The Runge-Kutta45 kernel does not work as expected. I'm using it for 2D advection using output from a 1/12th-degree NEMO run, and time steps explode so that particles end up all over the place. I've prepared a zip with some data and a notebook to illustrate the issue.

Both Parcels 2.4.2 and 3.0.2 exhibit this issue. In Parcels2, I get an OutOfBoundsError that shows that time steps grow to values of: 2457600.0s

The figure below is after 90 days of simple advection, using a single flow field snapshot, so particles should follow streamlines.

A suggested improvement would be for users to define a maximum timestep (max_dt). However, I find it strange that the timestep is blowing up so much anyway. Only implementing a maximum timestep doesn't guarantee that the RK45 algorithm works as it should, as it might reduce the algorithm to RK4.