Abstract:Obtaining system parameters and reconstructing the full flow state from limited velocity observations using conventional fluid dynamics solvers can be prohibitively expensive. Here we employ machine learning algorithms to overcome the challenge. As an example, we consider a moderately turbulent fluid flow, excited by a stationary force and described by a two-dimensional Navier-Stokes equation with linear bottom friction. Using dense in time, spatially sparse and probably noisy velocity data, we reconstruct the spatially dense velocity field, infer the pressure and driving force up to a harmonic function and its gradient, respectively, and determine the unknown fluid viscosity and friction coefficient. Both the root-mean-square errors of the reconstructions and their energy spectra are addressed. We study the dependence of these metrics on the degree of sparsity and noise in the velocity measurements. Our approach involves training a physics-informed neural network by minimizing the loss function, which penalizes deviations from the provided data and violations of the governing equations. The suggested technique extracts additional information from velocity measurements, potentially enhancing the capabilities of particle image/tracking velocimetry.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to obtain system parameters from limited velocity observation data and reconstruct the complete flow - field state of two - dimensional turbulence. Specifically, the researchers use machine - learning algorithms, especially Physics - Informed Neural Networks (PINNs), to overcome the high cost of traditional fluid - mechanics solvers in dealing with such problems. The main challenges in the paper include: 1. **Reconstructing a spatially dense velocity field from sparse and possibly noisy velocity data**: This involves using data points that are dense in time but sparse in space to infer the state of the entire flow field. 2. **Inferring pressure and driving force**: Through velocity data, infer the pressure field and driving - force field, where the pressure and driving force can be determined up to a harmonic function and its gradient respectively. 3. **Determining the unknown fluid viscosity and bottom - friction coefficient**: These parameters are crucial for describing fluid behavior but are usually not directly measurable. ### Specific problem description - **Data sparsity**: The velocity data collected in the experiment may be sparse in space, that is, there are no velocity measurements at certain positions. - **Data noise**: Noise is inevitably introduced in actual measurements, which will affect the reconstruction accuracy. - **Unknown parameters**: The viscosity and bottom - friction coefficient of the fluid are usually unknown and need to be inferred through the model. - **Physical constraints**: The Navier - Stokes equation and the incompressibility condition need to be satisfied during the reconstruction process. ### Solution The researchers propose a method based on Physics - Informed Neural Networks (PINNs) to train the neural network by minimizing the loss function. The loss function consists of two parts: 1. **Data loss** (\(L_{\text{data}}\)): Measures the deviation between the predicted velocity and the measured velocity. \[ L_{\text{data}}=\frac{1}{N_{\text{data}}U^{2}}\sum_{n = 1}^{N_{\text{data}}}\left|\hat{v}(r_{n},t_{n})-v_{d}(r_{n},t_{n})\right|^{2} \] where \(U^{2}\) is the mean - squared velocity based on the measured data. 2. **Physical loss** (\(L_{\text{eq}}\)): Measures the deviation between the predicted values and the Navier - Stokes equation and the incompressibility condition. \[ L_{\text{eq}}=\frac{1}{N_{\text{eq}}}\sum_{n = 1}^{N_{\text{eq}}}\left[e_{1}^{2}(r_{n},t_{n})+e_{2}^{2}(r_{n},t_{n})+e_{3}^{2}(r_{n})\right] \] where: \[ e_{1}=\frac{\partial_{t}\hat{v}+(\hat{v}\cdot\nabla)\hat{v}+\nabla\hat{p}+e^{\lambda_{1}}\hat{v}-e^{\lambda_{2}}\nabla^{2}\hat{v}-\hat{f}}{L/U^{2}} \] \[ e_{2}=\frac{\partial_{x}\hat{v}_{x}+\partial_{y}\hat{v}_{y}}{L/U} \] \[ e_{3}=\frac{\partial_{x}\hat{f}_{x}+\partial_{y}\hat{f}_{y}}{L^{2}/U^{2}} \] The final total loss function is: \[ L = L_{\text{data}}+\beta L_{\text{eq}} \] where \(\beta\) is a weighting coefficient used to balance the data loss and the physical loss. ### Results - **Effect of data density**

Inferring parameters and reconstruction of two-dimensional turbulent flows with physics-informed neural networks

Inferring Parameters and Reconstruction of Two-dimensional Turbulent Flows with Physics-informed Neural Networks

Flow Reconstruction with Uncertainty Quantification from Noisy Measurements Based on Bayesian Physics-Informed Neural Networks

Validation and parameterization of a novel physics-constrained neural dynamics model applied to turbulent fluid flow

Physics-constrained bayesian neural network for fluid flow reconstruction with sparse and noisy data

Physics-Based Deep Learning for Flow Problems

Reconstructing unsteady flows from sparse, noisy measurements with a physics-constrained convolutional neural network

Predicting Pressure Fields from Incomplete Velocity Fields Based on Deep Convolutional Neural Network

Reconstruction of the turbulent flow field with sparse measurements using physics-informed neural network

Turbulence model augmented physics-informed neural networks for mean-flow reconstruction

Mean flow reconstruction of unsteady flows using physics-informed neural networks

Explore missing flow dynamics by physics-informed deep learning: The parameterized governing systems

Reconstructing Turbulent Flows Using Physics-Aware Spatio-Temporal Dynamics and Test-Time Refinement

Turbulence model augmented physics informed neural networks for mean flow reconstruction

Explore missing flow dynamics by physics-informed deep learning: the parameterised governing systems

Inferring incompressible two-phase flow fields from the interface motion using physics-informed neural networks

Estimating divergence‐free flows via neural networks

Physics-informed deep-learning applications to experimental fluid mechanics

Graph Neural Networks for Aerodynamic Flow Reconstruction from Sparse Sensing

Physics-informed Machine Learning Approach for Augmenting Turbulence Models: A Comprehensive Framework

Bayesian inverse Navier-Stokes problems: joint flow field reconstruction and parameter learning