Researchers have developed WorldParticle, an AI model that represents materials as particles and predicts their motion, aiming to simulate a range of physical systems within one framework.

Unlike conventional simulations that require researchers to build equations and solvers for each material or process, WorldParticle learns general dynamics from particle positions and physical properties, while separating known external forces from its learned predictions.

The researchers report that the model can simulate cloth, elastic solids, fluids, granular materials and molecular dynamics, combining an explicit predictor for external forces with a learned correction step.

In tests across six physical phenomena, the team reported greater accuracy than numerical simulation methods; in a granular-collapse example, WorldParticle formed a realistic pile where some comparison methods left particles suspended in the air.

For changing protein structures, WorldParticle used 50-femtosecond time steps rather than the 0.5-femtosecond steps used in conventional simulations, a 100-fold increase in step size—not a demonstrated 100-fold reduction in total computing time.

To make large simulations more manageable, the researchers developed a technique that merges and compresses particle information while aiming to preserve the fine details needed for predictions.

The authors say the model can generalize to materials and conditions absent from training, and demonstrated interactive control, inverse design and learning from real-world manipulation data.

Potential uses include simulating soft robots moving across soil, fluids interacting with flexible structures, and exploring molecular configurations in protein research; the team hopes to build a broader scientific simulation pipeline.

The longer-term ambition is a foundation model that lets scientists simulate many scenarios without building each simulator from scratch, though WorldParticle remains a research model rather than a replacement for established tools.

Despite its shared architecture, WorldParticle still uses separately trained model parameters for different categories of physical behavior, so one model capable of handling every category remains a future goal.

Its reliability still needs to be established through tests of error accumulation over long simulations, performance at larger particle counts and behavior outside training conditions.

Researchers evaluating the system should compare its results with trusted numerical baselines and document the training data, tested conditions and relevant data and licensing details.