Chapter 4: The Death of the Cavern

You have seen the diagram. Everyone has seen the diagram.

A cone in cross-section, drawn in brown or grey. A vertical pipe running up its middle. And underneath, at the bottom of the pipe, a red or orange blob with a rounded outline: the magma chamber. Sometimes there is a little arrow pointing at it, and sometimes the label says reservoir of molten rock, and in the better textbooks there is a scale bar.

It is on the wall of every geography classroom in the world. It is the picture that comes into your head when someone says the word volcano. I drew it myself, aged about nine, in coloured pencil, and I suspect you did too.

And it is wrong. Not simplified — most teaching diagrams are simplified and that is fine — but wrong in its central claim, which is that beneath a volcano there is a discrete void containing liquid rock. There is no cavern. There has never been a cavern. The thing under a volcano is a different kind of object entirely, and working out what it actually is has been the principal achievement of volcanology over the last twenty-five years.

I ended the last chapter by saying that nobody ever observed the magma chamber — that it was a regularization choice, an assumption imported to make an underdetermined problem solvable, which then hardened into an image and got drawn in every textbook for a century. This chapter is about what replaced it, how the replacement was established, and — because I promised you the honesty would be mechanical — how much of the replacement is itself still inference.

Register: the mush model is now the mainstream position and rests on several independent lines of evidence, so I will treat its core as empirical. The details — melt fractions, lens lifetimes, extraction mechanisms — are actively argued and I will mark them theoretical as we go.

Why the cavern had to die

Three separate lines of evidence converged on the same conclusion, which is the pattern you want before you throw out a century of diagrams.

The geophysics never found it. As seismic imaging of volcanoes improved through the 1990s and 2000s, the expectation was that better resolution would sharpen the outline of the chamber. It did not. What the images kept showing beneath active volcanoes were broad, diffuse regions of moderately reduced seismic velocity — consistent with a little melt distributed through a lot of rock, and not at all consistent with a large body of liquid. The melt fractions being inferred were low: single figures to a few tens of per cent at best, over regions much larger than any drawn chamber. The better the instrument got, the more the cavern failed to appear.

The thermodynamics forbade it. This argument is the one I find most decisive, and it requires no imaging at all. Put a large body of liquid magma in the upper crust and ask how long it stays liquid. The answer, from straightforward heat-flow modelling, is: not long. Crust is cold. A magma body loses heat to it, and as it cools it crystallizes, and the timescale for a substantial upper-crustal body to freeze up is on the order of thousands to tens of thousands of years — geologically, an eyeblink — unless it is being continuously resupplied with fresh hot magma at a rate high enough to hold the crystallization off.

But we have volcanic systems that have been demonstrably active in the same place for hundreds of thousands of years. The cavern model requires those systems to have stored a lake of liquid rock for that entire time. The heat budget simply does not permit it. Whatever is down there, it cannot have been mostly liquid for most of its life.

The crystals gave it away. And then the petrological evidence, which is the most direct of the three. When you look closely at the crystals carried up in an erupted magma, a great many of them turn out to be older than the eruption — sometimes far older. They have complex growth histories, multiple episodes of resorption and regrowth, and radiometric ages that can predate the eruption by tens or hundreds of thousands of years.

These crystals were not grown in the liquid that carried them out. They were picked up. They had been sitting somewhere, in some kind of long-lived storage, and were entrained shortly before the eruption by melt moving through them.

Which tells you what the storage is made of. Not liquid. Crystals.

Mush

Here is the replacement picture.

Beneath a volcano is a region of rock that is hot, partially molten, and mostly crystalline: a dense framework of interlocking crystals with silicate liquid occupying the spaces between them. Not a chamber of liquid with crystals floating in it. A rigid sponge of crystals with melt in the pores.

The technical term is crystal mush, and I want to insist on how different an object this is from the diagram, because the difference is not cosmetic.

The key property is mechanical, and it has a threshold. When the crystal content of a magma is low, the crystals are suspended in liquid, they do not touch, and the whole thing behaves as a fluid — it can convect, it can flow, it can be erupted. As the crystal content rises, the crystals begin to interfere with one another, and somewhere in the region of half crystals by volume they lock into a continuous touching framework. Past that point the material has a yield strength. It is no longer a liquid containing solids; it is a solid containing liquid, and it will not flow, and — this is the part that matters for everyone living downhill — it cannot erupt.

