FLFeng Liu

Feng Liu

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Feng Liu has a total of 45 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • FPGA-Based Parallel Compressible Flow Solver Accelerator

    by Songpei Li, Yubo Li, Feng Liu, Liang Zhao

    Field-Programmable Gate Arrays (FPGAs), with their highly parallel computing capabilities and flexible hardware programmability, have emerged as a promising hardware platform for accelerating Computational Fluid Dynamics (CFD) simulations. To address the requirements of numerical solution for compressible flows, this study designs a compressible flow solver accelerator optimized for data-level parallelism and pipelining. The solver integrates custom modules for complex function evaluation to meet the hardware implementation requirements of complex function computations in CFD simulations. Numerical calculations are performed on the one-dimensional convection-diffusion problem and the Sod shock tube problem. The results demonstrate that the designed FPGA acceleration scheme yields accurate and reliable computational results, achieving an order-of-magnitude improvement in computational performance compared with the CPU platform.

  • Open Access

    Transferable reduced order modeling of physical systems via a dilated vector quantized variational autoencoder

    by Yidong Zhang, Hongwei Tang, Yubo Li, Anping Wu, Xing Guo, Feng Liu, Shaoqiu Xiao

  • An anisotropic mesh generation method based on streamline-tracking topology

    by Miaomiao Sun, Fazhi Tang, Yubo Li, Jun Huang, Feng Liu, Wanqiu Jiang, Qingfeng Wang

    Mesh generation is a critical step in the numerical simulation of computational fluid dynamics, with mesh quality directly affecting both computational efficiency and result accuracy. Traditional methods for mesh topology generation heavily rely on human expertise and heuristic algorithms, often lacking a rigorous mathematical foundation and theoretical underpinning. When confronted with complex boundaries, multi-scale features, and dynamic topological structures, these traditional methods frequently encounter issues such as poor mesh adaptability, intersecting mesh lines, localized distortions, degradation of mesh quality, and high costs associated with manual intervention. To address these challenges, this paper proposes an anisotropic mesh generation method based on streamline-tracking topology, aiming at achieving the automatic generation of high-quality three-dimensional mesh topologies. The core concept treats the mesh as a potential field, with the central task being the explicit formulation of this potential field. In this paper, the signed distance field, a scalar field, is obtained via Delaunay triangulation. By computing its gradient, an irrotational vector field (i.e., potential field) is constructed. Mesh topology is generated by tracking the gradient field along streamlines, which then guides the generation of anisotropic meshes. Notably, streamlines inherently do not intersect, effectively eliminating the common problem of mesh line crossings encountered in traditional mesh generation processes. Furthermore, experimental validation is conducted on several representative aerospace and biomedical geometries. The results demonstrate that the proposed method exhibits strong robustness, adaptability, and topological consistency, highlighting its promising potential for automatic, high-quality mesh generation of complex geometries.

  • Frequency-structured cross-fidelity errors in multiscale flow reconstruction

    by Chengcheng Yang, Hongwei Tang, Qingfeng Wang, Feng Liu, Jun Huang

    Coarse-grid, filtered, and downsampled flow fields often retain the dominant large-scale organization of a high-fidelity (HF) solution while losing localized gradients, interfaces, shear layers, corner vortices, and curved-wall details. We, therefore, examine cross-fidelity errors from a fluid-mechanics perspective and formulate HF recovery as a frequency-decoupled, multi-fidelity, physics-constrained reconstruction task. Fourier-enhanced multi-fidelity neural networks (FE-MFNN) combine a shared low-frequency backbone, a Fourier-enhanced cross-fidelity residual branch, adaptive Fourier-basis screening, and progressive physics-constrained training based on low-fidelity (LF) data, boundary information, and residuals of the governing partial differential equation (PDE). Direct LF–HF spectral diagnostics show that the Burgers and Allen–Cahn residuals contain substantially larger high-band energy fractions than their corresponding HF fields under the same 75% cumulative-energy cutoff. Computational fluid dynamics diagnostics reveal weak co-location between cavity velocity-residual energy and HF vorticity but a stronger association between cylinder pressure-residual energy and HF pressure-gradient structures. Across scalar PDE benchmarks, lid-driven cavity flow, and subsonic compressible flow around a cylinder, FE-MFNN improves high-frequency reconstruction in shock-like, interface-dominated, and shear-layer-dominated cases and remains competitive for curved-boundary compressible flow. These results support frequency-decoupled reconstruction as a physically interpretable strategy in the tested multiscale-flow settings. The reported frequency structure is an empirical diagnostic of the archived coarse-resolution and separately exported LF datasets, not a universal physical law.

