WorksComputer Science
Hyper-reduction methods for accelerating nonlinear finite element simulations: open source implementation and reproducible benchmarks
Abstract
Hyper-reduction methods have gained increasing attention for their potential to accelerate reduced order models for nonlinear systems, yet their comparative accuracy and computational efficiency are not well understood. Motivated by this gap, we evaluate a range of hyper-reduction techniques for nonlinear finite element models across benchmark problems of varying complexity, assessing the inevitable tradeoff between accuracy and speedup. More specifically, we consider interpolation methods based on the gappy proper orthogonal decomposition as well as the empirical quadrature procedure (EQP), and apply them to the hyper-reduction of problems in nonlinear diffusion, nonlinear elasticity and Lagrangian hydrodynamics. Our numerical results are generated using the open source libROM, Laghos and MFEM numerical libraries. Our findings reveal that the comparative performance between hyper-reduction methods depends on both the problem and the choice of time integration method. The EQP method generally achieves lower relative errors than interpolation methods and is more efficient in terms of quadrature point usage, resulting in a lower wall time for the nonlinear diffusion and elasticity problems. However, its online computational cost is observed to be relatively high for Lagrangian hydrodynamics problems. Conversely, interpolation methods exhibit greater variability, especially with respect to the use of different time integration methods in the Lagrangian hydrodynamics problems. The presented results underscore the need for problem specific method selection to balance accuracy and efficiency, while also offering useful guidance for future comparisons and refinements of hyper-reduction techniques.
Introduction
Partial differential equations (PDEs) are employed in numerous scientific fields to model the behavior of spatiotemporal systems. Despite their vast utility, only a few PDEs of practical interest have known analytical solutions under general conditions. As such, in most applications approximate PDE solutions are sought by various discretization methods, such as the finite differences method or the finite element method (FEM). In this study we consider the spatial discretization of time-dependent PDEs with algebraic constraints, resulting in systems of semi-explicit differential-algebraic equations (DAEs)
where denotes time and denotes physical parameters driving the system. The unknowns in the system include the differential variable , with a prescribed initial condition , and the algebraic variable . Moreover, is the system mass matrix, , are linear operator matrices, and is a constraint matrix. These matrices are assumed to be independent of physical parameters throughout the paper. Finally, and are nonlinear functions.
Time integration schemes such as multistep and Runge-Kutta methods can be used to solve the ODE part of the semi-explicit DAE system in (1), with solved algebraically. However, for many important physical systems the associated PDEs are nonlinear and require a fine mesh to resolve the solution, rendering the numerical modeling of such systems computationally expensive. This expense is compounded in multi-query applications such as PDE-constrained optimization [1, 2, 3, 4] where a large number of repeated PDE solves are necessary. For instance, design optimization of inertial confinement fusion experiments requires the determination of more than a dozen independent parameters. Depending on the fidelity of the physics model, each forward simulation takes a few minutes to hundreds of CPU hours to complete [5]. When employing a simple gradient based optimization algorithm, finite difference approximation of the objective function gradient requires as many radiation hydrodynamics simulations as the dimension of the parameter search space, resulting in prohibitively high computational cost for every parameter update.
For these reasons, it is of great interest to develop methodologies that accelerate the solving of large-scale nonlinear PDEs. One such methodology is reduced order modeling, where a full order model (FOM) is approximated with an accurate but computationally inexpensive reduced order model (ROM). ROMs can be divided into two main classes: data based and physics based. Purely data based methods such as dynamic mode decomposition (DMD) [6, 7] or neural network approximation [8, 9] have the benefit of not requiring prior knowledge of the PDE problem specification. However, they typically do not make accurate predictions in extrapolation and are not robust to noisy input data [10]. On the other hand, physics based ROM approaches incorporate prior knowledge into the ROM formulation and typically fare better in terms of generalization than purely data based approaches [10]. Physics based ROM approaches include physics informed neural networks [11], neural operators [12] and projection based reduced order models (pROMs), which are the class of ROMs studied in this paper. At a high level, a pROM projects the PDE onto a low dimensional subspace which approximates the PDE solution space. The projected PDE is then solved in subspace coordinates, leading to reduced computational cost. For a more comprehensive description of pROMs we refer the reader to the survey papers [13, 14].
Projection based ROMs have been shown to efficiently accelerate the solving of physics problems such as the Burgers equation [15, 16, 17, 18], the Navier-Stokes equations [19, 20], nonlinear diffusion problems [21, 22], convection-diffusion problems [23], the shallow water equations [24, 25], the Euler equations [26, 27, 28], wave problems [29, 30], Boltzmann transport problems [31] and the Schrödinger equation [32]. In addition, pROMs can enable substantial speed-up in design optimization problems such as lattice structure design [33], flutter avoidance wing shape optimization [34], topology optimization of wind turbine blades [35], reservoir simulations [36, 37, 38] and rocket nozzle shape design [39].
In the standard pROM formulation, the effects of the nonlinear terms and in the ROM subspaces are evaluated on the FOM mesh. This causes the pROM computational cost to scale with the FOM size, significantly reducing their efficiency. Hyper-reduction methods address this issue by approximating the nonlinear terms through evaluation at a reduced set of sample points, thereby achieving substantial computational speedup while maintaining a low approximation error. As indicated in [40], hyper-reduction methods can be broadly classified into two categories. The first category consists of interpolation or “approximate-then-project” methods, which interpolate the nonlinear terms at a subset of selected nodes before projecting them onto the reduced-order basis (ROB). This class of methods originates in the gappy proper orthogonal decomposition (POD) approach [41], which reconstructs high dimensional data using empirical basis functions. These methods rely on interpolation based approximations, where different sampling methods [42, 43, 44, 45, 46, 47] are introduced to select interpolation points for optimizing the tradeoff between accuracy and speedup. The second category comprises quadrature or “project-then-approximate” methods, which directly approximate the reduced nonlinear terms by constructing sparse quadrature rules. These methods can be viewed as data based quadrature techniques, where both the sample points and associated weights are optimized using training data. Examples include the seminal work by An et al. [48], the energy-conserving sampling and weighting method [49], the empirical cubature method [50] and the empirical quadrature procedure (EQP) [51, 52].
Given that hyper-reduction methods are being employed in applications across various domains, a comprehensive comparative analysis of their accuracy and efficiency is of significant interest. Existing works have provided valuable insights into specific hyper-reduction techniques, with recent efforts focusing on a single application at a time, such as nonlinear heat conduction [22, 53], magneto-mechanics [54, 55] and the Navier-Stokes equations [56].
In the present study we undertake a detailed empirical comparison of several hyper-reduction methods by applying them to a diverse selection of problem cases. More specifically, from the class of interpolation methods, we consider the Discrete Empirical Interpolation Method (DEIM) [43, 44], its subsequent development Q-DEIM [45], and S-OPT [47]. These methods were selected to capture distinct interpolation-based sampling strategies: DEIM serves as a widely adopted baseline, Q-DEIM improves the numerical conditioning of the interpolation operator through rank-revealing QR factorization, and S-OPT employs an optimization-based sampling procedure inspired by sensor placement and uncertainty quantification. For the class of quadrature methods, we consider EQP [52], which constructs sparse quadrature rules through non-negative least-squares optimization and has emerged as a widely used approach for hyper-reduction of nonlinear finite element simulations. Throughout this study, a two-stage workflow is employed to formulate and evaluate hyper-reduced projection-based ROMs. The workflow is split into an offline and an online stage, illustrated in Figure 1. First, the offline stage collects snapshots of FOM problem simulations, parameterized by . From the snapshots, the linear ROM basis is formed by POD. The FOM snapshots are additionally utilized in initializing the chosen hyper-reduction method. In the second online stage, the ROM is used to make accelerated predictions, again parameterized by . We apply the hyper-reduction methods in this study to three distinct and challenging nonlinear benchmark problems discretized by finite element methods: nonlinear diffusion, nonlinear elasticity, and Lagrangian hydrodynamics. The main features of our work are summarized below.

