Chapter 2: The Physics of the Cork

In June of 1991, on the Philippine island of Luzon, Mount Pinatubo produced the second largest eruption of the twentieth century. It put enough sulfur into the stratosphere to measurably cool the entire planet for the better part of two years. The column reached the stratosphere in a matter of hours. Pyroclastic flows filled valleys to depths of tens of metres and were still hot enough, years afterward, to flash rainwater into steam.

In June of 1991, on the island of Hawai‘i, Kīlauea was in the eighth year of an eruption that had been running more or less continuously since 1983, and people were driving out after dinner to look at it.

Same planet. Same month. Same fundamental process — mantle-derived melt arriving at the surface. One of them was a scheduled evening outing. The other one changed the global mean temperature.

This chapter is about why.

I want to be careful about the register. The mechanics I am going to describe — how gas comes out of magma, why some magma is stiff and some is runny, what happens when a foam shatters — are empirical and well established, reproduced in laboratories and matched against the deposits of hundreds of eruptions. The question of what actually pulls the trigger on any particular occasion is theoretical, contested, and probably has more than one answer. And the argument at the end of the chapter, about whether eruption timing is knowable even in principle, is where I will start being careful about how far the evidence carries.

Not a boiling pot

Let me first take away another picture, because it is nearly as widespread as the sea-of-magma one and it fails in a more interesting way.

The instinctive model of an explosive eruption is thermal. The magma gets hotter and hotter until it boils over, the way a pot does — heat drives the violence, and the violence is a kind of overflow.

It is not. Magma does not boil in that sense, and the temperature of an explosive magma is generally lower than that of a placid one. The energy that blows a mountain apart is not thermal energy in any direct way. It is stored chemical-mechanical energy — gas held in solution under pressure — and the mechanism is far closer to a bottle of champagne than to a saucepan.

Here is the physics, and it turns on a single dependency.

Magma at depth contains dissolved volatiles: water above all, and then carbon dioxide, sulfur species, chlorine, fluorine. The water is not present as steam. It is dissolved into the silicate liquid at the molecular level, chemically bound into the melt structure, in the same way that carbon dioxide is dissolved into a sealed bottle of sparkling water and is not, while the bottle is sealed, in the form of bubbles.

And the amount of water a silicate melt can hold in solution depends, overwhelmingly, on pressure. High pressure, high solubility. Drop the pressure and the melt can no longer hold what it is carrying, and the water comes out of solution as a separate gas phase. This is called exsolution, and it is the engine of every explosive eruption that has ever happened.

So the sequence is: magma rises, pressure falls, dissolved water exsolves into bubbles, and the bubbles — because gas is enormously compressible and expands as the pressure keeps dropping — take up more and more room.

That last clause is where people's intuitions fail, so let me be blunt about the arithmetic. A magma carrying a few per cent water by mass sounds unimpressive. A few per cent of anything sounds unimpressive. But mass is the wrong currency. What matters is volume, and when that water comes out of solution at shallow depth and expands toward atmospheric pressure, the gas it produces occupies a volume vastly greater than the melt that was carrying it — by factors in the hundreds or thousands.

The magma does not gain energy on the way up. It simply stops being able to keep what it was already holding, and what it was holding wants to be a great deal larger than the space available.

The moment a liquid becomes a spray

Bubbles have to start somewhere. In a melt they nucleate — often on crystal surfaces, which is why the crystal content of a magma turns out to matter for how it erupts — and then they grow, both by more water diffusing into them from the surrounding melt and by simple decompression as the whole assemblage keeps rising.

Picture what this does to the material. You begin with a liquid containing a few small bubbles. As it ascends you get a liquid containing many bubbles. Then a liquid that is mostly bubbles, separated by thinning films of melt: a foam. And then the films become too thin to survive the stresses acting on them, and the foam comes apart.

That moment has a name — fragmentation — and it is the single most important transition in volcanology.

Before fragmentation, you have a liquid with gas suspended in it. After fragmentation, you have a gas with liquid fragments suspended in it. The continuous phase and the dispersed phase swap places. And because the continuous phase is now a gas that is still violently expanding and no longer has to drag a viscous liquid along with it, the mixture accelerates catastrophically. Everything above the fragmentation level is travelling upward at speeds that can be a substantial fraction of the speed of sound. Everything below it is still, comparatively, oozing.

