Â鶹ÒùÔº

Researchers at Seoul National University and Kyung Hee University report a framework to control collective motions, such as ring, clumps, mill, flock, by training a physics-informed AI to learn the local rules that govern interactions among individuals.

The paper is in the journal Cell Reports Â鶹ÒùÔºical Science.

The approach specifies when an ordered state should appear from random initial conditions and tunes geometric features (average radius, cluster size, flock size). Furthermore, trained on published , the model uncovers interaction mechanisms observed in real flocks.

Collective motion is an emergent phenomenon in which many self-propelled individuals (birds, fish, insects, robots, even human crowds) produce large-scale patterns without any central decision-making. Each individual reacts only to nearby neighbors, yet the group exhibits coherent collective motion. Analyzing how simple local interactions give rise to such global order is challenging because these systems are noisy and nonlinear, and perception is often directional.

Learning local rules with physics-informed AI

To address these challenges, the team built neural networks that obey the laws of dynamics and are trained on simple pattern characteristics and, when available, experimental trajectories.

The neural networks infer two basic types of local interaction rules: distance-based rules that set spacing, velocity-based rules that align headings, as well as their combination. The team also showed that self-propelled agents following these rules reproduce intended target collective patterns with specified geometrical characteristics.

Examples include adjusting ring radius, cluster size in clumps, and rotational mode (either single or double) in mill; inducing continuous transitions among different collective modes; and achieving motions near obstacles and within confined areas.

The same framework can be fit to short segments of real trajectories by incorporating an anisotropic field of view, yielding interaction laws consistent with the leader-follower hierarchy observed in nature.

Opening new possibilities in collective behavior and robotics

By turning collective behavior into something that can be decoded, this approach offers practical engineering and scientific benefits. In robotics, it provides a blueprint for programming drone and ground-robot swarms to form and switch patterns on demand.

In the natural sciences, it helps quantitatively identify which local interactions are sufficient to explain observed flocking, enabling hypothesis testing about sensory ranges and alignment strength.

More broadly, the method could guide the design of active materials that self-assemble into target shapes and help generate realistic synthetic datasets for studying complex, decentralized systems.