Source-linked AI summary
Physics-Constrained Bayesian Neural Network for Fluid Flow Reconstruction with Sparse and Noisy Data
Luning Sun, Jian-Xun Wang
TL;DR
Sparse and noisy measurements make high-resolution flow reconstruction difficult. The paper uses a physics-constrained Bayesian neural network whose likelihood combines data and equation information, then evaluates it on idealized vascular flows. The method reconstructs mean flow fields accurately while estimating prediction uncertainty under different noise levels.
Problem
High-resolution flow fields must be reconstructed from measurements that are sparse and possibly noisy, but limited information makes this challenging in engineering applications.
Method
A Bayesian neural network learns from sparse velocity data while penalizing Navier-Stokes violations at unmeasured spatiotemporal points through a probabilistic likelihood.
Results
The PC-BNN accurately predicts mean flow fields and reasonably estimates prediction uncertainties across different data noise levels in idealized vascular-flow cases.
Takeaways & Limitations
Physical constraints significantly improve reconstruction from limited clean data, while the Bayesian formulation provides uncertainty estimates when measurements are noisy.
Takeaways & Limitations
The physical equations are treated as soft probabilistic constraints, with model confidence represented by a residual covariance matrix.
Abstract
from arXiv · showhide
In many applications, flow measurements are usually sparse and possibly noisy. The reconstruction of a high-resolution flow field from limited and imperfect flow information is significant yet challenging. In this work, we propose an innovative physics-constrained Bayesian deep learning approach to reconstruct flow fields from sparse, noisy velocity data, where equation-based constraints are imposed through the likelihood function and uncertainty of the reconstructed flow can be estimated. Specifically, a Bayesian deep neural network is trained on sparse measurement data to capture the flow field. In the meantime, the violation of physical laws will be penalized on a large number of spatiotemporal points where measurements are not available. A non-parametric variational inference approach is applied to enable efficient physics-constrained Bayesian learning. Several test cases on idealized vascular flows with synthetic measurement data are studied to demonstrate the merit of the proposed method.
1. Introduction
Sparse and noisy flow measurements make high-resolution flow reconstruction important but challenging across engineering and biomedical applications. Existing approaches use offline flow data or physics-based information, motivating a physics-constrained Bayesian neural network that incorporates uncertainty.
- Sparse, noisy measurements limit high-resolution flow reconstruction in applications including cardiovascular imaging and aerodynamic monitoring.MR imaging can have limited resolution and signal-to-noise ratio, while lidar and similar sensors are often sparsely located.
- Existing reconstruction methods compensate for incomplete measurements using either offline full-field flow databases or physics-based models.The methods are organized according to the type of additional information incorporated into reconstruction.
- POD, DMD, and related data-driven methods extract flow structures or learn mappings from sparse measurements to high-resolution fields.These methods rely on coherent structures, correlation features, or learned mappings between sparse and full-field representations.
- Deep learning offers efficient flow-reconstruction algorithms but commonly depends on sufficient offline training data that may be inaccessible.Neural networks have been used both to learn POD coefficients and to directly map sparse measurements to high-resolution flow fields.
- The proposed physics-constrained Bayesian neural network imposes equation-constrained training probabilistically and uses SVGD for efficient learning from limited data.Its likelihood combines measurement noise and model inadequacy, and its performance is demonstrated on idealized vascular flows with sparse, noisy velocity data.
2. Methodology
The method reconstructs high-resolution flow fields from sparse velocity data by combining neural-network approximation with Navier–Stokes constraints and Bayesian uncertainty modeling. Non-parametric variational inference trains an ensemble whose predictions provide mean fields and reconstruction uncertainty.
- Physics-constrained reconstruction: A pointwise DNN learns velocity and pressure from sparse velocity data while physical laws are enforced at unmeasured spatiotemporal collocation points.The approach avoids requiring extra offline databases of high-resolution flow fields and uses governing-equation residuals as additional information.
- Physics-constrained reconstruction: The physics-regularized objective combines sparse-data mismatch with penalties for violations of incompressible Navier–Stokes equations and boundary conditions.Pressure inlet/outlet conditions are enforced by construction, while no-slip wall conditions are imposed softly for complex geometries.
