Article
Mathematics
Peng Zhi, Yuching Wu, Cheng Qi, Tao Zhu, Xiao Wu, Hongyu Wu
Summary: The study investigates the role of deep learning in computational mechanics. A convolutional neural network technique based on modified loss function is proposed as a surrogate of the finite element method. Physics-informed neural networks are developed to solve boundary value problems based on elliptic partial differential equations. The proposed method is applied for the first time to solve boundary value problems with elliptic partial differential equations. The results show that the proposed surrogate-based approach is in good agreement with the conventional FEM and modification of the loss function improves the prediction accuracy of the neural network. The study demonstrates the potential of the deep learning approach as a significant surrogate model.
Article
Mathematics, Applied
Siping Tang, Xinlong Feng, Wei Wu, Hui Xu
Summary: In this paper, the authors propose a simplified neural network called polynomial interpolation physics-informed neural networks (PI-PINN) to solve nonlinear partial differential equations. By utilizing orthogonal polynomials to construct the neural network, the PI-PINN structure is shown to be effective in solving these equations. Numerical experiments and investigations on reverse problems demonstrate the accuracy and efficiency of the proposed approach.
COMPUTERS & MATHEMATICS WITH APPLICATIONS
(2023)
Article
Computer Science, Interdisciplinary Applications
Jared O'Leary, Joel A. Paulson, Ali Mesbah
Summary: This study proposes a framework for training artificial neural networks to learn the hidden physics within stochastic differential equations (SDEs). The framework propagates stochasticity through the known structure of the SDE and utilizes automatic differentiation and mini-batch gradient descent to establish the parameters of the neural networks. The results demonstrate the potential of this method in unraveling the hidden physics of multivariate stochastic dynamical systems.
JOURNAL OF COMPUTATIONAL PHYSICS
(2022)
Article
Engineering, Multidisciplinary
Zeng Meng, Qiaochu Qian, Mengqiang Xu, Bo Yu, Ali Riza Yildiz, Seyedali Mirjalili
Summary: In this study, the physics-informed neural network (PINN) is used as a black-box solution tool for complex limit state functions (LSFs) expressed as implicit partial differential equations (PDEs) in structural reliability analysis. The proposed PINN-FORM method combines PINN with the first-order reliability method (FORM) and achieves high accuracy in predicting both the solutions of PDEs and the reliability index within a single training process.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2023)
Article
Engineering, Multidisciplinary
Jinshuai Bai, Gui-Rong Liu, Ashish Gupta, Laith Alzubaidi, Xi-Qiao Feng, YuanTong Gu
Summary: Our study reveals that physics-informed neural networks (PINN) are often local approximators after training. This led to the development of a novel physics-informed radial basis network (PIRBN), which maintains the local approximating property throughout the training process. Unlike deep neural networks, PIRBN consists of only one hidden layer and a radial basis activation function. Under appropriate conditions, we demonstrated that PIRBNs can converge to Gaussian processes using gradient descent methods. Furthermore, we investigated the training dynamics of PIRBN using the neural tangent kernel (NTK) theory and explored various initialization strategies. Numerical examples showed that PIRBN is more effective than PINN in solving nonlinear partial differential equations with high-frequency features and ill-posed computational domains. Moreover, existing PINN numerical techniques such as adaptive learning, decomposition, and different loss functions can be applied to PIRBN. The reproducible code for all numerical results is available at https://github.com/JinshuaiBai/PIRBN.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2023)
Article
Operations Research & Management Science
Carlos J. Garcia-Cervera, Mathieu Kessler, Francisco Periago
Summary: This paper introduces the use of physics-informed neural networks to solve controllability problems in partial differential equations (PDEs). It derives error estimates for the generalization error of state and control based on classical observability inequalities and energy estimates for the PDE. These error bounds are applicable to any exact controllable linear PDE system in any dimension, providing a rigorous justification for the use of neural networks in this field. Preliminary numerical simulations of three different types of PDEs are conducted to demonstrate the performance of the proposed methodology.
