From Jos Stam's "Stable Fluids" paper in SIGGRAPH 99, p. 121-128.

The semi-Lagrangian method increases the accuracy and stability of fluid flow simulations. These flows are commonly computed on an Eulerian (e.g. 3-D voxel) grid, and the simulation, e.g. from the Navier-Stokes equations, are computed at each voxel location. The semi-Lagrangian method computes the velocity at a voxel location by tracing backward in time to find out where a particle at the voxel was before. This location before is likely between the velocities stored in the discrete grid, so the velocity is interpolated from the surrounding values.

Check out Caltech for a cool online demonstration of stable smoke using the semi-Lagrangian method.