New AI Simulates Sand and Gravel 12x Faster Than Current Methods
This could help engineers predict landslides and design safer buildings without costly experiments.
Sand isn't simple. A pile of grains can flow, jam, collapse, and settle in ways that challenge even the best physics software. Engineers who design dams, tunnels, silos, or landslide barriers normally need slow, high-precision simulations called material point methods. TRACE is a new AI trained on those simulations, and it learns to reproduce their behavior in a fraction of the time.
How does it work? TRACE treats every grain as a small point in a network, with lines connecting the grains that touch. Instead of just tracking where each grain has been, it tracks the ''memory'' of each contact—what forces, slips, and jams happened between those specific grains. When grains break apart and meet new neighbors, TRACE keeps that contact history alive. This lets the AI predict how the whole pile will evolve, including the final shape and how much energy it releases during collapse.
The results are impressive. In tests of 2D and 3D columns of grain falling over, TRACE cut long-term position errors by 31-62% and final pile-shape errors by 58-89% compared to existing AI simulators, with fewer parameters. It also ran 12.2 times faster (in 2D) and 8.9 times faster (in 3D) than the reference physics solver. Engineers could test hundreds of scenarios in the time it used to take for one.
The catch: AI is only as good as its training data, and TRACE has been tested on controlled column-collapse benchmarks, not on every real-world soil condition. But the code is open-source, so researchers can stress-test it against their own experiments. If it holds up, TRACE could make granular simulation fast enough for everyday design decisions—potentially preventing landslides, predicting sinkholes, and making industrial powder handling safer.
- TRACE is a new AI that simulates how sand, soil, and powders collapse or flow—up to 12.2x faster than traditional physics simulations.
- It remembers the history of each grain-to-grain contact, cutting long-rollout position error by 31-62% and final deposit error by up to 89%.
- Practical uses: safer design of dams, foundations, silos, tunneling, and better control of powders in food and drug manufacturing; code is open-source.
Why It Matters
Cheaper, faster simulation of grains means engineers can predict landslides, soil failures, and industrial hazards before they happen.