38 per cent of all eruptions in Iceland come from Grimsvötn. It is an amazing number: this hidden volcano, invisible and unreachable to all but the most hardy explorer, is among the most active volcanoes in the world. Not the most active: the volcano with the highest frequency of eruptions is Mayon, which has erupted 47 times (I think) in the last 500 years. Grimsvötn has ‘only’ had 64 known eruptions since 1200 AD, mainly identified through tephra layers around Vatnajokull. On average, it seems to erupt once every 13 years. But not all of its eruptions leave much tephra and more events may have been missed. The 2004 eruption left no tephra outside of the icecap and such an eruption would have been too small to have been counted in the 64! Even if only a third of its eruptions have been missed, it would erupt more often than Mayon. So perhaps Grimsvötn is the most frequently erupting volcano in the world? (Comments welcome!)
In fact this volcano is so dominant that the reported periodicity in Icelandic volcanoes (with a peak every 140 years or so) is almost entirely due to this one volcano.
So it is not a surprise that we are watching the Grimsvötn seismographs like a hawk chasing a mouse. (Or would that be the mouse watching the hawk?) And it is not only us. IMO has daily updates of number of earthquakes in the area. The total number during February was the highest since February 2011, and that was not long before its largest eruption in recent years. Could it be ..?
Carl and I have had a friendly tussle about this for several years. We tried to predict when its next eruption would be. Carl felt that it would be last year. Courageously, I decided to wait until after that before making a new prediction. I don’t mind Carl being right, but I don’t like me being wrong. Delaying allowed me to hedge my bets. But now the race is run, Carl owes me a beer (not Carlsberg, please), and I can safely take the risk of being wrong.
My approach has been to forecast the future (a risky endeavour: it is better to wait until the future has become the past and go for hind-casting) by looking at the seismic moment plots. These are being carefully maintained by IMO, and the upward trajectory implies the future is nigh. Dinosaurs of the world, take heed! The asteroid comes! To quote L’Internationale:
- La raison tonne en son cratère
C’est l’éruption de la fin
But it is not all plain sailing. Like a volcanic Dr Who, IMO has quietly been changing the past, and updated the plots, backdated. It is hard enough to predict the future without the past being changed. What is going on? Is this a momentous change? Or a micro tremor?
Changing the past
To see how the past has changed, let’s compare the old versus the new plots. Below they are shown together, one retrieved in September 2019 (top) and the other in January 2020 (bottom). The IMO-induced change happened at some date in between, and it caused the numbers to go up across the board, for all times. It is not just a scaling factor since the blue and green lines now cross over, which they did not uses to do. It is also not a complete overhaul since most features of the lines remain recognizable. So what happened? We could just ask them. But that is boring. Can we figure it out? (And while they are capable of changing the past, can we ask them to have a go at the UK Brexit referendum as well?)
It is easier to show the difference if we plot the old and the new together on the same plot. That is done below, for the post-2011 data. Black shows the old data, and red is the new. The left panel shows the raw numbers, obtained by digitizing the plot. In the plot, the numbers at the top axis indicate the day number, and the numbers at the bottom axis show the corresponding year.
The red (updated) line runs above the black (old) line at all dates. Did all the measured earthquakes become more powerful, back-dated? Perhaps. But let’s first see how the numbers changed, before asking the ‘why’ question. The right panel shows this: if we multiply the black (old) data by a factor larger than one, the two plots overlap very well. But the factor that we need to multiply it by changed with time. Before day 1800, the ratio is 1.5. Afterwards, it is 1.25. (These numbers are not very precise. I picked them because they seemed to work.) There are some small remaining differences but these may be related to how the plots were digitized, as that procedure is not perfect. And of course, my procedure of a simple scaling factor may not be fully accurate either.
But to come back to the question I did not want to ask: why? The easiest explanation would be that IMO is correcting for either incompleteness of the earthquake measurements, or for variable sensitivity of the seismograph network. The cumulative plot includes all earthquakes with a seismic moment magnitude larger than unity. (Note that the limit used in the daily earthquake count is different, and counts everything above 0.8.) This limit corresponds to a very weak earthquake, unnoticable by anyone without instruments. But (as an example), if the seismographs only detect earthquakes down to M=1.5, then some earthquakes will be missing. We can guess how many would be missing since there are known relations for how many weak earthquakes there typically are for every stronger one, and so we can correct for that. Another reason why earthquakes may be missed is if it was not possible to measure a location for the event. Again that is more likely to happen for weak earthquakes.
A second reason for why the numbers may have changed is if a different conversion is used between the measured strength and the seismic moment. Several different equations exist for this and IMO could have changed the definition they use. But this would not explain why the factor changed midway through the sequence. The change in scaling factor occured around day 1800. This is approximately late May or early June 2016. It is not a very accurate date, but it is plausible that at this time (the end of winter on the glacier) maintenance was being done on the seismographs.
How about the pre-2011 data? The same procedure works for those data as well. For the the build-up to the 2004 eruption (the green line on the first plot), the new and old line overlap well if I assume a scaling factor of 1.1 before day 1750, and a factor of 1.3 for quakes afterwards. Day 1750 corresponds to October 2003. For the red line, going towards the 2011 eruption (the big one) the adjustment is more drastic: the scaling factor is now 2.2 until day 1400 (October 2008), after which it goes down to 1.4. It may be a coincidence that in both cases the change happened in October, but it may also just be a good time for maintenance, the last time the instruments are easily reachable before the winter storms come. (It may also be a coincidence that 2.2 is close to 1.52..)
