Ideal fluid dynamics may be viewed as a continuum version of classical mechanics, which implies strong geometric properties of the equations of motion including symmetries, invariants, and variational principles. One of the key numerical challenges in numerical weather prediction is the question how to preserve these structures under spatial and temporal truncation. I will talk about a Lagrangian particle method that leads to a finite dimensional system of Newtonian equations of motion with a number of desirable conservation properties. Applications include the spherical shallow-water equations and a vertical slice model. The emphasis on conservative discretizations will be motived by a simple example which will be related to ensemble weather prediction.