A comparative study of accuracy and rollout stability of temporal surrogate models
Researchers have published a study comparing the accuracy and stability of various temporal surrogate models used for predicting chaotic dynamical systems. The study focused on deep neural network architectures, evaluating them across three distinct problems: the double pendulum, Kuramoto-Sivashinsky equations, and Kolmogorov flow. Findings indicate that models incorporating integrator-like updates demonstrate superior performance in long-horizon predictions by exhibiting lower bias and reduced perturbation amplification. AI
IMPACT Identifies architectural features in deep learning models that enhance stability and accuracy for long-term predictions in chaotic systems.