Gold nanoparticles on a substrate at high temperature lose mass. The standard expectation from classical Ostwald ripening is that small particles shrink and large particles grow — the Gibbs-Thomson effect drives mass from high-curvature (small) to low-curvature (large) surfaces. The prediction is size-dependent: small particles should shrink faster.
In situ electron microscopy with automated tracking reveals a different picture (arXiv:2603.20635). The average mass loss per nanoparticle is nearly size-independent. Small and large particles lose mass at roughly the same rate. The mechanism is substrate-mediated evaporation: atoms leave the particle to the substrate surface, diffuse along the substrate, and eventually desorb. The substrate provides a parallel evaporation channel that doesn't depend on particle curvature.
Beyond the mean behavior, individual nanoparticles exhibit enormous fluctuations. Two particles of the same initial size can follow wildly different mass trajectories — one shrinking steadily, another gaining mass from neighbors before eventually shrinking. A Langevin framework captures the stochastic dynamics. The fluctuations are not measurement noise but intrinsic to the nanoscale mass exchange.
The net mass loss suppresses conventional coarsening. In standard Ostwald ripening, large particles grow at the expense of small ones. Here, everybody loses mass to the substrate, and the collective loss overwhelms the curvature-driven transfer. The size distribution narrows instead of broadening — the opposite of coarsening.
The structural insight: stochasticity at the nanoscale is not a perturbation on a deterministic process. It's the dominant feature. The mean field theory (Ostwald ripening) predicts the wrong qualitative behavior because it averages over fluctuations that determine the outcome. Predicting what happens requires the full stochastic framework, not the average. The noise is the signal.