As far as I know this is fundamentally impossible.
Physics simulations are temporally dependent data, in the sense that each frame status depends heavily on the previous one.
To calculate what happens at any given moment you need to know what happened immediately before, so you must have exact geometric and kinetic information about every intervening element in the simulation to know where they were and calculate where they will be next.
Without that you can't have just pull out of thin air how a simulation looks at an advanced arbitrary point in time without knowing exactly how it was an instant before.
This makes it impossible for a computer to calculate "frame 100" from scratch without having frame 099, or 099 without 098, and so on till the beginning of the simulation from a known "rest position".
The only methods I've read about for distributed simulation calculations are systems where the simulation is divided up into spacial chunks and each node calculates one of the chunks, which can effectively spread across several workers.
This however requires that all intervenients talk to each other frequently, and maintain a certain level of synchronicity, since all nodes need to finish a certain frame for so they can move forward.