Chapter 3: The Instrument You Cannot Insert

We have landed instruments on the surface of Venus, where the temperature is high enough to melt lead and the atmospheric pressure is that of a kilometre of ocean, and they sent back photographs before they died. We have flown a spacecraft through the plume of a moon of Saturn and sampled it in flight. We have put a probe into the outer atmosphere of the Sun.

Nobody has ever measured the inside of a magma body four kilometres beneath a mountain.

I want that sentence to sit for a moment, because the whole of this part of the book is a consequence of it. The least accessible place in the solar system, in practical terms, is not far away. It is directly underneath a great many people, and we have never been inside one, and the reason is not funding or ambition or engineering conservatism. The reason is structural: the object of study destroys the instrument.

That constraint shapes everything the discipline believes, everything it can claim, and — as this chapter will show — a good deal of what it has cost.

What "cannot" means here

Let me be precise about the difficulty, because "we can't measure it" is the kind of statement that invites an engineer to say well, have you tried.

Magma is somewhere between about seven hundred and about twelve hundred degrees Celsius, depending on composition. That alone is not disqualifying; we have materials that survive those temperatures. It is corrosive, being a hot silicate liquid with dissolved halogens and sulfur, which is worse. It sits at pressures of hundreds of atmospheres. And it is not a static target — it moves, it convects, it is embedded in rock that deforms.

But the genuinely hard part is not surviving the magma. It is reaching it.

To place an instrument inside a magma body you have to drill to it, through kilometres of hot, fractured, hydrothermally altered rock, having first determined where it is to within the accuracy of a borehole — and, as this chapter will explain, our knowledge of where these bodies are and what shape they take is exactly the thing we are trying to improve. You would be drilling toward a target whose position is one of the unknowns.

And then there is the objection that stops most conversations: drilling into a pressurized, volatile-bearing magma body is not obviously a safe thing to do. The intuition that you might trigger something is a reasonable one. It happens to be wrong, and we know it is wrong for an unexpected reason, which I will come to shortly.

So the discipline is left in a position almost no other physical science occupies. Astronomers cannot touch a star, but they have photons arriving continuously from the object itself, carrying composition, temperature, velocity and magnetic field. Volcanologists have the surface above the object, the tremors it makes, the gas it exhales, and the material it eventually throws out. Everything else is inference.

Unzen, 3 June 1991

Mount Unzen is on the island of Kyushu, in southwestern Japan, above the town of Shimabara. In late 1990 it began erupting for the first time in roughly two centuries, and by the spring of 1991 it was doing the thing described at the end of the last chapter: extruding a lava dome of stiff, degassed, silica-rich magma at its summit.

Dome-building volcanoes are watched closely, because dome-building volcanoes collapse. Through May the dome grew and shed material, and pyroclastic flows ran down the eastern flank with increasing frequency and increasing reach. An evacuation zone had been established. Residents had left. And the flows had become, by the standards of such things, routine — which is a dangerous word to have earned.

Among the people on the mountain were Katia and Maurice Krafft, a French couple who had spent roughly a quarter of a century photographing eruptions, and who were, by a wide margin, the most experienced volcano cinematographers in the world. They were at Unzen specifically to film pyroclastic flows, and the reason they wanted that footage matters to the rest of this chapter: they had been building a body of material for hazard education, on the argument that people who have never seen what a pyroclastic flow does cannot be expected to take a warning about one seriously.

With them was Harry Glicken, an American volcanologist. Eleven years earlier Glicken had been the observer stationed at a monitoring post north of Mount St. Helens. He left the post for a meeting with his graduate adviser and another scientist, David Johnston, took his shift. On the morning of the eighteenth of May 1980, Johnston was at that post when the north flank of the mountain collapsed. His last transmission was that it was happening. His body was never found.

Glicken went on to work on debris avalanches — on precisely the phenomenon that had killed the man who took his place.

