Work

Recursive online learning for stable long-term prediction of turbulent flows

Year: 2026

Type: article

Abstract: Accurate long-term prediction of turbulent flows remains a key challenge in data-driven modeling owing to compounding errors during recursive rollout. We propose a recursive online learning (ROL) algorithm that incorporates simulation feedback into the learning process so as to reduce local prediction errors on surrogate-generated states encountered during long-horizon recursive rollout. We validated the ROL algorithm on both statistically stationary and extremely transient turbulent channel flows using several deep learning architectures. The results show that ROL consistently improves long-term stability and accuracy across the tested network architectures, even in a bypass-like transition. Importantly, ROL-enhanced models generalize robustly to strong noise perturbations and to larger domains beyond the training size while maintaining physically consistent turbulence dynamics. In addition, the trained models achieved speed-ups ranging from 33 × to 231 × relative to a graphics processing unit (GPU)-parallelized numerical solver. The proposed method is effective across the diverse network architectures tested, and enables fast and accurate inference, making it useful in scientific simulations that require robust long-horizon predictions.

Source: Engineering Applications of Artificial Intelligence

DOI: 10.1016/j.engappai.2026.116285

Authors: Junhyuk Kim, Yonggyun Yu

Institutions: Korea Atomic Energy Research Institute, Korea University of Science and Technology

Language: English

Topic: Model Reduction and Neural Networks

Subfield: Statistical and Nonlinear Physics

Field: Physics and Astronomy

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $3,200

Funders: National Research Council of Science and Technology, Ministry of Science and ICT, South Korea