Source-linked AI summary
Surrogate Modeling for Fluid Flows Based on Physics-Constrained Deep Learning Without Simulation Data
Luning Sun, Han Gao, Shaowu Pan, Jian-Xun Wang
TL;DR
Expensive CFD simulations limit real-time fluid-flow prediction, while sparse data challenges conventional deep-learning surrogates. The paper trains a physics-constrained DNN without labeled simulation data and reports excellent agreement with CFD benchmarks, including propagated uncertainties.
Problem
Computationally expensive CFD models limit the use of principled fluid-flow simulations for real-time predictions.
Method
A structured DNN enforces initial and boundary conditions while governing-PDE residuals drive training without labeled CFD simulation data.
Results
The surrogate achieves excellent agreement with CFD benchmarks, with reasonably small training losses and test errors across the evaluated scenarios.
Takeaways & Limitations
The results support physics-constrained DNN surrogates for fluid simulations without labeled data in the studied internal-flow settings.
Takeaways & Limitations
Earlier neural-network approaches were limited by the available NN techniques and computational power at that time.
Abstract
from arXiv · showhide
Numerical simulations on fluid dynamics problems primarily rely on spatially or/and temporally discretization of the governing equation into the finite-dimensional algebraic system solved by computers. Due to complicated nature of the physics and geometry, such process can be computational prohibitive for most real-time applications and many-query analyses. Therefore, developing a cost-effective surrogate model is of great practical significance. Deep learning (DL) has shown new promises for surrogate modeling due to its capability of handling strong nonlinearity and high dimensionality. However, the off-the-shelf DL architectures fail to operate when the data becomes sparse. Unfortunately, data is often insufficient in most parametric fluid dynamics problems since each data point in the parameter space requires an expensive numerical simulation based on the first principle, e.g., Naiver--Stokes equations. In this paper, we provide a physics-constrained DL approach for surrogate modeling of fluid flows without relying on any simulation data. Specifically, a structured deep neural network (DNN) architecture is devised to enforce the initial and boundary conditions, and the governing partial differential equations are incorporated into the loss of the DNN to drive the training. Numerical experiments are conducted on a number of internal flows relevant to hemodynamics applications, and the forward propagation of uncertainties in fluid properties and domain geometry is studied as well. The results show excellent agreement on the flow field and forward-propagated uncertainties between the DL surrogate approximations and the first-principle numerical simulations.
1. Introduction
The introduction motivates cost-effective fluid-flow surrogates by highlighting CFD’s computational burden and the difficulty of training data-fit models with sparse, expensive data. It proposes a physics-constrained, data-free DNN surrogate trained through governing-PDE residuals for parametric Navier–Stokes systems.
- Motivation: CFD simulations are computationally cumbersome for turbulent flows and complex or varying geometries, limiting real-time prediction and many-query analysis.These costs also constrain applications such as clinical diagnosis, surgery planning, aircraft optimization, and uncertainty quantification.
- Limitations: Existing projection-based ROMs face stability, robustness, code-intrusiveness, and limited speedup issues, while traditional data-fit models struggle with strong nonlinearities and high dimensionality.Data-fit surrogates are built from CFD solutions at selected parameter-space collocation points.
- Data scarcity: Sparse and potentially noisy labeled data from expensive CFD simulations or experiments prevent off-the-shelf DL models from fully exploiting their capacity in small-data regimes.Known governing equations can instead constrain or drive learning to compensate for insufficient data.
- Contribution: The paper develops a parametric, physics-constrained, data-free DNN surrogate for fluid simulations, trained solely by minimizing governing-PDE residuals without expensive CFD simulation data.The introduction frames this as a first attempt for parametric fluid surrogate modeling without simulation data and addresses the unclear applicability of physics-constrained learning to Navier–Stokes systems.