JOURNAL OF OPTIMIZATION THEORY AND APPLICATIONS
(2023)
Article
Engineering, Multidisciplinary
Yifei Zong, QiZhi He, Alexandre M. Tartakovsky
Summary: We propose a multi-component approach to improve the training of the physics-informed neural network (PINN) model for parabolic problems with sharply perturbed initial conditions. By considering the advection-dispersion equation (ADE) with a point (Gaussian) source initial condition, we find that the decay rate of perturbations in the initial condition can lead to large approximation errors in the PINN solution for d & GE; 2. In addition, localized large gradients in the ADE solution make the Latin hypercube sampling of the equation's residual, which is common in PINN, highly inefficient. We also demonstrate the sensitivity of the PINN solution of parabolic equations to the choice of weights in the loss function.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2023)
Article
Engineering, Multidisciplinary
David Dalton, Dirk Husmeier, Hao Gao
Summary: Modern computational soft-tissue mechanics models have the potential to provide unique patient-specific diagnostic insights, but their deployment in clinical settings has been limited by the high computational costs of conventional numerical solvers. In this study, we propose an emulation framework for soft-tissue mechanics using a Graph Neural Network (GNN) and physics-informed training, which allows for highly accurate and efficient predictions.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2023)
Article
Computer Science, Interdisciplinary Applications
Zhi-Yong Zhang, Hui Zhang, Li-Sheng Zhang, Lei -Lei Guo
Summary: In this work, a new method called symmetry-enhanced physics informed neural network (SPINN) is introduced to improve the accuracy of PINN for solving PDEs. The effectiveness of SPINN is tested through numerical experiments on various PDEs, and the results show that SPINN outperforms PINN in terms of accuracy with fewer training points and simpler neural network architecture.
JOURNAL OF COMPUTATIONAL PHYSICS
(2023)
Article
Engineering, Multidisciplinary
Weiheng Zhong, Hadi Meidani
Summary: This paper proposes a new class of physics-informed neural networks, called the Physics-Informed Variational Auto-Encoder (PI-VAE), for solving stochastic differential equations (SDEs) or inverse problems involving SDEs. The PI-VAE integrates a variational autoencoder (VAE) with the governing equations, using automatic differentiation to calculate derivatives and the Maximum Mean Discrepancy (MMD) as the loss function. Numerical experiments demonstrate the satisfactory accuracy and efficiency of the proposed method in approximating stochastic processes and solving forward, inverse, and mixed problems related to SDEs.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2023)
Article
Engineering, Multidisciplinary
Diab W. Abueidda, Qiyue Lu, Seid Koric
Summary: Deep learning and the collocation method are merged to solve partial differential equations describing structures' deformation, offering a meshfree approach that avoids spatial discretization and data generation bottlenecks.
INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING
(2021)
Article
Computer Science, Artificial Intelligence
Onur Bilgin, Thomas Vergutz, Siamak Mehrkanoon
Summary: This paper introduces a novel two-stream deep model based on the graph convolutional network (GCN) architecture and feed-forward neural networks (FFNN) for learning solutions to nonlinear partial differential equations (PDEs). The model incorporates both graph and grid input representations using two streams corresponding to GCN and FFNN models, respectively. The model is trained in two phases, where the parameters of each stream are trained separately and then the learned representation solutions are fed to fully connected layers for further learning.
Article
Mathematics, Applied
Natalia P. Bondarenko
Summary: This study focuses on the Sturm-Liouville operator with singular potentials of class W2-1 on a graph of arbitrary geometrical structure. It addresses the partial inverse problem of recovering the potential on a boundary edge from a subspectrum with known potentials on other edges. The study provides a uniqueness theorem, a reconstruction algorithm, global solvability, as well as local solvability and stability for this inverse problem.
MATHEMATICAL METHODS IN THE APPLIED SCIENCES
(2021)
Article
Computer Science, Artificial Intelligence
Zixue Xiang, Wei Peng, Xu Liu, Wen Yao
Summary: The paper introduces a method of defining the loss function through adaptive weights and demonstrates that the self-adaptive loss balanced physics-informed neural networks (lbPINNs) outperform PINNs in solving partial differential equations.
Article
Mathematics, Applied
Lu Lu, Raphael Pestourie, Wenjie Yao, Zhicheng Wang, Francesc Verdugo, Steven G. Johnson
Summary: Inverse design, such as topology optimization, is widely used in engineering for achieving targeted properties by optimizing designed geometries. The proposed physics-informed neural networks with hard constraints (hPINNs) can effectively solve topology optimization problems without the need for a large dataset, demonstrating smoother design outcomes compared to conventional methods.