An explosive eruption is, in the most literal sense, the moment a magma stops being a liquid and becomes a spray. Somewhere in the conduit there is a surface — a moving, unstable, physically real boundary — with bubbly liquid beneath it and a jet of hot gas and shattered glass above it. The whole subject can be organised around whether a given magma reaches that surface or not.

There is a second route to fragmentation which I find conceptually delicious, and which connects directly back to the last chapter. Silicate melt is viscoelastic: it responds as a liquid when deformed slowly and as a brittle solid when deformed fast. The material has a characteristic relaxation time, and if you deform it faster than it can relax, it does not flow — it breaks, like glass, because under those conditions it effectively is glass. In a rapidly accelerating conduit the strain rates can exceed that threshold, and the magma shatters as a solid would.

In Chapter 1 I said that whether something behaves as a solid or a fluid depends on the ratio of its relaxation time to the time you spend watching. That was how we got the mantle to flow while staying solid. Here is the same principle running the other way: hold a liquid and deform it fast enough, and it will shatter in your hands.

The one variable that decides everything

So why does Kīlauea pour and Pinatubo detonate?

Both magmas contain dissolved volatiles. The difference is what the melt does to those volatiles on the way up, and that comes down to a single physical property, which is viscosity — the melt's resistance to flowing.

Silicate melts are built from tetrahedra of silicon and oxygen. Those tetrahedra can link at their corners into chains, sheets and three-dimensional networks — a process called polymerization — and the more silica a melt contains, the more thoroughly polymerized it is, and the stiffer it becomes.

Basalt, the direct product of mantle melting, is comparatively poor in silica and only weakly polymerized. It is runny. Rhyolite, the end product of the distillation chain described in Chapter 1, is rich in silica and heavily polymerized. It is not runny at all. And the gap between them is not a matter of degree in any ordinary sense: the viscosity range spanned by natural magmas covers something like ten to twelve orders of magnitude. There are very few variables anywhere in the earth sciences with that kind of dynamic range. Comparing basalt to rhyolite is not comparing honey to treacle. It is comparing water to something that will barely move at all on a human timescale.

Two further complications, both of which matter.

First, dissolved water does not merely sit in the melt — it actively breaks the polymer network apart, and therefore lowers viscosity. Which sets up a nasty feedback. As magma rises and degasses, it loses the very thing that was keeping it fluid. Degassing stiffens the melt. A stiffer melt is worse at letting the remaining gas escape. That is a loop that runs in exactly the wrong direction.

Second, crystals. As magma cools it crystallizes, and once the crystal content passes roughly a third of the volume, the crystals begin to interfere with one another and the effective viscosity climbs steeply. A crystal-rich magma is a much less cooperative object than a crystal-poor one.

Now the payoff. In a low-viscosity basalt, bubbles can rise through the melt and escape. They can also link up into a connected network, making the magma permeable, so that gas drains away through the foam and out through fractures in the conduit walls. Gas that has left the system cannot fragment it. So basaltic eruptions tend to be effusive — lava flows, lava lakes, fire fountains — spectacular, sometimes destructive of property, and generally survivable by anyone who walks briskly.

In a high-viscosity rhyolite or dacite, none of that works. Bubbles cannot rise through the melt at any useful rate. Permeability develops badly and seals itself. The gas is stuck, and it keeps exsolving, and the pressure keeps building inside a material too stiff to relieve it — until fragmentation, and then everything happens at once.

Kīlauea's magma lets its gas go. Pinatubo's could not.

And notice where those magmas came from. Kīlauea sits over a hotspot, fed by decompression melting of the mantle: dry, hot, basaltic. Pinatubo sits on a subduction zone, and its magma is the product of flux melting — mantle melted by water driven out of an oceanic plate on its way down. Which is to say, as I put it at the end of the last chapter, that the volcanoes which kill people in large numbers are, at bottom, seawater coming back up. Water makes the magma in the first place, and then water is the thing that blows it apart.

The valve, and the volcanoes that do both

I should complicate the clean two-personality story, because if you leave this chapter thinking volcanoes come in two types you will misread most of what happens on real mountains.

Permeability is the hinge, and permeability is not a fixed property. It develops and it seals, sometimes repeatedly, in the same conduit over hours. A volcano that is degassing efficiently is quiet — you can watch a plume drift off the summit for years and nothing much happens. A volcano whose plumbing has sealed itself is charging.

