Researchers have built an AI system that can predict every plausible outcome of a physical event, not just one.
The method, called Bi-FORK, tackles what's known as a bifurcation: a point where a single starting condition can lead to multiple equally valid outcomes, the way a compressed beam might buckle left or right with no way to know which in advance. Most AI models used to simulate physics assume one input maps to one correct answer, so they either pick a single branch or blur the options into an average that matches neither. Bi-FORK instead generates full trajectories using a technique called latent flow matching, then uses repulsion-guided sampling to keep the different outcome branches distinct rather than collapsed together. The team tested it on buckling beams, mechanical metamaterials, and Allen-Cahn phase separation, a model of how mixtures split into separate zones, at resolutions up to 260,000 points, well beyond what earlier approaches could handle.
That scaling matters more than it sounds. Bifurcations show up constantly in structural engineering, materials design, and climate modeling, so a simulation tool that silently averages away the real uncertainty isn't just imprecise, it can be actively misleading. A model that keeps multiple valid futures on the table is a more honest match for how these systems actually behave.
This is a research paper, not a shipped product, and the real test is whether it holds up on the messier, noisier systems engineers actually lose sleep over, like aging bridges or degrading batteries, rather than the clean textbook cases studied here.