SIAM JOURNAL ON SCIENTIFIC COMPUTING
(2021)
Article
Engineering, Multidisciplinary
Masayuki Yano, Tianci Huang, Matthew J. Zahr
Summary: A globally convergent method using projection-based reduced-order models (ROMs) and trust-region methods is proposed to accelerate density-based topology optimization. By replacing large-scale finite element simulations with ROMs, the computational cost of objective function and gradient evaluations is significantly reduced. The trust-region method is introduced to ensure convergence to critical points of the original topology optimization problem, leading to accelerated convergence to optimal designs by up to an order of magnitude.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2021)
Article
Engineering, Multidisciplinary
Alexander Schein, Kevin T. Carlberg, Matthew J. Zahr
Summary: Model-reduction techniques aim to reduce computational complexity by enforcing dynamics in a low-dimensional subspace, but often result in reduced-order models violating intrinsic physical properties. Ensuring that reduced-order models preserve these properties can improve accuracy and stability in many applications.
INTERNATIONAL JOURNAL FOR NUMERICAL METHODS IN ENGINEERING
(2021)
Article
Multidisciplinary Sciences
Xin-Yang Liu, Jian-Xun Wang
Summary: Model-based reinforcement learning (MBRL) aims to improve sample efficiency by learning a predictive model of the environment, but the quality of the learned model is crucial for its performance. This study proposes a physics-informed MBRL framework that leverages prior knowledge of the environment's underlying physics to enhance the quality of the learned model and reduce interactions with the environment.
PROCEEDINGS OF THE ROYAL SOCIETY A-MATHEMATICAL PHYSICAL AND ENGINEERING SCIENCES
(2021)
Article
Chemistry, Physical
Ruiyang Li, Jian-Xun Wang, Eungkyu Lee, Tengfei Luo
Summary: This study introduces a data-free deep learning scheme, physics-informed neural network (PINN), for solving the phonon Boltzmann transport equation (BTE) with arbitrary temperature gradients. Numerical experiments suggest that the proposed PINN can accurately predict phonon transport under arbitrary temperature gradients and shows great promise for thermal design.
NPJ COMPUTATIONAL MATERIALS
(2022)
Article
Computer Science, Interdisciplinary Applications
Tianci Huang, Matthew J. Zahr
Summary: High-order implicit shock tracking is a new numerical method that accurately approximates solutions of conservation laws with non-smooth features. By optimizing the mesh and high-order approximation, this method can reliably simulate complex, high-speed, compressible flows.
JOURNAL OF COMPUTATIONAL PHYSICS
(2022)
Article
Computer Science, Interdisciplinary Applications
Charles J. Naudet, Johannes Toger, Matthew J. Zahr
Summary: This study investigates the effectiveness of Simulation-based Imaging (SBI) in predicting wall shear stress (WSS) compared to standard Magnetic Resonance Imaging (MRI) postprocessing techniques. The results show that SBI is not sensitive to the Reynolds number, improves with increased MRI data, and provides accurate reconstructions with minimal MRI voxels. However, the reconstruction quality decreases linearly with increased noise in the data. The study also reveals the flexibility of the CFD mesh used in SBI, although the reconstruction quality becomes more dependent on other parameters, particularly the resolution of the MRI data, for coarser meshes.
ENGINEERING WITH COMPUTERS
(2022)
Article
Mechanics
Pan Du, Xiaozhi Zhu, Jian-Xun Wang
Summary: Optimization and uncertainty quantification are increasingly important in computational hemodynamics. However, existing methods face challenges when dealing with complex 3D patient-specific shapes. This study proposes a deep learning surrogate modeling solution to enable rapid hemodynamic predictions.
Article
Computer Science, Interdisciplinary Applications
Marzieh Alireza Mirhoseini, Matthew J. Zahr
Summary: This work introduces a new approach called model reduction with implicit feature tracking to reduce the computational cost of solving convection-dominated partial differential equations (PDEs). The approach utilizes a nonlinear approximation manifold composed of a low-dimensional affine space and a space of bijections to overcome the limitations of traditional model reduction techniques. By minimizing the residual of the unreduced PDE discretization over the reduced nonlinear manifold, the proposed method achieves accurate approximations to convection-dominated problems with limited training.
JOURNAL OF COMPUTATIONAL PHYSICS
(2023)
Article
Engineering, Multidisciplinary
Deepak Akhare, Tengfei Luo, Jian-Xun Wang
Summary: Various manufacturing technologies are being developed to improve the manufacturing of composites owing to their low weight and high performance. Traditional first-principle modeling approaches are not accessible due to the complex physics involved. A physics-integrated neural differentiable (PiNDiff) model is developed to tackle the challenges of modeling the curing process of thick thermoset composite laminates with partially known physics.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2023)
Article
Biology
Mohammadreza Movahhedi, Xin-Yang Liu, Biao Geng, Coen Elemans, Qian Xue, Jian-Xun Wang, Xudong Zheng
Summary: A hybrid physics-informed neural network algorithm is proposed to infer 3D flow-induced tissue dynamics from sparse 2D images. The algorithm combines a differentiable fluid solver with a recurrent neural network model of soft tissue. The effectiveness of the algorithm is demonstrated on synthetic and experimental data, successfully reconstructing 3D vocal dynamics, aerodynamics, and acoustics from sparse 2D vibration profiles.
