Article
Chemistry, Multidisciplinary
Chun-Teh Chen, Grace X. X. Gu
Summary: Elastography is a medical imaging technique that measures tissue elasticity using ultrasound signals. Current methods suffer from low resolution and reliance on material incompressibility assumption. A new physics-informed deep-learning method is proposed, which reconstructs the Young's modulus field based on axial displacement alone, and removes the incompressibility assumption. The method provides accurate elasticity reconstruction and mitigates errors caused by material stiffness.
Article
Engineering, Multidisciplinary
Han Gao, Matthew J. Zahr, Jian-Xun Wang
Summary: Despite the promise of PINNs in solving forward and inverse problems, there are technical challenges that hinder complex and realistic applications. To address these challenges, a discrete PINN framework based on GCN and PDE is proposed to solve PDEs in both forward and inverse settings, offering improved scalability, boundary enforcement, and handling of irregular geometries.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2022)
Article
Computer Science, Interdisciplinary Applications
Qin Lou, Xuhui Meng, George Em Karniadakis
Summary: In this study, physics-informed neural networks (PINNs) are utilized to solve forward and inverse problems via the Boltzmann-BGK formulation, enabling simulations of both continuous and rarefied flows. The PINN-BGK model, consisting of three sub-networks, successfully approximates various benchmark flows and infers flow fields in the entire computational domain through minimizing residuals and mismatches. Results show that PINN-BGK can accurately infer velocity fields in the entire domain and transfer learning can accelerate the training process significantly.
JOURNAL OF COMPUTATIONAL PHYSICS
(2021)
Article
Computer Science, Interdisciplinary Applications
Guang Lin, Yating Wang, Zecheng Zhang
Summary: In this study, a multi-variance replica exchange stochastic gradient Langevin dynamics method is proposed to address the challenges of multiple local optima in optimization and multiple modal posterior distribution in inverse problems. The method reduces computational cost by using solvers with different fidelities and achieves faster and more accurate convergence.
JOURNAL OF COMPUTATIONAL PHYSICS
(2022)
Article
Engineering, Multidisciplinary
Mingyuan Yang, John T. Foster
Summary: In this paper, a multi-output physics-informed neural network (MO-PINN) is proposed to provide solutions with uncertainty distributions for both forward and inverse PDE problems with noisy data. It is demonstrated that MO-PINN can provide accurate predictions and its solutions are consistent with traditional methods. Additional statistical knowledge can be incorporated to improve the prediction if available.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2022)
Article
Computer Science, Interdisciplinary Applications
Lei Yuan, Yi-Qing Ni, Xiang-Yun Deng, Shuo Hao
Summary: Physics informed neural networks (PINNs) are a novel deep learning paradigm for solving forward and inverse problems of nonlinear partial differential equations (PDEs). This study proposes an auxiliary PINN (A-PINN) framework that bypasses the limitation of integral discretization by defining auxiliary output variables and using automatic differentiation. The A-PINN demonstrates higher accuracy compared to traditional PINN and shows good performance in solving various nonlinear IDE problems.
JOURNAL OF COMPUTATIONAL PHYSICS
(2022)
Article
Mechanics
Xiaojian Li, Yuhao Liu, Zhengxian Liu
Summary: Physics-Informed Neural Network (PINN) is a promising technique for solving partial differential equations (PDEs) in flow problems. An adaptive gradient descent algorithm (AGDA) is proposed based on the analysis of interaction mechanisms, which is validated through analytical PDEs and flow problems. The AGDA optimizer for PINN training achieves higher efficiency and robustness compared to the Adam optimizer and PCGrad optimizer, reducing training time and iteration numbers while maintaining specified accuracy.
Article
Computer Science, Interdisciplinary Applications
Jing Li, Alexandre M. Tartakovsky
Summary: We propose a PI-CKL-NN method for parameter estimation in differential equation (DE) models with sparse measurements. The method approximates space- or time-dependent parameters using KL expansions conditioned on parameter measurements and approximates states using deep neural networks (DNNs). Unknown weights in the KL expansions and DNNs are found by minimizing a cost function that enforces state measurements and DE constraints. Regularization is achieved by adding the l2 norm of conditional KL coefficients to the loss function. Compared to other physics-informed machine learning methods, our approach enforces statistical knowledge in addition to DE constraints and data for parameter estimation. Experimental results show that our method is more accurate than the PINN method, especially when parameter observations are very sparse.