Figure 1. ROM workflow stages
Comparison. We present a broad comparison of hyper-reduction methods, including both interpolation and quadrature methods.
Application. We evaluate each considered hyper-reduction method on three application domains with distinct challenges, revealing tradeoffs between accuracy and computational speedup.
Open source implementation. We employ an open source numerical implementation of all hyper-reduction methods and problem cases considered in this work, using
the libROM [57], Laghos [58] and MFEM [59] numerical libraries and providing the commands that can be used for their reproduction.
The remainder of the paper is structured as follows. § 2 introduces a unified variational framework to describe the example problems and discusses their finite element discretization. § 3 serves as an introduction to the pROM approach used for the linear part of the ROM formulation. § 4 presents the different hyper-reduction methods compared in this work, which are used to approximate the evaluation of nonlinear terms. § 5 describes the example problems and presents our numerical results, followed by a discussion in § 6. Finally, Appendix § 9 provides details on how to reproduce our numerical results.
Finite element discretization
Considering (1) in a variational setting, we seek and such that , where and are Hilbert spaces and forms a Gelfand triple (rigged Hilbert space). In addition, we ask that the following equations be satisfied for any and
where denotes the duality pairing between and and the scalar product on . Additionally, , and denote bilinear forms, while and are general nonlinear functions.
The spatial discretization of (2) is carried out using the finite element method, the result of which is the system of differential-algebraic equations presented in (1). To that end, let be a basis for a -conforming finite element space , with the coordinate representation of with respect to . We denote the finite element approximation of by
Similarly, for the algebraic variables let be a basis for a -conforming finite element space , with the coordinate representation of with respect to . The finite element approximation of is denoted by
Then the differential variable vector and the algebraic variable vector satisfy (1), wherein the matrix representations are defined entrywise by
and can often be computed exactly, and the vector representations are computed as a semi-discretization of the nonlinear functions acting on the test functions,
As in many applications, we assume that the nonlinear residual functionals admit local integral representations over subdomains of the spatial domain . More precisely, we assume the existence of nonlinear densities and such that the nonlinear vectors can be assembled from their action on the finite element test functions. Using numerical quadrature, these actions are approximated by
where and denote quadrature weights, and and the associated quadrature points.
Projection based reduction
As mentioned in § 1, solving (1) can be computationally expensive when the FOM dimension is very large. In such cases, the solution procedure can be accelerated by projecting the PDE to the affine subspaces and for the differential and algebraic variables respectively, where
with and denoting the prescribed offsets. The subspaces and are spanned by the sets of basis vectors and respectively
The differential and algebraic state vector are then approximated using the subspace basis vectors
and it suffices to solve for the vectors of generalized coordinates of the differential and algebraic variables
To achieve the desired computational acceleration, ROMs typically leverage an offline-online splitting paradigm. In the offline phase (§ 3.1), which is performed once, computationally expensive tasks such as high-fidelity snapshots generation and basis construction are executed. In contrast, the online phase (§ 3.2) utilizes these precomputed quantities to assemble and solve the low-dimensional system for any new parameter configuration or time step, at a significantly lower computational cost than the offline phase. The specific procedures and computational complexities of these two phases are detailed in the subsequent subsections.
Offline phase
The offline phase is a one-time procedure aimed at constructing the low-dimensional subspaces and . To computationally manipulate the finite element functions introduced in (7), we first map them to their corresponding algebraic coordinate vectors in the Euclidean space via the isomorphisms and . Specifically, the prescribed offsets are mapped as
To find an optimal choice for and , one can follow the Proper Orthogonal Decomposition (POD) procedure. High-fidelity FOM simulations are conducted over a selected training set of parameters. For notational simplicity, each different variable snapshot is identified by a single index , and represented by a triplet , where is the corresponding time instance and is the corresponding parameter where -th state snapshot is simulated. Similarly, each algebraic variable snapshot is represented by a triplet , for . From these snapshots, we collect the vector representation of the differential and algebraic variables and decentralize:
The reduced bases can be subsequently extracted by performing a thin singular value decomposition (SVD) on these snapshot matrices:
with the singular values arranged in descending order. The choice of the basis dimensions and governs a critical trade-off: larger dimensions improve the accuracy of the reduced-order model by retaining more physical information, whereas smaller dimensions minimize computational complexity to maximize speed-up. To balance this, the dimensions and are typically determined via an energy criterion, selecting the minimal number of singular vectors required to capture a user-defined percentage and of the total snapshot energy, defined by the ratio of the cumulative sum of the singular values, i.e.,
where and are the diagonal entries of and respectively. The basis matrices and are then constructed from the first and left singular vectors of and respectively, i.e.,
where each column represents an individual finite element basis function
for and .
Online phase
The reduced system of DAEs is derived by performing the Galerkin projection of (1) onto the introduced subspaces. With the basis matrices (11) in place, the approximate variables (8) can be written in vector notation
Invoking the approximation (12) and applying Galerkin projection directly to (1), we obtain the reduced order model of dimension
with
and
Note that the nonlinear terms and are the projections of the image of the vector valued functions and onto the column spaces of and respectively. The projected functions are defined entrywise as
which involves numerical quadrature of the shown integrand functions over the spatial domain.
The nonlinear terms and defined in (16) need to be evaluated at every state update in a time marching scheme, and involve lifting back to the finite element functions, which scales with the FOM sizes and , before taking the nonlinear actions and . This dominates the ROM computational cost and prevents it from achieving significant speedup. To overcome this issue, the evaluation of and is approximated by sampling and at a small number of evaluation points, giving rise to various hyper-reduction methods.
In the next section we focus on hyper-reduction methods that are based on interpolation or quadrature, presenting their formulation and computational implementation.
Hyper-reduction
As explained in the previous section, the projection based reduction of a nonlinear problems leads to nonlinear terms that retain their dependence on the FOM size parameters and . To remove this dependence, various hyper-reduction methods have been developed. The central idea of hyper-reduction is to replace the evaluation of nonlinear terms and defined in (16) by approximations and which can be evaluated from the ROM coefficients without the need of full lifting to by (12) during the online phase. Invoking these approximations to (13), we obtain the hyper-reduced order model of dimension
In what follows, we focus on the hyper-reduction of ; the hyper-reduction of can be performed analogously. The hyper-reduction methods considered in this paper can be classified into interpolation-based (Section 4.1) and quadrature-based (Section 4.2). For each class of methods, we will describe in detail the acceleration mechanism in the online phase, the additional precomputed quantities required in the offline phase, and the loss-minimizing design of hyper-reduction schemes.
Interpolation methods
The acceleration mechanism of interpolation methods is to evaluate appropriately sampled entries of the nonlinear term and interpolate to the other entries. Based on gappy proper orthogonal decomposition [41], the interpolation methods uses a reduced basis approximation of
where
is a basis matrix for approximating the nonlinear term , which can be obtained in a similar manner to (11) from additional snapshots of in the offline phase, and denotes the coordinates of with respect to . The approximation (18) is performed by interpolation of at sampled indices, with . More precisely, let be a set of sampled indices and the associated sampling matrix whose entries are given by . Using the above, we formulate a least squares problem for the coordinates
whose solution is given by
where superscript † denotes the Moore-Penrose pseudoinverse. Combining with (15), this yields
where the matrix is small, which can be precomputed and stored offline. In the online phase, it suffices to evaluate , i.e., the values of at the sampling indices , and multiply to the precomputed matrix . In view of (5), the evaluation of only requires the values of the nonlinear density in the support of the selected test functions , which is referred to as the sample mesh, and avoids the expensive full lifting.
In addition to the basis matrices and described in Section 3.1, the offline procedure of interpolation-based hyper-reduction methods involves the construction of additional basis matrices and , and the selection of sampling indices and , which heavily affects the tradeoff between accuracy and speed-up. Noting that the matrix defines a rank- oblique projection onto the column space of , an error bound for interpolation methods as a function of is given in [47 Theorem 3.1]
where the 2-norm is equal to the largest singular value of , or equivalently the reciprocal of the smallest strictly positive singular value of . The same theorem provides the alternative expression
where is given by
Different greedy sampling strategies for adaptively minimizing the error can be devised based on the expressions (20)–(22), three of which are considered in this work. DEIM sampling methods (§ 4.1.1) aim to lower the error bound in (20) by minimizing the 2-norm of . Q-DEIM sampling methods (§ 4.1.2) have the similar goal of minimizing the condition number of to minimize the error bound. Finally, the S-OPT sampling algorithm (§ 4.1.3) aims first to increase the non-singularity of to increase numerical stability, and second to increase the column orthogonality of in order to decrease the 2-norm of .
DEIM
The discrete empirical interpolation method (DEIM) was introduced in [43] for greedily selecting sampling indices from the linearly independent rows of such Algorithm 1 Oversampled DEIM sampling algorithm
Input: Basis matrix and desired number of sampled indices
Output: Set of sampling indices
for do
Construct
for do
Construct
Compute
if then return
end for
end for
that the 2-norm of is minimized. A variant of this algorithm that allows for oversampling () was introduced in [60], a simplified version of which is presented in Algorithm 1.
Q-DEIM
The Q-DEIM algorithm was introduced in [45] as a variant of DEIM for selecting the sampling indices in a way that produces a smaller estimate of the condition number when . In contrast to the greedy selection process employed in DEIM, the Q-DEIM algorithm relies on a rank revealing QR factorization of . A further development of Q-DEIM termed GappyPOD+E was proposed in [61], which achieves low projection errors with few oversampling points by maximizing the lower bound of the smallest singular value of . This algorithm is presented in Algorithm 2.
S-OPT
The majority of gappy POD-based methods for selecting are based on the principle of minimizing the largest singular value of the sampled matrix . This is referred to as E-optimality [62] in the optimal design community and suffers from the issue of numerical instability [47]. On the other hand, the S-OPT algorithm introduced in [47] pursues S-optimality, which results in better numerical stability than Algorithm 2 GappyPOD+E sampling algorithm
Input: Basis matrix and desired number of sampled indices
Output: Set of sampling indices
, where , , and is upper triangular.
Take , with
for do
Construct from
, where , , , , and is diagonal
Compute where
Compute
Compute
end for
sampling methods based on E-optimality. The core principle of S-OPT is to simultaneously promote the nonsingularity of the normal matrix and the column orthogonality of the sampled matrix by maximizing measure [63]
The S-OPT algorithm as described in [47 Algorithm 1] is presented in Algorithm 3. In the developed algorithm, the determinant computations are performed using a fast method that exploits the rank-one updates involved in each iteration [63 Section 3.3.3].
Quadrature methods
The acceleration mechanism of quadrature-based methods is to approximate the integrations in the finite element assembly by an appropriately selected subset of quadrature integration points. As shown in (16), the nonlinear term is defined as projection of the vector valued function onto the derived reduced subspace. Recalling the duality pairings correspond to integration of the nonlinear integrands (5),
Algorithm 3 S-OPT sampling algorithm
Input: Basis matrix and desired number of sampled indices .
Output: A set of indices .
for do
Construct
Construct
Construct
, with , ,
end for
for do
Construct
Construct
, with
end for
the quadrature-based hyper-reduction methods seek to form a sparse and re-weighted approximation of (16) on the reduced subspace as the test function space, i.e.,
where the new weights are non-negative, i.e., , and sparse, i.e., with . In the online phase, the evaluation of only requires the ROM finite element approximation and where it takes to compute the nonlinear density at selected quadrature points , again called the sample mesh, and thus avoids the expensive full lifting.
In addition to the basis matrices and described in § 3.1, the offline procedure of quadrature-based hyper-reduction methods involves finding two sets of sparse quadrature weights and , which heavily affects the tradeoff between accuracy and speed-up. The computational expense of calculating scales with the number of sampled quadrature points, making it desirable to minimize . At the same time, the set of nonzero quadrature weights affects the quality of the approximation in (24), which in turn has a significant impact on the downstream ROM accuracy (in analogy to the role of in interpolation methods). It is therefore desirable to obtain a vector of weights that minimizes while keeping the approximation error below a chosen tolerance. The task of computing has been formulated as a linear programming problem [52] and as an equivalent nonnegative least squares (NNLS) problem [64, 65]. In this work we follow the latter approach and determine by solving the NNLS problem
where is an accuracy constraints matrix with constraint rows and is the original quadrature weights vector. The entries of each row of correspond to the evaluation of the integrand at all FOM quadrature points for a given combination of snapshot quadruplets introduced in § 3.1 and ROM test functions which gives rise to constraints. More specifically, we define entrywise by: for , , ,
After forming , the NNLS problem (25) is solved using the Lawson-Hanson algorithm [66], which is recited in Algorithm 4. Even though the NNLS problem (25) does not attempt to explicitly minimize , we find that significantly sparse solutions are obtained in our numerical experiments using a solver tolerance , matching the behavior demonstrated in previous works [64, 65].
Before applying the algorithm, two conditioning operations are performed on , which we have found to increase the derived quadrature rule’s accuracy and decrease the number of iterations needed for the algorithm’s convergence. First, each row of is rescaled by dividing the row by its maximum absolute value. In our numerical experiments we have found that within each row, nonzero entries may differ considerably in magnitude. Rescaling all entries to take values within was found to improve the convergence of the Lawson-Hanson algorithm. Second, we compute the LQ decomposition of matrix and use matrix as the constraints matrix. It has been previously reported that using a constraints matrix with orthonormal rows (constraints) consistently reduces the number of iterations needed for the Lawson-Hanson algorithm to converge [67], and this indeed matches our own experience. Additional details on the implementation of the EQP method and the NNLS formulation can be found in [28]. Finally, we note that generally the resulting sparse quadrature rule can only be expected to yield accurate solutions in the neighborhood of the differential variable snapshots used for forming . If desired, additional snapshots can be incorporated into to make the derived quadrature rule more robust. However,
Algorithm 4 Lawson-Hanson algorithm
Input: , and the solver tolerance
Output:
1:
2:
3:
4:
5:
6: while and do
7:
8:
9:
10:
11:
12:
13:
14:
15: while do
16:
17:
18:
19:
20:
21:
22:
23:
24: end while
25:
26:
27:
28: end while
increasing the number of snapshots will increase the offline time spent solving the NNLS problem and may also increase the resulting number of nonzero entries of .
Sample mesh
The hyper-reduction methods tested in this work require the selection of the number of sampling points ( for gappy POD methods and for EQP). Ideally, this selection is done such that the nonlinear functions and are accurately approximated while the computation time is minimized. However, the relation between the number of sampling points and the resulting computation time is not always straightforward. Our implementation of the tested hyper-reduction methods relies on the construction of a sample mesh, which consists of the finite elements containing the sampled quadrature or interpolation points. Each finite element in our computational domain includes a prescribed number of FEM basis functions (degrees of freedom) and associated quadrature points. The interpolation hyper-reduction methods sample degrees of freedom, corresponding to the sampled indices . Their implementation requires the evaluation of the sampled degrees of freedom at all local quadrature points. On the contrary, the EQP method samples quadrature points, determined by the nonzeros resulting from solving the NNLS problem (25). Its implementation requires the evaluation of all local degrees of freedom at the subset of sampled quadrature points. Because of this difference in approach between the two classes of methods, the computational cost of the tested hyper-reduction methods depends on element-wise factors that go beyond the number of sampled points. In many cases, the EQP method can also be implemented using pointwise integration, effectively viewing the whole computational domain as one “global element” to bypass the need for a sample mesh [64]. This is possible when the nonlinear terms allow for a clean offline-online splitting, where the geometric mapping and basis function gradients are precomputed in a fixed reference configuration. However, for the Lagrangian hydrodynamics problems (§ 5.3), the computational mesh is continuously deforming with time, while the governing equations involve complex state-dependent terms such as artificial viscosity that rely on local spatial stencils. Specifically, robust shock-capturing via artificial viscosity necessitates the calculation of velocity gradients and characteristic length scales derived from the instantaneous configuration of the local mesh [58]. Evaluating these element-based terms requires the online reconstruction of local geometric quantities (e.g., Jacobians and velocity gradients) and neighborhood connectivity at each time step. Since these geometric quantities evolve dynamically and are inextricably linked to the mesh topology, they cannot be fully precomputed in a fixed reference configuration. While alternative FOM discretizations (e.g., purely pointwise artificial viscosity) might allow for a fully mesh-independent implementation, they often lack the robustness required for complex shock-physics. As a result, the standard formulations used in the present work require access to the local mesh structure. This complex relationship between the sample mesh and ROM cost can lead to situations where using fewer sampled points does not necessarily lead to lower computational cost. Anticipating our later discussion, Figure 2 illustrates such a scenario for the triple point expansion problem (§ 5.3.3), where the sampled elements are visualized as volumes and the sampled points as dots. In this particular example, although fewer points are sampled by EQP than by S-OPT, the points sampled by EQP are scattered across many high order elements. This means that the sample mesh obtained by EQP contains significantly more elements than the one obtained by S-OPT. In such a case, the computational overhead associated with loading the sampled elements and reconstructing their local geometry can outweigh the savings gained from reducing the total number of quadrature points. As a result, the hyper-reduction speed-up becomes heavily dependent on the spatial distribution of the sampled points and the spatial sparsity of the sampled elements, rather than just the number of sampled quadrature points. These observations help highlight a key message: the employed FOM discretization can be an important factor in determining the speedup that can be delivered by a chosen hyper-reduction method.

