Parametric Interpolation of Dynamic Mode Decomposition for Predicting Nonlinear Systems
New AI framework learns a single model to simulate fluid flows and electron beams across multiple parameters.
A team of researchers from institutions including Ohio State University and the University of Michigan has introduced a new AI framework called parameter-interpolated dynamic mode decomposition (piDMD). This method represents a significant advance in parametric reduced-order modeling (ROM) for predicting the behavior of complex, nonlinear physical systems. Unlike previous parametric DMD techniques that interpolate between separately trained models—a process that can fail with sparse data—piDMD learns a single, unified Koopman surrogate model that inherently accounts for how the system changes with different parameters. This approach embeds known parameter-affine structure directly into the core regression step of the Dynamic Mode Decomposition algorithm.
The team rigorously validated piDMD on three challenging physics benchmarks: fluid flow past a cylinder, electron beam oscillations in magnetic fields, and virtual cathode oscillations simulated with an electromagnetic particle-in-cell (EMPIC) method. In all cases, piDMD demonstrated superior performance, delivering accurate long-horizon predictions even with fewer training samples. It showed particular robustness in multi-dimensional parameter spaces, where existing interpolation-based methods often struggle. The framework's ability to predict system behavior at unseen parameter values without retraining makes it a powerful tool for computational scientists and engineers.
This development matters because simulating complex nonlinear systems—like plasma physics or aerodynamic flows—is computationally expensive. Traditional high-fidelity simulations can take days. Reduced-order models like DMD aim to create fast, approximate surrogates, but making them work across a range of conditions (parameters) has been a major hurdle. piDMD's novel architecture directly addresses this by building parameter dependence into the model's foundation, leading to a more robust and data-efficient tool for design, control, and real-time prediction in engineering and scientific applications.
- Learns a single parameter-affine Koopman model across multiple conditions, eliminating fragile post-training interpolation.
- Validated on complex physics simulations, achieving accurate predictions with less training data than prior methods.
- Shows improved robustness in multi-dimensional parameter spaces, a key challenge for simulating real-world systems.
Why It Matters
Enables faster, more data-efficient simulation of complex engineering systems like aerodynamics and plasma physics for design and control.