COMMUNICATIONS BIOLOGY
(2023)
Article
Computer Science, Interdisciplinary Applications
Tianshu Wen, Matthew J. Zahr
Summary: We propose a numerical method to efficiently solve large-scale optimization problems governed by nonlinear systems of equations, including discretized partial differential equations, using projection-based reduced-order models. The method constructs a hyperreduced model on-the-fly during the optimization process, avoiding the need for offline training and ensuring all snapshot information is collected. Numerical experiments demonstrate the global convergence and computational efficiency of the method, with significant speedups compared to standard optimization approaches.
JOURNAL OF COMPUTATIONAL PHYSICS
(2023)
Article
Computer Science, Interdisciplinary Applications
Tianci Huang, Charles J. Naudet, Matthew J. Zahr
Summary: High-order implicit shock tracking is a numerical method that approximates solutions of conservation laws by aligning computational mesh elements with non-smooth features. This paper introduces a robust framework for parameter-dependent lead shocks and demonstrates its effectiveness through various benchmark problems.
JOURNAL OF COMPUTATIONAL PHYSICS
(2023)
Article
Mechanics
Dariush Bodaghi, Jian-Xun Wang, Qian Xue, Xudong Zheng
Summary: In this study, the effects of antagonistic muscle actuation on fish fin ray propulsion were investigated using a computational flow-structure interaction (FSI) model. The results show that the phase of muscle actuation is a critical factor affecting its effects, and different phases result in different thrust and efficiency.
JOURNAL OF FLUID MECHANICS
(2023)
Article
Engineering, Multidisciplinary
Akshay J. Thomas, Mateusz Jaszczuk, Eduardo Barocio, Gourab Ghosh, Ilias Bilionis, R. Byron Pipes
Summary: We propose a physics-guided transfer learning approach to predict the thermal conductivity of additively manufactured short-fiber reinforced polymers using micro-structural characteristics obtained from tensile tests. A Bayesian framework is developed to transfer the thermal conductivity properties across different extrusion deposition additive manufacturing systems. The experimental results demonstrate the effectiveness and reliability of our method in accounting for epistemic and aleatory uncertainties.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Zhen Zhang, Zongren Zou, Ellen Kuhl, George Em Karniadakis
Summary: In this study, deep learning and artificial intelligence were used to discover a mathematical model for the progression of Alzheimer's disease. By analyzing longitudinal tau positron emission tomography data, a reaction-diffusion type partial differential equation for tau protein misfolding and spreading was discovered. The results showed different misfolding models for Alzheimer's and healthy control groups, indicating faster misfolding in Alzheimer's group. The study provides a foundation for early diagnosis and treatment of Alzheimer's disease and other misfolding-protein based neurodegenerative disorders using image-based technologies.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Jonghyuk Baek, Jiun-Shyan Chen
Summary: This paper introduces an improved neural network-enhanced reproducing kernel particle method for modeling the localization of brittle fractures. By adding a neural network approximation to the background reproducing kernel approximation, the method allows for the automatic location and insertion of discontinuities in the function space, enhancing the modeling effectiveness. The proposed method uses an energy-based loss function for optimization and regularizes the approximation results through constraints on the spatial gradient of the parametric coordinates, ensuring convergence.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Bodhinanda Chandra, Ryota Hashimoto, Shinnosuke Matsumi, Ken Kamrin, Kenichi Soga
Summary: This paper proposes new and robust stabilization strategies for accurately modeling incompressible fluid flow problems in the material point method (MPM). The proposed approach adopts a monolithic displacement-pressure formulation and integrates two stabilization strategies to ensure stability. The effectiveness of the proposed method is validated through benchmark cases and real-world scenarios involving violent free-surface fluid motion.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Chao Peng, Alessandro Tasora, Dario Fusai, Dario Mangoni
Summary: This article discusses the importance of the tangent stiffness matrix of constraints in multibody systems and provides a general formulation based on quaternion parametrization. The article also presents the analytical expression of the tangent stiffness matrix derived through linearization. Examples demonstrate the positive effect of this additional stiffness term on static and eigenvalue analyses.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Thibaut Vadcard, Fabrice Thouverez, Alain Batailly
