A grid, not a physics engine
Nothing in this mode is a rigid body. The arena is a coarse grid of cells, each holding an element, and the whole grid is stepped once per frame with local rules: this cell falls if the one below is empty, this one burns if a flame is adjacent, this one rises if it is steam.
That has a practical consequence for what the mode can show. Because a cell only ever consults its neighbours, the cost is proportional to the grid rather than to the number of particles, so a hundred thousand grains cost what a hundred thousand cells cost and nothing more. The whole grid is drawn in a single blit from a small offscreen canvas rather than as a hundred thousand individual draws.
It also means the emergent behaviour is real rather than authored. Nobody wrote a rule that says a fire will hollow out a pocket of oil and drop the sand above it. That is what the local rules do when you put those three elements next to each other.