The consequence is the single most important thing in this chapter. Most magma, most of the time, is not eruptible. It is locked into a mush framework, hot, ready in a chemical sense, and mechanically unable to go anywhere.

For an eruption to happen, some quantity of melt has to be extracted from that mush and assembled into a body that is liquid enough to move — a lens or a pocket of high melt fraction, formed transiently within the larger mush column. Those lenses are what erupt. They are small compared with the system that produced them, they are short-lived, and they are extremely difficult to detect, because a modest lens inside a large warm mush body barely changes the geophysical signal at all.

The old question — how big is the magma chamber? — turns out to be malformed. The useful questions are how much eruptible melt currently exists, whether it is connected, and what could mobilize more of it. Those are much harder questions and we are much worse at answering them.

How the extraction happens is theoretical and contested: compaction of the crystal framework under its own weight, gas-driven filter pressing, melt buoyancy through a permeable crystal network, or the mechanical disturbance of new magma arriving from below. There is a reasonable literature on each and no settled resolution.

The volume problem, deferred

There is an obvious objection to all this, and since it is the correct objection I want to raise it myself rather than let it sit unmentioned.

If magma cannot be stored as liquid, and eruptible lenses are small and transient, how do we account for eruptions that discharge hundreds or even thousands of cubic kilometres of crystal-poor melt in a matter of days?

That is a real tension, it is at the centre of an active argument, and it is the subject of Chapter 10. The short version is that the melt must be extracted from a mush volume very much larger than the erupted volume — either slowly, accumulating over a hundred thousand years or more, or rapidly, mobilized out of a long-dormant mush body on a timescale of decades. Different dating techniques give different answers, and the people holding each position are not fools. I will lay that out properly when we get to the calderas.

For now, note only that the mush model makes the problem harder rather than easier, and that a model which creates difficulties for itself is usually a sign of a model that is tracking something real.

How to read a tomographic image

Since much of the evidence above comes from seismic imaging, and since you are going to encounter such images, I want to give you the tools to read one sceptically. This section is more technical than most of the book. It is also, I think, the most practically useful thing in it.

What tomography does. Take earthquakes — thousands of them, occurring at various places and depths — and record their arrivals at a network of seismometers. Each ray path from a source to a receiver samples the rock along its route, and the travel time tells you the average speed along that route. Collect enough intersecting paths from enough directions and you can solve for the three-dimensional velocity structure of the volume they cross. It is the same mathematics as a medical CT scan.

Why it sees melt. Melt reduces seismic wave speeds. Crucially, it reduces shear-wave speed more than compressional-wave speed, for the reason established in Chapter 1: liquid has no rigidity, so anything that introduces liquid degrades the material's ability to transmit shear disproportionately. The ratio between the two velocities is therefore a more diagnostic indicator of melt than either alone. Attenuation — how much energy the wave loses — is also sensitive.

Now the four things to check before believing a picture.

One: where were the rays? You can only image where the rays went, and rays go from earthquakes to stations. Earthquakes are not distributed for your convenience; they happen where the rock is breaking. Regions with poor ray coverage appear in the final image looking exactly like regions with good coverage — coloured, confident, and meaningless. A responsible paper includes resolution tests showing which parts of the model the data actually constrains. Look for them.

Two: is it smeared? Errors in tomography tend to spread along ray paths, producing elongate artefacts oriented in the direction of the dominant ray geometry. An anomaly whose long axis happens to align with the ray coverage should be treated with suspicion.

Three — and this is the big one — what did the inversion prefer? The problem is underdetermined, as everything in the last chapter was, so it must be regularized, and the standard regularization penalizes roughness. The inversion is instructed to find the smoothest model compatible with the data.

Think about what that does. A small body with a high melt fraction and a large body with a low melt fraction can fit the same travel-time data. The smoothness penalty will systematically prefer the second. Tomographic images are therefore biased toward showing big diffuse anomalies rather than small intense ones — and they will do that even if the true structure is a set of sharp lenses.

This cuts both ways in a manner I want to be scrupulous about. Above, I used the failure of imaging to find high melt fractions as evidence against the cavern. That argument is weakened, though not destroyed, by the fact that the method is predisposed to smear intense features into diffuse ones. It is one reason the thermal argument matters so much: it is independent of imaging, and it does not care what the inversion preferred.

