PaperlayerView exact revision on arXiv

Source-linked AI summary

The GNAT method for nonlinear model reduction: effective implementation and application to computational fluid dynamics and turbulent flows

Kevin Carlberg, Charbel Farhat, Julien Cortial, David Amsallem

arXiv:1207.1349v1math.NAmath.APphysics.comp-ph

TL;DR

High-fidelity nonlinear CFD can be too expensive for time-critical analysis and large-scale computation. The paper develops GNAT as a fully discrete, residual-minimizing projection method with gappy-POD hyper-reduction and a sample-mesh implementation. It reports robust, accurate CFD reductions with substantially lower computational resource requirements.

  • Problem

    High-fidelity CFD simulations can require unaffordable computational resources or time, while reduced operators for general nonlinear systems remain costly to assemble.

  • Method

    GNAT operates on fully discrete equations, minimizes the discrete residual through Gauss–Newton Petrov–Galerkin projection, and uses gappy POD to sample tensor rows.

  • Results

    GNAT is demonstrated for parameterized moving-shock flow and large-scale CFD, with the work reporting orders-of-magnitude reductions in required computational resources.

  • Takeaways & Limitations

    The sample-mesh concept supports parallel online GNAT computations using the same flow solver and substantially fewer computational cores than the original CFD mesh.

  • Takeaways & Limitations

    GNAT’s reduced-order model is tied to the high-dimensional model’s time integrator and cannot generally be expressed as a low-dimensional ODE.

Abstract

from arXiv · show

The Gauss--Newton with approximated tensors (GNAT) method is a nonlinear model reduction method that operates on fully discretized computational models. It achieves dimension reduction by a Petrov--Galerkin projection associated with residual minimization; it delivers computational efficency by a hyper-reduction procedure based on the `gappy POD' technique. Originally presented in Ref. [1], where it was applied to implicit nonlinear structural-dynamics models, this method is further developed here and applied to the solution of a benchmark turbulent viscous flow problem. To begin, this paper develops global state-space error bounds that justify the method's design and highlight its advantages in terms of minimizing components of these error bounds. Next, the paper introduces a `sample mesh' concept that enables a distributed, computationally efficient implementation of the GNAT method in finite-volume-based computational-fluid-dynamics (CFD) codes. The suitability of GNAT for parameterized problems is highlighted with the solution of an academic problem featuring moving discontinuities. Finally, the capability of this method to reduce by orders of magnitude the core-hours required for large-scale CFD computations, while preserving accuracy, is demonstrated with the simulation of turbulent flow over the Ahmed body. For an instance of this benchmark problem with over 17 million degrees of freedom, GNAT outperforms several other nonlinear model-reduction methods, reduces the required computational resources by more than two orders of magnitude, and delivers a solution that differs by less than 1% from its high-dimensional counterpart.

1. Introduction

High-fidelity CFD is often too costly for time-critical applications, motivating nonlinear model-reduction methods. This work develops GNAT for CFD, combining discrete residual-minimizing projection with hyper-reduction and demonstrating robustness, accuracy, and efficient implementation.

  • High-fidelity CFD simulations can require unaffordable resources or completion times, limiting time-critical flow control, optimization, uncertainty quantification, and system identification.
  • Projection-based model reduction constructs a low-dimensional subspace offline and uses it online to compute approximate solutions in that subspace.
  • Assembling reduced operators can remain expensive because its cost scales with the large dimension of the underlying high-dimensional model.
  • Several existing nonlinear reduction approaches operate at continuous or semi-discrete levels, but many lack robustness for highly nonlinear unsteady CFD problems.
  • GNAT operates on fully discrete nonlinear equations, using a Petrov–Galerkin projection and function-sampling hyper-reduction designed around consistency and discrete-optimality conditions.
  • The work develops error bounds, introduces a parallelizable sample mesh, and demonstrates GNAT’s potential for CFD applications.

2. Problem formulation

The paper formulates parameterized nonlinear CFD as a high-dimensional state-and-output problem and targets fast, physics-based approximations. Its offline–online strategy constructs a surrogate from training simulations for time-critical evaluations across parameter inputs.

