying strengths locations. Metrics are introduced quantify model's accuracy in predicting wave strength location. results demonstrate achieves prediction error approximately 3.5% significantly outperforms reference ROM techniques, such as proper orthogonal decomposition (POD)-ROM isometric mapping (ISOMAP)-ROM, for training sample sizes.

Language: Английский

---

Enhancing airfoil design optimization surrogate models using multi-task learning: Separating airfoil surface and fluid domain predictions

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(3)

Published: March 1, 2025

Computational fluid dynamics is essential for airfoil design optimization. Typically, it involves numerous numerical procedures such as grid generation, boundary condition setup, and simulations, leading to high computational costs extended research periods, which pose a long-standing challenge aerodynamic development. Recently, the data-driven deep learning method has emerged new approach, significantly reducing time. However, these models have difficulties maintaining desired accuracy, particularly when balancing surface characteristics with internal volume features. In this study, we introduce novel utilizing multi-task (MTL) handle predictions interconnected yet distinct tasks. By employing multi-head neural network architectures advanced MTL optimization strategies, our approach effectively resolves inherent conflicts between domain predictions. Our demonstrates significant improvement in predictive accuracy of both flow fields force coefficients. Extensive experiments were conducted using an open-source dataset that includes field data various shapes under different flight conditions. The results indicate MTL-based surrogate model outperforms existing models, providing more reliable efficient tools practical applications engineering.

Language: Английский

---

Prediction of Subsonic Cascade Flow Fields by Physics-informed Graph Neural Networks with Unbalanced Data

Aerospace Science and Technology, Journal Year: 2025, Volume and Issue: unknown, P. 110173 - 110173

Published: March 1, 2025

Language: Английский

---