Statistical Reduced Order Models Bring Sensor Data and Simulations Together at Lower Cost

Simulations of physical systems are rarely perfect, and neither are the sensors that measure them. When engineers try to fuse the two—combining a numerical model of, say, an acoustic field with sparse, noisy readings from real sensors—a mismatch almost always remains. Their approach, published in Computational Science and Engineering, is called statistical reduced order modelling, or statROM, and it addresses a long-standing bottleneck: solving a full finite element model at every frequency of interest before conditioning it on data quickly becomes computationally prohibitive.

The foundation of the work is the statistical finite element method, or statFEM, a recently proposed data assimilation technique. In statFEM, the sensor data are modelled statistically as the sum of three ingredients: the numerical finite element solution, a model discrepancy term that captures the gap between simulation and reality, and measurement noise. All of these are treated as Gaussian processes, which allows the assimilation to proceed through Bayesian principles. Uncertainty is propagated through the finite element model to produce a Gaussian prior for the state of the system, and conditioning that prior on the measurements yields a posterior that both fits the data better and provides consistent estimates of uncertainty, including the effects of model misspecification. The hyperparameters of the model error covariance are learned from the data itself via an empirical Bayes procedure before the update is performed.

Where statFEM struggles is with parameter-dependent problems. The Helmholtz equation, which governs time-harmonic acoustic wave propagation, is the canonical example: its solution depends on the angular frequency, and in applications such as structural health monitoring, a model must be updated across an entire frequency range each time new measurements arrive. Because the resulting finite element systems can be very large—expensive to solve even once—computing a sweep over many frequencies multiplies the cost to the point of infeasibility. The standard remedy is model order reduction: a projection-based reduced order model (ROM) approximates the large system by a much smaller one, typically through techniques such as proper orthogonal decomposition or Krylov subspace methods.

The catch is that a ROM is never a perfectly accurate surrogate. Its approximation error has a well-known mathematical structure, governed by the residual equation, but in a naive application of statFEM to a ROM prior, that error is simply lumped into the general model error term. This inflates the estimated model discrepancy and produces posterior variance that is larger than necessary, while the systematic bias introduced by the reduced approximation goes uncorrected. Hermann and colleagues recognized that because the ROM error can be estimated without ever solving the full order system, it deserves separate, explicit treatment in the statistical data model.

Their solution combines two ideas. First, the ROM prior itself is built with a moment matching Krylov subspace method—specifically an adaptive order, structure-preserving second-order rational Arnoldi-type procedure—so that the reduced transfer function matches the Taylor series coefficients of the full order model up to a chosen order around selected expansion frequencies. Second, an adjoint-based error estimator quantifies the ROM approximation error at a modest number of points throughout the computational domain. The adjoint approach is the key to affordability: instead of comparing the ROM solution to the full order solution at every sample, which would defeat the purpose of reduction, the estimator reuses the LU decomposition already computed during the offline construction of the ROM and solves only reduced adjoint systems.

The estimated ROM error is then interpolated with a Gaussian process to cover the entire domain, not just the sensor locations, and inserted into the statFEM data model as a distinct term with a non-zero mean. This is what distinguishes statROM from the standard approach. Because the ROM-induced bias appears explicitly in the data model, it can be corrected directly: the observations are adjusted by the estimated error before the Bayesian update, so the reduced prior no longer has to be stretched to fit uncorrected data. The consequence, shown in the paper’s numerical experiments, is a smaller estimated model error, a much tighter posterior variance, and a posterior mean closer to the ground truth. The authors report that the proposed method achieves better accuracy and faster convergence throughout a given frequency range compared to applying standard statFEM to a ROM prior without the error correction.

The evidence comes from two numerical examples with artificial, i.e. simulated, sensor data. The first is a one-dimensional Helmholtz problem with a random forcing and a random heterogeneous material coefficient, discretized with 100 elements, designed to expose convergence properties in detail. Synthetic data were generated from a fine reference solution with 1000 degrees of freedom, observed at 11 equally spaced sensors with 20 noisy readings each. With only five matched moments around an expansion frequency of 100 Hz, the ROM prior deviates visibly from both the data and the finite element prior; increasing to fifteen matched moments reduces the error to machine precision at the evaluation frequency. In the posterior convergence study, even a simple version of the adjoint estimator with twelve training points delivered faster convergence than the standard statFEM update on the ROM prior, and higher-fidelity estimators accelerated convergence further.

The second example scales the test to a two-dimensional acoustic scattering problem with substantially more degrees of freedom: an unstructured mesh on a square domain containing a circular scatterer with a homogeneous Dirichlet boundary, absorbing outer boundaries, and a stochastic incident wave as the source. Here the authors also introduced a deliberate model misspecification by altering the source term used to generate the ground truth, superimposing a sine wave perpendicular to the incoming wave. Across three sensor configurations—5 sensors with 20 observations each, 30 sensors with 50, and 80 sensors with 200—the proposed statROM approach consistently produced lower posterior error than the standard ROM-based statFEM, approaching the accuracy of the full order statFEM method. When the ROM was made highly accurate (twenty matched moments at a frequency near the expansion point), the methods converged to similar errors, since there was little bias left to correct.

The computational savings are concrete. In the 2D example, the average solve time per sample was 0.001938 seconds for the full order model versus 0.0006217 seconds for the ROM, including matrix reduction, solving and projection back to full space. With 256 parameter samples and an error estimator using 200 training points, the total prior runtime was 5.8377 seconds for the statROM pipeline against 6.4382 seconds for the full order approach—meaning 256 finite element solves were replaced by 456 cheaper reduced solves while still coming out ahead. The authors note the method is most valuable for ROMs of medium quality: for very poor ROMs the error estimator is unreliable, and for very accurate ones the error is too small to matter.

Limitations remain, and the authors are candid about them. The ROM currently applies primarily to frequency sweeps rather than general parameter variation; memory requirements for the large covariance matrices become a concern in bigger systems, particularly if parallelization over the frequency grid is pursued; and the QMC prior construction carries approximation errors from limited sample sizes and the Gaussian assumption. The source code, implemented in Python with FEniCSx, is publicly available on GitHub and Zenodo and reproduces both examples, though it currently handles only the given Helmholtz equation on the given geometries. As next steps, the team points toward parameterizing the ROM, sparse precision matrix techniques to tame memory demands, and a practical application in aerospace acoustics—where fast, uncertainty-aware updating of frequency-dependent models from sparse sensor data could prove genuinely transformative.

Subject of Research: Statistical reduced order modelling for data assimilation in frequency-dependent acoustic simulations

Article Title: Statistical reduced order modelling for the parametric Helmholtz equation

Article References: Hermann, L., Bollhöfer, M., & Römer, U. (2025). Statistical reduced order modelling for the parametric Helmholtz equation. Computational Science and Engineering, 2(1), Article 3. https://doi.org/10.1007/s44207-025-00008-1

Image Credits: AI Generated

DOI: 10.1007/s44207-025-00008-1

Keywords: statFEM, reduced order modelling, Helmholtz equation, data assimilation, Gaussian processes, Bayesian updating, Krylov subspace methods, moment matching, adjoint error estimation, uncertainty quantification, acoustics, finite element method