Figure 2. Sample mesh comparison for the triple point expansion problem. Left: S-OPT sample mesh with 22 sampled elements and sampled quadrature points (in dark blue). Right: EQP sample mesh with 178 sampled elements and 150 sampled quadrature points each for the energy field (in green) and the velocity field (in red).
Numerical results
The inherent tradeoff for any hyper-reduction method between minimizing computational time and maximizing approximation accuracy can be quantified with a Pareto front.
Definition 5.1 (Pareto front) Given a finite solution set and objective functions , the Pareto front is defined as the set of solutions for which no satisfies both
In this study, we consider the case , with relative online phase computation time and relative error between FOM and ROM solutions as the two Pareto front objective functions. It is expected that the Pareto front, and thus the optimal choice of hyper-reduction method, will vary for different problems.
The hyper-reduction methods introduced in § 4 are applied to five problems: nonlinear diffusion (§ 5.1), nonlinear elasticity (§ 5.2) and three Lagrangian hydrodynamics problems: Sedov blast wave (§ 5.3.1), Taylor-Green vortex (§ 5.3.2) and triple point expansion (§ 5.3.3). Detailed settings for these problems are summarized in Table 1, including mesh size, finite element spaces , FOM size , the ODE solver, final time , and the number of time steps for FOM runs. For each considered problem case, the physical quantities are parametrized with . Figure 3 visualizes the final-time FOM solutions in various benchmark problems.
| Problem | Mesh | FE space | FOM size | ODE solver | ||
| Nonlinear diffusion | BE | |||||
| Nonlinear elasticity | RK4 | |||||
| Sedov blast wave | RK4<br>RK2Avg | <br> | ||||
| Taylor-Green vortex | RK4<br>RK2Avg | <br> | ||||
| Triple point expansion | RK4<br>RK2Avg | <br> |
Table 1. Summary of simulation setups.

Figure 3. Visualization of final-time fields from FOM simulations. The first column shows the nonlinear diffusion problem (top, pressure) and the nonlinear elasticity problem (bottom, velocity). The remaining columns correspond to Lagrangian benchmark problems, showing internal energy per unit mass (top) and velocity magnitude (bottom) for the Sedov blast wave (second column), the Taylor-Green vortex (third column), and the triple-point expansion problem (fourth column).
In the results that follow we use the error between FOM and ROM solutions to quantify the error introduced by the different hyper-reduction methods for each problem case. In addition to the hyper-reduction error, there are two additional sources of error included in the reported error values: (1) the projection error introduced by restricting the dynamics to a linear subspace of the original space; (2) for predictive simulation runs, the sampling error introduced when sampling each problem’s parameter space. The reason the FOM-to-ROM error can be employed to perform a fair comparison between different hyper-reduction methods is that the projection and sampling errors are constant among the different hyper-reduction methods for each problem case. This is because: (1) the same linear subspaces are used across all simulations within each problem case; (2) the same parameter values are sampled for training and used for out-of-sample testing for each problem. As such, even though the magnitude of the reported FOM-to-ROM errors includes additional error sources, their relative difference depends only on the employed hyper-reduction methods, allowing for fair comparisons within each problem case.
Nonlinear diffusion
We consider the nonlinear diffusion equation in
where denotes the pressure, denotes the flux, denotes an external source function, and denotes the initial condition on the pressure. The system is parametrized through the initial condition .
To fit the nonlinear diffusion equation (27) into the general variational framework (2) introduced in § 2 we identify the following components. The dependent variables are given by and , with associated vector spaces , and with the standard inner product , leading to the trivial Gelfand triple . The bilinear forms , and are defined as
for any , and , . Finally, the duality pairing involving the nonlinear functions and are given by
for any , and .
To generate our numerical results we consider a two dimensional problem defined on the unit square . The initial condition for pressure is parametrized by through
where , with conductivity and source . We perform the spatial discretization using a mixed finite element formulation, with the finite element spaces chosen such as to satisfy the inf-sup condition for stable mixed formulations. In particular, the variables and are approximated in and . We choose the lowest-order Raviart-Thomas elements and , where denotes the space of piecewise constant scalar functions on the mesh, and is the lowest-order Raviart-Thomas space consisting of vector fields whose normal components are continuous across element interfaces. Backward Euler (BE) method is used for temporal discretization with a fixed time step size up to final time .
Solutions at all the time steps with parameters are collected as snapshots for the predictive case and for the reproductive case in the offline stage. We set for the online stage in both the reproductive and the predictive case. The number of snapshots per training parameter is and . We perform both reproductive and predictive simulation runs, with a range of residual energy fractions summarized in Figure 4. In the figure, the basis dimensions satisfying different prescribed energy fractions are reported for each field, with the relation defined in (10).

