In 1934, H. Frenzel and H. Schultes were trying to speed up photographic development by running an ultrasound transducer through the developer bath. They weren't trying to make light. They noticed strange dots appearing on the film — tiny spots that shouldn't have been there. The bubbles created by the transducer were emitting light. Nobody was looking for this. It was a side effect of looking for something else.
The phenomenon is sonoluminescence: light emitted by collapsing bubbles. The basic mechanism, worked out over the following decades, goes like this. An acoustic standing wave at roughly 26–40 kHz creates a pressure field that cycles between positive and negative. During the low-pressure phase, dissolved gas nucleates into a bubble that expands from a few microns to roughly 50 microns in radius. When the pressure reverses, the bubble wall accelerates inward. The collapse is essentially adiabatic — so fast that no heat escapes — and the gas compresses catastrophically. The bubble wall eventually exceeds the speed of sound in the gas. At minimum radius, the bubble has compressed a volume that was 50 microns across down to roughly half a micron — a volume reduction on the order of 10^9. The light flash occurs at this moment.
The flash lasts between 35 and 200 picoseconds. A picosecond is 10^-12 seconds. The acoustic cycle it belongs to has a period of 38 microseconds — 38,000 nanoseconds, 38,000,000 picoseconds. The light flash occupies roughly 0.001% of that cycle. The bubble spends essentially all of its time doing nothing remarkable, and then does something extraordinary for a duration that makes femtosecond laser pulses look leisurely. Seth Putterman at UCLA noted that sonoluminescence is "the only known method of producing picoseconds of light without complex, expensive lasers." It is also, as far as anyone can tell, the most energetically concentrated spontaneous event in ordinary fluid dynamics: sound at centimeter scales collapsing to light at nanometer scales, a focusing of energy density by roughly twelve orders of magnitude.
The striking thing — the thing I keep returning to — is that nobody can see it happen. Not directly. The hot zone at peak compression has a radius smaller than a bacterium and lasts a fraction of a nanosecond. There is no instrument that can observe the interior of the collapse as it occurs. Every temperature measurement is indirect: you look at the spectrum of the light flash and fit it to a plasma emission model; you inject a molecule whose decomposition products depend on temperature in a known way; you measure ionic emission line ratios from metal compounds added to the liquid. These methods give numbers ranging from 6,000 K to over 20,000 K, and they disagree with each other by factors of 2–10, partly because each method averages over different volumes and timescales.
The theoretical models push higher. Converging shock-wave models, which posit that the supersonic bubble wall launches an inward-traveling spherical shock that intensifies as it converges on the center, predict temperatures that diverge toward infinity at the focus — the Guderley scaling for self-similar implosion. In practice, real physical limits intervene before infinity, but the implication is that the true peak temperature at the center may be far higher than any spectroscopic measurement can detect. The plasma at the core is optically thin: it doesn't absorb enough of its own radiation to reach thermal equilibrium with the light it's emitting, so the spectrum encodes the emission from the whole volume, averaged, not just the hottest point.
Putterman and Weninger wrote in their 2000 Annual Review paper: "neither the imploding shock nor the plasma has been directly observed." This was not a complaint about the state of instrumentation at the time. It was a statement about the fundamental limits of observation for an event of these dimensions. The converging shock wave that the models predict is smaller than the wavelength of visible light for most of its inward travel. The plasma lasts less time than it takes light to cross a cell nucleus. You cannot look at it. You can only look at what it leaves.
What it leaves is the spectrum of the flash. The spectrum is featureless and continuous — no emission lines from particular atoms or molecules, just a smooth curve that rises steeply into the ultraviolet until it hits the absorption cutoff of water at around 200 nanometers. This is more consistent with bremsstrahlung from a weakly ionized plasma than with the thermal radiation of a hot blackbody: a genuine blackbody would fall off exponentially at short wavelengths, while free-electron radiation in a thin plasma decays much more gradually. The spectrum says "plasma" without showing you the plasma.
There is one detail that I find particularly good. A bubble driven into single-bubble sonoluminescence starts out as an air bubble — nitrogen, oxygen, argon, and traces of other gases in the proportions they appear in dissolved air. After many acoustic cycles in the stable luminescent state, the bubble has become almost pure argon. The nitrogen and oxygen react: at 10,000+ K, N2 and O2 break into radicals, recombine into NO and other species, and dissolve back into the surrounding water. Argon is a noble gas. It cannot react. There is nowhere for it to go. Cycle by cycle, the reactive gases disappear and the argon accumulates, until the bubble's chemistry has self-selected for the one component that survives the conditions it creates. The bubble purifies itself. This takes hours — the bubble might undergo millions of cycles before it reaches equilibrium — and it happens silently, with no external intervention, just chemistry running to its natural endpoint under conditions the bubble generates.
The Taleyarkhan controversy of 2002 is worth noting briefly not because it ended well but because of what it illustrates about inference in extreme conditions. Rusi Taleyarkhan published a paper in Science claiming to have observed nuclear fusion inside collapsing bubbles filled with deuterated acetone — 2.45 MeV neutrons, tritium buildup, signals correlated with the sonoluminescence pulse. D-D fusion requires roughly 10^8 K. The spectroscopic temperature estimates for ordinary sonoluminescence are around 10,000–20,000 K. The claim required the collapse to be five thousand times hotter than anything anyone had measured. An immediate independent replication at Oak Ridge found neutron counts consistent with random coincidence, at least three orders of magnitude below genuine fusion. Later attempts at replication found nothing. What Taleyarkhan had found, eventually, was a misconduct investigation and a stripped professorship. The gap between "inferred from spectrum" and "physically confirmed" left room for someone to imagine a different inference. The event is too brief and too small. The room for interpretation is large.
The bubble's timing, as a coda, is more precise than the oscillator driving it. The period between successive flashes varies by no more than 40 picoseconds. The electronic oscillator generating the 26 kHz driving field varies by more than that. The bubble, despite its violent collapse and chaotic internal dynamics, self-organizes into temporal regularity finer than the signal that created it. This also has never been fully explained. It is a property of the collapsed state, not of the driving field. The system settles into an order that the input doesn't contain.