Four: how was velocity converted to melt fraction? This step is rarely emphasized and carries enormous uncertainty. The effect of melt on seismic velocity depends on the geometry of the melt — whether it sits in thin films along grain boundaries or in isolated rounded pockets. Thin films are elastically devastating; a very small amount of melt in that configuration produces a large velocity drop. Rounded inclusions of the same volume produce far less. So the same measured anomaly can be interpreted as a small melt fraction in films or a much larger one in pockets, and the number quoted in the abstract depends on a modelling assumption about geometry that nobody has observed directly.

None of this makes tomography untrustworthy. It makes it interpretable, which is a different and more demanding condition. When you next see a cross-section of a volcano with a red blob in it, the question to ask is not "is there magma there" but "what family of structures is compatible with this data, and which member of that family am I being shown."

The system is a column

The other thing that changed, alongside mush, was the vertical extent.

The old picture had a chamber at some depth — a few kilometres, usually — connected to the surface by a conduit and to nothing much below. Current thinking replaces this with a transcrustal magmatic system: a vertically extensive column of mush and melt distributed through the entire crust, from the mantle boundary upward, with local zones of higher melt fraction at various levels and only the topmost of them functioning as the immediate source of an eruption.

Magma in such a system is not stored in one place. It is processed through a column — rising, stalling, cooling, crystallizing, mixing with what is already there, being remobilized by the next batch from below, and slowly changing composition all the way up. The distillation described in Chapter 1 does not happen at a point. It happens continuously along a vertical assembly line tens of kilometres long.

Two consequences worth carrying forward.

For hazard, the shallow reservoir is a staging area, not the system. Its state is set by what is happening beneath it, and what is happening beneath it is largely invisible — a deep source produces surface deformation that is broad, low-amplitude, and easy to mistake for nothing.

For the argument of Part III, this is exactly what I need. A transcrustal system is not an event. It is a standing structure, operating continuously over hundreds of thousands to millions of years, of which eruptions are the occasional overflow. That is a description of a process, and processes are the kind of thing that can build a planet.

The eruption nobody was watching

Which brings me to Hunga Tonga–Hunga Haʻapai, and to an awkward observation about how the discipline allocates its attention.

On the fifteenth of January 2022, a submarine volcano in the Kingdom of Tonga produced the most powerful explosion ever recorded by instruments. The eruption column reached something like fifty-seven kilometres, punching into the mesosphere — the highest column ever measured, and the largest since Krakatoa in 1883. An atmospheric pressure wave propagated outward and circled the globe more than once, registering on barometers everywhere, including on amateur weather stations in suburban gardens. A tsunami was recorded in every ocean basin. The submarine setting meant seawater had direct access to the magma, and the violence of that interaction is a large part of why the eruption behaved as it did — and why it injected an extraordinary quantity of water vapour, rather than mainly sulfur, into the stratosphere.

For a chapter about the difficulty of observing volcanoes, this event is remarkable for one reason: almost every instrument humanity possesses saw it at once. Geostationary weather satellites captured the column at high cadence. The global hydroacoustic network heard it through the ocean. Seismometers, barometers, tsunami gauges, lightning detection networks, and GNSS receivers measuring disturbances propagating through the ionosphere all recorded aspects of the same event simultaneously.

A single event observed that densely constrains models in ways that decades of sparse observation cannot, because it removes the usual excuse. When you have one instrument you can always attribute a discrepancy to the instrument. When fifteen independent physical systems record the same eruption from different angles, the models have nowhere to hide.

And now the awkward part. Hunga Tonga was a remote submarine volcano with minimal local monitoring, and nobody was watching it in the sense that matters. It was not on a shortlist. The best-observed eruption in human history was observed by global systems built for other purposes, largely by accident, while the volcano itself had almost nothing on it.

I do not raise this to score a point. Monitoring resources are finite and are, quite properly, concentrated where populations are. But it is worth noticing that our richest dataset came from a volcano we were not studying, and that the reason we have it is that the twenty-first century happens to be blanketed in instruments pointed at everything else.

Models, and the temptation of a beautiful figure

The last piece of Part II's methodological account is simulation, because increasingly the way a volcanologist explores an inaccessible system is to build one in a computer.

The physics is genuinely hard and the modelling is genuinely good. A conduit model solves the coupled equations for a multiphase mixture — melt, crystals, exsolving gas — ascending a pipe, with viscosity changing as water leaves the melt, permeability developing and sealing, and a fragmentation criterion somewhere in the middle. Out of it you get eruption rates, column heights, the conditions under which a column collapses into pyroclastic flows. These models reproduce real eruptions well enough to be useful, and they have taught the field things that could not have been learned any other way.