Figure 4. Reduced basis dimensions with respect to the residual energy fraction for the nonlinear diffusion simulations.
Figure 5 illustrates how the ROM error varies with the number of sampled points for different hyper-reduction methods across a range of residual energy fractions. In every figure panel, the Pareto front between error of the pressure field, on the vertical axis, and number of sampled points on the horizontal axis, is shown for the different hyper-reduction methods. We note that the Pareto front points are optimal with respect to computational time and error, not necessarily with respect to the number of sampled interpolation/quadrature points. Every figure panel also includes a dashed reference line indicating the error of the ROM without hyper-reduction for the corresponding residual energy fraction to gauge the accuracy cost of hyper-reduction.

Figure 5. ROM error with varying sampled points for the nonlinear diffusion problem. The dashed line indicates the error of the ROM without hyper-reduction (‘noHR’).
As the residual energy fraction (or equivalently, ROM basis dimensions) increases, the errors of both interpolation methods and EQP converge toward the error of the ROM without hyper-reduction. At all methods in the reproductive case reach a relative error of about , whereas for the predictive case the lowest relative error is approximately . In both reproductive and predictive runs the error levels of the interpolation methods seems to correlate with and not with the number of sampled interpolation points. On the contrary, the EQP error levels demonstrate a tendency to decrease as the number of quadrature points increases, especially when is large. The lowest error levels obtained by each method are approximately of the same order of magnitude for this problem case. However, the EQP method seems to generally require a smaller number of quadrature points to reach a given error level.
The tradeoff between computational cost and relative error is made explicit in Figure 6, which depicts the Pareto fronts obtained for each hyper-reduction method among all values of . In the figure, the relative -error between ROM and FOM is plotted on the vertical axis and the relative wall time, i.e. the fraction between ROM wall time and FOM wall time for the online phase, is plotted on the horizontal axis. The overall Pareto front is drawn as a line connecting the most efficient configurations across all methods. In the reproductive case the overall Pareto front consists entirely of EQP method simulations. The lowest error level achieved is about , at which point EQP records a relative online time of approximately , while the interpolation methods record about . In the predictive scenario, although EQP shows greater efficiency for higher error levels, a tradeoff appears as the error levels decrease. More specifically, at error levels below or when the relative online time exceeds , the interpolation methods emerge on the overall Pareto front.

Figure 6. Pareto fronts across all hyper-reduction methods for the nonlinear diffusion problem, illustrating the tradeoff between relative error and online phase wall time.
To summarize, for the nonlinear diffusion problem, the Pareto optimal method is EQP in the reproductive case as well as in the predictive case for error levels above . For lower predictive error levels, interpolation methods also appear on the Pareto front, with no method consistently outperforming the others. The observed lower error levels in the predictive scenario compared to the reproductive one could be explained by the larger number of basis vectors retained for the same energy residual as shown in Figure 4. Furthermore, the samples taken in the offline stage seem to describe the system conditions tested in the online stage well. If one were to sample further away from its online stage value, then the obtained error would be expected to rise.
Nonlinear elasticity
We consider the dynamics of a visco-hyperelastic solid of homogeneous neo-hookean material undergoing finite strains, i.e. with both geometric and material nonlinearities. The state of the solid is described as a function of time and reference geometry at initial time . The system evolves according to the equations
where and denote the solid’s deformation and velocity fields, while and are the constant reference density and viscosity of the material in its initial configuration. In addition, denotes the material derivative and the normal vector on the boundary ; the spatial differential operators act with respect to the reference coordinates . The divergence of the first Piola-Kirchhoff stress tensor can be written as a nonlinear function , . Finally, denotes the initial condition on the velocity. The system is parametrized through the initial condition .
To write the nonlinear elasticity equations (28) in the general variational form (2) we make the following identifications. We denote the differential variable , and define the space and Hilbert space . There is no algebraic variable in this problem case. The inner product and bilinear form are given by
Finally, the nonlinear function is defined by its action
where denotes the dual of with respect to .
Our numerical results are generated for a two dimensional problem defined on the rectangular domain . The velocity initial condition is parametrized by through
with . The stress tensor is derived from a strain energy density function that involves the shear modulus and bulk modulus of the material and depends nonlinearly on the deformation gradient and its determinant , assuming finite strains,
where is a reference volumetric scaling and , yielding the stress tensor
For the spatial discretization of the nonlinear elasticity equations we use a single finite element space of dimension to approximate the displacement and velocity fields. Therefore, the differential variable is approximated in . We choose , where denotes the space of continuous piecewise tensor-product polynomials of degree at most 2 in each coordinate direction, and is its -component vector-valued counterpart. Therefore . Fourth-order explicit Runge-Kutta (RK4) method is used for temporal discretization with a fixed time step size up to final time .
Solutions and intermediate Runge-Kutta stages at all the time steps with parameters are collected as snapshots for the predictive case and for the reproductive case in the offline stage. In both cases we set for the online stage. The number of snapshots per training parameter is . Rather than constructing a single POD basis for the product space as described in § 3.1, we exploit its natural decomposition and compute independent reduced bases for each state component. We partition the differential variable into the velocity and displacement variables of equal full dimension. The reduced bases for the two variables are constructed separately from samples of FOM displacements and velocities. The dimensions of the reduced variables are denoted for the reduced displacement dimension and for the reduced velocity dimension. More precisely, the finite element representations of the velocity and displacement snapshots are assembled into two snapshot matrices, , and a thin singular value decomposition is performed separately on each matrix. The reduced basis dimensions are selected independently using the energy criterion (10), yielding the basis matrices , . The overall differential variable ROM subspace is then defined as the direct product of the corresponding reduced subspaces. Equivalently, the resulting basis functions are
supported on exactly one component of the product space, i.e.,
The overall reduced state dimension is then given by the sum of the two reduced dimensions. The reduced basis dimensions obtained for each tested residual energy fraction are shown in Figure 7.

Figure 7. Reduced basis dimensions used for each residual energy fraction in the nonlinear elasticity simulations.
Figure 8 and Figure 9 present an analogous comparison to that in Figure 5, with the displacement and velocity error results shown separately. The plotted points represent Pareto-optimal configurations for each hyper-reduction method, where Pareto optimality is determined with respect to computational time and the combined -product norm of the error in the state . The results demonstrate that for all methods the relative errors decrease as the number of interpolation/quadrature points increases. We note that the error levels for displacement are noticeably smaller than those for velocity. For displacement errors, in both the reproductive and predictive scenarios, EQP generally attains a given error level with fewer quadrature points than the interpolation methods. However, as the error approaches that of the ROM without hyper-reduction, the performance gap diminishes, and the interpolation methods may achieve comparable or even better efficiency. The interpolation methods exhibit relatively similar performance among themselves, and, with sufficiently many sampled points, all method converge toward the error of ROM without hyper-reduction. For the velocity field, interpolation methods reach the error level of ROM without hyper-reduction across all-scenarios, whereas EQP fails to do so in the reproductive scenario at low residual error fractions. In all cases, interpolation methods outperform EQP near the error level of the ROM without hyper-reduction. Overall, EQP tends to be more efficient at relatively high error levels, whereas interpolation methods become more efficient as higher accuracy is sought.

Figure 8. ROM displacement error with varying sampled points for the nonlinear elasticity problem.

Figure 9. ROM velocity error with varying sampled points for the nonlinear elasticity problem.
Figure Figure 10 displays the Pareto fronts obtained for each hyper-reduction method with respect to the displacement and velocity state errors and relative online wall time. For the displacement field, EQP composes the overall Pareto front across most error ranges in both scenarios, with Q-DEIM becoming Pareto optimal at the lowest error levels. The lowest attained relative error is approximately in the reproductive scenario and in the predictive scenario. For the velocity field, EQP similarly dominates the Pareto front in the reproductive scenario, except near error where interpolation methods are the highest performing, and at the lowest error levels where Q-DEIM and DEIM are Pareto optimal. Q-DEIM achieves the lowest error slightly above at a relative online time of approximately . In the predictive scenario for the velocity field, no single method consistently outperforms the others. The lowest attained error is approximately at relative online time , again by Q-DEIM. While no single method consistently dominates for the velocity field, EQP dominates the overall Pareto front for displacement errors across both scenarios, making it the more attractive choice when displacement accuracy is the primary concern.