This is why viscous magma so often builds a dome: a pile of degassed, stiff lava extruded slowly at the vent, essentially a plug. Domes are unstable objects. They grow, oversteepen, and collapse — and when a hot, gas-charged dome collapses, the sudden unloading depressurizes the magma behind it and it fragments explosively, sideways, into the collapse scar.

Merapi does this. Soufrière Hills does this. And so does Unzen, which is where three people we will meet properly in the next chapter were standing on the third of June, 1991, twelve days before Pinatubo.

What actually pulls the trigger

Register: theoretical from here to the end of the section. The mechanisms below are all plausible, several are individually well evidenced, and there is no consensus that any one of them is primary.

Everything so far explains what happens once magma is ascending. It does not explain why the ascent starts on a particular Tuesday after four hundred years of nothing. That is the trigger problem, and it is unsolved.

The leading candidates:

Recharge. A batch of hot, mafic, gas-rich magma arrives from below and is injected into a cooler, more evolved reservoir. It heats the resident magma, adds volatiles, disturbs it mechanically, and pressurizes the system. The evidence for this is unusually good — you can read it directly in the erupted products, in banded pumice where two magmas were mingled but not fully mixed, and in crystals whose outer rims record a sudden change in the conditions they were growing in. Better still, the rate at which chemical gradients blur inside those crystals is a clock: measure how far a zoning boundary has diffused and you can estimate how long the crystal sat in the new conditions before it was erupted. The answers are often startlingly short — weeks, sometimes days.

Second boiling. A reservoir can pressurize itself with no new input at all. As magma cools it crystallizes, and the minerals that crystallize first are largely anhydrous — they take no water into their structures. So the water that was dissolved in the melt is progressively concentrated into the shrinking volume of remaining liquid, until that liquid becomes saturated and the water exsolves. Gas appears, volume increases, pressure rises. The system detonates itself by sitting still and getting colder, which is a genuinely strange thing for a system to do.

Buoyancy and roof failure. Melt is less dense than the rock above it. If enough accumulates, the buoyancy overpressure alone can exceed the strength of the roof, and the failure is then a structural event rather than a chemical one. This mechanism scales badly and interestingly — it becomes more important the larger the reservoir gets, which is why it reappears in Chapter 10 when we get to the calderas.

External forcing. And then the untidy category: things outside the magmatic system that appear, statistically, to nudge it. Large tectonic earthquakes are followed by nearby eruptions somewhat more often than chance would predict. Deglaciation unloads the crust and appears to have been followed by a sharp increase in Icelandic eruption rates. Some volcanoes show weak correlations with tidal cycles, with seasonal loading, with heavy rainfall. Individually these signals are small and several are disputed. Collectively they suggest a system sitting so close to its threshold that quite modest external stresses can matter.

That last observation is, I think, the most important thing in this section, and it points at the chapter's conclusion. If a volcano can be tipped by rainfall, the volcano was already almost going.

The honest summary is that we have a good list and no ranking. Different volcanoes probably work differently; the same volcano probably works differently on different occasions; and the deposits, which are our main evidence, record what happened rather than what would have happened otherwise. I do not think anyone can currently tell you what triggers eruptions in general, and I would be suspicious of anyone who says otherwise without a great many qualifications attached.

What actually kills people

Now the part that matters most, and where the popular picture is not merely incomplete but inverted.

Ask someone how a volcano kills and they will describe lava. Rivers of it, chasing people down a hillside. It is the image every disaster film reaches for.

Lava flows almost never kill anyone. They are, in the overwhelming majority of cases, slow. They destroy houses, farms, roads, whole towns — Paricutín took two — and they are economically devastating and emotionally unbearable to watch. But they typically advance at a walking pace or less, and human beings can, and do, simply leave. The lava flow is the hazard that takes everything except your life.

There is one significant exception, and it is worth naming here because it returns in Chapter 13. Nyiragongo, in the eastern Democratic Republic of the Congo, produces lava of such unusually low silica content — and therefore such extraordinarily low viscosity — that its flows have been recorded moving at highway speeds. It is the volcano where the cinematic version is roughly true, and there is a city of well over a million people at the bottom of it.

So what does the killing?

Pyroclastic density currents. The dominant cause of volcanic death in the historical record, and the subject of the next section.

