PRCT-Net: A Transfer-Learning Physics-Informed Convolutional Transformer for Fast Surrogate Prediction of Confined Dual-Cylinder 3D Free-Surface Two-Phase Flow
Jul 2026· Journal of Fluids Engineering· 0 citations
TL;DR
The results demonstrate that the proposed PRCT-Net substantially enhances computational efficiency while maintaining high prediction accuracy for two-phase flow problems, and supports single-stage training and sustained deployment, enabling robust inference across a wide range of unseen operating conditions.
Abstract
The supporting foundations of offshore engineering structures, such as offshore platforms and sea-crossing bridges, are continuously exposed to wave-induced loads. A thorough understanding of their hydrodynamic characteristics is essential for ensuring operational stability and extending service life. Traveling wave phenomena in such scenarios are representative gas-liquid two-phase flow problems. Conventional finite volume methods for simulating gas-liquid two-phase flows generally impose strict requirements on mesh resolution and time-step size. Accurate solutions become particularly difficult to obtain when the computational domain exhibits complex topological features. Furthermore, changes in boundary conditions or computational domains necessitate repeated simulations, which prevents the efficient reuse and long-term deployment of numerical models.
To address these limitations, this study developed a physics-informed neural network framework that incorporated a Res-CNN-based (residual-connected convolutional neural network) geometric encoder and a transformer attention mechanism into a transfer learning architecture, referred to as PRCT-Net (physics-informed residual-connected convolutional transformer network). This study proposes a computational approach for three-dimensional gas-liquid two-phase flow problems. The effectiveness of the proposed method were validated by simulating traveling-wave flow fields around dual cylinders under multiple operating conditions. The results demonstrate that, compared with conventional finite volume methods, the proposed PRCT-Net substantially enhances computational efficiency while maintaining high prediction accuracy for two-phase flow problems. In addition, the model supports single-stage training and sustained deployment, enabling robust inference across a wide range of unseen operating conditions.
Efficient prediction of dynamic-boundary flow fields in fluid–structure interaction (FSI) remains difficult because of strong nonlinearity, moving geometries, and the high cost of conventional computational fluid dynamics (CFD). To address this, we propose dynamic boundary flow net (DBF-Net), a point-based deep learning approach designed to replace the CFD solver in a Python-based FSI loop for efficient unsteady flow prediction around deforming structures. DBF-Net adopts a hierarchical encoder–decoder architecture for unstructured point clouds: Set Abstraction modules extract multi-scale local geometric and flow features, Feature Propagation layers recover high-resolution fields, and a DynamicFusionUnit models temporal evolution across successive time steps. The trained network is coupled with a Python-based FSI solver to replace the CFD solver in the fluid loop. The method is assessed on three transonic aeroelastic cases: forced oscillation of a two-degree-of-freedom Isogai airfoil, aeroelastic response of the same airfoil at different speed indices, and FSI of a flexible three-dimensional AGARD 445.6 wing. The results show that DBF-Net predicts the unsteady flow and coupled aeroelastic response with good accuracy. In the tested cases, the proposed framework provides about a 67× speedup for the two-dimensional simulations and up to 2.3×103 acceleration for the three-dimensional simulations relative to CFD-based FSI. These results indicate that DBF-Net is a promising surrogate for efficient aeroelastic analysis of systems with dynamic boundaries.
Hongjie Zhou, Tingwei Ji, Changdong Zheng et al.· The Physics of Fluids· 0 citations
We present a conditional latent generative framework for parametric rarefied flows that separates neural-field representation, latent transport, and frozen physics adaptation. Neural-field auto-decoders compress discrete-velocity cavity solutions and direct simulation Monte Carlo cylinder solutions into shared coordinate decoders. Train-only principal-component charts support conditional flow matching (FM) and diffusion without a deterministic condition-to-latent backbone, and structured low-rank adapters correct selected decoder outputs while the upstream pipeline remains frozen. On two steady benchmarks, the frozen pipelines interpolate out-of-sample conditions with cavity kinetic relative $L_1$ errors at the $10^{-5}$ level and cylinder per-field area-weighted RMSEs of 0.038 (density), 0.041 (temperature), and below 0.01 (velocities). For the cavity, physics adaptation reduces the matched-grid Bhatnagar--Gross--Krook diagnostic by 28.65% while preserving field accuracy; for the cylinder, the analytic wall map enforces no-penetration exactly and, jointly with the learned FM adapter, reduces the inlet violation to 0.277 and the global mass-balance ratio to 0.963 of the frozen values with negligible field-error change. A five-seed controlled comparison with deterministic condition-to-chart multilayer perceptrons shows that, although the generative pipelines do not surpass the compact MLP in point accuracy on these single-valued steady problems, the results validate sampling-based conditional transport on the shared representation as an effective steady surrogate, with a natural route to multivalued or stochastic solution families.
