Skip to content
PaperFren

Turbulence

Can we infer temperature in turbulent convection from velocity alone?

Toscano JD, Käufer T, Wang Z, et al. · Science advances · 2025

Open access · cc by · source: Europe PMC

By forcing a neural network to obey the equations of fluid flow and heat transport, the authors recovered the temperature field of a real turbulent convection experiment from velocity measurements alone to within about 4%.

Study at a glance

Design
Computational / modelling — Physics-informed Kolmogorov-Arnold network trained on Lagrangian particle-tracking velocities from a water-glycerol Rayleigh-Bénard cell, validated against thermochromic-liquid-crystal temperature measurements.
N
No participant count; the data are 282 experimental snapshots with about 3000 particle data points each from one convection run.
Population
Turbulent Rayleigh-Bénard convection in a hexagonal cell (water-glycerol, Prandtl number 10.6)
Outcome
Error of reconstructed velocity and inferred temperature versus measurements; boundary-layer profiles, dissipation-rate and velocity-gradient statistics

Structured fields used in claim comparison tables when every cited study has a complete layer.

Key findings

Reconstructed velocities on unseen data had relative errors of about 10% per component, and the inferred temperature matched the measured temperature with a relative error of 3.62% without ever seeing temperature during training. The inferred fields reproduced thermal plumes, regions of negative local heat flux, a thermal boundary-layer thickness consistent with scaling theory, and the teardrop-shaped Q-R velocity-gradient distribution typical of turbulence. Dissipation-rate statistics agreed qualitatively with earlier point measurements and simulations, though the model smoothed out extreme events; adding a few temperature observations reduced that smoothing.

Methodology

The team heated a hexagonal cell of water-glycerol from below and cooled it from above, tracking temperature-sensitive liquid-crystal particles in 3D with three cameras to record both velocity and temperature. They then trained a physics-informed network (built on Chebyshev Kolmogorov-Arnold networks) on half of the velocity data only, requiring it to satisfy the Navier-Stokes and energy equations in a pressure-free vorticity form. Extra training tricks, including attention-based resampling of high-error points and staged training from simple to full physics, were added, and the held-out velocities and all measured temperatures were used for validation.

Limitations

All results come from a single run at one Rayleigh and Prandtl number in a thin slab of one cell, so generalisation to other regimes or geometries is untested. The model systematically underestimates high-magnitude events such as strong heat-flux bursts, and the temperature measurements themselves are biased by the limited sensitivity range of the liquid crystals, especially near the plates. Comparisons with literature dissipation statistics are qualitative and at different parameters. The paper is primarily a method demonstration rather than a new discovery about convection physics.

How this study connects

Role on claims

Each row is a claim on a concept or method page where this paper supports, challenges, or qualifies the statement. Roles are hand-checked — not a model guess.

  • Plumes and gradient statistics are encoded in the velocity field, but extremes are the hardest part to reconstruct.

    Coherent structures in turbulent convection can be recovered from velocity alone: a physics-informed network trained only on 3D particle-tracking velocities inferred the temperature field of a Rayleigh-Benard experiment with about 3.6% relative error, reproducing thermal plumes and the teardrop-shaped Q-R velocity-gradient distribution, though it smoothed extreme dissipation events.

    Evidence for the claim as stated.

  • The studies probe very different regimes: marginal wall-bounded turbulence near transition, fully developed magnetised plasma turbulence, buoyancy-driven convection at Prandtl number 10.6, and idealised isotropic turbulence. Intermittency appears in more than one of them, but their scaling laws are not interchangeable.

    Evidence for the claim as stated.

Open questions

Tensions this paper is part of

From concept pages' “where studies disagree.” Disagreement means the same question; scope means different assays, populations, or outcomes.

Related papers in this topic

Same topic cluster — not a recommendation engine.