Researchers propose a new way to train AI models to solve physics equations — by learning how uncertainty flows, not predicting single outcomes

A research team argues that most AI approaches to solving complex physics equations are built on the wrong foundation: they try to predict specific outcomes, when they should be modeling how uncertainty and change move through physical systems. This shift in how you train the model means it could handle longer predictions, built-in uncertainty estimates, and designs that respect actual physics constraints.

For years, AI researchers have borrowed tricks from language and image models to tackle physics equations, but those tricks assume you're predicting a single best answer — which doesn't match how physics actually works. Physics equations describe how systems evolve continuously over time; uncertainty isn't a side problem, it's central. This paper argues the abstraction itself is wrong, and proposes training models on transport dynamics instead. If this shift takes hold, it could make AI-based physics solvers usable for longer timescales and harder problems, which matters because exact solutions to complex physics equations are prohibitively expensive to compute. The question is whether this reframing actually works at scale outside the research setting.