↑ Vocaro Guide to MPM

Games & Commercial Applications

For the algorithm see Theory; for how to build/optimize it see Implementation.

This document surveys where MPM appears in products and game-adjacent work, as distinct from explaining the algorithm itself. It asks:

The entries are organized by type of use and strength of evidence. Dates provide a useful chronology, but being earliest is not treated as a measure of technical importance.

MPM is well-established in offline film visual effects (VFX) and engineering simulation, increasingly seen in real-time tech demos and research, and present in games through fluid middleware, sandbox projects, playable demonstrations and prototypes. The amount of public technical detail varies sharply among those uses, so the survey distinguishes confirmed methods from product statements and visual evidence.

1. Adoption landscape

MPM is production-proven, with adoption strongest in film and engineering and narrower roles in games. The landscape breaks into five areas, ordered by how battle-tested each is. Commercial DCC adoption dates from Houdini 20.5 in July 2024:

MPM is mature in film visual effects and engineering, available in commercial DCC tools, active in research, and used selectively in game middleware, sandboxes and prototypes

The public record includes commercial middleware used in shipped games, products with an MPM demonstration or subsystem, playable sandbox projects, research prototypes, and patents. These are different kinds of adoption, so the entries below state both the source class and the scope that the source actually establishes.

The dates mark public demonstrations, releases or disclosures; they do not rank the work or imply that every implementation remained unchanged afterward.

Date Work What the public record establishes
2020 Space Simulation Toolkit A developer-confirmed MPM demonstration tied to a paid 2D sandbox; the released and current solver scope is undocumented.
~2021–2022 onward Zibra Liquids Commercial MLS-MPM fluid middleware used in an undisclosed number of shipped Unity and Unreal games.
2023 Echoes of Somewhere demo and AR sculpting experiments A public game demo built around Zibra-driven liquid control, plus hand-tracked liquid-sculpting interaction studies.
2024 Liquid Crystal, Hydrolab and PB-MPM Playable all-continuum sandbox builds, an AR/VR liquid-puzzle prototype, and EA SEED’s BSD-licensed real-time research solver.

Nearby methods often confused with MPM

A granular or fluid appearance does not identify the underlying solver. These examples use other methods or have no supporting MPM evidence:

Game What it actually uses
Noita Custom cellular-automata engine (“Falling Everything”)
Teardown Voxel-based physics
Falling-sand sandboxes (e.g. The Powder Toy, Sandspiel) Cellular automata
Fluidity (2010) and similar water-puzzle games Other methods; predate the relevant MPM graphics work
Black Myth: Wukong (2024), whose sand and snow clips are widely shared Unreal Engine 5 Chaos / Niagara. No MPM claim from the developer or supporting technical evidence. Listed because “the sand physics are wild” clips circulate widely

No public MPM method claim accompanies NetEase’s Where Winds Meet or Naraka: Bladepoint, and the GDC vault carries no production MPM talk for either title.

The shared lesson is that granular or fluid appearance does not identify an MPM method. Cellular automata, voxels and engine-native effects can produce similar imagery at much lower cost.

Documented uses and projects

The cases below cover standalone products, middleware, applications built on middleware, experiments, and corporate research activity.

Evidence-class rule. Each case below identifies the kind of source supporting it. A storefront or promotional video proves what was marketed and when; a developer statement proves what that developer said; a technical manual can establish exposed controls; a patent establishes claims and ownership; and none silently substitutes for source code, an independent benchmark, or a solver-boundary diagram.

3. Evidence limits and terminology

The survey covers what is publicly documented, which is not the same as everything that exists. Three limits govern how its entries should be read:

  1. Studios rarely publish solver internals. A small or mid-sized studio could have shipped MPM-based effects without documenting or marketing the method, and nothing in the public record would show it.
  2. A product-level MPM statement is not a solver-boundary diagram. Space Simulation Toolkit is a real commercial-product connection, but the available sources do not say whether one mode, every material, the launch build, or the current build uses MPM. “Core physics” and “an MPM mode” are different claims.
  3. Deployment scope matters. A localized MPM effect, a fluid-middleware subsystem, and a world built around one multi-material solve are all uses of MPM, but they have different engineering costs and product roles. The survey identifies the scope instead of ranking one as the more important use.

4. Why game use remains selective

MPM appears in narrower game roles than in film, engineering and DCC tools. Several structural reasons explain that distribution:

4.1 What ships instead — the incumbent water & snow stack

The last bullet deserves specifics, because “what a studio would use instead” is the actual competition, and it is very good and very cheap. Two techniques own the water and snow space, and neither is a continuum solver; a third, a 2D grid fluid built for smoke and fire, is routinely stretched to cover water and mud as well:

Water → 2D shallow-water height fields. Not cellular automata — a real physics solver, just a 2.5D one (full method in Related Methods §1.5). The reference implementation is Chentanez & Müller’s Real-time Simulation of Large Bodies of Water with Small Scale Details (SCA 2010): shallow water equations on a height field, with particles spawned where the height field cannot cope (breaking waves, waterfalls, splashes) and reabsorbed on landing. The cost is startling — complete scenes at ~4–18 ms/frame on a 2010-era GTX 480.