Figure 10. Pareto fronts for displacement (top) and velocity (bottom) across all hyper-reduction methods in the nonlinear elasticity problem.
The Pareto fronts highlight the tradeoff between computational efficiency and accuracy. For the velocity field, the reproductive scenario allows accuracy to be improved progressively as computational efficiency is sacrificed, from relative errors above at relative online time below to about at relative online time above . In contrast, the predictive scenario cannot reduce the relative error below approximately , even at the highest computational cost. On the other hand, in the case of structural dynamics the displacement field can be of greater interest, and in this case a small relative error is achieved with the EQP at a significantly higher speedup. For instance, a relative error can with the EQP method be achieved with approximately speedup in the reproductive case and speedup in the predictive case.
Lagrangian hydrodynamics
We consider the hydrodynamics of a compressible fluid governed by the Euler equations in a Lagrangian reference frame. Similarly to the nonlinear elasticity case, the medium is described as a function of time and reference coordinates of the material particles at initial time . In the Lagrangian description of the medium’s motion, the physical quantities are expressed as a function of its initial position and time . The system of the Euler equations of compressible fluid dynamics in a Lagrangian reference frame is written as follows [68]
where denotes the material derivative, the density of the fluid, and the position and velocity of the particles, the internal energy per unit mass and the deformation stress tensor. In the equations, the spatial differential operators act with respect to the current coordinates . We consider an isotropic stress tensor where denotes the thermodynamic pressure and the artificial viscosity stress, both of which depend on the differential variables nonlinearly. The thermodynamic pressure is described by the equation of state and can be expressed as a function of the density and internal energy. In what follows, we focus on the case of polytropic ideal gas with an adiabatic index , which yields the equation of state
The system is prescribed with an initial condition and a boundary condition , where is the outward normal unit vector on the domain boundary.
The semidiscrete Lagrangian formulation involves a computational mesh that deforms based on the fluid velocity. Using the Reynolds transport theorem [68], the mass conservation in (30) is satisfied by
where denotes the Jacobian of the deformation gradient. Moreover, any integrals over the current domain involve the transformation to the reference domain using the determinant of the transformation gradient .
To write the remaining equations in (30) in the variational form (2) we define the following quantities. We denote the differential variable by and define the vector space . There is no algebraic variable . We define the Hilbert space with inner product
and denoted by its dual with respect to . The bilinear form is given by
while the action of the nonlinear function by
Following [58] we adopt a spatial discretization for (30) using a kinematic space of dimension for approximating the position and velocity functions, and a thermodynamic space of dimension for approximating the energy. The density can be eliminated, and the equation of mass conservation can be decoupled from (30). Therefore, the differential variable is approximated in . We choose and , where denotes the space of continuous piecewise tensor-product polynomials of degree at most 2 in each coordinate direction, so that is the corresponding -component vector-valued space, and denotes the space of discontinuous piecewise tensor-product polynomials of degree at most 1 in each coordinate direction. Therefore, . Each problem case is time integrated using two methods: the averaged second-order explicit Runge Kutta (RK2Avg) method for conserving the total energy [58, 69] and the standard fourth-order explicit Runge Kutta (RK4) method. In either method, adaptive time step control is employed for numerical stability. We refer the reader to [58] for additional details.
We perform numerical simulations of three problem cases: Sedov blast wave, Taylor-Green vortex and triple point expansion, introduced in the following sections. For each problem we focus on reproductive simulations, where the constructed reduced models are tested using initial conditions that were part of their training data.
Solutions and intermediate Runge-Kutta stages at all the time steps and parameters are collected as snapshots. We make two modifications to the offline basis construction procedure outlined in § 3.1. First, we employ the time windowing technique, where the simulation time interval is decomposed into a number of time sub-intervals (windows)
Instead of building a single ROM for the whole time interval , multiple ROMs are constructed for piecewise approximating the local dynamics. In practice, the number of time windows used is determined by prescribing the number of snapshots to be collected per window. The snapshots are clustered into groups by the time they are simulation into the corresponding time window. Second, as in the previous example, we assemble finite element representations of the velocity, internal energy, and mesh position snapshots for the time window into three snapshot matrices, , , , and create the basis matrices , , which defines the product ROM basis matrix by
where . This separate basis construction allows flexible separate state update necessary in the RK2Avg scheme. For each problem, the number of snapshots and residual energy fraction are such as to ensure an adequate basis size for each window. Finally, we also employ an adaptive timestepping strategy to dynamically adjust the timesteps during each simulation [26, 58].
The parameters used for the reduced simulations of each problem are summarized in Table 2, including the number of time windows () and the total number of snapshots () and reduced basis dimensions ( and ) summed over all time windows, i.e., and A detailed development of the reduction process using interpolation hyper-reduction methods is given in [26] and using quadrature methods in [28].
| Problem | ODE solver | |||||
| Sedov blast wave | 6 | RK4 | 15 | 1451 | 118 | 37 |
| Sedov blast wave | 6 | RK2Avg | 15 | 754 | 151 | 40 |
| Taylor-Green vortex | 6 | RK4 | 2 | 1064 | 8 | 22 |
| Taylor-Green vortex | 6 | RK2Avg | 2 | 532 | 8 | 23 |
| Triple point expansion | 6 | RK4 | 2 | 379 | 11 | 10 |
| Triple point expansion | 6 | RK2Avg | 2 | 190 | 12 | 10 |
Table 2. Parameters used for the reduced simulations of Lagrangian hydrodynamics problems.
Sedov blast wave
The Sedov blast wave problem is a three dimensional standard shock hydrodynamic test, involving the release of energy from a localized source and the formation of a propagating shock wave [70]. We consider a localized source of internal energy initially deposited at the origin of a three dimensional cube, in analogy to the two-dimensional test in [58]. The computational domain is the unit cube with wall boundary conditions on all surfaces, . We prescribe initial velocity and initial density . The initial energy is given by a delta function supported at the origin. In our implementation, the delta function energy source is approximated by setting the internal energy to zero in all degrees of freedom except at the origin, with . The adiabatic index in the ideal gas equation of state is set to .
Figure 11 illustrates how the ROM error varies with the number of sampled interpolation/quadrature points, averaged over the number of time windows, across different hyper-reduction methods. Our comparisons focus on the achieved ROM error between the two ODE solvers (RK4 and RK2Avg) under a fixed residual energy fraction as given in Table 2. To obtain the presented results we increase the number of sampled interpolation points for interpolation methods and the number of nonzero sparse quadrature weights for EQP within each time window. The figure compares the errors for each of the three state variables: position, velocity and energy. The position field generally exhibits the smallest error, while the ordering of the energy and velocity errors depends on the case. For both solvers, the EQP method is the most efficient in terms of the number of quadrature points, achieving a given error level with approximately one order of magnitude fewer quadrature points. For the RK4 solver, EQP attains lower errors than the interpolation methods for the position and velocity fields; however, for the energy field, EQP does not achieve the lowest error levels, with the best interpolation results being nearly one order of magnitude smaller. Among the interpolation methods, DEIM generally demonstrates the most efficient performance in terms of the number of sampled interpolation points. For the RK2Avg solver, the interpolation methods achieve lower errors than EQP across all fields considered. Among them, S-OPT shows a slight advantage for the velocity field at lower error levels, whereas the position and energy fields exhibit comparable performance.

Figure 11. ROM error with varying sampled points for the Sedov blast wave problem.
Comparing the two solvers, EQP shows little difference in error between RK4 and RK2Avg, with RK4 yielding slightly smaller errors for the velocity and energy fields. In contrast, the interpolation methods exhibit more pronounced solver dependence. For the velocity and position fields, RK2Avg enables an improvement of approximately one order of magnitude in the attainable error, reaching error levels of approximately for the position field and for the velocity field. For the energy field, however, RK2Avg results in errors that are roughly one order of magnitude larger than those obtained with RK4, with minimum errors around . Overall, while EQP is highly efficient in terms of the number of sampled points, the achievable error levels across methods vary with the solver and state component.
Figure 12 reveals the tradeoff between approximation accuracy and time, showing the Pareto fronts for -product norm of the error in the state , where all differential variables are aggregated into a single error metric rather than shown separately. For the RK4 solver, the EQP method dominates the overall Pareto front at error levels below , with a relative online time of approximately .

Figure 12. Pareto fronts across hyper-reduction methods for the Sedov blast wave problem.
Among the interpolation methods, DEIM achieves the best performance and occupies the upper segment of the overall Pareto front, reaching its lowest error level near at a relative online time of about . In the case of the RK2Avg solver, the interpolation methods constitute the Pareto-optimal set, with no single method dominating the others. The lowest velocity error is obtained by DEIM, achieving an error level of at a relative online time of about . Although the interpolation methods require considerably more sampled interpolation points than EQP, their relative online times are comparable or even smaller. This could be due to computational overhead stemming from the EQP sample mesh implementation, as discussed in the introduction of § 5. We also note the solver dependent performance of each hyper-reduction method. For EQP, the error and relative online time are similar across solvers. For S-OPT, both solvers exhibit gradual tradeoffs: RK4 spans error levels between and over relative online times from to , whereas RK2Avg reaches lower error levels between and at smaller relative online times from to . For DEIM and Q-DEIM, both methods exhibit a relatively steep tradeoff under RK4, with the error dropping from to over a narrower relative online time range of approximately to . Under RK2Avg, however, DEIM and Q-DEIM follow a more gradual tradeoff, similar to that observed for S-OPT. Overall, RK2Avg paired with interpolation methods yields the most favorable performance across the results considered.
Taylor-Green vortex
The Taylor-Green vortex problem is a three dimensional standard benchmark test for the incompressible inviscid Navier-Stokes equations. In this work we manufacture a smooth solution by extending the steady state Taylor-Green vortex solution to the compressible Euler equations, in analogy to the two-dimensional test in [58]. Our computation domain is the unit cube with wall boundary conditions on all surfaces, . Denoting , the initial velocity is given by
and the initial thermodynamic pressure by
with initial density . Based on the above, the initial energy is determined by the equation of state (31) with adiabatic index set to .
Figure 13 presents the ROM errors across hyper-reduction methods with respect to the average number of interpolation/quadrature points. With the RK4 solver, the EQP method requires fewer quadrature points than the interpolation methods to reach a given error level across all variables. More specifically, EQP achieves roughly an order of magnitude lower velocity error than the interpolation methods while using fewer quadrature points. Among the interpolation methods, S-OPT tends to exhibit the most efficient performance in terms of the number of quadrature points required for a given velocity error. With the RK2Avg solver, however, a different trend is observed. We note that the EQP method is converging to the no-hyper-reduction ROM performance level for all sampled values of . Although EQP requires fewer quadrature points to reach near no-hyper-reduction ROM performance, its velocity error levels remain higher than those of the interpolation methods. For approximately quadrature points, the interpolation methods achieve lower velocity errors than EQP. The minimum velocity error obtained with the interpolation methods is approximately one order of magnitude smaller than that of EQP, albeit with a larger number of sampled interpolation points. Overall, EQP demonstrates greater efficiency with the RK4 solver, whereas with RK2Avg the interpolation methods yield higher accuracy in velocity.

