Hybrid reduced order model cuts simulation costs both offline and online

Engineers building digital twins of physical systems face a stubborn trade-off: the surrogate models that make real-time simulation possible are expensive to create, expensive to run, or both. A new study published in Advanced Modeling and Simulation in Engineering Sciences proposes a way out of this dilemma with a two-stage reduced order modeling strategy that is computationally efficient both during its construction and during its use. The work, by Rahul Dhopeshwar, Harshit Bansal and Karen Veroy, was published on 6 August 2026 as an open access research article.

Reduced order models, or ROMs, are compressed mathematical stand-ins for full-scale simulations of parametrized partial differential equations. Instead of solving a problem with thousands or millions of finite element degrees of freedom, a ROM projects the physics onto a low-dimensional basis of representative solution shapes. This makes them key enablers for digital twins, which need fast, repeated evaluations to support real-time monitoring, simulation-based decision-making and optimization of complex systems. The authors focus on model-driven surrogates for data-poor scenarios, where machine learning approaches trained on large datasets are not practical.

The difficulty arises with nonlinear problems. In a linear setting, the reduced equations can be assembled cheaply because contributions can be precomputed. For nonlinear problems, however, evaluating the nonlinear operators still requires operations in the high-dimensional finite element space, and these cannot be precomputed. Hyper-reduction techniques address this by approximating the nonlinear terms with a small set of interpolation points, but most of them rely on Newton solvers whose iterations must repeatedly evaluate the approximate nonlinearities. Regression-based alternatives, which learn the mapping from parameters to reduced coefficients using methods such as neural networks or Gaussian processes, are fast online but demand many full-order solutions during training. A comparison study cited by the authors confirms the bottleneck: hyper-reduction is offline-efficient but online-inefficient, while regression is the reverse.

The new approach, called a two-stage reduced order model or 2sROM, chains the two families of methods so that each covers the other’s weakness. In the first offline stage, the researchers build an intermediate model using the Reduced Basis Empirical Quadrature Procedure, RBEQP, a hyper-reduction technique based on an empirical quadrature rule. The reduced basis is constructed with a greedy algorithm that iteratively adds full-order solutions at the parameter values worst approximated by the current basis, guided by a cheap residual-based error estimator. Because full-order solves are performed only at greedily selected parameters, this stage needs far fewer expensive simulations than snapshot-based methods like Proper Orthogonal Decomposition.

The empirical quadrature procedure is central to the efficiency of this intermediate model. Rather than using the full-order quadrature rule, whose number of points scales with the finite element dimension, the method finds a sparse set of quadrature points and weights that keeps the reduced solution accurate within a tolerance. This is formulated as a constraint minimization problem seeking the fewest non-zero weights subject to a manifold accuracy constraint, solved in practice through a relaxed linear programming form or an iterative non-negative least squares approach with early stopping, which yields sparse rules. The authors also introduce a small number of smoothing iterations, in which the RBEQP problem is solved with the current quadrature rule and the resulting solutions are used to reevaluate the rule, letting the quadrature and the state approximations coevolve. In practice, two to three such iterations proved sufficient.

In the second offline stage, the intermediate RBEQP model becomes a data factory. Instead of running costly full-order simulations to generate the large training set that Gaussian process regression requires, the researchers use the cheap intermediate model to produce reduced basis coefficients across a much larger training parameter set, sampled with scrambled Sobol sequences. The regression is trained with an anisotropic Matérn 5/2 kernel with automatic relevance determination, which assigns each parameter dimension its own characteristic length scale and suits the limited regularity of the reduced coefficients. The result is a Gaussian process model whose online predictions involve no assembly costs and no Newton iterations, making the final model fast to evaluate for any new parameter value.

The team tested the framework on two nonlinear model problems. The first is steady-state heat conduction on the unit square with a temperature-dependent conductivity and four material parameters, solved on a mesh with 1472 degrees of freedom. The second is a neo-Hookean hyperelastic cantilever beam with a hole under self-weight, parametrized by Young’s modulus, Poisson ratio, weight density and gravity angle, with 3708 degrees of freedom. Accuracy was measured as the root mean squared relative H1 error over 1600 test samples, and all computations ran on a laptop with an Intel Core i7-10750H CPU and 32 GB of RAM, using the FEniCS library for full-order simulations and GPy for the regression.

The results show a clear separation of labor between the stages. For the heat conduction problem, the offline cost of 2sROM is dominated by the RBEQP stage, so enlarging the reduced basis raises offline time much more than increasing the number of GPR training samples, which are generated cheaply by the intermediate model. Compared at similar error levels, 2sROM consumed significantly less training time than PODGPR, the regression-based baseline trained directly on full-order snapshots, with the savings especially pronounced at small greedy batch sizes. In the online phase, 2sROM behaved like PODGPR and was much faster than RBEQP, which still carries the overhead of assembly and Newton iterations. For the beam problem, with the offline budget fixed at 1500 seconds, 2sROM trained its GPR component on 1600 parameter samples against 800 for PODGPR; at small basis dimensions 2sROM showed higher errors because its training data came from less accurate intermediate solutions, but as the basis grew it achieved significantly lower errors thanks to its denser training set.

The headline numbers come from the conclusion. To reach an average error below 3 percent on the nonlinear heat conduction problem, 2sROM required about 4 times less offline training time than PODGPR and ran about 7 times faster online than RBEQP, with similar results for the hyperelasticity problem. Among the three methods compared, RBEQP achieved the lowest error but was the most expensive to evaluate, while 2sROM offered what the authors describe as a good balance between accuracy and speed. The authors note that the framework is most useful for problems with a fast-decaying Kolmogorov N-width, meaning those representable with a low-dimensional basis, or where full-order snapshots are very costly to compute.

The practical payoff, the researchers argue, lies in multi-query settings such as inverse problems and data assimilation, where numerous simulations at different parameter values are required and both construction time and prediction time matter. They suggest the approach could accelerate parameter estimation and predictive process control in autonomous manufacturing processes such as sintering, and plan to extend the strategy to more complex and time-dependent problems. For now, the study demonstrates on two canonical nonlinear benchmarks that a carefully staged combination of greedy hyper-reduction and Gaussian process regression can deliver the offline economy of one method family and the online speed of the other, without giving up accuracy.

Subject of Research: A hybrid two-stage reduced order modeling strategy for parametrized nonlinear partial differential equations that combines greedy hyper-reduction with Gaussian process regression for offline and online computational efficiency

Article Title: Two-stage ROM: a hybrid reduced order model for offline and online efficiency

Article References: Dhopeshwar, R., Bansal, H., & Veroy, K. (2026). Two-stage ROM: a hybrid reduced order model for offline and online efficiency. Advanced Modeling and Simulation in Engineering Sciences, 13(1), Article 17. https://doi.org/10.1186/s40323-026-00322-z

Image Credits: AI Generated

DOI: 10.1186/s40323-026-00322-z

Keywords: reduced order models, digital twins, hyper-reduction, Gaussian process regression, empirical quadrature, nonlinear partial differential equations, greedy algorithm, heat conduction, hyperelasticity, computational efficiency, surrogate modeling, finite element method