Machine learning guesses at chance on the fate of a deterministic, chaos-free Illinois cell model

Determinism fixes the outcome. In a deterministic system the final state is set by the starting configuration and the update rules ^[5], so any run of this model can be forecast simply by running it. The claim in the Nature Communications paper is narrower than "unpredictable" sounds, and more interesting. It is that a system with no chaos can still offer no shortcut from the starting grid to the fate ^[3][4]. Chaos is the usual cause of that kind of trouble, because tiny deviations in the start get magnified ^[6].

To test whether the difficulty was real, the team gave several machine learning algorithms a binary task: from the initial configuration, say whether a run would end static or in a moving wave ^[2]. That split puts the rectilinear and spiral waves into one class ^[15]. The algorithms were right about half the time. Phys.org's account calls that no better than random guessing ^[2].

The denominator matters here. Fifty percent is chance only if static and moving fates occur about equally often. If one fate dominated, a classifier that always guessed it would beat 50%. Either way, the trained models found nothing in the starting grid they could use ^[2]. The test is also limited to the algorithms tried, and showing that these classifiers fail is a weaker claim than showing that every method would.

The second finding is the one I find more exciting. As a run proceeds, its dynamics build topological structure that eventually becomes a reliable predictor of the final fate ^[1]. On this account, predictability is something the system acquires partway through. I think the classifier test is the firmer half of the work as reported. The summary does not identify the topological features or say how early in a run they become informative.

The work began with biology. "When I started my group 11 years ago," Hyun Youk, an Illinois physics professor, said, "we started working on models of living systems to understand how complex dynamics can arise from simple deterministic rules, specifically systems of living cells that interact with each other to form spatial patterns." ^[7][14] In 2020 his group searched computationally for ways cells could communicate by secreting and sensing molecules ^[8]. "We found communication modes that matched how cells in nature form the very same types of spatial patterns," he said ^[8].

That match is in the types of pattern, and it comes from simulated cells. The simulated grid also wraps around at its edges, so a pattern that leaves on the right reappears on the left, a choice the team made for ease of simulation ^[10].