That paper is not merely academic: Fluid Flux (Imaginary Blend / Krystian Komisarek), a commercial Unreal Engine water system released April 2022, is a direct implementation of it — the developer said so explicitly when asked, linking the paper. Its shipped envelope is a useful benchmark for what “good enough water” costs today:

Fluid Flux (UE4.26–UE5.5+)
Method 2D SWE height field + Niagara particles
Simulation cost ~0.5–0.7 ms/frame (dev’s figure, RTX 2060 @ 120 fps)
Max sim resolution 1024×1024, ~1 m/pixel → ~1 km² (coastline 2048², ~20 km²)
Features Wet–dry shorelines, waterfalls, advected foam, caustics, wetness, underwater, buoyancy/boats, wave breaks, ocean blending
Hard limits Heightfield projection only — no caves, no planets; domains can’t move or rotate at runtime; console builds lose precision to 16-bit floats
Effort Developer states 15 months to first release, 3+ years R&D total

It is also externally validated as physics, not just as an effect: a 2025 peer-reviewed digital twin of a river-experiment facility used Fluid Flux as its solver and compared the results against a real-scale flow experiment. A game engine’s water plugin standing up to hydraulic-engineering scrutiny is a fair measure of how mature this incumbent is.

Smoke, fire and fog → a 2D Eulerian grid, stretched to cover water. The other widely licensed real-time fluid for the same engine is FluidNinja LIVE (Andras Ketzer; first listed in December 2020, rebuilt as LIVE-2 for UE 5.6 and later in September 2026). Its own manual settles the method question: it is a grid Navier–Stokes solver — density and velocity advected through a fixed 2D grid, with pressure and divergence buffers, run as GPU simulation stages — and one model serves gases and liquids alike. It is not MPM and not a particle method: particles enter only as interaction sources and leave only as decoration pushed by the grid’s velocity. Three design choices define the product:

What that buys is cost. The vendor reports 0.2 ms per frame for a 512×512 smoke grid and 1.2 ms for a 2048×2048 creek on an RTX 3080, with rendering (2–3 ms) rather than simulation dominating the frame. The manual also lists more than a dozen commercial releases said to use it, while noting that most such usage is under non-disclosure and is inferred from screenshots. Measured against MPM, it makes the height field’s trade more sharply still: a 2D field carries no material identity, no conserved volume and no constitutive response, so the sand, mud and snow on its feature list are appearances rather than granular or viscoplastic behavior — and for a smoke plume or a footprint trail, that is exactly the right engineering choice.

Layered simulation is the shipped pattern. An indie voxel engine handling oceans will typically run a bulk voxel fluid sim for the water body, wave and tide motion applied as oscillating “target height” masks, a connectivity flag marking which water is joined to the ocean (so rock pools disconnect as the tide drops and keep their water), and a separate, higher-resolution, surface-only 2D ripple simulation running near the player for wakes and interaction. Nobody runs one solver at one resolution — the budget goes where the camera is. That is the same demote/promote logic the height-field papers use, arrived at independently.

BFS makes the same representation split explicit for snow: a thin height-map layer for broad coverage, VFX snow squares for weather, and physical particles for interaction. The developer identifies the physical path as modified PBD, not MPM, and describes a future transition in which frozen particles align to and become cellular-automata voxels. The conversion remains a plan rather than a published result, but the three-level presentation is already useful product evidence: persistent cover, airborne weather and interactive bulk snow do not have to share one representation.

Snow, sand and mud → displacement textures + parallax occlusion mapping. Deformable ground in shipped games is overwhelmingly not simulated at all (full method in Related Methods §2.3): a mask written by simple analytic primitives at the feet, a diffusion pass to soften it, and a parallax shader that marches the view ray through a height texture so the surface reads as displaced without any geometry. Indie implementations of exactly this run comfortably in real time on laptop hardware, and are genuinely convincing to look at.

The critique of them is the more useful artifact, because it is a public catalog of where the illusion breaks, written by players rather than researchers: no compaction (a repeatedly-walked path never packs down into a faster route), no cohesion (the thin pillar of undisturbed snow that survives between a walker’s legs cannot exist), no conserved volume (displaced material doesn’t have to go anywhere), no bulk failure (no post-holing, no collapse), and — most tellingly — deep examples read as sand, flour, or mud rather than snow, with repeated objection to bodies moving through waist-deep material unimpeded.

Where the two approaches actually differ. For shallow decorative deformation or a plain river, the incumbents look excellent at a fraction of the cost, and remain the sensible engineering choice. What they do not produce is the list above: compaction, cohesion, conserved volume, bulk failure, depth-correct resistance, material mixing, and phase change. Those are consequences of a constitutive model rather than features that can be bolted onto a height field or a displacement texture, so refining the incumbent stack does not converge on them. The trade is straightforward: MPM produces those behaviors and costs considerably more per frame.

5. Real-time enablers

Three developments expand the range of MPM workloads that can run interactively (all detailed in Theory and Implementation):

MLS-MPM, PB-MPM and GPU optimization each address a real-time blocker, together putting real-time MPM within reach on consumer hardware

These techniques support several product shapes: specialized middleware, localized effects, interactive sandboxes, and unified multi-material simulations. The appropriate scope depends on platform, particle count, material model and frame budget. A permissive reference license governs copying that source and nothing more — the underlying techniques carry their own patents, which is a separate question with its own document (Patents).

The frontier keeps moving (2024–2026): XPBI (SIGGRAPH Asia 2024 — real-time continuum inelasticity in a PBD pipeline), CK-MPM (SIGGRAPH 2025 — compact-kernel performance), convex MPM↔︎rigid coupling pushed to interactive rates on GPU (Zong et al. 2024; Yu et al. 2025, the Drake/robotics line), and PB-MPM derivatives (e.g. an affine-projection stabilizer, The Visual Computer 2025). Most remain research or demonstrations, while commercial use is concentrated in the narrower roles cataloged above. Full citations in References.

6. EA SEED and PB-MPM

SEED (Search for Extraordinary Experiences Division) is EA’s internal applied-research lab, and the group that published PB-MPM:

Sources for this document