A Goal-Oriented Adaptive Sampling Procedure for Projection-Based Reduced-Order Models with Hyperreduction

Calista Biondic, Siva Nadarajah · 2025

Read A Goal-Oriented Adaptive Sampling Procedure for Projection-Based Reduced-Order Models with Hyperreduction by Calista Biondic, Siva Nadarajah (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2505-00712.txt.

Introduction

Calista Biondic and Siva Nadarajah improve how engineers pick training runs for fast stand-in models of aircraft flow, in a 2025 paper. Full air-flow simulations are costly, so designers use reduced models built from a sample of full runs. For nonlinear problems those stand-ins can lose their speed edge, since parts must be recomputed at every step.

Hyperreduction trims that recomputation by sampling only key spots in the flow. The team's adaptive method chooses the next full run based on a target output, like lift, rather than blind coverage. Goal-driven sampling spends each costly run where it most cuts error in the quantity that matters. Aircraft design, with its many what-if questions, is the driving use case.

Catalog record

CREATORCalista Biondic, Siva Nadarajah
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE119 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2505-00712

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Goal-Oriented Adaptive Sampling Procedure for Projection-Based Reduced-Order Models with Hyperreduction Calista Biondic, Siva Nadarajah Abstract Projection-based reduced-order models (PROMs) have demonstrated accuracy, reliability, and robustness in approximating high-dimensional, differential equation-based computational models across many applications. For this reason, it has been proposed as a tool for high-querying parametric design problems like those arising in modern aircraft design. Since aerodynamic simulations can be computationally expensive, PROMs offer the potential for more rapid estimations of high-fidelity solutions. However, the efficiency can still be tied to the dimension of the full-order model (FOM), particularly when projected quantities must be frequently recomputed due to non-linearities or parameter dependence. In the case of Petrov-Galerkin models, the projected residual and Jacobian are re-evaluated at every Newton iteration, thereby limiting the anticipated cost improvements. Hyperreduction is one of the tools available to approximate these quantities and address this issue. This work tests the energy-conserving sampling and weighting (ECSW) method as a potential approach for hyperreduction. It will be incorporated into the work in a previous article [ 1 ] which had developed an adaptive sampling procedure for building a reduced-order model (ROM) with a controlled functional error. The impacts of hyperreduction on computational cost and accuracy will be studied using the NACA0012 airfoil. Journal: Computers & Fluids 1 Introduction Many fields of design and engineering are increasingly reliant on computational models. In aircraft design, for instance, computational fluid dynamics (CFD) has become an essential tool for accelerating analysis and design processes and helps inform experimental testing decisions. Advancements in higher-order methods, faster solution algorithms, and parallel high-performance computing have resulted in access to high-fidelity solutions. These innovations have been instrumental in advancing the design of more fuel-efficient and environmentally friendly aircraft. In certain applications, the high computational costs and storage requirements of these models render them computationally intractable. For example, the time needed to produce high-fidelity CFD solutions poses a significant challenge in high-querying problems, such as aerodynamic shape optimization (ASO), where the model must be evaluated at numerous design parameter combinations, and in online real-time scenarios, like active control, where solutions must be computed within seconds. This limitation has driven substantial interest in methods to reduce the complexity and cost of large-scale computational models. Reduced-order modeling has been proposed as a possible solution to accelerate the process of computing new CFD solutions while preserving the fidelity achieved by higher-order methods. Reduced-order models can broadly be classified under two types: non-intrusive and intrusive models. Non-intrusive models treat the problem as a black box and are independent of the full-order equations. There are a multitude of challenges that arise when using these models, particularly in large-scale dynamic systems. These include violations of physical constraints, high costs associated with generating large training data sets, and lack of confidence indicators [ 2 ] . Examples of non-intrusive models include proper orthogonal decomposition coupled with interpolation [ 3 ] , Isomap [ 4 , 5 ] , Kriging [ 6 , 7 ] , and neural networks [ 2 , 8 , 9 , 10 ] . Intrusive models operate under the assumption that the solution of a large-scale system of dimension N can be represented on a much lower dimensional subspace of dimension n induced by parameter variation. The low-dimensional representation is found by projecting the governing equations onto the reduced-order subspace, which is why these models are often referred to as projection-based reduced-order models (PROMs). The two most commonly used approaches are the Galerkin projection [ 11 , 12 , 13 ] and the least-squares Petrov-Galerkin projection (LSPG) [ 14 , 15 , 16 ] . These models are often preferable to non-intrusive models as they are physics-informed, making them more robust, and they offer accessible error metrics [ 17 ] , which help predict the expected accuracy of the model online. Noting that this work focuses on solving fluid flows problems, least-squares Petrov-Galerkin (LSPG) PROMS have been demonstrated to be both numerically stable and accurate in the context of convection-dominated laminar flows and turbulent flow problems, where Galerkin PROMs often exhibit instability [ 18 ] . Therefore, an LSPG framework will be implemented in this work. The reduced-order subspace used in the projection is obtained by collecting solution snapshots and compressing this information in a reduced-order basis that spans the sampled solution space. Efficient construction of this basis can be a challenge, in particular, appropriate snapshot selection can directly impact the quality of the approximation. Some of the available sampling approaches include uniform sampling, Halton sequence method, Latin hypercube sampling, and greedy sampling [ 19 , 20 , 5 , 21 , 22 , 23 ] . The shortcomings of these approaches are detailed in [ 1 ] . They are compared with a new goal-oriented adaptive sampling approach developed in [ 1 ] , [ 17 ] , [ 24 ] , and [ 25 ] . The adaptive sampling procedure was designed to reduce the output error of an LSPG projection-based reduced-order model (ROM) to within a prescribed tolerance. The sampling procedure took advantage of dual-weighted residual (DWR) error indicators to estimate the error between the ROM and FOM at specific design parameter locations, which were then used to estimate the output error across the entire parameter space and determine new snapshot locations. It was shown to be more effective at building the reduced-basis than the other sampling approaches previously mentioned. The remaining major challenge not addressed in [ 1 ] is that the computational cost of the online stage of PROMs still scales with the dimension of the FOM, N . In parametric and/or non-linear problems, the parametric vectors and matrices such as the residual and Jacobian must be re-evaluated at each non-linear iteration and then projected onto the reduced-order basis. This can be computationally expensive, and the cost would scale with the dimension of the FOM. If not addressed, the expected time-saving benefits of the PROM cannot be fully realized. There are generally two approaches for addressing this computational bottleneck and approximating these quantities: exact and inexact [ 26 ] . Both share the fundamental strategy that underlies most PROM computations, of breaking down the evaluation of these quantities into offline and online components. Exact methods are applicable to specific classes of problems, two examples being parametric, linear FOMs admitting an efficient parameter-affine representation, and non-linear FOMs with low-order polynomial dependence of the internal force vector on the solution and a time-independent external force vector. These methods are often referred to as exact precomputation-based methods, as they compute the reduced matrices and vectors in two parts: the first part addresses the computational bottleneck and can be precomputed offline, while the second part, whose computational complexity scales with integer powers of the reduced-order subspace n , can be processed online and in real-time [ 26 ] . The problems addressed in this work, however, do not fall into the classes suitable for precomputation. Therefore, inexact or approximate reconstruction methodologies must be used. Hyperreduction is an inexact method that has been developed for addressing more general cases which can be both linear or non-linear, parametric or non-parametric [ 26 ] . These approaches introduce an additional layer of approximation to evaluate the high-dimensional quantities in

Read the complete text