Yilun Qi, Guangqin Zhang, Xu Wang et al.· 0 citations
Deep learning surrogates for RANS flow prediction around airfoils face two persistent bottlenecks. A single neural architecture cannot simultaneously resolve sharp near-wall boundary layers and smooth far-field potential flow. Additionally, force prediction is undermined by the numerical instability of computing wall-normal velocity gradients from continuous-field approximations. Both of these points are addressed with a DD-RNO (domain-decomposed routed neural operator), combining a spectral geometry encoder with two physics-guided innovations: (a) a differentiable domain routing mechanism that partitions the flow field into inviscid, boundary-layer, and wake regimes---dispatching query points to specialized regional decoders, and (b) learned canonical quadrature (LCQ), which replaces unstable pressure integration with flow-conditioned, learned integration weights that predict lift and drag directly from surface pressure. On the AirfRANS benchmark, DD-RNO cuts velocity field mean-square error (MSE) by 17$x$ ($u_x$) and 12$x$ ($u_y$) over the strongest baseline, widening to 23$x$ under out-of-distribution Reynolds extrapolation---evidence that the routing mechanism generalizes with the physics it encodes rather than merely fitting the training distribution. LCQ reduces drag MSE by 7.5$x$ relative to conventional pressure integration and raises drag rank correlation from $\rho = 0.250$ to $\rho = 0.997$. Ablations confirm that both components are indispensable to performance: removing domain routing increases velocity error by 8.2$x$, and removing LCQ increases relative drag error more than 40-fold. At ~144 ms per sample---a 10,000$x$ speedup over conventional RANS solvers---DD-RNO offers a surrogate accurate and fast enough for real-time aerodynamic design and optimization loops.
Abstract. In recent years, there has been a growing interest in applying deep learning techniques to Computational Fluid Dynamics (CFD), particularly for accelerating simulations and enhancing data resolution. This work explores the use of Convolutional Neural Networks (CNNs) to reconstruct high-resolution CFD fields from low-resolution inputs. A U-Net architecture was selected due to its proven ability to capture both global context and fine-scale features through its encoder–decoder structure combined with skip connections, facilitating detailed reconstructions. The study focuses on two canonical CFD problems: a bump-in-a-duct configuration and a hydrogen combustion chamber. These cases were chosen to encompass both non-reactive and reactive flow regimes and to test the generalization capability of the model under different physical conditions. The network was trained to predict high-resolution pressure, axial velocity, and heat of reaction fields from their corresponding low-resolution versions, thus performing a super-resolution task. Due to the limited size of the available dataset, several regularization techniques were employed to improve model robustness, including Dropout layers and Batch Normalization. Additionally, different data normalization strategies were tested, such as Z-normalization and Min–Max scaling, to enhance training stability and ensure faster convergence. A multi-network approach was also developed to separately handle the prediction of pressure fields and that of axial velocity and heat release rate, mitigating the risk of feature mixing and allowing each network to specialize in learning specific patterns. The datasets were generated using two solvers: a custom Fortran-based code solving the non-dimensional Euler equations, and the commercial CFD software Ansys Fluent for more complex reactive flow simulations. This combination enabled the creation of a multifidelity database, covering a wide range of fluid dynamic behaviors and ensuring better generalization capabilities during training and validation phases. The training process involved careful hyperparameter optimization, early stopping strategies, and loss function selection to prevent overfitting and maximize prediction quality. Quantitative assessments demonstrated prediction errors on the order of a few percent, while qualitative analyses confirmed the ability of the models to accurately reconstruct the flow structures, even in complex reacting flows. The final phase of the project focused on identifying potential future developments of the proposed framework, particularly exploring the capability of the U-Net architecture to address a broader class of problems. One promising direction involves extending the methodology to correct two-dimensional solutions in order to approximate three-dimensional flow fields. This approach consists of training the network with the 2D planes extracted from the full 3D simulations. Furthermore, techniques such as data augmentation could be implemented to improve the model’s generalization and robustness. The main contribution of this work includes the dataset generation, the design and implementation of the U-Net architecture, the training framework, and the performance evaluation. The results highlight the promising role of CNN-based surrogate models in enhancing CFD simulation capabilities and accelerating future engineering workflows by enabling fast and accurate predictions across different fidelity levels.
Sebastiano Geremia Isidoro Marci· Materials Research Proceedin...· 0 citations
The results demonstrate that PINN achieves more accurate and stable full-field vibration reconstructions than conventional PINNs, particularly under conditions involving high-frequency modes, and highlights the potential of hybrid data-physics neural frameworks as an efficient and reliable approach for solving complex PDE-governed dynamical systems.
Hailong Liu, S. Hedayatrasa, Yunpeng Zhu et al.· e-Journal of Nondestructive...· 0 citations
Neural surrogates are widely used in scientific machine learning for fast prediction of three-dimensional (3D) thermo-fluid fields. However, generating training data using conventional numerical solvers often incurs substantial computational and storage costs. We propose to train an attention graph neural network by minimizing the finite-volume method (FVM) residuals of the governing equations. These residuals are evaluated directly on the mesh, requiring no labeled data. We evaluate the trained surrogates against computational fluid dynamics (CFD) references and a data-supervised baseline across four scenarios. On the two steady-state benchmarks, the FVM-loss model achieves an all-field normalized root-mean-square error (nRMSE) of 2.3-2.8%. It demonstrates close agreement with the CFD references, including the buoyancy-energy coupling. On the two parametric transient cases, the FVM-loss model outperforms the supervised baseline in terms of accuracy, while avoiding the data-generation cost entirely. These results indicate that the FVM loss can provide a practical training signal for neural surrogates and reduce the model development cost.
Tianyu Li, Zhi-Ying Cao, Qingang Zhang et al.· 0 citations