On the afternoon of the third of June, 1991, the dome at Unzen collapsed and produced a pyroclastic flow considerably larger than the ones that had preceded it. It travelled further than expected, and the dilute ash-cloud surge above the dense basal flow — the component described in the last chapter, the one that detaches and is not constrained by valley walls — spread laterally across ground that people had judged to be outside the runout.

Forty-three people were killed. Among them were the Kraffts and Glicken. Among them were also a number of journalists who had entered the restricted zone to photograph the eruption, the taxi drivers who had brought them and were waiting, and firefighters who had gone in to bring people out.

I want to resist the shape this story wants to take. It is not a martyrdom, and the people who died were not all engaged in science; several were engaged in journalism, and several more in trying to remove the journalists. What it is, precisely, is a misjudged runout distance — a wrong estimate of how far a flow would reach, made by some of the most experienced people alive, about a volcano that had been producing smaller versions of the same flow for weeks.

And a misjudged runout distance is not a lapse of nerve or of care. It is the exact failure mode this chapter is about. They could not see the dome's interior, or the pressure behind it, or the volume that was about to fail. Nobody could. They were doing what everyone in the discipline does, which is inferring an inaccessible system from its surface behaviour, and the inference was wrong by a few hundred metres in the only direction that mattered.

Twelve days

On the fifteenth of June, 1991, two and a half thousand kilometres away, Mount Pinatubo erupted.

Pinatubo had not erupted in something like five centuries. It was not, in early 1991, on anybody's list of urgent problems; it was heavily vegetated, unmonitored, and locally regarded as a mountain rather than a volcano. When the unrest began in April, a joint team from the Philippine Institute of Volcanology and Seismology and the United States Geological Survey had roughly two months to establish a monitoring network, work out what the volcano was doing, build an alert system from nothing, and persuade several hundred thousand people — including the population of a large American air base — to leave.

The persuasion was the hard part, and it is the part that concerns this chapter.

Consider the problem honestly. You are asking a farming population, and a military command, to abandon homes and installations on the basis of a probability, about a mountain that has done nothing in living memory or in anybody's grandparents' memory, and the thing you are warning them about is a phenomenon they have never seen and for which ordinary language has no useful referent. "Pyroclastic flow" means nothing. "Hot ash" sounds survivable. Every intuition a reasonable person has — shelter indoors, get behind something solid, wait it out — is not merely useless here but actively fatal.

What the team had was the Kraffts' film.

The footage the couple had spent their careers gathering had been assembled into a hazard education video for the International Association of Volcanology and Chemistry of the Earth's Interior. Copies were left at nearly every briefing during the Pinatubo emergency. It was shown to civil defence officials and to military commanders. It was carried out to communities as part of the information campaign.

It worked because it did the one thing no chart and no probability could do: it made the abstraction visible. You cannot argue someone out of a comfortable intuition with a number. You can, sometimes, dislodge it by showing them footage of the thing.

The evacuations happened. The USGS and PHIVOLCS estimate that the forecasts saved somewhere between five thousand and twenty thousand lives. Roughly four hundred people died — most of them, as it happens, from roofs collapsing under ash made heavy by a typhoon that arrived at the same time as the eruption, which is the sort of coincidence that no forecasting system is designed to absorb.

What the chain actually shows

I have told these two events together because the connection between them is the closest thing this book has to a thesis about knowledge.

The Kraffts spent twenty-five years getting close to pyroclastic flows in order to record them, because the phenomenon cannot be understood at a distance and cannot be conveyed without images. That proximity killed them. Twelve days later, the record they had produced by being that close was used to empty a mountain, and somewhere between five and twenty thousand people who would otherwise have stayed put went somewhere else.

I do not want to make this tidy, and I especially do not want to make it redemptive in the way that eulogies do. Their deaths did not purchase the Pinatubo evacuation; the film already existed, and they would have been just as useful alive. There is no ledger here in which the trade comes out even.

What the sequence does show is the structure of the problem. The information that makes a warning credible can only be obtained near the thing that makes warnings necessary. That is not a moral claim about courage. It is a statement about where the data is, and it applies to the whole discipline: to the gas chemist standing in a plume, to the field team installing a seismometer on an active flank, to anyone who has to go and look because looking is the only method available.