Figure 13. ROM error with varying sampled points for the Taylor-Green vortex problem.
Figure 14 focuses on the tradeoff between computational time and -product norm of the error in the state across different hyper-reduction methods. With the RK4 solver, the Pareto-optimal configurations progress from DEIM to S-OPT and finally to EQP as the error level decreases. The lowest total error is achieved by EQP, reaching approximately with a relative online time slightly below . With RK2Avg the overall Pareto front is dominated by the interpolation methods. They attain a total error of about at a relative online time of , and around at . Among them, Q-DEIM generally forms the Pareto front. To compare the performance of each method depending on the solver, all interpolation methods perform better with RK2Avg than with RK4. For EQP, the minimum obtained total error is slightly smaller with the RK4 solver, while the corresponding relative online times are of a similar magnitude. Overall, the interpolation methods exhibit shorter relative online times across both solvers. However, when using RK4 the lowest error is obtained by EQP at around , making it the preferred option for accuracy below . The lowest total error among all configurations is attained by Q-DEIM with the RK2Avg solver, reaching the level.

Figure 14. Pareto fronts across hyper-reduction methods for the Taylor-Green vortex problem.
Triple point expansion
The triple point expansion problem is a three dimensional hydrodynamic shock test that features two materials in three different states. More specifically, we consider a three state, two material, 2D Riemann problem which generates vorticity, analogous to the two dimensional test in [58]. The computational domain is with wall boundary conditions. The initial velocity is given by . Denoting , the initial density is given by
and the initial thermodynamic pressure by
The adiabatic index is set to
with the initial energy then determined by (31).
Beginning with Figure 15 we see that EQP achieves the error level of ROM without hyper-reduction with both solvers. With RK4 this error is obtained with quadrature points, while with RK2Avg a comparable error level is reached with roughly half as many quadrature points. Contrary to EQP, the accuracy of the interpolation methods varies significantly between the two solvers. With RK4 these methods are less efficient, exhibiting minimum velocity errors around , more than two orders of magnitude larger than EQP. With RK2Avg their performance improves considerably, reaching minimum velocity errors of about , which represents a marked improvement over RK4, albeit still higher than EQP. The interpolation methods typically require between and sampled interpolation points, exceeding those required by EQP. Overall, EQP attains low error levels using fewer quadrature points across both solvers.

Figure 15. ROM error with varying sampled points for the triple point expansion problem.
Figure 16 shows that, although the interpolation methods use more sampled points than EQP, their relative online times are comparable or even smaller for the triple-point problem. As discussed in § 5.3.1, this discrepancy between the number of quadrature points and relative online wall time could be due to differences in sample meshes between EQP and interpolation-based methods; see Figure 2. For both solvers, the Pareto optimal configurations are initially composed of interpolation methods, with EQP taking over for lower error levels. At the lowest error levels, EQP achieves total errors of approximately with relative online times of about for RK2Avg and for RK4. On the contrary, the interpolation methods exhibit lower relative online times and reduced accuracy. With the RK4 solver they constitute the Pareto optimal set when the relative online time is below , corresponding to total errors around . With RK2Avg they remain Pareto optimal when the relative online time is below , with total errors of or higher. Consistent with earlier observations, these methods perform significantly better with RK2Avg than with RK4. Overall, EQP remains the most accurate method across both solvers, whereas interpolation methods offer a clear accuracy–efficiency tradeoff.