  • 2.1. Parameterized nonlinear CFD problem: The CFD state is a typically large vector of discrete conserved fluid variables governed by a nonlinear semi-discretized flux and source-term function.
  • 2.1. Parameterized nonlinear CFD problem: Input parameters may include shape, free-stream, design, and analysis variables, while outputs can include lift, drag, and other quantities.
  • 2.1. Parameterized nonlinear CFD problem: The formulation focuses occasionally on finite-volume discretizations over dual CFD meshes, while the concepts extend to finite-difference and stabilized finite-element methods.
  • 2.2. Objective: time-critical analysis: Fast evaluation is defined either by sufficiently short time for near-real-time analysis or by sufficiently low core-time resources for many-query applications.
  • 2.2. Objective: time-critical analysis: Directly solving the high-dimensional state equations and computing outputs can be prohibitively time- and resource-intensive when the CFD dimension is sufficiently large.
  • 2.2. Objective: time-critical analysis: An offline stage solves the model at training parameters and uses the resulting data to construct a surrogate for rapid online evaluations.
  • 2.2. Objective: time-critical analysis: Unlike data-fit input–output surrogates, model reduction approximately solves physics-based state equations for online inputs before computing outputs.

3. Overview of the GNAT model-reduction method

GNAT constructs fully discrete reduced models through residual-minimizing Petrov–Galerkin projection and gappy-POD hyper-reduction. Its design emphasizes discrete optimality, consistency, and online computations that avoid scaling with the high-dimensional CFD dimension.

  • Computational strategy: GNAT organizes the original high-dimensional model and two increasingly approximated reduced models into a three-tier hierarchy.Each reduced model is built from snapshots of the preceding model, snapshot compression, and an approximation exploiting the compressed snapshots.
  • Projection: GNAT seeks the solution in a low-dimensional affine trial subspace, representing the state increment with generalized coordinates.The trial dimension n_w is much smaller than the high-dimensional dimension N.
  • Projection: Residual minimization yields a discrete-optimal Petrov–Galerkin projection that minimizes the tier I residual over the trial subspace.The fully discrete framework makes the test basis depend on the discrete residual and time integrator.
  • Consistency: Consistency requires snapshot bases built from appropriate solution-increment data rather than only state snapshots.An alternative procedure using differences from the initial state also gives consistency under certain conditions.
  • Hyper-reduction: GNAT uses gappy POD to approximate residual and Jacobian tensors by evaluating only selected rows.The sample indices define the sampled quantities used in hyper-reduction.
  • Hyper-reduction: The online GNAT cost is typically very small because the required computations do not scale with the high-dimensional CFD dimension N.This addresses the nonlinear least-squares bottleneck that affects projection-based nonlinear model reduction.

3.5. Computation of outputs

GNAT computes outputs after the reduced simulation using only the state entries and basis rows that affect the desired quantities. This is efficient for spatially local outputs but not generally for spatially global ones.

  • Output reconstruction: GNAT computes outputs in post-processing from generalized coordinates and the needed rows of the initial condition and state basis.The output calculation accesses the subset K of state indices affecting the output.
  • Output reconstruction: Local outputs such as pointwise flow variables, lift, and drag require O(n_kn_w) operations when n_k ≪ N.The operation count is small because only the K rows of the basis are used.
  • Output reconstruction: Spatially global outputs do not generally satisfy the small-support condition needed for efficient output computation.Their dependence on many state entries can prevent the same reduced operation count.
  • Offline preparation: GNAT preprocessing constructs state, residual, and Jacobian POD bases from tier I and tier II training simulations.The snapshot procedures determine which quantities are collected during training.
  • Offline preparation: The method enforces n_J ≥ n_w and n_i ≥ n_R, n_J to support full rank and unique gappy-POD approximations.It also determines sample-index sets I, J, and K and constructs the associated sample meshes.

Online stage

The GNAT online stage consists of running the reduced simulation and then computing the requested outputs.

  • Online workflow: GNAT first applies the online ROM simulation algorithm for the specified online inputs.This produces the generalized coordinates needed for subsequent output evaluation.
  • Online workflow: GNAT then applies the output-computation algorithm to obtain the desired outputs.Output evaluation follows the reduced simulation as a separate online-stage step.

4. Error bounds

The paper develops global error bounds for fully discrete nonlinear model reduction with backward Euler and uses them to justify GNAT’s residual-minimization and POD choices. Several bound components are computable or estimable from reduced quantities and assumptions.

  • Bound construction: Global state-error bounds are developed for discrete nonlinear reduction methods using backward-Euler time discretization.The bounds are intended to expose the design advantages of GNAT.
  • Assumptions: The error analysis assumes inverse Lipschitz continuity, backward-Euler integration, and a high-dimensional solution satisfying an absolute residual tolerance.These assumptions define the stated scope of the bounds.
  • Bound consequences: Minimum-residual Petrov–Galerkin reduction minimizes b_n and therefore the tightest error bound in the stated recurrence.The reduced state is selected by minimizing the discrete residual over the trial subspace.
  • Bound consequences: With Φ_R = Φ_J, GNAT minimizes the projected residual term appearing in both c_n and d_n.Under this choice, GNAT iterations are equivalent to Gauss–Newton minimization of the projected residual norm.
  • Bound consequences: POD construction of Φ_R minimizes the average residual projection error over the residual snapshot set.That projection error appears as the final term of d_n.
  • Computability: The tightest bound is computable when the Lipschitz constant a is known or estimated, because b_n requires only Newton tolerance and residual norms.The projected residual is computable from sampled residual quantities, while the orthogonal projection error can be estimated.

