Neural SINDy: robust discovery of governing equations using neural ODEs and sparse identification
Nonlinear Dynamics · 2026-11 · 114 (21)
DOI: 10.1007/s11071-026-13122-2 ↗Kurz gefasst
Fachartikel in der Zeitschrift Nonlinear Dynamics, online erschienen am 06.11.2026. Verfasst von Min Jae Jeon und 2 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0). Gefördert von: National Research Foundation of Korea (NRF), funded by MSIT, National Research Foundation of Korea (NRF), funded by the Ministry of Education.
Abstract (Originalsprache)
Identifying governing equations from observed trajectories is challenging when time derivatives are not directly accessible and numerical differentiation becomes unreliable under sparse or noisy measurements. Neural ordinary differential equations (Neural ODEs) learn continuous-time dynamics without explicit derivative information but do not directly provide interpretable symbolic models, whereas sparse identification of nonlinear dynamics (SINDy) yields interpretable governing equations yet relies critically on accurate derivative estimates. We propose Neural SINDy (nSINDy), a two-stage framework that combines Neural ODE–based state–derivative reconstruction with sparse regression. A Neural ODE is first trained on observed trajectories to learn a continuous-time vector field and construct derivative estimates; when observations are noisy, the learned dynamics are further used to reconstruct smoothed state trajectories. These state–derivative data are then used within SINDy to recover sparse and mechanistically interpretable governing equations without directly observed derivatives. Experiments on the Lotka–Volterra, SIR, and enzyme kinetics systems show that, across repeated training runs, nSINDy achieves higher exact-structure recovery rates than conventional SINDy under limited sampling and measurement noise and produces trajectories more consistent with the reference dynamics. In addition, we evaluate nSINDy on the chaotic Rucklidge and five-dimensional Lorenz–96 systems, where it similarly reconstructs the underlying structures from sparse datasets. Furthermore, when applied to the real lynx–hare population dataset, nSINDy successfully recovers a Lotka–Volterra-type structure, showing close agreement with the Ensemble SINDy (E-SINDy) baseline. These results support nSINDy as a practical framework for interpretable data-driven equation discovery when derivative observations are unavailable or unreliable.
Förderung
- National Research Foundation of Korea (NRF), funded by MSIT
- National Research Foundation of Korea (NRF), funded by the Ministry of Education
Zitieren
APA
Jeon, M. J., Lee, D., & Byun, J. H. (2026). Neural SINDy: robust discovery of governing equations using neural ODEs and sparse identification. Nonlinear Dynamics, 114(21). https://doi.org/10.1007/s11071-026-13122-2
Vancouver
Jeon MJ, Lee D, Byun JH. Neural SINDy: robust discovery of governing equations using neural ODEs and sparse identification. Nonlinear Dynamics. 2026;114(21). doi:10.1007/s11071-026-13122-2
BibTeX
@article{Jeon2026,
author = {Jeon, Min Jae and Lee, Davin and Byun, Jong Hyuk},
title = {Neural SINDy: robust discovery of governing equations using neural ODEs and sparse identification},
journal = {Nonlinear Dynamics},
year = {2026},
volume = {114},
number = {21},
doi = {10.1007/s11071-026-13122-2}
}
EUVIMED