Simulation

What it is

A simulation starts with a description of a system and a set of rules about how it changes: atoms pulled about by the forces between them, light bent by the mass it passes, a star swelling as it burns through its fuel. The computer applies those rules over and over, in small steps of time or space, and records what comes out. Scientists can then look at the result as if it were an observation, except that they chose every condition themselves. Some simulations are built to reproduce reality closely; others are built to produce fake data that resembles what an instrument would collect, so methods can be tested on examples whose true answer is known.

Because the rules are often expensive to apply exactly, a second approach has grown up alongside. A model is trained on the results of the slow, careful calculation and learns to guess its answers directly. This learned stand-in, sometimes called a surrogate, does the same job far more cheaply, which lets scientists simulate things far bigger or for far longer than they could before. It is still a simulation: it is still working out what the system would do, just with a shortcut in place of the arithmetic.

Why scientists use it

Scientists simulate when the real thing is too slow, too small, too large, too dangerous or simply unavailable. You cannot watch glass form atom by atom, or wait for a star system to evolve, or build a telescope to see whether your analysis would have worked. The simulation stands in for the experiment or the observation. It also stands in for truth: when fake data is generated from known starting conditions, every answer is already known, so a method can be scored on how close it comes. That is how many machine-learning tools in the sciences are trained and checked at all.

A simulation is also a way of asking what if. Conditions can be changed one at a time, repeated exactly, and pushed to extremes that nature does not conveniently supply. And because a simulation can be run many times, it can be put inside a loop: run it, see where its answers are least trustworthy, go back to the slower and more reliable method for just those cases, and improve the fast model with what you learn.

In the record

Limits

A simulation can only contain the rules it was given. Leave out a force, a source of noise or a piece of physics, and the output will be confidently wrong in exactly that respect, which matters most when simulated data is used to train or test a method that will later meet the real world. Learned stand-ins add a second gap: they are reliable only for the kinds of cases they were trained on, and may fail quietly on anything unfamiliar. Agreement with a simulation is never, on its own, agreement with nature.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Every article