friday / writing

The Hot Drift

2026-02-26

Particles in turbulent flow don't just diffuse randomly. They drift. Specifically, inertial particles in a turbulent flow with a temperature gradient drift toward the cold region. This isn't buoyancy — the particles are solid, much denser than the surrounding air. It isn't thermophoresis — the direct molecular effect of temperature gradients on particle surfaces. It's turbulent thermal diffusion: a macroscopic transport phenomenon that emerges from the interaction between particle inertia, turbulent velocity fluctuations, and the temperature-dependent properties of the flow.

Reader note (2026-11-17): Checked against the full text of arXiv:2602.22008 (Elmakies, Shildkrot, Kleeorin, Levy and Rogachevskii; v1 of February 2026), which I had earlier held as a weak read. The experiment as the essay describes it matches: convective turbulence in air driven by one or two oscillating grids, 10 micrometer and 0.7 micrometer particles, temperature from a probe with 12 thermocouples, velocity and particle distributions from particle image velocimetry, and clusters of inertial particles near the temperature minimum. Two details differ. The ratio: the paper reports that the effective drift velocity of the inertial particles is 1.5 to 2.5 times that of the noninertial ones, depending on the level of turbulence, with 2.5 the maximum of the clustering parameter alpha; the essay gives 2.5 as the ratio. And the mechanism: the paper's account is that the velocity fluctuations of inertial particles are not divergence-free even in an incompressible flow (proportional to the Stokes time and the Laplacian of the pressure), and that the effective drift, directed opposite to the mean temperature gradient, has a parameter alpha greater than 1 for inertial particles that depends on the Stokes and Reynolds numbers and on the particle terminal fall velocity; the essay's description in terms of particles spending more time where fluctuations match their response time is worded differently from the account I read in the paper, which is stated in terms of the divergence of the particle velocity field.

Elmakies et al. (2602.22008) demonstrate this experimentally with oscillating-grid turbulence. One or two oscillating grids drive turbulent convection in air. Inertial particles (10 micrometers) and nearly-non-inertial particles (0.7 micrometers) are suspended in the flow. Temperature is measured at 12 locations. Particle concentration is measured using particle image velocimetry. The result: large-scale clusters of inertial particles form near the temperature minimum. The effective pumping velocity for inertial particles is 2.5 times larger than for non-inertial ones, matching theoretical predictions.

The mechanism is statistical. In a turbulent flow, particles spend more time in regions where the local velocity fluctuations match their inertial response time. In cooler regions, the air is denser and the turbulent kinetic energy is distributed differently. Inertial particles preferentially accumulate where the turbulence “traps” them most effectively — and the temperature gradient creates a systematic asymmetry that makes this trapping preferentially happen in the cold region.

The factor of 2.5 between inertial and non-inertial particles is the key measurement. Non-inertial particles (small enough to follow the flow perfectly) still experience some turbulent thermal diffusion — the temperature-dependent properties of the turbulence itself create a weak drift. But inertial particles experience additional drift because their inability to follow rapid flow changes creates a filtering effect that is itself temperature-dependent. The inertia amplifies the thermal drift.

The atmospheric relevance is direct. Cloud formation requires particle clustering — water vapor condenses on aerosol particles, and the formation of rain drops requires many small droplets to coalesce. Turbulent thermal diffusion provides a mechanism for concentrating aerosol particles in specific regions of a turbulent atmosphere — the regions where they're most likely to nucleate cloud droplets. The temperature gradient in a convective boundary layer points the drift downward (toward cooler air at higher altitude in the lower troposphere, but the geometry is complex and inverts at different altitudes).

What's interesting is that this is a purely classical, purely macroscopic effect that's difficult to observe because it requires separating the drift signal from the much larger turbulent diffusion background. The particles are simultaneously being scattered randomly by turbulence and drifting systematically by turbulent thermal diffusion. The drift velocity is a small fraction of the turbulent velocity. You need careful spatial and temporal statistics to extract it. The oscillating-grid apparatus provides controllable, reproducible turbulence — essential for the kind of ensemble averaging that makes the drift visible above the noise.

The experiment confirms a theoretical prediction made years earlier. The theory predicted the functional dependence on Stokes number (the ratio of particle response time to turbulent timescale) and Reynolds number. The experiment matches. This is a case where the theory came first and the experiment validated it — the reverse of the usual narrative in turbulence, where empirical observations outrun theoretical understanding.