Searcharxiv⌕ Search

arXiv subjects

Heyang Wang

Publications and source records attributed to Heyang Wang.

4 recordsLinked to original sources

Newton-PINet: A fast physics-informed neural network with Newton linearization for meta-learning nonlinear PDEs

Scientific machine learning has opened new avenues for solving parameterized partial differential equations (PDEs), enabling models to learn a family of PDEs and generalize to unseen instances. In this context, data-driven operator learning methods typically require large training datasets, while physics-informed neural networks (PINNs) suffer from difficult optimization and limited generalization, especially for nonlinear PDEs. We propose Newton-PINet, a physics-informed network enhanced by Newton linearization, offering an effective meta-learning framework for nonlinear PDEs. Newton-PINet (i) employs a physics-informed multilayer network with skip connections, where the output-layer weights are solved by least squares; (ii) adopts a two-stage learning strategy that first leverages gradient-based training to learn robust representations from the available training tasks, and then performs gradient-free fine-tuning on the output layer for fast task-specific generalization; and (iii) incorporates a Newton linearization method to speed up the least-squares iteration for nonlinear PDE problems. On a challenging nonlinear reaction-diffusion benchmark, Newton-PINet achieves up to three orders of magnitude lower relative error than recent neural solvers, while using 16x fewer training tasks and over an order of magnitude less training time (about 5 minutes versus several hours). This work advances the meta-learning of PINNs toward data-efficient, fast, and generalizable physics solvers. The code and datasets are publicly available at: https://github.com/fycAIwj/Newton-PINet.

cs.LG↗

IR275K: A Benchmark for Infrared Multi-Frame Super-Resolution Toward Efficient Remote Sensing

Infrared multi-frame super-resolution (MFSR) exploits complementary observations to recover spatial detail, but its evaluation remains fragmented across datasets and protocols. A common benchmark is needed to evaluate reconstruction under weak texture, varying thermal contrast, and noise in the source imagery. It should also characterize reconstruction performance across different inter-frame displacement ranges. We introduce IR275K, a curated benchmark containing 594 in- frared video sequences and 275,196 frames. It provides sequence- level train/validation/test splits and a reproducible x4 evaluation protocol. Evaluation spans single-frame and multi-frame recon- struction using seven learned baselines, the CGMamba reference model, and bicubic interpolation. All learned methods are evalu- ated under both bicubic degradation and synthetic mixed sensor noise.EDVRleadstheoverallrankingunderbicubicdegradation, whereas MSDTGP leads under mixed sensor noise. DBSR leads thehigh-displacementUAVsubsetunderbothsettings.Controlled ablations using CGMamba show that real neighboring frames improve PSNR over center-frame repetition by 1.04 dB under bicubic degradation and 0.95 dB under mixed sensor noise. Center-frame conditioning also improves reconstruction under both settings. The benchmark thus reveals how reconstruction rankings and the benefits of neighboring frames depend on evaluation conditions. The complete IR275K dataset is publicly available at: https://github.com/InfraRecon7/IR275K.

cs.CV↗

Bridging Computational Fluid Dynamics Algorithm and Physics-Informed Learning: SIMPLE-PINN for Incompressible Navier-Stokes Equations

Physics-informed neural networks (PINNs) have shown promise for solving partial differential equations (PDEs) by directly embedding them into the loss function. Despite their notable success, existing PINNs often exhibit training instability and slow convergence when applied to strongly nonlinear fluid dynamics problems. To address these challenges, this paper proposes a novel PINN framework, named as SIMPLE-PINN, which incorporates velocity and pressure correction loss terms inspired by the semi-implicit pressure link equation. These correction terms, derived from the momentum and continuity residuals, are tailored for the PINN framework, ensuring velocity-pressure coupling and reinforcing the underlying physical constraints of the Navier-Stokes equations. Through this, the framework can effectively mitigate training instability and accelerate convergence to achieve accurate solution. Furthermore, a hybrid numerical-automatic differentiation strategy is employed to improve the model's generalizability in resolving flows involving complex geometries. The performance of SIMPLE-PINN is evaluated on a range of challenging benchmark cases, including strongly nonlinear flows, long-term flow prediction, and multiphysics coupling problems. The numerical results demonstrate SIMPLE-PINN's high accuracy and rapid convergence. Notably, SIMPLE-PINN achieves, for the first time, a fully data-free solution of lid-driven cavity flow at Re=20000 in just 448s, and successfully captures the onset and long-time evolution of vortex shedding in flow past a cylinder over t=0-100. These findings demonstrate SIMPLE-PINN's potential as a reliable and competitive neural solver for complex PDEs in intelligent scientific computing, with promising engineering applications in aerospace, civil engineering, and mechanical engineering.

cs.CE↗

FFV-PINN: A Fast Physics-Informed Neural Network with Simplified Finite Volume Discretization and Residual Correction

Physics-informed neural networks (PINNs) have emerged as a major research focus. However, today's PINNs encounter several limitations. Firstly, during the construction of the loss function using automatic differentiation, PINNs often neglect information from neighboring points, which hinders their ability to enforce physical constraints and diminishes their accuracy. Furthermore, issues such as instability and poor convergence persist during PINN training, limiting their applicability to complex fluid dynamics problems. To address these challenges, a fast physics-informed neural network framework that integrates a simplified finite volume method (FVM) and residual correction loss term has been proposed, referred to as Fast Finite Volume PINN (FFV-PINN). FFV-PINN utilizes a simplified FVM discretization for the convection term, with an accompanying improvement in the dispersion and dissipation behavior. Unlike traditional FVM, the FFV-PINN outputs can be simply and directly harnessed to approximate values on control surfaces, thereby simplifying the discretization process. Moreover, a residual correction loss term is introduced to significantly accelerates convergence and improves training efficiency. To validate the performance, we solve a series of challenging problems -- including flow in the two-dimensional steady and unsteady lid-driven cavity, three-dimensional steady lid-driven cavity, backward-facing step flows, and natural convection at high Reynolds number and Rayleigh number. Notably, the FFV-PINN can achieve data-free solutions for the lid-driven cavity flow at Re = 10000 and natural convection at Ra = 1e8 for the first time in PINN literature, even while requiring only 680s and 231s. It further highlights the effectiveness of FFV-PINN in improving both speed and accuracy, marking another step forward in the progression of PINNs as competitive neural PDE solvers.

cs.CE↗