JOURNAL OF COMPUTATIONAL PHYSICS
(2022)
Article
Engineering, Multidisciplinary
Jeremy Yu, Lu Lu, Xuhui Meng, George Em Karniadakis
Summary: The article introduces a new method called gradient-enhanced physics-informed neural networks (gPINNs) to improve the accuracy of PINNs. gPINNs leverage the gradient information of the PDE residual to enhance the loss function. Experimental results demonstrate that gPINNs outperform traditional PINN with fewer training points.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2022)
Article
Engineering, Electrical & Electronic
Tao Shan, Zhichao Lin, Xiaoqian Song, Maokun Li, Fan Yang, Shenheng Xu
Summary: In this paper, a new physics-constrained approach is proposed to solve 2-D inverse scattering problems (ISPs) by extending physics-informed supervised residual learning (PhiSRL) with Born approximation (BA). PhiSRL aims to solve ISPs iteratively by applying the convolutional neural networks (CNNs) to learn the update rules of reconstructions. PhiSRL is employed to invert lossy scatterers by introducing BA to linearize ISPs and reduce the computational burden of forward modeling. Numerical and experimental results validate the effectiveness of the proposed approach.
IEEE TRANSACTIONS ON ANTENNAS AND PROPAGATION
(2023)
Article
Mathematics
Meijun Zhou, Gang Mei, Nengxiong Xu
Summary: In this paper, a novel approach that couples the smoothed finite element method (S-FEM) and physics-informed neural networks (PINNs) is proposed. The approach utilizes S-FEM to generate high-fidelity datasets for PINN inversion, thus improving the accuracy of PINN in solving forward problems. Computational results show that the proposed coupling of S-FEM and PINN exhibits high precision and convergence in solving both forward and inverse problems.
Article
Mechanics
Ruizhi Zhai, Deshun Yin, Guofei Pang
Summary: This article leverages deep learning approaches to solve forward and inverse problems of two-dimensional laminar flows for power-law fluids. The proposed Power-Law-Fluid-Net (PL-Net) framework includes a surrogate model for the forward problems and utilizes a small set of measurement data for the inverse problems. The incorporation of hard boundary conditions and constitutive parameters improves the accuracy of the numerical solutions.
Article
Optics
Siddhartha Mishra, Roberto Molinaro
Summary: The study introduces a machine learning algorithm based on physics informed neural networks for simulating and solving radiative transfer problems, demonstrating superior accuracy and efficiency through empirical experiments.
JOURNAL OF QUANTITATIVE SPECTROSCOPY & RADIATIVE TRANSFER
(2021)
Article
Computer Science, Artificial Intelligence
Xu Liu, Wen Yao, Wei Peng, Weien Zhou
Summary: In this paper, a novel Bayesian physics-informed extreme learning machine (BPIELM) is proposed to solve forward and inverse linear partial differential equation (PDE) problems with noisy data. BPIELM not only quantifies uncertainty arising from noisy data and provides more accurate predictions, but also has a considerably lower computational cost compared to PINN.
Article
Engineering, Multidisciplinary
Chen Xu, Ba Trung Cao, Yong Yuan, Guenther Meschke
Summary: Recently, the physics-informed neural networks (PINNs), a class of machine learning methods, have been widely used in solving scientific computing problems by embedding physical laws into the loss function. This paper presents a multi-task learning method that utilizes uncertainty weighting to improve the efficiency and accuracy of PINNs for inverse problems in linear elasticity and hyperelasticity. The approach is demonstrated through the prediction of external loads in engineering structures based on limited displacement monitoring points, and shows robustness and better performance compared to traditional analysis methods.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2023)
Article
Engineering, Geological
Yared W. Bekele, Trond Kvamsdal, Arne M. Kvarving, Steinar Nordal
INTERNATIONAL JOURNAL FOR NUMERICAL AND ANALYTICAL METHODS IN GEOMECHANICS
(2016)
Article
Computer Science, Interdisciplinary Applications
Yared W. Bekele, Hiroyuki Kyokawa, Arne M. Kvarving, Trond Kvamsdal, Steinar Nordal
COMPUTERS AND GEOTECHNICS
(2017)
Proceedings Paper
Engineering, Geological
H. Kyokawa, Y. W. Bekele
Computer Methods and Recent Advances in Geomechanics
(2015)