相关参考文献
注意:仅列出部分参考文献,下载原文获取全部文献信息。
Article
Engineering, Electrical & Electronic
Youzuo Lin et al.
Summary: The goal of seismic inversion is to obtain subsurface properties from surface measurements. Seismic images are crucial for various applications such as subsurface energy exploration, earthquake early warning, and estimating subsurface contaminant transport pathways. These properties influence the transmission of seismic waves, and forward models can predict surface measurements based on known physics.
IEEE SIGNAL PROCESSING MAGAZINE
(2023)
Article
Engineering, Multidisciplinary
Min Zhu et al.
Summary: Deep neural operators can learn nonlinear mappings between infinite-dimensional function spaces. They provide a new simulation paradigm in science and engineering as surrogate solvers of partial differential equations. However, pure data-driven neural operators and deep learning models are usually limited to interpolation scenarios, and extrapolation may lead to large errors and failure. In this study, we systematically investigate the extrapolation behavior of DeepONets and propose a bias-variance trade-off strategy for extrapolation. We also develop a complete workflow and propose reliable learning methods that guarantee safe prediction under extrapolation conditions.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2023)
Article
Geochemistry & Geophysics
Hani Alzahrani et al.
Summary: Data-driven artificial neural networks offer advantages in geophysical problems, particularly in seismic velocity model building. However, the challenge of training generalization needs to be addressed, and the influence of training model structures on test data should be considered.
Article
Geochemistry & Geophysics
Weiqiang Zhu et al.
Summary: Full-waveform inversion (FWI) is an accurate imaging method for modeling the velocity structure. However, its strong nonlinearity can lead to optimization being trapped in local minima. In this study, we propose a neural-network-based FWI method that integrates deep neural networks and FWI to overcome this issue.
Article
Water Resources
Gege Wen et al.
Summary: Numerical simulation of multiphase flow plays a vital role in geoscience applications. The U-FNO neural network architecture, based on the Fourier neural operator (FNO), offers a superior and efficient solution for solving multiphase flow problems. It outperforms traditional simulators in accuracy, speed, and data utilization.
ADVANCES IN WATER RESOURCES
(2022)
Article
Engineering, Multidisciplinary
Lu Lu et al.
Summary: This article investigates the performance of two neural operators and develops new extensions to make them more suitable for complex industrial applications. The experimental results show that DeepONet and FNO perform similarly in relatively simple settings, but FNO's performance deteriorates significantly in complex geometries.
COMPUTER METHODS IN APPLIED MECHANICS AND ENGINEERING
(2022)
Article
Mathematics, Applied
Pengzhan Jin et al.
Summary: This paper studies operator regression via neural networks for multiple-input operators and proposes a novel neural operator, MIONet, which can learn multiple-input operators. The paper demonstrates the accuracy and application of prior knowledge in computational examples.
SIAM JOURNAL ON SCIENTIFIC COMPUTING
(2022)
Article
Environmental Sciences
Fangda Li et al.
Summary: In this study, a hybrid network (AG-ResUnet) is proposed to estimate velocity models from common source point gathers, using fully convolutional layers, attention mechanism, and residual unit. Experimental results demonstrate the high effectiveness and efficiency of AG-ResUnet in transfer learning and noisy data inversion, providing generalized and robust velocity prediction with rich structural details.
Article
Computer Science, Artificial Intelligence
Beichuan Deng et al.
Summary: This article presents the analysis of approximation rates of operator learning in Chen and Chen (1995) and Lu et al. (2021), where continuous operators are approximated by a sum of products of branch and trunk networks. The study considers the rates of learning solution operators from both linear and nonlinear advection-diffusion equations with or without reaction. The findings show that the approximation rates depend on the architecture of branch networks as well as the smoothness of inputs and outputs of solution operators.
Article
Physics, Multidisciplinary
Lu Lu et al.
Summary: This paper proposes a multifidelity neural operator based on deep neural networks, which can reduce the demand for high-fidelity data and achieve smaller errors in solving heat transport problems. By combining with genetic algorithms and topology optimization, it enables fast solvers and inverse design for the phonon Boltzmann transport equation (BTE).
