Flow field reconstruction from sparse sensor measurements with physics-informed neural networks

Physics of Fluids, Год журнала: 2024, Номер 36(7)

Опубликована: Июль 1, 2024

In the realm of experimental fluid mechanics, accurately reconstructing high-resolution flow fields is notably challenging due to often sparse and incomplete data across time space domains. This exacerbated by limitations current tools methods, which leave critical areas without measurable data. research suggests a feasible solution this problem employing an inverse physics-informed neural network (PINN) merge available with physical laws. The method's efficacy demonstrated using around cylinder as case study, three distinct training sets. One was velocity from domain, other two datasets were limited obtained domain boundaries sensors wall. coefficient determination (R2) mean squared error (RMSE) metrics, indicative model performance, have been determined for components all models. For 28 model, R2 value stands at 0.996 associated RMSE 0.0251 u component, while v registers 0.969, accompanied 0.0169. outcomes indicate that method can successfully recreate actual field considerable precision more than cylinder, highlighting PINN's potential effective assimilation technique mechanics.

Язык: Английский

---

Using diffusion models for reducing spatiotemporal errors of deep learning based urban microclimate predictions at post-processing stage

Physics of Fluids, Год журнала: 2025, Номер 37(3)

Опубликована: Март 1, 2025

Computational fluid dynamics (CFD) is a powerful tool for modeling turbulent flow and commonly used urban microclimate simulations. However, traditional CFD methods are computationally intensive, requiring substantial hardware resources high-fidelity Deep learning (DL) models becoming popular as efficient alternatives, less computational to model complex non-linear interactions in A major drawback of DL that they prone error accumulation long-term temporal predictions, often compromising their accuracy reliability. To address this shortcoming, study investigates the use denoising diffusion probabilistic (DDPM) novel post-processing technique mitigate propagation models' sequential predictions. this, we employ convolutional autoencoder (CAE) U-Net architectures predict airflow around cubic structure. The DDPM then applied model's refining reconstructed fields better align with statistical results from large-eddy Results demonstrate that, although deep provide significant advantages over numerical solvers, susceptible predictions; however, utilizing step enhances by up 65% while maintaining three times speedup compared solvers. These findings highlight potential integrating transformative approach improving reliability learning-based simulations, paving way more scalable modeling.

Язык: Английский

---

A novel attention enhanced deep neural network for hypersonic spatiotemporal turbulence prediction

Physics of Fluids, Год журнала: 2024, Номер 36(5)

Опубликована: Май 1, 2024

High Reynolds number turbulent flow of hypersonic vehicles exhibits multi-scale structures and non-equilibrium high-frequency characteristics, presenting a significant challenge for accurate prediction. A deep neural network integrated with attention mechanism as reduced order model is proposed, which capable capturing spatiotemporal characteristics from high-dimensional numerical data directly. The leverages encoder–decoder architecture where the encoder captures high-level semantic information input field, Convolutional Long Short-Term Memory learns low-dimensional characteristic evolution, decoder generates pixel-level multi-channel field information. Additionally, skip connection structure introduced at decoding stage to enhance feature fusion while incorporating Dual-Attention-Block that automatically adjusts weights capture spatial imbalances in turbulence distribution. Through evaluating time generalization ability, effectively evolution characteristics. It enables rapid prediction high over reasonable accuracy maintaining excellent computational efficiency.

Язык: Английский

---

An attention-enhanced Fourier neural operator model for predicting flow fields in turbomachinery Cascades

Physics of Fluids, Год журнала: 2025, Номер 37(3)

Опубликована: Март 1, 2025

Flow field information within cascades is crucial for refined turbomachinery design. Currently, this primarily obtained through experimental methods or numerical simulations, both of which are complex and time-consuming. Data-driven deep learning approaches offer a potential solution rapid flow evaluation. However, existing learning-based prediction models exhibit certain limitations in accuracy generalization, particularly regions with high gradients, often the primary sources aerodynamic losses. To address these issues, study develops high-precision cascade model, A-FNO, based on Galerkin-type self-attention mechanism Fourier Neural Operator (FNO). A-FNO designed newly proposed FNO, has demonstrated excellent performance solving partial differential equations. This extends its application to problems. mitigate FNO predicting areas steep gradient changes, we incorporate capture dependencies between different field, thereby enhancing FNO's ability express details. Experimental results demonstrate that significantly improves surrounding boundary layer. The maximum relative error velocity predictions 5%, pressure 2%, temperature 1%.

Язык: Английский

---

A prediction of urban boundary layer using Recurrent Neural Network and reduced order modeling

Building and Environment, Год журнала: 2025, Номер unknown, С. 112804 - 112804

Опубликована: Март 1, 2025

Язык: Английский

---

Physics-informed neural network (PINNs) for convection equations in polymer flooding reservoirs

Physics of Fluids, Год журнала: 2025, Номер 37(3)

Опубликована: Март 1, 2025

This paper realizes the application of physics-informed neural network (PINN) in polymer flooding reservoir model, achieving high-precision calculations water saturation and concentration distributions a one-dimensional channel. The investigates impacts different PINN structures, forms governing equations used, strength artificial viscosity added to on computational performance PINN, especially accuracy. Three numerical examples are implemented this paper, with high-fidelity solution serving as benchmark. results show that, when total number grid parameters is similar, PINN-1, which estimates both using single network, exhibits significantly better than PINN-2, two separate networks. simplification equation for can improve training accuracy PINN. addition enhance improvement effect first increases then decreases coefficient increases. research provides reference subsequent development high-accuracy proxy models engineering.

Язык: Английский

---

WCNS3-MR-NN: A machine learning-based shock-capturing scheme with accuracy-preserving and high-resolution properties

Journal of Computational Physics, Год журнала: 2025, Номер unknown, С. 113973 - 113973

Опубликована: Март 1, 2025

Язык: Английский

---

Extended multiphysics-informed neural network for conjugate heat transfer problems

International Journal of Heat and Mass Transfer, Год журнала: 2025, Номер 246, С. 127098 - 127098

Опубликована: Апрель 10, 2025

Язык: Английский

---

An Image and State Information-Based PINN with Attention Mechanisms for the Rapid Prediction of Aircraft Aerodynamic Characteristics

Aerospace, Год журнала: 2025, Номер 12(5), С. 434 - 434

Опубликована: Май 13, 2025

Prediction of aircraft aerodynamic parameters is crucial for design, yet traditional computational fluid dynamics methods remain time-consuming and labor-intensive. This work presents a novel model, the image state information-based attention-enhanced physics-informed neural network (ISA-PINN), which significantly improves prediction accuracy. Our model incorporates following innovations: designed attention module dynamically extracts hidden features from pattern data while selectively focusing on relevant dimensions target information. Meanwhile, image-information fusion combines multi-scale geometric derived images to enhance overall By embedding equations, maintains physical consistency enhancing interpretability. Extensive experiments validate effectiveness our rapid parameter prediction, achieving significant reduction in error that performance by 29.25% RMSE 37.99% MRE compared existing methods. A 6.12% increase test set confirms model’s robust generalization ability.

Язык: Английский

---

Using physics-informed derivative networks to solve the forward problem of a free-convective boundary layer problem

Scientific Reports, Год журнала: 2025, Номер 15(1)

Опубликована: Май 28, 2025

Язык: Английский

---