2. Methodology
The methodology builds a parametric, physics-constrained FC-NN surrogate for incompressible Navier–Stokes flows, encoding boundary conditions by construction and training without CFD simulation data. The network minimizes governing-equation residuals to produce rapid flow-field predictions for specified parameters.
- Problem formulation: The surrogate approximates velocity and pressure as functions of time, spatial coordinates, and variable parameters describing fluid properties, inlets/outlets, and geometry.The governing formulation uses incompressible Navier–Stokes equations with prescribed initial and boundary conditions.
- Network architecture: A structured fully-connected neural network encodes initial and boundary conditions by construction while mapping time, coordinates, and parameters to velocity and pressure states.A particular solution contributes the encoded conditions to the raw FC-NN outputs.
- Prediction: After offline training, the DNN provides rapid online flow-field predictions for any given parameter set and is expected to conform to conservation laws and specified conditions.Evaluating a trained FC-NN involves a few matrix multiplications and costs far less than a CFD simulation.
- Physics-constrained training: Training minimizes Navier–Stokes residuals through a physics-based loss, eliminating the need for CFD simulation data or numerical discretization for each parameter.The approach avoids solving the equations with traditional numerical methods for every parameter setting.
- Derivative computation: Automatic differentiation computes the first- and second-order derivatives required to form the PDE residuals in the training loss.The paper describes automatic differentiation as avoiding truncation and round-off errors associated with numerical differentiation.
3. Numerical Results
Numerical experiments on idealized 2D vascular flows show that the data-free physics-constrained DNN surrogate reproduces flow fields and uncertainty distributions well. The surrogate also enables rapid many-query uncertainty propagation without simulation training data.
- Experimental setup: The study evaluates circular pipe, stenotic, and aneurysmal flows, including uncertainty propagation from fluid properties and domain geometry to velocity, pressure, and wall shear stress.Only steady-state solutions are considered, so the initial-condition constraint is neglected.
- Flow-field accuracy: DNN predictions agree closely with analytical or CFD benchmarks across pipe, stenotic, and aneurysmal flows, capturing velocity, pressure, flow deceleration, and wall-shear-stress changes.For pipe flow, predicted velocity profiles almost exactly match analytical solutions across viscosity samples, while stenotic pressure drops and aneurysmal WSS variations are accurately captured.
- Uncertainty propagation: Propagated uncertainty distributions generally match CFD or analytical benchmarks, with nearly overlapping viscosity PDFs; geometry uncertainty is captured overall but has a slightly shifted peak density.The center-velocity distribution is non-Gaussian in the stenotic case, and aneurysmal viscosity-uncertainty PDFs almost coincide with CFD benchmarks.
- Computational efficiency: 500 Monte Carlo samples require about 10 CPU seconds with the surrogate, yielding a 2000-times speedup over CFD-based uncertainty propagation.Each forward DNN evaluation takes less than 2 × 10^-2 CPU seconds, compared with about 20,000 CPU seconds for the CFD-based test-case sampling.
4. Discussion
The discussion finds hard boundary-condition enforcement more reliable than soft enforcement for variable-radius flows, while activation adaptivity has a smaller effect. Physics-constrained, data-free learning avoids simulation-data overhead and remains promising despite acknowledged limitations.
- Boundary-condition enforcement: Hard boundary constraints outperform soft constraints for stenosis and aneurysm flows because soft enforcement poorly imposes vessel-wall no-slip conditions near bottlenecks.Circular pipe flow performs well with either approach, but variable-radius cases expose the soft constraint’s weakness.
- Boundary-condition enforcement: Soft boundary enforcement produces incorrect flow patterns and fails to capture the nonlinear centerline pressure profile compared with hard constraints and CFD.For the stenotic flow, soft-constraint predictions with λ = 1, 10, 100, and 1000 are compared against hard constraints and the CFD benchmark.