5. Implementation of online computations and post-processing

GNAT online computations use purpose-built sample meshes containing only the nodes and geometrical entities needed for residual, Jacobian, state, or output evaluation. This preserves the existing CFD solver’s distributed implementation while reducing mesh size and computational cost.

  • 5.1. Sample mesh concept: Online GNAT CFD computations use only mesh information associated with index sets I and J during the main computation.The residual and Jacobian use I, while the fluid state uses J.
  • 5.1. Sample mesh concept: The online sample mesh contains nodes in I and J together with geometrical entities required by the finite-volume stencil.Its connectivity is imposed by the spatial discretization stencil rather than by an independent geometrical interpretation.
  • 5.1. Sample mesh concept: Output computation uses a separate mesh containing nodes in K and the geometrical entities needed to evaluate the requested outputs.For lift and drag, this is the wet surface mesh plus any solver-required entities.
  • 5.1. Sample mesh concept: The sample-mesh approach reuses the high-dimensional CFD solver for online reduced computations and supports the same distribution and parallelization strategy.Because the sample mesh is typically a tiny fraction of the original, it can require fewer cores and core-hours.
  • 5.2. Underlying node sampling algorithm: Algorithm 3 selects residual-sampling nodes to reduce gappy-POD residual-projection error while allowing overdetermined least-squares systems and separate residual and Jacobian bases.It operates on mesh nodes, balancing conservation equations without unnecessarily enlarging the sample mesh.
  • 5.2. Underlying node sampling algorithm: Sample-node selection must include boundary information and ensure every input variable affects the residual at at least one sampled node.The method requires at least one inlet or outlet sample node when such a boundary exists.

6. Applications

GNAT is evaluated on a parameterized moving-discontinuity problem and on turbulent Ahmed-body flow. The experiments show accurate predictions alongside substantial computational savings, while sample-mesh size and consistency affect convergence and performance.

  • Parameterized inviscid Burgers’ equation: GNAT is applied to a Burgers’-equation problem with a moving shock in a predictive scenario using changed online inputs.The study uses Godunov’s finite-volume scheme and a 4001-node grid producing a CFD model of dimension N = 4000.
  • Parameterized inviscid Burgers’ equation: 1.26% relative time-averaged discrepancy accompanies a GNAT prediction that closely matches the high-dimensional reference solution.The GNAT solution shows transient oscillations at t = 2.5 that dissipate over time.
  • Parameterized inviscid Burgers’ equation: 1167 times longer was required by the high-dimensional CFD solution than by the online GNAT solution for the Burgers’ problem.Both solutions were computed using a single processor.
  • GNAT performance assessment: CR improvement exceeds 230 for GNAT models using either snapshot-collection procedure 0 or 1.The sample-mesh implementation enables ROM simulations on as few as 4 cores; procedure 1 could support 438 predictive simulations for the core-hours of one high-dimensional computation.
  • GNAT performance assessment: Procedure 1 reproduces the Ahmed-body drag history almost perfectly and pressure contours nearly identically, including the complex wake region.Procedure 0 is less accurate, while procedure 1 also requires fewer Newton iterations per time step and runs faster.
  • Effect of node sampling and interpolation vs. least-squares approximation: Less than 1% discrepancy is obtained for the drag-coefficient time history across all three tested sample meshes.Adding sample nodes improves Newton convergence on average, but the smallest sample mesh is fastest; 253-node interpolation gives the best overall CPU performance in this case.
  • Effect of node sampling and interpolation vs. least-squares approximation: Both collocation approaches become nonlinearly unstable after a few flow-simulation time steps, whereas the DEIM-like approach performs poorly.These comparisons expose weaknesses of alternative hyper-reduction techniques for this highly nonlinear problem.

7. Conclusions

The paper develops GNAT with sample-mesh parallel implementation and global error bounds, then demonstrates parametric robustness and large-scale CFD acceleration while preserving accuracy.

