Learning interpretable dynamics of stochastic complex systems from experimental data

Inferring stochastic differential equations (SDEs) from empirical data for complex networked systems remains challenging. This paper introduces the Langevin Graph Network Approach (LaGNA), which outperforms five state-of-the-art methods in learning hidden SDEs. Applications to bird flock movement and tau pathology diffusion in mouse brains demonstrate LaGNA's ability to infer governing equations, providing insights into flocking dynamics and tau spread, enabling early prediction of tau occupation and revealing distinct pathology dynamics in mutant mice.

Ting-Ting Gao, Baruch Barzel, Gang Yan