StabOp: A Data-Driven Stabilization Operator for Reduced Order Modeling

Spatial filters have played a central role in large eddy simulation and, more recently, in reduced order model (ROM) stabilization for convection-dominated flows. Nevertheless, important open questions remain: in under-resolved regimes, which filter is most suitable for a given stabilization or closure model? Moreover, once a filter is selected, how should its parameters, such as the filter radius, be determined? Addressing these questions is essential for the reliable design and performance of filter-based stabilization strategies. To answer these questions, we propose a novel strategy that differs fundamentally from current filter-based approaches: we replace traditional spatial filters with a data-driven stabilization operator (StabOp) that yields accurate results for a given resolution, quantity of interest, and stabilization strategy. Although the new StabOp can be used for both classical discretizations and ROMs, and for different types of filter-based stabilization or closure, for clarity we focus on ROMs and the Leray stabilization. To build the new StabOp, we postulate its model form as a linear, quadratic, or nonlinear mapping, and then solve a PDE-constrained optimization problem to minimize a given loss function. Using the resulting StabOp in the Leray ROM (L-ROM) yields a new stabilized ROM, StabOp-L-ROM. To assess the StabOp-L-ROM, we compare it with the L-ROM and the standard ROM in numerical simulations of four flows: 2D flow past a cylinder at Re=500, lid-driven cavity at Re=10000, 3D flow past a hemisphere at Re=2200, and minimal channel flow at Re=5000. Our numerical results show that the StabOp-L-ROM can be orders of magnitude more accurate than the classical L-ROM tuned with an optimal filter radius in the predictive regime. Furthermore, while the new StabOp smooths the input flow fields, its smoothing mechanism differs from that of classical spatial filters.

💡 Research Summary

This paper introduces a data‑driven stabilization operator, called StabOp, as a fundamentally new alternative to traditional spatial filters used in reduced‑order model (ROM) stabilization for convection‑dominated flows. The authors focus on the Leray ROM (L‑ROM) as a representative filter‑based stabilization technique, but the StabOp concept is presented as generally applicable to any filter‑based ROM closure or stabilization.

The construction of StabOp proceeds in three steps. First, a model form is postulated: a linear mapping, a quadratic mapping, or a nonlinear mapping realized by a neural network. Second, a PDE‑constrained optimization problem is formulated, where the objective (loss) function measures the discrepancy between a quantity of interest (QoI) computed from the ROM and the same QoI obtained from the full‑order model (FOM). Typical QoIs include kinetic energy, vortex shedding frequency, or time‑averaged statistics. Third, the optimization is carried out in the low‑dimensional ROM coefficient space, yielding the parameters of the chosen mapping (e.g., matrix entries, tensor coefficients, or neural‑network weights). Because the optimization variables live in the ROM subspace, the computational cost is negligible compared with solving the full Navier–Stokes equations.

Once trained, StabOp replaces the traditional spatial filter in the Leray ROM. In the classic L‑ROM, a differential filter with radius δ smooths the advecting velocity field; δ must be tuned empirically, and an inappropriate choice leads either to residual spurious oscillations (δ too small) or over‑smoothing (δ too large). StabOp‑L‑ROM eliminates this hyper‑parameter: the learned operator automatically produces the optimal amount of smoothing for the given resolution and QoI, effectively “learning” the best filter radius and even a more sophisticated transformation than a simple low‑pass filter.

The authors test StabOp‑L‑ROM on four benchmark problems that are known to be challenging for under‑resolved ROMs: (i) 2‑D flow past a cylinder at Re = 500, (ii) 2‑D lid‑driven cavity at Re = 10 000, (iii) 3‑D flow past a hemisphere at Re = 2 200, and (iv) 3‑D minimal channel flow at Re = 5 000. For each case, they compare three models: the standard Galerkin ROM (G‑ROM), the classical L‑ROM with an optimally tuned filter radius, and the new StabOp‑L‑ROM. Results show that StabOp‑L‑ROM achieves orders‑of‑magnitude reductions in the L² error of the velocity field (often 10–100× smaller) and reproduces key statistical quantities (energy spectra, vortex shedding frequency, mean pressure drop) with high fidelity. In the predictive regime—i.e., when the ROM is run beyond the training window—the StabOp‑L‑ROM remains accurate, whereas the classical L‑ROM deteriorates noticeably.

Spectral analysis of the filtered fields reveals that StabOp’s smoothing mechanism differs from that of traditional differential or projection filters. While classical filters act mainly as low‑pass operators, the learned StabOp selectively damps only those modes that contribute most to the error, preserving energetically important structures. This data‑driven selectivity explains the superior performance.

In summary, the paper makes three major contributions: (1) it reframes the filter‑selection and filter‑parameter‑tuning problems as a data‑driven operator‑learning task; (2) it provides a practical, low‑cost algorithm to train such operators directly in the ROM space; and (3) it demonstrates that the resulting StabOp‑L‑ROM dramatically outperforms the best possible classical Leray ROM across a range of challenging flow configurations. The authors suggest future work on extending StabOp to other closure models (e.g., variational multiscale, approximate deconvolution), handling parametric variations, and integrating the approach into real‑time control and optimization frameworks.