First two computers are initialized to the same configuration of the symbol gas...
The algorithm for running the simulation is made so that no matter how long it is run of simulated time, the final position of the symbols is deterministic based on the starting values. Though it is very psuedo random where every symbol will be after a fair amount of time, because they are all bouncing off of each other and off the walls. One computer has a message it wants to send to the other computer such as...
"this is a secret message..."
It first sends the position of the symbol "t" rounded to some accuracy, then the simulation is run for some predetermined amount of time such as 1000 seconds of simulation time, then it sends the new position of the symbol "h" and repeats for the whole message.
To decode the other computer looks for which symbol is at the first position sent from the other computer, then runs the simulation for the predetermined amount of time then which symbol is at the second position sent, etc..
I think because of the simulated entropy that a third party listening in on the message, but not knowing what the initial parameters of the simulation were, will never be able to decode a message. Even if they have an encoded message and a plaintext decoding of the same message, because only positions are given and not vectors or the predetermined time scale it would be impossible to generalize to decode other messages.
I think it's better than other methods because all have a sort of formula using a key that might be very complex but then there's always the possibility of finding the inverse. Here there is no formula for where one of the symbols will be after a certain amount of time and the simulation simply has to be run to find where a symbol will be so there is no inverse either.