Physics aware analytics for accurate state prediction of dynamical systems

Mandal, Ankit and Tiwari, Yash and Panigrahi, Prasanta K. and et al, . (2022) Physics aware analytics for accurate state prediction of dynamical systems. Chaos, Solitons & Fractals, 164. pp. 1-9. ISSN 0960-0779

Text
Chaos_Solitons_&_Fractals.pdf - Published Version
Restricted to Registered users only
Download (3MB) | Request a copy

Abstract

It has been successfully demonstrated that synchronisation of physical prior, like conservation laws with a conventional neural network significantly decreases the amount of training necessary to learn the dynamics of non-linear physical systems. Recent research shows how parameterisation of Lagrangian and Hamiltonian using conventional Neural Network's weights and biases are achieved, and then executing the Euler-Lagrangian and Hamilton's equation of motion for prediction of future state vectors proved to be more efficient than conventional Neural Networks in predicting non-linear dynamical systems. In classical mechanics, the Lagrangian formalism predicts the trajectory by solving the second-order Euler-Lagrangian equation following all the symmetries of the system while, the Hamiltonian mechanics explicitly uses the Poisson bracket of canonical position and canonical momentum to arrive at two independent first-order equations of motion. For a physical system that follows energy conservation, we obtain the momentum taking the first derivative of the Lagrangian which supplements the energy conservation via Hamiltonian Mechanics. The dynamics of the physical system converges when the generalised coordinates transform to suitable action angle variables using Poisson bracket formalism conserving the energy. In this work, we have demonstrated the prediction by integrating Hamiltonian and Lagrangian neural network leading to faster convergence with the actual dynamics from less number of training epochs. Hence, integration of HNN(Hamiltonian Neural Network) and LNN(Lagrangian Neural Network) outperforms their autonomous implementation. We have tested our proposed model in three conservative non-linear dynamical systems and found promising results. © 2022 Elsevier Ltd

[error in script]

IITH Creators:

IITH Creators	ORCiD

Item Type:

Article

Additional Information:

This research received no specific support from governmental, private, or non-profit funding agencies. The content and writing of the paper are solely the responsibility of the authors. Ankit Mandal is thankful to DST for the INSPIRE fellowship. The author MPal wishes to thank ABB Ability Innovation Centre, India for their support in this work.

Uncontrolled Keywords:

Action angle variable & Poisson bracket; Hamiltonian Neural Networks; Lagrangian Neural Networks; Neural networks; Physics inspired learning

Subjects:

Physics

Divisions:

Department of Physics

Depositing User:

. LibTrainee 2021

Date Deposited:

28 Sep 2022 14:39

Last Modified:

28 Sep 2022 14:39