- Bayesian formulation: Bayesian learning represents uncertainty from measurement noise and model inadequacy through a likelihood built from data and equation constraints.The physical-equation residual is modeled probabilistically, with its covariance controlling confidence in the physical model; data covariance parameters are learned from observations.
- Uncertainty estimation: After training, forward propagation of the DNN ensemble yields high-resolution mean velocity predictions and variance fields reflecting reconstruction uncertainty.The ensemble estimates prediction statistics using Monte Carlo methods and the law of total variance.
- Non-parametric variational inference: Stein variational gradient descent uses particles to approximate the posterior without specifying a variational distribution family.Its gradient moves particles toward high-density posterior regions, while a repulsive force preserves diversity and can capture multi-modal posteriors.
3. Result
Across idealized stenotic and aneurysm-bifurcation flows, physics-constrained learning improved reconstruction from sparse data, while PC-BNN additionally handled noise and quantified uncertainty.
- Experimental design: The experiments used two idealized vascular geometries with synthetic sparse measurements sampled from fully resolved CFD solutions, including noise-free deterministic and noisy Bayesian cases.Both reconstructed mean fields and uncertainties were investigated for different data-noise levels.
- Case 1: deterministic stenotic reconstruction: 5.6% relative error was achieved with Navier-Stokes constraints in stenotic flow, versus 11.3% with divergence-free constraints and 24.8% for purely data-based learning.The Navier-Stokes-constrained reconstruction was almost identical to the CFD benchmark.
- Case 2: deterministic aneurysm-bifurcation reconstruction: 13.7%, 12.1%, and 12.8% relative errors for u, v, and P were obtained with Navier-Stokes constraints, compared with 35.1%, 40.5%, and 69.9% for purely data-based learning.Only six sparse slices of CFD velocity data were used, and the physics-constrained model captured pressure patterns without pressure training data.
- Case 4: Bayesian aneurysm-bifurcation reconstruction: PC-BNN accurately captured the noisy bifurcation flow, and its standard-deviation field indicated larger uncertainty at the left outlet region.The mean reconstruction error and uncertainty increased as the noise level grew.
4. Discussion on the role of data and constraints in deep learning
The discussion examines whether sparse labeled data improve physics-constrained learning compared with using physical constraints alone. Adding sparse data accelerated learning and produced lower test error in the reported comparison.
- Comparison setup: The comparison evaluates the benefit of additional sparse labeled data after prior work showed that physics-constrained deep learning can solve Navier-Stokes flows without labeled data when boundary conditions are imposed properly.The reported test errors concern the physics-constrained learning comparison.
- Role of data: Sparse data-based physics-constrained learning decreased test error much faster than data-free physics-constrained learning.Figure 5 compares MSE histories for u, v, and P predictions.
- Role of data: About one order of magnitude lower test error was obtained with sparse data-based learning than with purely physics-constrained learning after the same number of training epochs.The comparison supports adding some labeled data to equation-constrained training.
5. Conclusion
This work presents a physics-constrained Bayesian deep learning framework for reconstructing high-resolution flow fields from sparse and possibly noisy data. Numerical experiments on idealized vascular geometries show accurate reconstruction with estimated uncertainties across different noise levels.
- The framework constructs a likelihood from measurement uncertainty and model inadequacy to impose physics constraints during Bayesian learning.Stein variation gradient descent enables efficient Bayesian learning.
- PC-BNN reconstructs flow fields while estimating uncertainties, particularly when measurements are corrupted by noise.The method predicts the mean flow field and estimates prediction uncertainties corresponding to different data noise levels.
- Numerical experiments evaluate the method on multiple flow-reconstruction cases with idealized vascular geometries and synthetic data.
- Physical-model constraints significantly improve reconstruction results from limited clean data.
Appendix
Table 1 summarizes PC-BNN reconstruction errors and uncertainties for u velocity fields reconstructed from sparse, noisy data. Both quantities are normalized by the corresponding CFD benchmark solution.
- The reported uncertainty is measured by std norm, with both mean and maximum values presented.
- Table 1 reports mean errors and uncertainties for reconstructed u velocity fields using PC-BNN with sparse, noisy data.
- Both reconstruction errors and standard deviations are normalized by the corresponding CFD benchmark solution.