FNM: a natural-law simulator

Streaming has long carried pixels and sound from strong machines to weak ones. FNM does the same for physics: a thin client sends a scene, a fleet of parallel solvers computes how nature would move it, and the coordinates stream back — rigid bodies, liquids, heat, gases, and chemical reactions, all solved off the device.

AAIC built FNM for a customer. The clips are its diagnostic viewer replaying returned state — coloured coordinate axes and a heads-up source count — not a production UI.

Fig. 1Four carts, four torques. Rigid-body kinematics: four wheeled chassis are released on an incline, each given a different drive torque, and their paths are returned across a two-second window. The client sent four bodies and their drive parameters; the solver returned where each one sits, frame by frame.

The need

The pattern is everywhere except where it matters most. Cloud gaming renders a frame on a datacentre GPU and streams the picture to a phone that could never draw it; music services keep the library on a server and send only what is playing. In each case the heavy thing stays where the power and storage are, and a thin stream reaches the device.

Physics has almost no equivalent. Rigid-body engines such as Project Chrono resolve mechanics and contact, but not the wider set of natural laws a real scene runs on — fluids that displace and drag, heat that flows toward equilibrium, gases that expand, chemistry that reacts. A game or a training simulator that wants those either bakes them offline or fakes them with shaders. FNM was built to close that gap: to make physics something a thin client can request rather than compute.

What crosses the wire

The contract is deliberately narrow. A Unity client sends the scene: where each object sits, its shape, and its physical properties — mass and density, temperature, viscosity, conductivity, whatever the material and the simulated law require. It sends no code and runs no physics.

The solvers

Behind the wire is not one engine but several, each a fast, parallelised solver for one family of law: rigid-body kinematics and contact, incompressible liquid dynamics, heat transfer, gas dynamics, and reaction chemistry. They share the scene and the clock, so a single hand-off can touch more than one at once — ice melting in a warm glass is rigid motion, buoyancy, heat flow, and phase change resolved together. The clips below are the returned state, replayed in the diagnostic viewer.

Fig. 2Displacement. A solid is lowered into a filled cylinder and the liquid rises to make room for it. Six sources — the object, the free surface, the vessel — over a seven-second run: liquid dynamics coupled to a moving rigid boundary.

Fig. 3Viscous drag. Liquid is released above a thin plate; the droplets fall, meet the surface, and are slowed by viscous friction against it. The orange vector is the returned velocity. Eight sources, tracked over roughly twelve seconds.

Fig. 4Melting. Ice bodies sit in a warmer liquid — buoyant rigid motion, heat transfer, and phase change integrated together across a six-hundred-second run. Fourteen sources drift and shrink as the solids give up mass; one scene, handed to several solvers at once.

Fig. 5Thermal equilibration. Liquids are poured together at different starting temperatures and left to settle across a simulated hour. Twenty sources; the colours track the returned temperature converging toward a common value. Heat transfer, run to steady state.

Fig. 6Bulbs in parallel. A battery drives a ladder of bulbs wired in parallel between two bus bars. The electrical solver returns the per-branch state and the geometry carries it over a four-and-a-half-second window — the same request-and-return loop, applied to current rather than motion.

A plot of one tracked angle in degrees over time across eight sources, with the run's range labelled. — Fig. 7Telemetry. Not everything returned is a coordinate. Alongside the geometry, FNM streams scalar property telemetry — here one tracked angle per source across eight sources, the run’s full range called out (min −178.6°, max 270.0°). It is state a client can read without rendering.

Cost of the computation

The point of the split is that the expensive half never touches the device. Integration — the timestepping of each law, the contact resolution, the coupling between solvers — runs parallelised on the strong machine and reduces to a compact trajectory table before anything is sent. What crosses the wire scales with the number of sources and the length of the timeline, not with the difficulty of the physics. A device replaying a six-hundred-second melt does as little work as one replaying a two-second roll.

A scene and its materials go in; the coordinates of how nature moves them come back. The device never integrates a law — it replays a table. The heavy work stays where the power is, and physics becomes something a thin client can request rather than compute.