Everything else in this part of the book is about the strategies volcanology has developed to reduce how often anyone has to do that.

The three times we hit it by accident

Which brings us to the exception, and to one of the more surprising empirical results in recent volcanology.

On at least three occasions, drilling operations have struck magma without intending to. In 2005, a geothermal well on the lower east rift zone of Kīlauea encountered dacitic melt at a depth of around two and a half kilometres. In 2009, a well drilled at Krafla in northern Iceland — aiming for supercritical fluids at four and a half kilometres — hit rhyolitic magma at roughly two. In the same year, drilling at Menengai in the Kenyan rift encountered magma as well.

None of these were planned. In each case the drill was going somewhere else and found melt on the way.

Three things came out of them, and all three matter.

The first is scientific. The magma quenched against the drill string and the borehole, and quenched magma is glass, and glass is a sample. Material came up. For the first time, anyone had a piece of magma whose depth, temperature and pressure of origin were known rather than inferred, because it had been collected from a hole of measured depth rather than reconstructed backwards from an eruption. The Krafla encounter in particular produced rhyolite glass from a body nobody had known was there — which is worth sitting with, since Krafla was among the most intensively studied and instrumented volcanic systems in the world. The geophysics had not seen it.

The second is engineering. The Krafla well was not abandoned. It was completed and produced superheated steam at temperatures and enthalpies far above any conventional geothermal well, which is the entire basis for the supercritical geothermal programme discussed in Chapter 14. A drilling accident became an energy result.

The third is the one that reframes the whole problem. Nothing happened. No eruption was triggered. No explosion, no runaway, no propagating fracture. In each case the magma quenched locally against the cold intrusion of the borehole and sealed it, and the volcano carried on as before.

That is a significant finding, because the fear of triggering something is the objection that has always sat in the way of deliberate magma drilling. Three accidental encounters do not constitute a proof of safety, and I do not want to present them as one. But they convert the question from "is this unthinkable" to "under what conditions is this manageable" — and that is the difference between a taboo and an engineering problem. It is why there is now a serious proposal to drill into the Krafla magma body on purpose and instrument it, which is the one intervention anywhere in this book that I think is both feasible and genuinely transformative.

We may yet get an instrument in. Not by defeating the constraint, but by accepting that the instrument will be destroyed and designing the experiment around losing it.

The instruments the volcano delivers

In the meantime, the discipline does something cleverer, and it is the part of this chapter I most want you to take away.

If you cannot send an instrument down, you can use the ones that are sent up.

Magma crystallizes as it cools and ascends. Crystals grow layer by layer, and the composition of each layer is set by the conditions in the melt at the moment it formed. When conditions change — a hotter batch arrives, pressure drops, the water content shifts — the crystal records the change as a compositional band. A crystal of plagioclase or olivine from an erupted rock is, quite literally, a stratigraphic record of the reservoir it grew in, and it is legible under an electron microprobe.

That is a data logger. The volcano built it, ran it for years or centuries at depth in a place we cannot reach, and then delivered it to the surface at high speed and free of charge.

There is better. As a crystal grows it sometimes traps a small bead of the surrounding liquid inside itself, sealing it off — a melt inclusion, a few tens of microns across. Because it is enclosed in a rigid crystal, that bead is pressure-isolated from everything that happens afterward. When the magma ascends and degasses and loses its volatiles, the inclusion does not. It still holds the water and carbon dioxide that were dissolved in the melt at the moment of trapping. Measure the volatile content of the inclusion and you have measured the volatile content of the magma at depth — and because volatile solubility depends on pressure, that gives you a depth.

A melt inclusion is a pressure gauge that was lowered into the reservoir, sealed, and returned to the surface. We did not have to insert it.

