A novel method for predicting fluid–structure interaction with large deformation based on masked deep neural network

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(2)

Published: Feb. 1, 2024

Traditional fluid–structure interaction (FSI) simulation is computationally demanding, especially for bi-directional FSI problems. To address this, a masked deep neural network (MDNN) developed to quickly and accurately predict the unsteady flow field. By integrating MDNN with structural dynamic solver, an system proposed perform of flexible vertical plate oscillation in fluid large deformation. The results show that both field prediction structure response are consistent traditional system. Furthermore, method highly effective mitigating error accumulation during temporal predictions, making it applicable various deformation Notably, model reduces computational time millisecond scale each step regarding part, resulting increase nearly two orders magnitude speed, which greatly enhances speed

Language: Английский

---

Multi-fidelity prediction of fluid flow based on transfer learning using Fourier neural operator

Physics of Fluids, Journal Year: 2023, Volume and Issue: 35(7)

Published: July 1, 2023

Data-driven prediction of laminar flow and turbulent in marine aerospace engineering has received extensive research demonstrated its potential real-time recently. However, usually large amounts high-fidelity data are required to describe accurately predict the complex physical information, while reality, only limited available due high experimental/computational cost. Therefore, this work proposes a novel multi-fidelity learning method based on Fourier neural operator by jointing abundant low-fidelity under transfer paradigm. First, as resolution-invariant operator, is first gainfully applied integrate directly, which can utilize simultaneously. Then, framework developed for current task extracting rich knowledge assist modeling training, further improve data-driven accuracy. Finally, three application problems chosen validate accuracy proposed model. The results demonstrate that our effectiveness when compared with other models 99% all selected field problems. Additionally, model without 86%. Significantly, simple structure precision fluid problems, provide reference construction subsequent

Language: Английский

---

Porosity prediction through well logging data: A combined approach of convolutional neural network and transformer model (CNN-transformer)

Physics of Fluids, Journal Year: 2024, Volume and Issue: 36(2)

Published: Feb. 1, 2024

Porosity, as a key parameter to describe the properties of rock reservoirs, is essential for evaluating permeability and fluid migration performance underground rocks. In order overcome limitations traditional logging porosity interpretation methods in face geological complexity nonlinear relationships, this study introduces CNN (convolutional neural network)-transformer model, which aims improve accuracy generalization ability prediction. CNNs have excellent spatial feature capture capabilities. The convolution operation can effectively learn mapping relationship local features, so better correlation well log. Transformer models are able complex sequence relationships between different depths or time points. This enables model integrate information from times, prediction accuracy. We trained on log dataset ensure that it has good ability. addition, we comprehensively compare CNN-transformer with other machine learning verify its superiority Through analysis experimental results, shows task introduction will bring new perspective development technology provide more efficient accurate tool field geoscience.

Language: Английский

---

Mitigating spectral bias for the multiscale operator learning

Journal of Computational Physics, Journal Year: 2024, Volume and Issue: 506, P. 112944 - 112944

Published: March 19, 2024

Language: Английский

---

Latent Neural PDE Solver: a reduced-order modelling framework for partial differential equations

Journal of Computational Physics, Journal Year: 2025, Volume and Issue: unknown, P. 113705 - 113705

Published: Jan. 1, 2025

Language: Английский

---

Deep learning-based reduced-order modeling for parameterized convection-dominated partial differential equations

Physics of Fluids, Journal Year: 2025, Volume and Issue: 37(2)

Published: Feb. 1, 2025

Reduced-order modeling of fluid flows has been an active area research. It approximates the evolution physical systems in time terms coherent patterns and structures that generally consist a dimensionality reduction mechanism dynamical model reduced state space. This paper proposes deep learning-based reduced-order composed β-variational autoencoder, multilayer perceptron, transformer architectures for problems governed by parameterized convection-dominated partial differential equations. In our approach, autoencoder is utilized as mechanism, trained to predict future system, perceptron applied learn relationship between different parameter values latent space representations. Therefore, system can be obtained online phase. The proposed method tested on several benchmark equations, such Burgers' equation, traffic flow problem, shallow water Navier–Stokes equation. results demonstrate applicability effectiveness

Language: Английский

---

Multi-scale time-stepping of Partial Differential Equations with transformers

Computer Methods in Applied Mechanics and Engineering, Journal Year: 2024, Volume and Issue: 426, P. 116983 - 116983

Published: April 13, 2024

Developing fast surrogates for Partial Differential Equations (PDEs) will accelerate design and optimization in almost all scientific engineering applications. Neural networks have been receiving ever-increasing attention demonstrated remarkable success computational modeling of PDEs, however; their prediction accuracy is not at the level full deployment. In this work, we utilize transformer architecture, backbone numerous state-of-the-art AI models, to learn dynamics physical systems as mixing spatial patterns learned by a convolutional autoencoder. Moreover, incorporate idea multi-scale hierarchical time-stepping increase speed decrease accumulated error over time. Our model achieves similar or better results predicting time-evolution Navier–Stokes equations compared powerful Fourier Operator (FNO) two transformer-based neural operators OFormer Galerkin Transformer. The code data are available on https://github.com/BaratiLab/MST_PDE.

Language: Английский

---

Physics informed token transformer for solving partial differential equations

Machine Learning Science and Technology, Journal Year: 2024, Volume and Issue: 5(1), P. 015032 - 015032

Published: Feb. 9, 2024

Abstract Solving partial differential equations (PDEs) is the core of many fields science and engineering. While classical approaches are often prohibitively slow, machine learning models fail to incorporate complete system information. Over past few years, transformers have had a significant impact on field Artificial Intelligence seen increased usage in PDE applications. However, despite their success, currently lack integration with physics reasoning. This study aims address this issue by introducing Physics Informed Token Transformer (PITT). The purpose PITT knowledge embedding PDEs into process. uses an equation tokenization method learn analytically-driven numerical update operator. By tokenizing derivatives, transformer become aware underlying behind physical processes. To demonstrate this, tested challenging 1D 2D operator tasks. results show that outperforms popular neural has ability extract physically relevant information from governing equations.

Language: Английский

---

Data-driven multifidelity surrogate models for rocket engines injector design

Data-Centric Engineering, Journal Year: 2025, Volume and Issue: 6

Published: Jan. 1, 2025

Abstract Surrogate models of turbulent diffusive flames could play a strategic role in the design liquid rocket engine combustion chambers. The present article introduces method to obtain data-driven surrogate for coaxial injectors, by leveraging an inductive transfer learning strategy over U-Net with available multifidelity Large Eddy Simulations (LES) data. resulting preserve reasonable accuracy while reducing offline computational cost data-generation. First, database about 100 low-fidelity LES simulations shear-coaxial operating gaseous oxygen and methane as propellants, has been created. experiments explores three variables: chamber radius, recess-length oxidizer post, mixture ratio. Subsequently, U-Nets were trained upon this dataset provide approximations temporal-averaged two-dimensional flow field. Despite fact that neural networks are efficient non-linear data emulators, purely approaches their quality is directly impacted precision they upon. Thus, high-fidelity (HF) created, made 10 simulations, much greater per sample. amalgamation low HF during transfer-learning process enables improvement model’s fidelity without excessive additional cost.

Language: Английский

---

A Hybrid CNN-Transformer Surrogate Model for the Multi-Objective Robust Optimization of Geological Carbon Sequestration

Advances in Water Resources, Journal Year: 2025, Volume and Issue: unknown, P. 104897 - 104897

Published: Jan. 1, 2025

Language: Английский

---