  • The multi-region conservation-enhanced physics-informed neural network for solving shock waves

    by Yubo Li, Hengli Xiong, Feng Liu

    Supersonic flow is characterized by strong nonlinear and multi-scale physical phenomena, such as shock waves. Despite their success in low-speed simulations, traditional physics-informed neural networks (PINNs) struggle with strong discontinuities in high-speed flows. Lacking effective constraints for local steep gradients, they suffer from reduced accuracy, numerical instability, and a heavy reliance on costly observational data. To address these limitations, this paper proposes a multi-region conservation-enhanced physics-informed neural network (MRCE-PINN), integrated with a pressure-based adaptive gradient guidance (APGG) strategy. This approach obviates the requirement for additional observational data by partitioning the computational domain into discrete sub-regions and embedding local integral conservation constraints of mass, momentum, and energy, thereby providing rigorous physical feedback for network optimization. Concurrently, the APGG strategy is employed to dynamically track high-gradient regions within the flow field to apply targeted constraints. The proposed method is validated under supersonic conditions through three benchmark cases: two-dimensional oblique shock waves, expansion waves, and conical models. Numerical results demonstrate that by strengthening local physical conservation, the method effectively compresses the thickness of the shock wave transition layer and suppresses numerical diffusion near wave fronts and geometric singularities. Comparative evaluations confirm that MRCE-PINN outperforms traditional PINN, conservative physics-informed neural networks, and physical information neural network embedded with multi-scale Fourier features in terms of shock-capturing sharpness, resolution of local flow features, and global prediction stability.

  • Robust reduced-order modeling for shock-dominated compressible flows over a range of parameters

    by Xiaoxin Yin, Hongwei Tang, Jun Huang, Feng Liu

    This paper focuses on the reduced-order modeling of strongly compressible flows over a range of parameters based on the combination of proper orthogonal decomposition (POD) and Kolmogorov–Arnold networks (KAN). Precisely, the present method first utilizes POD to extract the dominant modal features of the flow field. Then, the powerful high-dimensional nonlinear mapping capability of KAN is leveraged to reconstruct the complex dynamic behavior of shock waves. The results of several typical hypersonic flows imply that the proposed reduced-order model (ROM) significantly improves computational efficiency while demonstrating higher prediction accuracy than most existing intelligent ROM methods. More importantly, the present ROM can also effectively predict the flow fields for many unseen flow parameters. These results highlights the generalization performance of the proposed method and its remarkable advantages in predicting complex compressible flows.


  • FluxNet: Accelerating flux calculation via coarse-to-fine neural modeling

    by Hao Jiang, Yubo Li, Yinglong Zheng, Zhigang Chen, Feng Liu, Jun Huang, Tao Zhu