And there is a clock. Chemical gradients inside a crystal do not stay sharp; atoms diffuse across the boundary between one zone and the next, blurring it, at a rate that depends on temperature and is measurable in the laboratory. So the sharpness of the boundary encodes elapsed time. Find a crystal whose rim records the arrival of new hot magma, measure how far that boundary has diffused, and you can say how long the crystal sat in the reservoir after the recharge event before it was erupted. This technique — diffusion chronometry — is how we know that the interval between a recharge event and an eruption is sometimes a matter of weeks, and sometimes of days.

I find this genuinely wonderful, and I want to be clear that it is not a consolation prize. In several respects the crystal record is better than a downhole instrument would be, because it has been running continuously for the entire history of the reservoir rather than for the duration of a funded project. What it cannot do is tell you about the magma that has not erupted yet — which is, unfortunately, the magma anybody actually needs to know about.

What the surface tells you, and what it hides

For the magma that is still down there, we are left with the signals it produces at the surface. Each one sees something. Each one is blind to something, and the blindness is the part usually left out.

Seismicity. Moving magma fractures rock, and fracturing rock radiates. Volcano-tectonic earthquakes indicate brittle failure; long-period events and tremor are generally read as the resonance of fluid-filled cracks. Seismicity is the workhorse of volcano monitoring and it is superb at telling you that something is moving. It is blind to movement that does not break anything — magma ascending aseismically through pathways already open — and it cannot, by itself, distinguish magma from water.

Deformation. Magma accumulating at depth pushes the ground above it up, by centimetres, in patterns that can be measured with tiltmeters, continuous GPS, and satellite radar interferometry that can resolve millimetres of motion over a whole mountain. This is powerful and it is the closest thing to a direct measurement of volume change we have.

It is also the cleanest illustration of the problem at the heart of this chapter, so let me spell it out. The standard first-order model for interpreting inflation — published by Kiyoo Mogi in 1958, from measurements of how Sakurajima subsided after its 1914 eruption — treats the source as a pressurized point in a homogeneous elastic half-space. Fit it to your surface data and it returns a depth and a volume change.

It also has degeneracies, and they are worth naming exactly rather than gesturing at, because a loose version of this argument is easy to overstate and I would rather not. Pressure and source volume cannot be separated at all — only their product is recoverable, which is why these results are always quoted as a volume change and never as a pressure. The width of the uplift bowl does constrain depth, and with dense coverage it constrains it decently; but real networks are sparse and noisy, and the resulting errors on depth and volume change are large and move together. And the shape of the source is not determined either: a sphere, a sill and a dike can all produce broadly similar surface patterns.

So you are not measuring the depth of the magma body. You are measuring a surface bulge compatible with a family of magma bodies, and choosing among them requires information the deformation data does not contain.

Gravity. Repeated high-precision gravity surveys detect changes in subsurface mass. Combine this with deformation and you get something neither gives alone: if the ground inflates and the mass increases, magma has arrived; if the ground inflates and the mass does not, the volume change is gas. That distinction is close to the most operationally useful thing in the monitoring toolkit, and it exists only because two blind instruments are blind in different directions.

Gas. Sulfur dioxide flux tracks how much magma is degassing. More diagnostically, the ratio of carbon dioxide to sulfur dioxide is a depth proxy: carbon dioxide is far less soluble and exsolves deeper, so a rise in the ratio suggests new magma arriving from below. Gas monitoring is direct chemistry from the system itself, which is rare and precious. It is also weather-dependent, plume-dependent, and frequently unavailable exactly when the volcano is too dangerous to approach.

Electrical methods. Magnetotellurics images subsurface electrical conductivity, and silicate melt is conductive. It sees things seismology misses. It cannot easily tell melt from hot saline water, which is a serious ambiguity in a hydrothermal system.

Thermal. Satellite and ground-based infrared measure surface heat output. Immediate, global in coverage, and almost entirely uninformative about depth.

The reframe

Put all of that together and a particular character emerges, which I think is the honest description of what volcanology actually is.

There are two kinds of problem here. The forward problem asks: given a magma body of specified shape, depth and pressure, what signals appear at the surface? That is well-behaved physics with a unique answer, and computers do it well.

