OnsagerNet: Learning stable and interpretable dynamics using a generalized Onsager principle

Machine learning is becoming an increasingly popular method for building mathematical models from observations of natural processes. Here, the central challenge to impart structure into model parameterization enforce physical relevance, yet retain a degree generality so that large variety dynamics can be learned. We introduce novel methodology, based on data-driven extension classical Onsager principle, strikes balance between these competing aspects. demonstrate its efficacy by quantitatively accurate and qualitatively faithful reduced order Rayleigh-B\'enard convection equations.