Summary: This contribution presents a methodology for detecting isolated branches of periodic solutions to nonlinear mechanical equations. The method combines harmonic balance method-based solving procedure with the Melnikov energy principle. It is able to predict the location of isolated branches of solutions near families of autonomous periodic solutions. The relevance and accuracy of this methodology are demonstrated through academic and industrial applications.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Weisheng Zhang, Yue Wang, Sung-Kie Youn, Xu Guo
Summary: This study proposes a sketch-guided topology optimization approach based on machine learning, which incorporates computer sketches as constraint functions to improve the efficiency of computer-aided structural design models and meet the design intention and requirements of designers.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Leilei Chen, Zhongwang Wang, Haojie Lian, Yujing Ma, Zhuxuan Meng, Pei Li, Chensen Ding, Stephane P. A. Bordas
Summary: This paper presents a model order reduction method for electromagnetic boundary element analysis and extends it to computer-aided design integrated shape optimization of multi-frequency electromagnetic scattering problems. The proposed method utilizes a series expansion technique and the second-order Arnoldi procedure to reduce the order of original systems. It also employs the isogeometric boundary element method to ensure geometric exactness and avoid re-meshing during shape optimization. The Grey Wolf Optimization-Artificial Neural Network is used as a surrogate model for shape optimization, with radar cross section as the objective function.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
C. Pilloton, P. N. Sun, X. Zhang, A. Colagrossi
Summary: This paper investigates the smoothed particle hydrodynamics (SPH) simulations of violent sloshing flows and discusses the impact of volume conservation errors on the simulation results. Different techniques are used to directly measure the particles' volumes and stabilization terms are introduced to control the errors. Experimental comparisons demonstrate the effectiveness of the numerical techniques.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Ye Lu, Weidong Zhu
Summary: This work presents a novel global digital image correlation (DIC) method based on a convolution finite element (C-FE) approximation. The C-FE based DIC provides highly smooth and accurate displacement and strain results with the same element size as the usual finite element (FE) based DIC. The proposed method's formulation and implementation, as well as the controlling parameters, have been discussed in detail. The C-FE method outperformed the FE method in all tested examples, demonstrating its potential for highly smooth, accurate, and robust DIC analysis.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Mojtaba Ghasemi, Mohsen Zare, Amir Zahedi, Pavel Trojovsky, Laith Abualigah, Eva Trojovska
Summary: This paper introduces Lung performance-based optimization (LPO), a novel algorithm that draws inspiration from the efficient oxygen exchange in the lungs. Through experiments and comparisons with contemporary algorithms, LPO demonstrates its effectiveness in solving complex optimization problems and shows potential for a wide range of applications.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Jingyu Hu, Yang Liu, Huixin Huang, Shutian Liu
Summary: In this study, a new topology optimization method is proposed for structures with embedded components, considering the tension/compression asymmetric interface stress constraint. The method optimizes the topology of the host structure and the layout of embedded components simultaneously, and a new interpolation model is developed to determine interface layers between the host structure and embedded components.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Qiang Liu, Wei Zhu, Xiyu Jia, Feng Ma, Jun Wen, Yixiong Wu, Kuangqi Chen, Zhenhai Zhang, Shuang Wang
Summary: In this study, a multiscale and nonlinear turbulence characteristic extraction model using a graph neural network was designed. This model can directly compute turbulence data without resorting to simplified formulas. Experimental results demonstrate that the model has high computational performance in turbulence calculation.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Jacinto Ulloa, Geert Degrande, Jose E. Andrade, Stijn Francois
Summary: This paper presents a multi-temporal formulation for simulating elastoplastic solids under cyclic loading. The proper generalized decomposition (PGD) is leveraged to decompose the displacements into multiple time scales, separating the spatial and intra-cyclic dependence from the inter-cyclic variation, thereby reducing computational burden.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)
Article
Engineering, Multidisciplinary
Utkarsh Utkarsh, Valentin Churavy, Yingbo Ma, Tim Besard, Prakitr Srisuma, Tim Gymnich, Adam R. Gerlach, Alan Edelman, George Barbastathis, Richard D. Braatz, Christopher Rackauckas
Summary: This article presents a high-performance vendor-agnostic method for massively parallel solving of ordinary and stochastic differential equations on GPUs. The method integrates with a popular differential equation solver library and achieves state-of-the-art performance compared to hand-optimized kernels.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2024)