Abstract

Neural ODEs are a widely used, powerful machine learning technique inparticular for physics. However, not every solution is physical in that it isan Euler-Lagrange equation. We present Helmholtz metrics to quantify thisresemblance for a given ODE and demonstrate their capabilities on severalfundamental systems with noise. We combine them with a second order neural ODEto form a Lagrangian neural ODE, which allows to learn Euler-Lagrange equationsin a direct fashion and with zero additional inference cost. We demonstratethat, using only positional data, they can distinguish Lagrangian andnon-Lagrangian systems and improve the neural ODE solutions.

 

Quick Read (beta)