  • Contributions: GNAT combines a sample-mesh implementation, global state-space error bounds, and demonstrations on parametric and highly nonlinear CFD problems.The applications include inviscid Burgers’ flow with moving shocks and turbulent flow over the Ahmed body.
  • CFD demonstration: GNAT reduces computational resources by more than two orders of magnitude for an Ahmed-body benchmark with over 17 million unknowns.The passage also states that GNAT preserves accuracy and outperforms several other nonlinear model-reduction methods.
  • Implementation: The sample-mesh concept eases implementation of GNAT’s online stage on parallel computing platforms.It is specifically introduced to support the distributed implementation of the method.

Appendix A. Proof of consistent snapshots for the state POD basis

The appendix proves that, under stated assumptions and without POD truncation, the Petrov–Galerkin projection reproduces the high-dimensional states for training inputs.

  • Assumptions: The proposition considers state snapshots formed either from deviations from the initial state or from successive state differences.These alternatives are W1 and W2, respectively, and either set is used to compute the state POD basis.
  • Assumptions: The proof assumes Gauss–Newton solves the nonlinear least-squares problem and converges to a stationary point in a level set containing only one stationary point.It also assumes that the reduced-order model uses the same initial condition as the high-dimensional model.
  • Result: The consistency result is ˜wn(µtrain) = wn(µtrain) for n = 0, . . . , nt.Thus, the Petrov–Galerkin ROM without hyper-reduction computes the same training-input states as the original CFD model.
  • Proof: The reduced residual is related to the full residual through the shared state history established inductively during the proof.The argument uses the equality of prior reduced and full states to identify the residual expressions.
  • Proof: For an untruncated POD basis, the snapshot spans contain the relevant state updates, allowing the full-order solution’s zero residual to satisfy the reduced projection condition.The proof applies this reasoning to both W1 and W2 snapshot constructions.

Appendix B. Further discussion of the various snapshot-collection procedures

The appendix links GNAT’s convergence behavior to the relationship between its residual and Jacobian bases, explaining why shared bases are preferred.

  • Shared bases: When ΦJ = ΦR and ΦR has orthonormal columns, GNAT inherits the convergence properties of Gauss–Newton.Under these conditions, the GNAT iterations are equivalent to Gauss–Newton iterations for nonlinear residual minimization.
  • Shared bases: Procedures 0 and 1 use the shared-basis configuration that motivates the corresponding snapshot-collection procedures.The stated rationale is the equivalence with Gauss–Newton under the shared-basis conditions.
  • Different bases: Procedures 2 and 3 may lack robustness because their different residual and Jacobian bases make the least-squares problem match quantities from different subspaces.The passage associates this mismatch with possible convergence difficulties.

Appendix C. Error bounds for the solution computed by a discrete nonlinear model reduction method

The appendix derives global solution-error bounds by relating residual discrepancies to local approximation errors, propagated state errors, and gappy POD approximation terms.

  • Error propagation: The residual difference between full and approximate solutions includes the current-state error, nonlinear-function discrepancy, and previous-step error.This decomposition begins the derivation of the local approximation-error bound.
  • Error propagation: Introducing an inverse Lipschitz constant transforms the residual relation into a bound on the local approximation error.With zero initial approximation error, the local bound yields a global-in-time result.
  • Gappy POD term: The gappy POD projector provides a computable residual component, ∥[ZΦR]+ Z ˜Rn( ˜wn+1)∥, within one global error bound.This quantity can be evaluated directly by GNAT.
  • Gappy POD term: The appendix also derives an upper bound for the gappy POD approximation error using the sample matrix and residual basis.The construction uses the QR factorization ZΦR = QR.
  • Global bound: Because bn ≤ cn ≤ dn, the resulting global bound combines the successive error-bound components at the n-th time step.The bound applies for 1 ≤ n ≤ nt.

Appendix D. Error bound for the gappy POD approximation

This appendix bounds the error of a gappy POD approximation by analyzing its projection matrix and the orthogonal projection error. The derivation relates the projection norm to the inverse triangular factor from a thin QR factorization.

  • The analysis considers approximating g ∈ R^N with a POD basis, sample indices, and a sample matrix formed from those indices.The number of sample indices satisfies n_i ≥ n_f.
  • The gappy POD projection matrix is P = Φ_g R^-1 Q^T Z, where ZΦ_f = QR is a thin QR factorization.The matrices Q and R have dimensions n_i × n_f and n_f × n_f, respectively.
  • The approximation decomposes as Pg = P(e + g*), with e denoting the difference between g and its orthogonal projection g*.Because g* lies in range(Φ_f), Pg* = g*, yielding (I − P)g = (I − P)e.
  • For a nontrivial projection matrix, ||I − P||_2 = ||P||_2, so the error bound can be expressed through the norm of P.The appendix then substitutes the projection-norm identity into the preceding bound.
Loading 1207.1349v1…