- Boundary-condition enforcement: λ = 1000 prescribes the boundary condition well but leaves a large PDE residual, whereas λ = 1 reduces x-momentum loss to O(10^-4) but fails to impose the BC.The competing equation and boundary losses make optimization difficult, and none of the tested λ values produces a physical stenotic flow pattern.
- Activation functions: Adaptive Tanh activation converges faster than Swish during the first 400 epochs, but all activation-function cases ultimately converge to the same loss order.With soft BCs, adaptive activations cannot prevent bad local minima, and the final loss is over one-order larger than with hard BC constraints.
- Activation functions: The study concludes that boundary-condition enforcement matters more than activation-function adaptivity for accuracy in physics-constrained data-free learning.The trainable activation hyperparameters are also more likely to reach optimized values when boundary conditions are enforced hard.
- Overall assessment: Physics-constrained DNNs learn parametric Navier–Stokes solutions without simulation data, avoiding offline CFD overhead while offering a mesh-free surrogate-modeling approach.Each CFD simulation takes about 40 CPU seconds, and the authors describe PDE-constrained DNN surrogate modeling as showing strong promise despite current limitations.
5. Conclusion
The paper presents a labeled-data-free structured DNN surrogate for parametric Navier–Stokes fluid flows, enforcing conditions architecturally and training through conservation-law violations. Tests on cardiovascular flows show excellent agreement with CFD while supporting uncertainty propagation across varying viscosities and geometries.
- Contribution: The proposed DNN surrogate models fluid simulations without labeled CFD data, addressing the expense and scarcity of simulation-generated training data.Each labeled datum requires an expensive CFD simulation, limiting efficient surrogate development.
- Method: A structured DNN automatically satisfies initial and boundary conditions and trains solely by minimizing mass and momentum conservation-law violations.The method approximates solutions of parametric Navier–Stokes equations without using simulation data.
- Experiments: The method was tested on circular pipe, stenotic, and aneurysmal flows relevant to cardiovascular applications, with viscosity and geometry variations.Uncertainties in these parameters were propagated through the trained surrogate and validated against CFD benchmarks.
- Results: Comparisons indicate excellent agreement between the physics-constrained DNN surrogates and CFD simulations, including velocity and pressure solutions without labeled training data.The surrogate accurately parameterizes solutions across varying viscosity and geometries and supports uncertainty propagation with enormous MC samples.
- Evaluation: The study also compared hard and soft IC/BC enforcement, adaptive activation functions, and labeled-data-free learning against traditional data-driven learning for accuracy and efficiency.The results demonstrate the method’s merit and promise for surrogate fluid models without CFD simulation data.
Appendix A
Appendix A reports the hardware used to train both the physics-constrained, data-free and purely data-driven models, and identifies performance tables for stenosis and aneurysm cases with varying viscosity and geometry.
- Training setup: Both the physics-constrained, data-free model and purely data-driven model were trained on a single NVIDIA 1080TI GPU card.The passage specifies the shared training hardware.
- Performance tables: Table 3 presents training and testing performance for the stenosis case with varying viscosity and geometry.The table is identified by its stated scope.
- Performance tables: Table 4 presents training and testing performance for the aneurysm case with varying viscosity and geometry.The table is identified by its stated scope.
Appendix B
Appendix B examines convergence histories for trainable adaptive-activation parameters under hard and soft boundary-condition enforcement. The parameters converge more readily with hard enforcement than with soft enforcement, while the legend distinguishes three layers.
- Appendix B: Adaptive-activation parameters are more likely to converge with hard boundary-condition enforcement than with soft enforcement.Hard enforcement corresponds to panels (a)–(e), whereas soft enforcement corresponds to panels (f)–(j).
- Appendix B: Figure 12 reports convergence histories for trainable adaptive-activation parameters under hard and soft boundary-condition enforcement.The parameters are β for Swish and n_a for Tahn.
- Appendix B: The legend labels a1, a2, and a3 as three different layers.These labels identify the layers represented in the convergence histories.