But I want to be exact about what they depend on, because the failure mode here is specific and seductive.

Every one of those models needs, as input: a viscosity law (which depends on composition, dissolved water, crystal content and temperature, each with its own uncertainty); a permeability law relating gas escape to bubble connectivity (poorly constrained by experiment); a fragmentation criterion (which is a modelling choice, not a measured quantity); conduit geometry (unknown — the conduit is underground); and initial conditions at depth (unknown, for all the reasons in Chapter 3).

Run such a model and it will produce an answer of arbitrary precision, rendered in colour, at whatever resolution you can afford. The output will look far more certain than the inputs justify. This is not a criticism of the modellers, who are as a rule scrupulous about stating their assumptions; it is a warning about what happens to a figure once it leaves the paper it was published in and starts appearing in presentations.

The correct use of these models is not to predict what a specific volcano will do on a specific day. It is to establish sensitivity: to find out which parameters the outcome actually depends on, and how strongly, and therefore where measurement effort should go. A model that tells you the answer hinges almost entirely on a quantity nobody has measured has done its job — arguably better than one that produces a confident number.

Where this leaves us

Let me draw Part II's account together, because the next chapter is going to test it against a town.

We cannot instrument the object. We infer it from surface signals, and the inference is formally underdetermined — many subsurface configurations fit the same data. The images we produce are shaped by the assumptions required to produce them at all, which is how a picture nobody ever observed ended up in every textbook for a century. And the object itself turns out to be far stranger and less tractable than the picture: not a chamber of liquid but a crystal mush column tens of kilometres tall, mostly unerupatible, occasionally assembling transient lenses of melt whose formation we cannot reliably detect.

That is the honest state of knowledge. It is a great deal of knowledge — I hope this part of the book has conveyed that the indirect methods are ingenious rather than merely regrettable — and it is nowhere near enough to tell you when.

So now put a scientist in front of a civil authority, with that state of knowledge, and a population downslope, and a decision that has to be made this week.

That is Chapter 5, and it is the chapter this book was written to arrive at.

Draft notes — verification status

Standing convention. The mush model's core is mainstream and safe in substance; the numbers and attributions are the exposure. Hunga Tonga's column height and the disputed radiative sign were resolved during the thesis-paper verification pass.

Already verified — do not re-check:

• Hunga Tonga–Hunga Haʻapai eruption date (15 January 2022); column height ~57 km, reaching the mesosphere; highest recorded since Krakatoa 1883; large stratospheric water vapour injection.
• That the radiative sign is disputed — earlier work reported net warming, more recent work slight cooling. Not asserted in this chapter; keep it that way here and handle it in Chapter 12.

Pending verification:

• The rheological lockup threshold — the crystal fraction at which a mush develops yield strength and becomes uneruptible (commonly given as ~50–55 vol%, sometimes framed as a 40–70% "rheological window"). The draft says "somewhere in the region of half by volume"; tighten only with a source.
• Thermal lifetime of an upper-crustal magma body without recharge ("thousands to tens of thousands of years"). Check against the modelling literature; the figure is strongly volume- and depth-dependent.
• Typical melt fractions inferred from tomography beneath active volcanoes ("single figures to a few tens of per cent").
• Antecryst ages exceeding eruption ages by "tens or hundreds of thousands of years" — verify the range and cite a specific well-documented system rather than generalizing.
• Attribution of the crystal mush framework: Bachmann & Bergantz (2004) on rhyolite extraction from crystal-rich residues; Hildreth, Cashman, Sparks, Marsh as principal contributors. Confirm before naming anyone in the text — the draft currently names nobody, which may be the safer choice.
• Cashman, Sparks & Blundy (2017), "Vertically extensive and unstable magmatic systems," as the standard reference for the transcrustal model.
• The claim that Vs is reduced more than Vp by melt, and that Vp/Vs is the more diagnostic indicator. Directionally certain; confirm the standard framing.
• Melt geometry effects on elastic properties — the thin-film versus isolated-pocket contrast and its magnitude.
• Hunga Tonga instrument list: geostationary satellite cadence, IMS hydroacoustic detection, ionospheric TEC disturbance, lightning rates, global barometric Lamb wave and number of circuits. Verify each; the draft asserts several.
• The characterization of Hunga Tonga's pre-eruption monitoring as minimal — this is the chapter's pointed claim and must be right.
• Whether "most powerful explosion ever recorded by instruments" is defensible as stated, and against what metric.