Adaptable Hamiltonian neural networks

The rapid growth of research in exploiting machine learning to predict chaotic systems has revived a recent interest Hamiltonian neural networks (HNNs) with physical constraints defined by Hamilton's equations motion, which represent major class physics-enhanced networks. We introduce HNNs capable adaptable prediction nonlinear systems: training the network based on time series from small number bifurcation-parameter values target system, HNN can dynamical states at other parameter values, where not been exposed any information about system these values. architecture differs previous ones that we incorporate an input channel, rendering parameter-cognizant. demonstrate, using paradigmatic systems, as few four bestows ability state entire interval. Utilizing ensemble maximum Lyapunov exponent and alignment index indicators, show our parameter-cognizant successfully route transition chaos. Physics-enhanced is forefront area research, provide approach understanding broad applications.2 MoreReceived 25 February 2021Accepted 13 May 2021DOI:https://doi.org/10.1103/PhysRevResearch.3.023156Published American Physical Society under terms Creative Commons Attribution 4.0 International license. Further distribution this work must maintain attribution author(s) published article's title, journal citation, DOI.Published SocietyPhysics Subject Headings (PhySH)Research AreasChaosMachine learningPhysical SystemsHamiltonian systemsNonlinear Dynamics