Lahars. Volcanic mudflows: a slurry of ash, debris and water with the consistency of wet concrete, moving fast and channelled down river valleys. They can be generated by an eruption melting snow and ice, or by rain remobilizing loose ash months or years afterward. They do not require a large eruption. They travel far beyond any reasonable-looking danger zone, because they follow drainage, and human settlements also follow drainage. The single deadliest volcanic event of the late twentieth century was a lahar, generated by a modest eruption, that reached a town more than forty kilometres from the summit. That is Chapter 5.

Tephra fall. Ash and pumice raining out of a column. Rarely lethal by itself in the open, but wet ash is heavy, and the commonest way it kills is roof collapse onto the people sheltering underneath.

Sector collapse. Volcanoes are structurally poor mountains — piles of weak, hydrothermally altered rubble — and they fail. Mount St. Helens in 1980 removed its own north flank as a landslide, which is what unroofed the pressurized body of magma inside and produced the lateral blast. The collapse was not a consequence of the eruption. It was the trigger for it.

Gas. Volcanic gas can kill without any eruption at all. In 1986, Lake Nyos in Cameroon released an enormous volume of carbon dioxide that had accumulated in its deep water from a volcanic source below. Carbon dioxide is denser than air. It flowed downhill into inhabited valleys and asphyxiated well over a thousand people, and a great many animals, in a single night. Nothing erupted. Nothing burned. The mechanism was a lake exhaling.

Tsunami. Flank collapse into water, or a caldera-forming eruption at sea, displaces ocean. A large share of the deaths attributed to Krakatoa in 1883 were drownings, and the same mechanism operated at Anak Krakatau in 2018 and at Hunga Tonga in 2022.

And then the indirect toll — crop failure, famine, the reordering of a climate — which is large enough to need its own chapter and gets one in Part IV.

The thing to be afraid of

I have deferred the pyroclastic density current until now because it deserves undivided attention.

Take the mixture that exists above the fragmentation surface: hot gas, ash, pumice, rock fragments. Send it up in a column. That column is buoyant only if it entrains and heats enough surrounding air to become lighter than the atmosphere. If it does not — because the eruption rate is too high, or the vent has widened, or the mixture is too dense — the column collapses under its own weight and comes back down.

What lands is a suspension of hot particles in gas that is, taken as a whole, denser than air. Denser-than-air fluids flow downhill. So it flows downhill.

That is a pyroclastic density current, and its properties are worth stating flatly. Temperatures range from a couple of hundred degrees Celsius to well over six hundred. Speeds are commonly comparable to a fast car and can be considerably higher. It has structure: a dense basal flow that hugs the ground and follows valleys, and above that a dilute, turbulent, expanded ash cloud — a surge — which is much less constrained by topography. The surge is what kills people who believed they were on the safe side of a ridge, because the surge can detach from the flow beneath it, climb over obstacles, and travel across water.

Its dynamic pressure knocks down masonry buildings. Its temperature is instantly lethal to anything breathing. Death is thermal and it is essentially immediate; the bodies at Herculaneum record a population that was killed by heat before anything else could happen to them.

I want to state the practical consequence as plainly as I can, because it governs everything in Part II of this book.

There is no sheltering from a pyroclastic density current. There is no structure to get behind, no protocol, no equipment. The only variable that matters is whether you are inside the runout or outside it, and the only intervention available is to not be there.

Which means the entire discipline of volcanic risk management reduces, in the end, to a single problem: getting people out of an area, in advance, on the basis of a forecast that is necessarily probabilistic, before anything has visibly happened. Every question about instruments and models and alert levels is downstream of that one.

And it means the people whose job is to characterize these flows have to get close to them.

On the third of June, 1991, at Unzen, a dome collapsed and generated a pyroclastic flow that went further than the people watching it had expected. Forty-three of them died, including two French volcanologists who had spent their working lives filming exactly this phenomenon, and an American who had survived St. Helens by having swapped shifts.

That is the next chapter.

Why the timing may not be knowable

Let me close with the limit, because it is a real one and I would rather set it up honestly now than have it arrive as an excuse in Chapter 5.

We can say a great deal about how eruptions work. We can identify the mechanisms, run them in laboratories, model them numerically, and read them backwards out of deposits with considerable confidence. What we cannot do is tell you when.

The naive expectation is that this is a resourcing problem — more instruments, denser networks, better models, and the forecasts will sharpen. Some of that is true and monitoring has improved enormously. But there are reasons to think part of the difficulty is structural rather than instrumental, and they are worth setting out.

