AI turns boiling sounds into measurements of hidden bubbles

Water can be cool enough to remain liquid through most of a channel while boiling against its hot metal wall. Tiny surface cavities start producing vapour bubbles. A bubble grows, detaches and gives cooler liquid access to the surface again. Much of the useful cooling happens in this restless layer.[2]

At Los Alamos’s Isotope Production Facility (IPF), cooling is part of making radioisotopes. A 100-MeV proton beam irradiates targets in a water-cooled holder; the resulting isotopes serve medicine, research and industry. Chemical processing subsequently separates the desired products. The target system must accommodate the energy delivered by the beam as well as the nuclear reactions it induces.[3]

A 2023 study developed laboratory experiments and simulations for that cooling problem. Radiation and shielding prevent direct in-beam observation of the target. The team reproduced relevant conditions outside the accelerator and used measured bubble behaviour to help calculate a boiling curve, connecting wall temperature with heat removal.[4]

A camera teaches an underwater microphone

The acoustic study, published on 25 March 2026, used TARDIS-β, an electrically heated test loop representing one IPF cooling subchannel. Camera observations supplied bubble labels for learning from hydrophone recordings.[1]

A convolutional neural network processed spectrograms to estimate bubble size, departure frequency, waiting time and the density of active bubble-forming sites, alongside heat flux and wall superheat.[1]

A spectrogram displays how the strength of different sound frequencies changes with time. It makes an acoustic recording into a pattern that an image-processing network can analyse. The learning task links that pattern to a physical measurement, so the output has units and a place in an engineering calculation.

The main training/validation split divided acoustic samples 70:30 within known operating conditions. The CNN’s advantage over simpler regressors appeared across broader intermediate conditions, rather than consistently at the fixed IPF setting.[1]

Why the size and timing matter

The RPI wall-boiling model used in Ansys CFX divides heat transfer into contributions from evaporation, ordinary convection and quenching—the fresh liquid heating after a bubble departs. Bubble departure size, frequency, site density and waiting time help determine those contributions. Detailed bubble physics occurs below the resolution of the larger flow simulation, so these inputs fill in processes its grid cannot directly resolve.[2]

A simple geometric example shows why counting bubbles is insufficient. Two spherical bubbles with diameters in a ratio of two have volumes in a ratio of eight. At the same vapour density, the larger contains eight times as much vapour. A stream of large bubbles and an equally frequent stream of small ones therefore represent different amounts of evaporated material. This is an illustration, not a measured result from the experiment.

Boiling curves from flow simulations supplied with acoustic estimates agreed with image-fed simulations within a ±25% reference band. This tested physical consistency between input methods, not independent experimental validation of critical heat-flux prediction.[1]

Separate transfer tests changed water temperature, flow and microphone position. Performance held reasonably near training conditions, with temperature shifts of about 20°C; larger departures reduced reliability.[1]

A physical model makes the bubble measurements useful. An engineer can ask which part of heat removal comes from producing vapour, which from renewed contact with liquid, and which from the surrounding flow. The bubbles become inputs to an account of where the energy goes.[2]

Report an error or suggest a correction