    The finite volume method (FVM) has emerged as a dominant numerical framework in computational fluid dynamics for simulating complex flow phenomena, such as aerospace external aerodynamics and internal combustion engine combustion, owing to its exceptional geometric adaptability and numerical stability. Flux calculation as the core component of FVM, directly dictates the performance of large-scale simulations. However, traditional flux splitting schemes, like AUSMPW+ (advection upstream splitting method with pressure-weighted improvements, a classic upwind scheme), suffer from limited computational efficiency on high-resolution grids due to extensive conditional branching and data dependencies. To address this, we introduce FluxNet, an end-to-end flux calculation accelerator based on a lightweight neural network, designed to overcome the bottlenecks of conventional numerical methods. FluxNet employs a coarse sampling-fine prediction framework, where training data are generated on coarse grids to constrain input parameter ranges and reduce data outliers. To ensure physical consistency, this is coupled with an error-guided progressive iterative sampling method, which suppresses error accumulation through dynamic error feedback during iterative computations, ensuring physical consistency. We conduct numerical experiments on canonical test cases, including both one-dimensional and three-dimensional scenarios. The results demonstrate that FluxNet achieves a nearly fourfold acceleration compared to AUSMPW+ while maintaining an average relative error below 0.615%—and its mean absolute percentage error is less than 1%, satisfying engineering-grade accuracy requirements. This study presents a data-driven paradigm for efficient flux calculation in FVM, with potential for multi-scale modeling to expand its applicability to complex engineering scenarios.


  • Universal heterogeneous parallel optimization for large-scale hypersonic flow simulations on unstructured meshes

    by Ye Zhang, Zhengyu Tian, Hang Yu, Feng Liu, Xiao She

    The development of hypersonic vehicles presents severe challenges to computational fluid dynamics (CFD) simulation efficiency, particularly for unstructured meshes where traditional central processing unit (CPU) architectures lack scalability and graphics processing unit (GPU) implementations require further optimization. This paper constructs a multi-level parallel acceleration framework for a three-dimensional unstructured solver targeting heterogeneous architectures. Profiling reveals the intrinsic constraint of memory access on non-independent parallel kernel functions. We enhance memory efficiency across multiple dimensions, including data layout reconstruction, mesh reordering, and kernel fusion. A decoupled reordering strategy partitioning the domain into inner-halo-padding regions is designed to enable overlap of multi-GPU computation and communication while preserving data locality. Benefiting from the generality of these optimizations, the framework is easily portable to other heterogeneous platforms like deep computing units (DCUs). Tests demonstrate speedups of approximately 1600x on GPU and over 500x on DCU compared to CPU implementations, enabling efficient simulations reaching hundred-million-cell scale with excellent scalability and cross-platform capability. The proposed framework offers a reusable paradigm for optimizing high-performance unstructured CFD software, enhancing hypersonic aerodynamic assessment efficiency.

  • Generalization of parametric and aerodynamic modeling for the overall geometric features of two-stage swept-back lifting body vehicles

    by Zhouchi Liu, Yubo Li, Anping Wu, Hongwei Tang, Tianhang Xiao, Feng Liu

    At the stage of overall shape design for high-speed aircraft, it is challenging to determine how the overall geometric features affect aerodynamic characteristics. Developing a set of geometric characteristic parameters that can effectively represent the aerodynamic characteristics of the aircraft with strong generalization ability is crucial to addressing this issue. Taking the lifting body as the research object, we propose the ratio of the projected area in the Z and Y directions to the theoretical no backward sweeping Z and Y projected area (SZ/SZT, SY/SYT) to characterize the existence of multistage swept-back features among the lifting bodies. In this study, we investigated the effect of parameter SY/SYT on aerodynamic characteristics through a two-dimensional flow model based on shock wave and expansion wave theories. To validate the effectiveness of the proposed parameters, neural networks were employed for aerodynamic modeling of aircraft, with a focused investigation on the effects of parameters SZ/SZT and SY/SYT on network generalization capability and prediction accuracy. A parametric approach was adopted to generate lifting-body configurations, and an aerodynamic database was established through engineering estimation methods. The results demonstrate that the proposed characteristic parameters SZ/SZT and SY/SYT reduce the prediction error by 34% and decrease the training data requirement by approximately one-third while maintaining equivalent error levels. Systematic parameter influence analysis elucidates the mapping relationships between key geometric characteristic parameters of lifting-body configurations and their aerodynamic performance.