Problem
Research questions and friction points this paper is trying to address.
Complex Model Simulation
Computational Efficiency
Model Simplification
Generative adversarial reduced order modelling
🤖 AI Summary
To address the low fidelity and long-term prediction instability of reduced-order models (ROMs) for high-dimensional nonlinear dynamical systems, this work introduces generative adversarial networks (GANs) into the ROM framework for the first time, enabling data-driven nonlinear manifold embedding. The proposed method jointly optimizes projection basis learning, physics-informed regularization, and temporal adversarial training—overcoming the representational limitations of traditional linear approaches (e.g., POD) in strongly nonlinear regimes. Evaluations across diverse fluid dynamics and reaction–diffusion systems demonstrate that the method improves long-term prediction accuracy by 30–50% over classical ROMs, reduces model dimensionality by approximately two orders of magnitude, and strictly enforces physical consistency (e.g., conservation laws and thermodynamic constraints). By unifying deep generative modeling with physical priors, this approach establishes a new paradigm for high-fidelity, low-dimensional, and physically interpretable dynamical system modeling.
Innovation
Methods, ideas, or system contributions that make the work stand out.
GAROM
Generative Adversarial Networks
Reduced Order Modeling
🔎 Similar Papers
No similar papers found.