The inverse problem asks the question anyone actually cares about: given these surface signals, what is the magma body? And inverse problems of this kind are what mathematicians call ill-posed. The solution is not unique — many different subsurface configurations produce the same observations, as the deformation example shows exactly. And it is unstable: small errors in the data can produce large changes in the inferred model.

To get any answer at all you have to add assumptions from outside the data — smoothness, simple geometry, a plausible range of physical properties. This is called regularization, and it is unavoidable, and it is honest practice. But it has a consequence that every reader of this book should carry forward:

Every published picture of the inside of a volcano is a choice among possibilities that the data does not distinguish between. Constrained by observation, yes. Determined by it, no.

That is not a complaint about the field. It is what the field is. Volcanology is not the observation of magmatic systems. It is inference about magmatic systems from their shadows, conducted under a permanent instrumentation ban that nature imposed and we did not choose.

And it explains something that would otherwise be very strange indeed — how the discipline's central image of what a volcano contains could have been wrong for a century, sitting in every textbook, drawn in every diagram, without anyone being able to say it was wrong.

Nobody ever saw the magma chamber. It was a regularization choice.

That is the next chapter.

Draft notes — verification status

Standing convention. Several claims in this chapter were independently resolved during the verification pass for the accompanying thesis paper and should be carried across rather than re-checked.

Already verified — do not re-check:

Unzen, 3 June 1991: 43 fatalities; Katia and Maurice Krafft and Harry Glicken among the dead.
The Kraffts' footage compiled into the IAVCEI hazard film Understanding Volcanic Hazards; copies left at nearly all Pinatubo briefings; used with civil defence officials and military commanders and in the community information campaign.
USGS/PHIVOLCS estimate of 5,000–20,000 lives saved at Pinatubo; approximately 400 deaths.

Pending verification:

Glicken and Johnston at Mount St. Helens: that Glicken was the assigned observer at the Coldwater II post, that he left for a meeting with his adviser, and that Johnston substituted. Widely reported and almost certainly correct in outline, but check the detail before printing it — this is the kind of story that acquires tidiness in retelling.
Unzen's dormancy before 1990 ("roughly two centuries") and the 1990–91 eruption chronology.
The composition of the 43 dead at Unzen — the draft asserts journalists, taxi drivers and firefighters were among them. Verify the breakdown; the point being made depends on it.
Pinatubo's repose interval before 1991 ("something like five centuries").
Typhoon Yunya's coincidence with the 15 June climax, and the attribution of most deaths to ash-loaded roof collapse.
The three accidental magma encounters: Puna/lower east rift zone of Kīlauea (2005, dacite, ~2.5 km); Krafla IDDP-1 (2009, rhyolite, ~2.1 km against a 4.5 km target); Menengai (2009). Confirm dates, depths and compositions individually — the outline earlier said "Kīlauea Iki" and that appears to be wrong; the 2005 encounter was in the lower east rift zone.
The claim that no eruption or hazardous response was triggered in any of the three encounters.
IDDP-1's completion as a superheated steam producer and its enthalpy relative to conventional geothermal wells.
Whether Krafla's magma body had genuinely not been detected by prior geophysics — the draft leans on this and it should be checked, not assumed.
Melt inclusion sizes ("a few tens of microns") and the standard claim of volatile retention on ascent.
Diffusion chronometry timescales ("weeks, sometimes days") — same flag as Chapter 2; the published range is wide and system-dependent.
RESOLVED, AND CORRECTED. Mogi (1958), derived from Sakurajima's post-1914 subsidence. The earlier draft claimed a small shallow source and a much larger deep one produce identical uplift patterns — that overstates it. The real degeneracies: pressure and source volume are inseparable (only the product ΔV is recoverable, which is why results are always quoted as a volume change); depth and ΔV errors are large and correlated under sparse, noisy coverage; source geometry (sphere/sill/dike) is undetermined; and the homogeneous half-space assumption biases inferred depth. Corrected here and in the Appendix.
Venus lander survival times; Parker Solar Probe corona passage; Cassini plume transits — all used only as rhetorical comparisons in the opening, but they should still be right.