How one AI design learns the motion of water, sand and cloth

Water sloshing against a flexible valve can bend it; that bending can change where the liquid goes. Predicting such exchanges often means bringing together different simulation tools. WorldParticle explores a way to reduce that custom work: a shared AI design that learns the motions of materials as different as water, sand and cloth. Soft robots and material design are among the possible beneficiaries, rather than established uses of the research. MIT’s account describes those ambitions.

Researchers at MIT, the University of British Columbia and other institutions developed the system. MIT highlighted it on November 9, but the research first appeared as a preprint on May 14, revised May 20. MIT says it will be presented at SIGGRAPH Asia, whose conference runs December 1–4 in Kuala Lumpur. The institutional account, paper and accompanying demonstrations concern the same study; they are not independent replications.

The appealing idea begins with dots. Imagine marking a sheet of fabric with thousands of them. Each dot has a position and a speed, while the fabric connects it to neighboring dots. Water can also be represented by moving points, though without those permanent fabric connections. Sand and atoms fit into this broad bookkeeping scheme too. The points represent different things, but they give the computer a common way to describe motion. The paper explains how that shared representation accommodates the materials’ differences.

That does not make cloth behave like water. Information about material properties, surrounding surfaces and connections still matters. A dot on cloth needs information about the sheet it belongs to, not merely whichever dots happen to be closest after a fold.

WorldParticle also keeps an explicit role for known physics. At each step, it first calculates how supplied external forces would advance the particles. A learned component then corrects their positions and velocities to account for interactions. For a falling sheet, the distinction is between advancing its downward motion and working out how the connected material responds as it encounters a surface. The AI learns part of the calculation rather than dispensing with physics altogether. The method combines nearby interactions with a compressed summary that lets distant parts of the material influence one another.

There is an essential qualification to “one AI design.” The researchers trained a separate model for each of six categories. They reused the architecture—the arrangement of computational components—but each category received its own training. A single trained model covering all six remains future work. That is still useful progress: changing materials need not require redesigning the underlying AI machinery. Training and limitations.

The tests make that distinction tangible. In water-like flow, the researchers varied viscosity, or resistance to flowing, and the placement of boat and island obstacles. In the cloth experiment, a sheet fell onto a fixed sphere while its stiffness and density varied. Test sequences were withheld from training. The question was whether the trained systems could handle different settings within those studied problems—not whether they could predict anything dropped into an unfamiliar world. Experimental details.

The team’s water, sand and cloth animations are simulations. They help show what the predictions look like, but attractive motion is only one level of evidence. The researchers also measured differences between predicted particle positions and velocities and reference trajectories generated by established numerical simulators. They report improvements over several competing AI simulators, including more stable motion over long sequences. The paper nevertheless acknowledges some loss of material-specific accuracy compared with specialized numerical methods. Results and comparisons.

That leaves another question before engineering use: is the prediction accurate enough for the decision being made? A believable splash and a dependable estimate of how equipment will behave require different validation. These experiments support the shared design’s promise; they do not establish the reliability of a particular valve or robot.

The molecular demonstration offers a particularly useful lesson in reading performance claims. The project page describes predictions of protein side-chain rotations—the motion of groups attached to the protein backbone—at intervals of 50 femtoseconds, compared with reference simulation steps of 0.5 femtoseconds. A femtosecond is one-millionth of one-billionth of a second.

The arithmetic is straightforward: 50 ÷ 0.5 = 100. Each prediction advances molecular time by an interval 100 times larger. That ratio alone does not establish a 100-fold computing speedup: elapsed computing time also depends on the work required for each step and the comparison conditions. Nor do those side-chain motions demonstrate a whole protein finding its folded shape. WorldParticle’s breadth is interesting enough without turning a larger stride through simulated time into a claim the experiment did not establish.