Figure 16. Pareto fronts across hyper-reduction methods for the triple point expansion problem.
Conclusion
We introduced a variational framework for describing PDEs solved by hyper-reduced models. Five example problems were stated in terms of this framework and solved using four hyper-reduction methods: three gappy POD-based interpolation methods and the EQP quadrature method. We compared the relative performance of these hyper-reduction methods by computing the Pareto optimal front with respect to speed-up and accuracy. We would like to remark that the scope of this comparison is limited to the selected methods and benchmark problems considered herein. While the presented results provide insight into the relative performance of the studied interpolation-based and quadrature-based methods, they should not be interpreted as a general comparison between these two classes of hyper-reduction approaches.
For the nonlinear diffusion and elasticity problems, we considered two types of simulations: (1) reproductive, where the reduced model is tested for parameter values that were used in its training; (2) predictive, where the reduced model predicts solutions for parameter values not used during its training. In the Lagrangian hydrodynamics problems, we considered only reproductive simulations. Our numerical experiments demonstrate that the performance of the tested hyper-reduction methods depends on the problem and time integration method used, with no single method being universally superior.
In the diffusion problem, the EQP method was generally more efficient in terms of relative online time in the reproductive scenario, reaching similar error levels to the interpolation methods but with higher speed-up. In the predictive scenario, the differences among the methods were less pronounced, with EQP and interpolation methods displaying similar performance at lower error levels.
A similar trend appeared for the nonlinear elasticity problem, both for the displacement and velocity fields: EQP was generally more efficient in the reproductive scenario and less dominant in the predictive one. The Q-DEIM method achieved the lowest errors across both scenarios and for both displacement and velocity, at the expense of a low relative speed-up of only 40%.
For the Lagrangian hydrodynamics problems, the performance of the interpolation methods was strongly influenced by the employed time integration method: interpolation methods performed consistently better with RK2Avg than with RK4. On the other hand, EQP did not demonstrate a clear dependency on the selected time integration method.
For interpolation methods, each selected mesh element includes all associated interpolation points, whereas EQP selects quadrature points individually from the whole mesh. This can explain why EQP achieved higher efficiency with respect to the number of quadrature points in most of our numerical results. However, due to computational overhead in the creation of the sampled mesh elements, this advantage did not translate to the expected efficiency in relative online time in our Lagrangian hydrodynamics test cases. The testing of these three hydrodynamics problems focuses on reproductive simulations to keep the comparisons as simple and clear as possible.
As seen in the other problems, predictive simulations tend to have larger error than reproductive ones, with more error components. In particular, there is approximation of the solution using a fixed reduced basis, in addition to approximation of the hyper-reduced terms. The larger number of error components makes a straightforward comparison of hyper-reduction methods difficult for more complex problems such as the Lagrangian hydrodynamics ones. Testing on reproductive simulations enables comparison of the hyper-reduction approximation error. It also simplifies comparison of computational time, by avoiding the need to balance the choice of basis size and number of hyper-reduction samples for fast and accurate approximation of the predictive solution.
Overall, our findings demonstrate that the EQP method can provide a significant speed-up at high accuracy levels for a broad range of problems. Nevertheless, interpolation methods can achieve a higher speed-up or better accuracy in certain cases. The relative performance of the studied hyper-reduction methods depends on the problem case, choice of time integration method, and whether the tested simulation is reproductive or predictive. These results highlight the need for problem-specific selection of hyper-reduction strategies when choosing between interpolation and quadrature methods.
Acknowledgments
This work was supported in part by the U.S. Department of Energy, Office of Science, Office of Advanced Scientific Computing Research, Scientific Discovery through Advanced Computing (SciDAC) program through the LEADS SciDAC Institute under Project Number SCW1933 at Lawrence Livermore National Laboratory (LLNL). LLNL is operated by Lawrence Livermore National Security LLC, for the U.S. Department of Energy, National Nuclear Security Administration under Contract DE-AC52-07NA27344. IM release number: LLNL-JRNL-2013738. Axel Larsson was supported by the U.S. National Science Foundation Grant award number 2122269. Part of this work was performed while Axel Larsson, Minji Kim and Chris Vales were research interns at LLNL supported by the Mathematical Sciences Graduate Internship (MSGI) program administered by the U.S. National Science Foundation. Chris Vales also acknowledges support from the U.S. Department of Energy under grant DE-SC0025101. The authors thank Adrian Humphry and Masayuki Yano from the University of Toronto for their valuable feedback on an earlier version of this manuscript and for providing a piece of the code used in our implementation of the empirical quadrature procedure.
Disclaimer
This document was prepared as an account of work sponsored by an agency of the United States government. Neither the United States government nor Lawrence Livermore National Security LLC nor any of their employees makes any warranty, expressed or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States government or Lawrence Livermore National Security LLC. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States government or Lawrence Livermore National Security LLC and shall not be used for advertising or product endorsement purposes.
Command line options
To facilitate the reproduction of our results, we present the command line options used for each problem. Examples are provided for the predictive scenarios in § 5.1 and 5.2 and for the reproductive scenario with the RK2Avg solver in § 5.3. Each problem corresponds to a specific executable: mixed_nonlinear_diffusion in libROM, nonlinear_elasticity_global_rom in libROM and laghos in Laghos. Due to ongoing development in the repository, the presented commands are subject to change. The following commands are compatible with the recent commits of the master branch of MFEM 1, the master branch of libROM 2, and the rom branch of Laghos 3.
In order to use the ROM capability of Laghos, the user has to navigate to the rom subdirectory.
The conventions used in the examples are as follows. For Sections A.1 and A.2, the following apply to the online phase:
EQP method: use the options
-eqp -ns; the number of quadrature points can be restricted with-maxnnls.Interpolation methods: specify the hyper-reduction method via
-hrtypeand the number of sampling points with-nsr.
For Section A.3:
-s 7corresponds to the RK2Avg solver and-s 4to RK4; RK2Avg is used in the examples.EQP method: requires the
-eqpoption in the merge phase; in the preprocessing and online phases use-hrsamptype eqpand-lqnnls.Interpolation methods: specify the hyper-reduction type in the preprocessing phase with
-hrsamptype; the number of sampling points is controlled by-sfacvand-sface.
Nonlinear diffusion
The following commands are used for the nonlinear diffusion problem in § 5.1. The working directory is assumed to be libROM/examples/prom.
Offline phase (predictive scenario):
./mixed_nonlinear_diffusion -p 1 -offline -id 0 -sh 0.25
./mixed_nonlinear_diffusion -p 1 -offline -id 1 -sh 0.15
./mixed_nonlinear_diffusion -p 1 -offline -id 2 -sh 0.35
Merge phase:
./mixed_nonlinear_diffusion -p 1 -merge -ns 3
Create the FOM comparison data:
./mixed_nonlinear_diffusion -p 1 -offline -id 3 -sh 0.3
Online phase for EQP:
./mixed_nonlinear_diffusion -p 1 -online -rrdim 6 -rwdim 3 -id 3 -sh
0.3 -eqp -ns 3
Online phase for S-OPT:
./mixed_nonlinear_diffusion -p 1 -online -rrdim 6 -rwdim 3 -id 3 -sh
0.3 -nldim 6 -nsr 100 -hrtype sopt
Nonlinear elasticity
The following commands are used for the nonlinear diffusion problem in § 5.2. The working directory is again assumed to be libROM/examples/prom.
Offline phase (predictive scenario):
./nonlinear_elasticity_global_rom -offline -dt 0.01 -tf 5.0 -s 14 -
vs 100 -sc 0.9 -id 0
./nonlinear_elasticity_global_rom -offline -dt 0.01 -tf 5.0 -s 14 -
vs 100 -sc 1.1 -id 1
Merge phase:
./nonlinear_elasticity_global_rom -merge -ns 2 -dt 0.01 -tf 5.0
Create the FOM comparison data:
./nonlinear_elasticity_global_rom -offline -dt 0.01 -tf 5.0 -s 14 -
vs 100 -sc 1.0 -id 2
Online phase for EQP:
./nonlinear_elasticity_global_rom -online -dt 0.01 -tf 5.0 -s 14 -vs
100 -eqp -ns 2 -rvdim 3 -rxdim 2 -hdim 1 -sc 1.0
Online phase for S-OPT:
./nonlinear_elasticity_global_rom -online -dt 0.01 -tf 5.0 -s 14 -vs
100 -hyp -hrtype sopt -rvdim 3 -rxdim 2 -hdim 5 -nsr 100 -sc
1.0
Lagrangian hydrodynamics
The following commands are used for the Lagrangian hydrodynamics simulations in § 5.3. The working directory is assumed to be Laghos/rom
Problem specification: First, we present the commands used to set up each problem.
Sedov blast wave problem:
./laghos -m data/cube01_hex.mesh -p 1 -pt 211 -s 7 -tf 0.2Taylor-Green vortex problem:
./laghos -m data/cube01_hex.mesh -p 0 -cfl 0.1 -s 7 -tf 0.1Triple-point expansion problem:
./laghos -m data/box01_hex.mesh -p 3 -cfl 0.5 -s 7 -tf 0.5
ROM simulation: Next, we present the command line options used in each stage of the ROM simulation for the Sedov blast wave problem. These options should be appended to the problem setup command for Sedov, except for the merge phase, which is executed independently. The same procedure applies to the other problems by replacing the problem specification accordingly. Below, we provide example commands for EQP and S-OPT.
Offline phase:
-offline -writesol -romsns -rpar 0 -rostype interpolate -sdim 10000 -sample-stages
Merge phase for EQP (executed independently):
./merge -nset 1 -romsns -romos -rostype interpolate -nwinsamp 50 -ef 0.999999 -eqpMerge phase for S-OPT (executed independently):
./merge -nset 1 -romsns -romos -rostype interpolate -nwinsamp 50 -ef 0.999999
Hyper-reduction preprocessing phase for EQP:
-online -romhrprep -romsns -rostype interpolate -nwin 15 -hrsamptype eqp -lqnnlsHyper-reduction preprocessing phase for S-OPT:
-online -romhrprep -romsns -rostype interpolate -nwin 15 -hrsamptype sopt -sfacv 2 -sface 2
ROM online phase for EQP:
-online -romhr -romsns -rostype interpolate -nwin 15 -hrsamptype eqp -lqnnlsROM online phase for S-OPT:
-online -romhr -romsns -rostype interpolate -nwin 15 -sfacv 2 -sface 2
Postprocessing phase (solution reconstruction and error evaluation):
-restore -romsns -rostype interpolate -nwin 15 -soldiffReferences
- [1]Tarantola, A.: Inverse Problem Theory and Methods for Model Parameter Estimation. SIAM, Philadelphia (2005)DOI
- [2]Biegler, L.T., Ghattas, O., Heinkenschloss, M., Bloemen Waanders, B.: Large-scale PDE-constrained optimization: an introduction. In: Large-scale PDE-constrained Optimization, pp. 3–13. Springer, Berlin (2003)
- [3]Biegler, L.T., Ghattas, O., Heinkenschloss, M., Keyes, D., Bloemen Waanders, B.: Real-time PDE-constrained Optimization. SIAM, Philadelphia (2007)
- [4]Gunzburger, M.D.: Perspectives in Flow Control and Optimization. SIAM, Philadelphia (2002)
- [5]Anirudh, R., Archibald, R., Asif, M.S., Becker, M.M., Benkadda, S., Bremer, P.-T., Budé, R.H., Chang, C.-S., Chen, L., Churchill, R., et al.: 2022 review of data-driven plasma science. IEEE Trans. Plasma Sci. (2023)DOI
- [6]Schmid, P.J.: Dynamic mode decomposition of numerical and experimental data. J. Fluid Mech. 656, 5–28 (2010)
