learned-material-surrogates opportunity asserted
Learned material surrogates at world scale
Use learned and GPU-parallel physics surrogates to bring per-particle fluids, gases, and destruction from small arenas to large persistent worlds.
thesis
Noita, Teardown, BeamNG, and Oxygen Not Included show players love worlds where every particle or voxel is simulated, but compute caps these simulations to small areas or 2D. Graph-network simulators and neural operators learn fluid, cloth, and deformation dynamics at large speedups, and GPU-parallel simulators make the remaining numerical work cheap. Combining a learned surrogate for the far field with exact simulation near the player could extend material simulation to whole maps.
design sketch
A colony or sandbox world where water, heat, gas, and structural stress are simulated everywhere at coarse resolution by a learned surrogate and exactly within a radius of active play, with consistent hand-off at the boundary. Players can flood a valley or collapse a mountain and see effects propagate realistically across the map.
validation experiment
Train a graph-network surrogate on a 2D falling-sand fluid simulator and measure mass conservation error and speedup over 10,000 steps at four times the training domain size. Success: mass error under 1 percent with a speedup above 10 times.
key risks
- Surrogates break conservation laws, producing exploitable infinite water or energy.
- Boundary artifacts between surrogate and exact zones.
- Determinism across GPUs for multiplayer.
falsifiable prediction
A game will use a learned surrogate model to simulate fluid, gas, or destruction at runtime.
Released after 2026-10-24 with at least 1,000 Steam user reviews or a console release; developer documentation or a technical talk confirms that a trained neural model computes physical material behavior during play, not only visual effects.