The first is that precursors indicate unrest, not commitment. Earthquakes, swelling, gas emission: all of these tell you magma is moving. They do not tell you it will arrive. Intrusions that stall in the crust and never reach the surface are common — arguably the normal outcome. Many episodes of vigorous unrest end in nothing at all. A signal that is followed by an eruption a minority of the time is a genuinely difficult thing to act on, and the people who have to act on it are the subject of Chapter 5.

The second is non-uniqueness, which we met in the last chapter and will meet again. The same surface signal can be produced by different subsurface configurations. You are inferring a system you cannot see from its shadow, and more than one shape casts that shadow.

The third is threshold behaviour. A system sitting near a failure threshold, where the final increment that tips it can be arbitrarily small, is not one whose exact timing can be forecast from bulk properties — because the deciding factor may be the orientation of a single fracture, or the strength of one section of roof, or whether a dike happens to intersect a pre-existing weakness. Those are not quantities anyone is going to measure. The observation from the trigger section — that some volcanoes appear responsive to rainfall — is the diagnostic. A system that can be tipped by weather is a system whose precise timing is not encoded in anything we can watch.

I want to be careful here, because this argument is easy to overstate, and overstated versions of it are used to excuse things that should not be excused. Chaos in a system does not mean nothing can be said. Probabilities can be estimated. Windows can be bounded. Populations can be moved. Pinatubo was forecast well enough to save somewhere between five and twenty thousand people, and that forecast was made with 1991 instruments by people who had been on the mountain for two months. The limit is on precision, not on usefulness.

But it does mean that the honest product of volcanology is a probability distribution and never a date — and that a discipline whose central object cannot be instrumented, whose signals are ambiguous, and whose systems may be genuinely threshold-driven is going to have to be unusually disciplined about what it claims.

How it manages that, and what it has cost, is the subject of Part II.

Draft notes — verification status

Standing convention. The mechanics in this chapter — solubility's pressure dependence, exsolution, fragmentation, the silica–viscosity relationship, permeability, PDC structure — are settled textbook physics and are safe in substance. As with Chapter 1, the exposure is in numbers, named events, and casualty figures.

Pending verification:

• Pinatubo 1991 as "second largest eruption of the twentieth century" (usually ranked behind Novarupta/Katmai 1912) and the ~0.5 °C global cooling over roughly two years.
• Kīlauea's Puʻu ʻŌʻō eruption as continuous from 1983 — confirm the framing is accurate for June 1991 specifically.
• Volume expansion factor on exsolution at shallow depth ("hundreds or thousands"). Deliberately given as a range; source before tightening.
• Typical dissolved water content by mass in arc magmas (commonly cited as ~4–6 wt%); not quoted in the draft — add only if checked.
• Fragmentation threshold as a gas volume fraction (~70–80%) — omitted from the text deliberately. Verify before including.
• Viscosity range across natural magmas ("ten to twelve orders of magnitude"), and representative values for basalt and rhyolite.
• Crystal content threshold at which effective viscosity climbs steeply (~30–40 vol%).
• Diffusion chronometry timescales for pre-eruptive recharge ("weeks, sometimes days") — check against the published range, which varies widely by system.
• Deglaciation and Icelandic eruption rates — the commonly cited order-of-magnitude increase after the last deglaciation.
• Statistical association between large tectonic earthquakes and subsequent nearby eruptions — this is genuinely contested; confirm the draft's hedging is strong enough.
• Nyiragongo lava flow velocities and the silica/alkali composition (foidite) responsible; Goma population.
• PDC temperature range (200–600+ °C) and velocity range.
• Herculaneum thermal-death evidence — this literature has been revised more than once; check the current position before asserting it.
• Lake Nyos 1986: death toll (~1,700–1,800) and the limnic overturn mechanism.
• Krakatoa 1883: proportion of deaths attributable to tsunami rather than to the eruption directly.
• Mount St. Helens 1980: the sequence in which the sector collapse unroofed the cryptodome and produced the lateral blast.
• Unzen, 3 June 1991: 43 fatalities; the Kraffts and Harry Glicken. Already verified in the thesis paper — carry the confirmed figure across rather than re-checking.
• Pinatubo lives-saved range (5,000–20,000). Already verified. Keep consistent with Chapter 5 and the thesis paper.
• Armero: referenced here obliquely as "more than forty kilometres from the summit" — verify the distance, and reconcile the death toll across all documents (still outstanding from the Introduction's notes).

Epistemology, and its cost