- [7]Brunton, S.L., Proctor, J.L., Kutz, J.N.: Discovering governing equations from data by sparse identification of nonlinear dynamical systems. Proc. Natl. Acad. Sci. USA 113(15), 3932–3937 (2016)DOI
- [8]Lagaris, I.E., Likas, A., Fotiadis, D.I.: Artificial neural networks for solving ordinary and partial differential equations. IEEE Trans. Neural Networks 9(5), 987–1000 (1998) https://doi.org/10.1109/72.712178
- [9]Merwe, R., Leen, T.K., Lu, Z., Frolov, S., Baptista, A.M.: Fast neural network surrogates for very high dimensional physics-based models in computational oceanography. Neural Networks 20(4), 462–478 (2007) https://doi.org/10.1016/j.neunet.2007.04.023
- [10]Karniadakis, G.E., Kevrekidis, I.G., Lu, L., Perdikaris, P., Wang, S., Yang, L.: Physics-informed machine learning. Nat. Rev. Phys. 3(6), 422–440 (2021)DOI
- [11]Raissi, M., Perdikaris, P., Karniadakis, G.E.: Physics-informed neural networks: a deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations. J. Comput. Phys. 378, 686–707 (2019)DOI
- [12]Kovachki, N., Li, Z., Liu, B., Azizadenesheli, K., Bhattacharya, K., Stuart, A., Anandkumar, A.: Neural operator: learning maps between function spaces with applications to PDEs. J. Mach. Learn. Res. 24(89), 1–97 (2023)DOI
- [13]Gugercin, S., Antoulas, A.C.: A survey of model reduction by balanced truncation and some new results. Int. J. Control 77(8), 748–766 (2004)
- [14]Benner, P., Gugercin, S., Willcox, K.: A survey of projection-based model reduction methods for parametric dynamical systems. SIAM Rev. 57(4), 483–531 (2015)DOI
- [15]Veroy, K., Prud’homme, C., Patera, A.T.: Reduced-basis approximation of the viscous burgers equation: rigorous a posteriori error bounds. C.R. Math. 337(9), 619–624 (2003) https://doi.org/10.1016/j.crma.2003.09.023
- [16]Choi, Y., Carlberg, K.: Space–time least-squares Petrov–Galerkin projection for nonlinear model reduction. SIAM J. Sci. Comput. 41(1), 26–58 (2019)arxiv.org/abs/1703.04560
- [17]Choi, Y., Coombs, D., Anderson, R.: SNS: a solution-based nonlinear subspace method for time-dependent model order reduction. SIAM J. Sci. Comput. 42(2), 1116–1146 (2020)
- [18]Carlberg, K., Choi, Y., Sargsyan, S.: Conservative model reduction for finite-volume models. J. Comput. Phys. 371, 280–314 (2018)DOI
- [19]Xiao, D., Fang, F., Buchan, A.G., Pain, C.C., Navon, I.M., Du, J., Hu, G.: Non-linear model reduction for the Navier–Stokes equations using residual DEIM method. J. Comput. Phys. 263, 1–18 (2014)DOI
- [20]Burkardt, J., Gunzburger, M., Lee, H.-C.: POD and CVT-based reduced-order modeling of Navier–Stokes flows. Comput. Methods Appl. Mech. Eng. 196(1-3), 337–355 (2006)DOI
- [21]Hoang, C., Choi, Y., Carlberg, K.: Domain-decomposition least-squares Petrov-Galerkin (DD-LSPG) nonlinear model reduction. Comput. Methods Appl. Mech. Eng. 384, 113997 (2021)
- [22]Fritzen, F., Haasdonk, B., Ryckelynck, D., Schöps, S.: An algorithmic comparison of the hyper-reduction and the discrete empirical interpolation method for a nonlinear thermal problem. Math. Comput. Appl. 23(1), 8 (2018)
- [23]Mojgani, R., Balajewicz, M.: Lagrangian basis method for dimensionality reduction of convection dominated nonlinear flows. arXiv:1701.04343 (2017)arxiv.org/abs/1701.04343
- [24]Zhao, P., Liu, C., Feng, X.: POD-DEIM based model order reduction for the spherical shallow water equations with Turkel-Zwas finite difference discretization. J. Appl. Math. (2014)DOI
- [25]Stefanescu, R., Navon, I.M.: POD/DEIM nonlinear model order reduction of an ADI implicit shallow water equations model. J. Comput. Phys. 237, 95–114 (2013)
- [26]Copeland, D.M., Cheung, S.W., Huynh, K., Choi, Y.: Reduced order models for Lagrangian hydrodynamics. Comput. Methods Appl. Mech. Eng. 388, 114259 (2022) https://doi.org/10.1016/j.cma.2021.114259
- [27]Cheung, S.W., Choi, Y., Copeland, D.M., Huynh, K.: Local Lagrangian reduced-order modeling for the Rayleigh-Taylor instability by solution manifold decomposition. J. Comput. Phys. 472, 11655 (2023) https://doi.org/10.1016/j.jcp.2022.111655
- [28]Vales, C., Cheung, S.W., Copeland, D.M., Choi, Y.: Machine-precision energy conservative reduced models for Lagrangian hydrodynamics by quadrature methods. arXiv:2508.21279 (2025) https://doi.org/10.48550/arXiv.2508.21279
- [29]Glas, S., Patera, A.T., Urban, K.: A reduced basis method for the wave equation. Int. J. Comput. Fluid Dyn. 34(2), 139–146 (2020)DOI
- [30]Cheung, S.W., Chung, E.T., Efendiev, Y., Leung, W.T.: Explicit and energy-conserving constraint energy minimizing generalized multiscale discontinuous Galerkin method for wave propagation in heterogeneous media. Multiscale Model. Simul. 19(4), 1736–1759 (2021)arxiv.org/abs/2009.00991
- [31]Choi, Y., Brown, P., Arrighi, B., Anderson, R., Huynh, K.: Space-time reduced order model for large-scale linear dynamical systems with application to Boltzmann transport problems. J. Comput. Phys. 424, 109845 (2021)arxiv.org/abs/1910.01260
- [32]Cheng, M.-C.: A reduced-order representation of the Schrödinger equation. AIP Adv. 6(9), 095121 (2016)DOI
- [33]McBane, S., Choi, Y.: Component-wise reduced order model lattice-type structure design. Comput. Methods Appl. Mech. Eng. 381, 113813 (2021)DOI
- [34]Choi, Y., Boncoraglio, G., Anderson, S., Amsallem, D., Farhat, C.: Gradient-based constrained optimization using a database of linear reduced-order models. J. Comput. Phys. 423, 109787 (2020)arxiv.org/abs/1506.07849
- [35]Choi, Y., Oxberry, G., White, D., Kirchdoerfer, T.: Accelerating design optimization using reduced order models. arXiv:1909.11320 (2019)arxiv.org/abs/1909.11320
- [36]Jiang, R., Durlofsky, L.J.: Implementation and detailed assessment of a GNAT reduced-order model for subsurface flow simulation. J. Comput. Phys. 379, 192–213 (2019)
- [37]Yang, Y., Ghasemi, M., Gildin, E., Efendiev, Y., Calo, V., et al.: Fast multiscale reservoir simulations with POD-DEIM model reduction. SPE Journal 21(06), 2–141 (2016)DOI
- [38]Wang, M., Cheung, S.W., Chung, E.T., Vasilyeva, M., Wang, Y.: Generalized multiscale model for fractured vuggy carbonate reservoirs. J. Comput. Appl. Math. 366, 112370 (2020)arxiv.org/abs/1810.12419
- [39]Amsallem, D., Zahr, M., Choi, Y., Farhat, C.: Design optimization using hyper-reduced-order models. Struct. Multidiscip. Optim. 51(4), 919–940 (2015)DOI
- [40]Grimberg, S., Farhat, C., Tezaur, R., Bou-Mosleh, C.: Mesh sampling and weighting for the hyperreduction of nonlinear Petrov-Galerkin reduced-order models with local reduced-order bases. Int. J. Numer. Methods Eng. 122(7), 1846–1874 (2021)arxiv.org/abs/2008.02891
- [41]Everson, R., Sirovich, L.: Karhunen–Loève procedure for gappy data. J. Opt. Soc. Am. A 12(8), 1657–1664 (1995)DOI
- [42]Barrault, M., Maday, Y., Nguyen, N.C., Patera, A.T.: An ‘empirical interpolation’ method: application to efficient reduced-basis discretization of partial differential equations. C.R. Math. 339(9), 667–672 (2004) https://doi.org/10.1016/j.crma.2004.08.006
- [43]Chaturantabut, S., Sorensen, D.C.: Nonlinear model reduction via discrete empirical interpolation. SIAM J. Sci. Comput. 32(5), 2737–2764 (2010)DOI
- [44]Carlberg, K., Farhat, C., Cortial, J., Amsallem, D.: The GNAT method for nonlinear model reduction: Effective implementation and application to computational fluid dynamics and turbulent flows. J. Comput. Phys. 242, 623–647 (2013)DOI
- [45]Drmac, Z., Gugercin, S.: A new selection operator for the discrete empirical interpolation method—improved a priori error bound and extensions. SIAM J. Sci. Comput. 38(2), 631–648 (2016)DOI
- [46]Drmac, Z., Saibaba, A.K.: The discrete empirical interpolation method: canonical structure and formulation in weighted inner product spaces. SIAM J. Matrix Anal. Appl. 39(3), 1152–1180 (2018)
- [47]Lauzon, J.T., Cheung, S.W., Shin, Y., Choi, Y., Copeland, D.M., Huynh, K.: S-OPT: a points selection algorithm for hyper-reduction in reduced order models. SIAM J. Sci. Comput. 46(4), 474–501 (2024)DOI
- [48]An, S.S., Kim, T., James, D.L.: Optimizing cubature for efficient integration of subspace deformations. ACM Trans. Graph. 27(5), 1–10 (2008) https://doi.org/10.1145/1409060.1409118
- [49]Farhat, C., Chapman, T., Avery, P.: Structure-preserving, stability, and accuracy properties of the energy-conserving sampling and weighting method for the hyper reduction of nonlinear finite element dynamic models. Int. J. Numer. Methods Eng. 102(5), 1077–1110 (2015)DOI
- [50]Hernandez, J.A., Caicedo, M.A., Ferrer, A.: Dimensional hyper-reduction of non-linear finite element models via empirical cubature. Comput. Methods Appl. Mech. Eng. 313, 687–722 (2017)
- [51]Patera, A.T., Yano, M.: An LP empirical quadrature procedure for parametrized functions. C.R. Math. 355(11), 1161–1167 (2017) https://doi.org/10.1016/j.crma.2017.10.020
- [52]Yano, M., Patera, A.T.: An LP empirical quadrature procedure for reduced basis treatment of parametrized nonlinear PDEs. Comput. Methods Appl. Mech. Eng. 344, 1104–1123 (2019) https://doi.org/10.1016/j.cma.2018.02.028
- [53]Bhattacharyya, S., Tao, J., Gildin, E., Ragusa, J.C.: Hyper-reduction techniques for efficient simulation of large-scale engineering systems. Arch. Comput. Methods Eng., 1–43 (2025)DOI
- [54]Brands, B., Davydov, D., Mergheim, J., Steinmann, P.: Reduced-order modelling and homogenisation in magneto-mechanics: a numerical comparison of established hyper-reduction methods. Math. Comput. Appl. 24(1), 20 (2019)DOI
- [55]Delagnes, T., Henneron, T., Clenet, S., Fratila, M., Ducreux, J.: Comparison of hyper-reduction methods combined with pod: Model order reduction of a squirrel cage induction machine in nonlinear case. IEEE Trans. Magn. 60(6), 1–10 (2024)DOI
- [56]Romor, F., Stabile, G., Rozza, G.: Explainable hyper-reduced order models on nonlinearly approximated solution manifolds of compressible and incompressible navier-stokes equations. J. Comput. Phys. 524, 113729 (2025)arxiv.org/abs/2308.03396
- [57]Choi, Y., Arrighi, W.J., Copeland, D.M., Anderson, R.W., Oxberry, G.M., USDOE NNSA: libROM (2019). https://www.osti.gov//servlets/purl/1505575
- [58]Dobrev, V.A., Kolev, T.V., Rieben, R.N.: High-order curvilinear finite element methods for Lagrangian hydrodynamics. SIAM J. Sci. Comput. 34(5), 606–641 (2012)DOI
- [59]MFEM: Modular Finite Element Methods. mfem.org. https://doi.org/10.11578/dc.20171025.1248
- [60]Carlberg, K., Bou-Mosleh, C., Farhat, C.: Efficient non-linear model reduction via a least–squares Petrov–Galerkin projection and compressive tensor approximations. Int. J. Numer. Methods Eng. 86, 155–181 (2011)DOI
- [61]Peherstorfer, B., Drmac, Z., Gugercin, S.: Stability of discrete empirical interpolation and gappy proper orthogonal decomposition with randomized and deterministic sampling points. SIAM J. Sci. Comput. 42(5), 2837–2864 (2020)arxiv.org/abs/1808.10473
- [62]Pukelsheim, F.: Optimal Design of Experiments. SIAM, Philadelphia (2006)
- [63]Shin, Y., Xiu, D.: On a near optimal sampling strategy for least squares polynomial regression. J. Comput. Phys. 326, 931–946 (2016)DOI
- [64]Du, E., Yano, M.: Efficient hyperreduction of high-order discontinuous Galerkin methods: element-wise and point-wise reduced quadrature formulations. J. Comput. Phys. 466, 111399 (2022) https://doi.org/10.1016/j.jcp.2022.111399
- [65]Sleeman, M.K., Yano, M.: Goal-oriented model reduction for parametrized time-dependent nonlinear partial differential equations. Comput. Methods Appl. Mech. Eng. 388, 114206 (2022) https://doi.org/10.1016/j.cma.2021.114206
- [66]Lawson, C.L., Hanson, R.J.: Solving Least Squares Problems. SIAM, Philadelphia (1995)
- [67]Humphry, A., Yano, M.: Efficient hyperreduction for large-scale problems: exploiting reducible constraint manifolds in empirical quadrature procedure. Int. J. Numer. Meth. Eng. 126(23), 70204 (2025) https://doi.org/10.1002/nme.70204
- [68]Aris, R.: Vectors, Tensors, and the Basic Equations of Fluid Mechanics. Dover, New York (1989)
- [69]Sandu, A., Tomov, V., Cervena, L., Kolev, T.: Conservative high-order time integration for Lagrangian hydrodynamics. SIAM J. Sci. Comput. 43(1), 221–241 (2021)DOI
- [70]Sedov, L.I., Volkovets, A.: Similarity and Dimensional Methods in Mechanics. CRC Press, Boca Raton (2018)