PHYSICAL REVIEW RESEARCH
(2022)
Article
Geochemistry & Geophysics
Bin Liu et al.
Summary: Velocity model inversion, an important task in seismic exploration, has seen advancements with the development of deep learning methods. A method for constructing realistic structural models has been proposed to improve inversion results and provide practical guidance for optimal strategies in realistic situations.
Article
Computer Science, Interdisciplinary Applications
Zhiping Mao et al.
Summary: This study introduces a novel efficient approach using neural networks to predict fluid properties changes in high-speed flow. By training and utilizing DeepONet, rapid and accurate predictions can be achieved, maintaining good predictive performance even when sparse measurements are available.
JOURNAL OF COMPUTATIONAL PHYSICS
(2021)
Article
Geochemistry & Geophysics
Jian Sun et al.
Summary: This paper discusses the importance of determining subsurface elastic property models and introduces different methods, including deep learning and physics-guided approaches, to address this issue. Experimental results show that the hybrid-trained network outperforms traditional fully data-driven networks.
Article
Chemistry, Physical
Chensen Lin et al.
Summary: This study explores the use of deep operator networks (DeepONet) to learn dynamics at different scales and demonstrates their accuracy and effectiveness in predicting multirate bubble growth dynamics. Results show that properly trained DeepONets can accurately predict macroscale bubble growth dynamics and outperform long short-term memory networks with only a few new measurements.
JOURNAL OF CHEMICAL PHYSICS
(2021)
Article
Multidisciplinary Sciences
Sifan Wang et al.
Summary: The physics-informed DeepONets framework can effectively learn the solution operator of arbitrary PDEs even in the absence of paired training data. It is able to predict the solution of various parametric PDEs up to three orders of magnitude faster compared to conventional solvers, setting a new paradigm for modeling and simulation of nonlinear and nonequilibrium processes in science and engineering.
Article
Computer Science, Interdisciplinary Applications
Shengze Cai et al.
Summary: This article introduces a new data assimilation framework, DeepM & Mnet, for simulating multiphysics and multiscale problems using pre-trained neural networks. The framework utilizes DeepONets to quickly and accurately build complex models based on very few measurements.
JOURNAL OF COMPUTATIONAL PHYSICS
(2021)
Review
Geochemistry & Geophysics
Siwei Yu et al.
Summary: Deep learning has attracted increasing attention in the geophysical community as a new data-driven technique, showing potential in accurately predicting complex system states and alleviating the curse of dimensionality. Various DL approaches have been applied and studied in geosciences, with promising future research directions including unsupervised learning, multimodal DL, among others.
REVIEWS OF GEOPHYSICS
(2021)
Review
Physics, Applied
George Em Karniadakis et al.
Summary: Physics-informed learning seamlessly integrates data and mathematical models through neural networks or kernel-based regression networks for accurate inference of realistic and high-dimensional multiphysics problems. Challenges remain in incorporating noisy data seamlessly, complex mesh generation, and addressing high-dimensional problems.
NATURE REVIEWS PHYSICS
(2021)
Article
Computer Science, Artificial Intelligence
Lu Lu et al.
Summary: This study extends the capabilities of neural networks with the introduction of the deep operator network (DeepONet), which can be used to learn various operators, including explicit and implicit operators. Different formulations of the input function space were studied and their effect on generalization error for 16 diverse applications was examined.
NATURE MACHINE INTELLIGENCE
(2021)
Article
Geochemistry & Geophysics
Jian Sun et al.
Article
Geochemistry & Geophysics
Zhongping Zhang et al.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2020)
Article
Engineering, Electrical & Electronic
Yue Wu et al.
IEEE TRANSACTIONS ON COMPUTATIONAL IMAGING
(2020)
Article
Computer Science, Information Systems
Yuxiao Ren et al.
Review
Geochemistry & Geophysics
J. Virieux et al.
Article
Computer Science, Artificial Intelligence
Z Wang et al.
IEEE TRANSACTIONS ON IMAGE PROCESSING
(2004)
Article
Geochemistry & Geophysics
KH Lee et al.