Two cubes suspended in an acoustic field form a coupled oscillator with a twist. The center-of-mass mode (both cubes moving together vertically) and the hinge mode (the angle between them opening and closing) are coupled non-reciprocally: the hinge affects the center of mass more than the center of mass affects the hinge.
This asymmetry stabilizes three distinct dynamical attractors — three qualitatively different long-term behaviors that the pair can settle into, depending on initial conditions. The non-reciprocity isn't engineered; it emerges from the geometry of the acoustic radiation pressure acting on non-spherical objects. Cubes, unlike spheres, have orientational degrees of freedom that couple to translational ones in asymmetric ways.
Note (2026-11-11): checked against the paper's full text. (1) The essay says the data-driven model 'predicts behavior over hundreds of oscillation cycles, far beyond the training window'. In the paper the hundreds of cycles are the persistence of the three attractors in the measured data; the model is fitted to the attractor shape, convergence and oscillation timescale of those same long records, and I found no test on a held-out time window or an extrapolation beyond the fitted data. The paper's claim is that fitting long-time features gives stable long-timescale predictions. (2) The essay explains the non-reciprocity as arising from acoustic radiation pressure on non-spherical objects, with the cubes' orientational degrees of freedom coupling asymmetrically to translation. The paper attributes the clustering to acoustic scattering forces that depend strongly on particle shape; the non-reciprocal cross-coupling term is found by fitting a minimal model and is supported by finite-element simulations that capture only static forces and torques. The orientational-coupling account is the essay's. (3) I did not check which direction of the coupling is stronger, so the essay's 'the hinge affects the center of mass more than the reverse' is unverified here.
The model was extracted from data, not derived from first principles. The researchers used trajectory measurements to fit a minimal dynamical system that captures the essential coupling. The data-driven model predicts behavior over hundreds of oscillation cycles — far beyond the training window.
Acoustic levitation is usually treated as a positioning tool: hold something in place with sound. Here it's a dynamical laboratory. The levitated cubes aren't static — they're active oscillators whose emergent dynamics arise from the interaction between shape and sound. The geometry of the particle determines the symmetry of the coupling, and the coupling determines the attractor landscape.