Reduced Order Model (ROM)

ROMs project the high-dimensional state space of a full-order model (FOM)—often millions of degrees of freedom—onto a low-dimensional subspace defined by a reduced basis. Proper Orthogonal Decomposition (POD) identifies the dominant energy-containing modes from simulation snapshots, while Galerkin projection enforces that the residual of the governing equations is orthogonal to this subspace. This preserves the essential physics while collapsing the computational complexity from hours to milliseconds.

Unlike purely data-driven black-box surrogates, projection-based ROMs retain the structure of the governing partial differential equations (PDEs). By projecting the Navier-Stokes, Maxwell's, or heat transfer operators onto the reduced basis, the ROM maintains conservation laws and stability properties. This is critical for digital twin applications where violating Kirchhoff's laws or thermal balance would invalidate the simulation for operational decision-making.

ROMs exploit a two-phase computational strategy. The offline phase is computationally expensive and performed once: solving the FOM for multiple parameter snapshots, constructing the reduced basis, and pre-computing parameter-independent operators. The online phase evaluates the ROM for new parameters with negligible cost. This decomposition is essential for real-time digital twin queries where a grid operator cannot wait for a full finite element solve.

ROMs are constructed to be parametric, meaning they can interpolate or extrapolate to new operating conditions not explicitly included in the training snapshots. Techniques like greedy basis sampling and empirical interpolation ensure the reduced basis spans the solution manifold across varying thermal loads, voltage levels, or material properties. This enables a single ROM to serve as a general-purpose surrogate for an entire operational envelope.

A rigorous ROM provides a posteriori error estimators that bound the deviation from the high-fidelity truth solution without requiring a full FOM solve. Residual-based error estimates quantify the violation of the governing equations in the reduced space. For digital twin applications in critical infrastructure, these certificates of accuracy are non-negotiable for operator trust and regulatory compliance.

For nonlinear systems like electromagnetic transients or turbulent thermal convection, standard Galerkin projection still scales with the FOM dimension due to nonlinear operator evaluation. Hyper-reduction techniques—such as the Discrete Empirical Interpolation Method (DEIM) or gappy POD—sample the nonlinearity at a sparse set of spatial points. This decouples the online computational cost entirely from the FOM size, achieving true real-time performance.

A data-driven method for extracting an optimal set of orthogonal basis functions—POD modes—from high-fidelity simulation snapshots or experimental data. These modes capture the dominant energetic structures of the system, allowing the ROM to represent complex spatio-temporal dynamics with a drastically reduced number of degrees of freedom.

• Identifies coherent structures that contain the most variance
• Often paired with Galerkin projection to derive the reduced system of ordinary differential equations
• Widely used for fluid dynamics and thermal convection problems in grid assets

An equation-free, data-driven technique that decomposes time-resolved data into spatio-temporal coherent modes, each associated with a specific oscillation frequency and growth/decay rate. Unlike POD, DMD extracts modes that have inherent temporal dynamics, making it ideal for analyzing transient stability and oscillatory instabilities in power systems.

• Produces a linear reduced-order model of nonlinear dynamics
• Directly identifies unstable eigenvalues relevant to grid oscillation detection
• Requires only sequential snapshot data, no access to the underlying governing equations

A mathematical technique for projecting the governing partial differential equations (PDEs) of a high-fidelity model onto a low-dimensional subspace spanned by reduced basis functions. The result is a system of ordinary differential equations that preserves the underlying physics while being orders of magnitude cheaper to solve.

• Ensures the ROM residual is orthogonal to the reduced subspace
• Preserves stability and conservation properties when applied correctly
• Commonly used with POD-derived basis functions for structural and thermal analysis

A class of techniques designed to overcome the computational bottleneck that arises when evaluating nonlinear terms in projection-based ROMs. Methods such as the Discrete Empirical Interpolation Method (DEIM) or gappy POD sample only a small subset of mesh points to approximate nonlinear functions, restoring the ROM's speed advantage.

• Essential for problems with non-polynomial nonlinearities like electromagnetic saturation
• Reduces complexity from full-order mesh size to a small number of interpolation points
• Enables true real-time execution of nonlinear parametric models

A broader category of computationally cheap models that approximate the input-output mapping of an expensive high-fidelity simulation. While ROMs are a subset of surrogates that explicitly reduce the state-space dimension, other surrogates include Gaussian Processes, Polynomial Chaos Expansions, and neural networks trained on simulation data.

• Used for uncertainty quantification and design optimization
• Can be non-intrusive, treating the high-fidelity solver as a black box
• Often combined with ROMs in multi-fidelity frameworks for grid planning

An approach that strategically combines a small number of expensive high-fidelity simulations with a large number of cheap ROM evaluations to maximize predictive accuracy under a fixed computational budget. This is critical for uncertainty propagation in grid scenarios where thousands of Monte Carlo samples are required.

• Low-fidelity ROM provides the trend; high-fidelity model corrects the bias
• Uses co-kriging or hierarchical stochastic collocation for fusion
• Enables probabilistic power flow with full transient dynamics considered