Predicting the impossible
So now we know why the past is not what it used to be. How about the future? Has that changed too? And can we turn the improved (hopefully) past into an improved prediction? Before doing this, we first need to do a reality check. If you want to learn to predict the future, it is advisable to start with predicting the past.
I am using an equation that describes cumulative failure, such as metal fatigue. The idea is that the increasing pressure in the magma reservoir causes microfractures, which grow and make each next fracture easier to make. Obviously this may be a poor model for volcanoes that keep all pressure inside until they suddenly blow up, a bit like some people I know. (Those are often also the people with the most interesting thoughts and ideas: bottling things up can give highly developed ejecta.) As far as I know, a similar equation has been used to model volcanoes using number of earthquakes, but it has not been done with cumulative seismic moment. It seemed an obvious thing to try. I did this first in 2017 and at the time it seemed to work. Those fits suggested an eruption in 2020 or 2021. So far, that is not wrong.
However, this approach is certainly not infallible. It can be thrown off by many things, for instance a larger earthquake (a fault failure). Grimsvötn is located along the main Iceland rift, and a rifting event would also overrule any events inside Grimsvötn.
The equation I used is shown below. In the equation m stands for the cumulative moment, A0 is some constant (a scaling factor), t is the elapsed time, and t0 is the time when the seismic moment m goes off the scale. This is the last possible time for failure! In practice, you would expect failure to happen a bit earlier.
We first try out the equation on the developments leading up to the 2004 and 2011 eruptions. The data, m versus t, is read off from the plots above, and I have fitted it with the above equation. The fit predicts the appropriate values for A0 and t0. At early times it is hard to calculate these separately: you can get very similar results if you increase A0 and decrease t0. (Mathematically, you measure the ratio between A0 and t0.) As the line curves upward, the values become better determined. But you don’t get fully accurate values until it is too late.
And here are the results for the Grimsvötn history. The equation can fit the overall shape of the curves very well. But there are deviations. Between 2002 and mid-2003, there were fewer quakes than would have been expected. The curve rapidly recovered the missing seismic moment in the second half of 2003, building towards the eruption in 2004. The left fit has constants of A0=1.45 and t0=2200 days. The run-up to the 2011 eruption can also be fitted well, but now there was an apparent excess of earthquakes during 2009, followed by a less active period before it rejoined the model curve mid-2010. Here the constants are A0=2.0 and t0=2600 days. Note that the build-up was slower than in 2004.
So predicting the past seems to be possible. At the very least, the model curve more or less worked, and Grimsvötn followed a similar curve and erupted at a similar point on the curve (corresponding to 4.8 on the axis) on both occasions. Using the new past, can we now predict the future?
The post-2011 curve shows a slow increase, becoming faster from 2015 onwards. The main problem with fitting this curve is the largish earthquake that happened during 2016, which increased the cumulative moment in one step from 1.7 to over 2. There is some doubt whether this single event should count towards Grimsvötn. Thomas has pointed out that it happened on a different fault system which happens to be just inside the box that IMO uses to identify Grimsvötn quakes. So perhaps we should take a cue from IMO, change the past, and delete this earthquake.
I tried two fits, one with this earthquake removed from consideration (shown on the left), and one with it included (shown on the right). With it included, I cannot get a decent fit. The curve is distinctly different in shape for what the model produces. So I am happy for it to be left out as a distraction. If the model doesn’t fit the data, the data must be wrong!
My best attempt is therefore the plot on the left. It has constants of A0=1.8 and t0=3900 days. If I assume that an eruption will happen when the seismic moment reaches the level ‘4.8’, as it did before, Grimsvötn will blow around day 3630, or 5 May 2021. The faster fit on the right, which includes the deleted quake, gives an eruption at day 3455, or 12 Dec 2020.
These precise dates dramatically overstate the accuracy of the fits. More reasonable is to take a three month uncertainty either way. Based on both models, I predict that the eruption will happen sometime in the period from September 2020 to August 2021. Within that period, Spring 2021 is the most likely.
There are quite a lot of assumptions here. It assumes that Grimsvötn will continue to behave as it has done before. It assumes that there won’t be a sudden event which breaks the fault, opens the floodgate and gives Carl ammunition to withhold my beer. It assumes that IMO will not change the past again. And it assumes that Grimsvötn doesn’t just change its mind – volcanoes don’t like being predictable!
Most importantly, it is still a bit too early for a unique fit. By giving less weight to the earlier parts of the curve, I can also make a slower fit, as shown below. This one pulls the trigger on day 3772, which is 24 September 2021. So we could even move the eruption to the autumn of 2021. Time will tell.
One question is left unanswered. There have been several periods where the cumulative seismic plot stopped increasing as fast as it should, after which it quickly recovered the lost amount. Why is this? Was Grimsvötn hibernating for a few years? It turns out that the cause is simple and interesting, but also hard to explain.
There have been two significant seismic episodes around Vatnajokull since 2011. At VC we know these well! And they correlate nicely with the Grimsvötn seismic interruptions.
The first of these was (of course) the Holuhraun eruption. Grimsvötn became very quiet from 2012 to mid 2014, corresponding to the run-up to Holuhraun. And ss soon as the eruption started, Grimsvötn began to recover from its seismic holiday. Apparently, the increasing pressure in Bardarbunga had calmed down Grimsvötn, and as soon as the break-out began, Grimsvötn recovered its stress.
The second recent event was the seismic crisis in Öræfajökul. While Öræfajökul was building up in 2018, Grimsvötn entered a quiet period. Grimsvötn recovered at the time that this crisis peaked; afterwards, while Grimsvötn’s actvity accelerated, Öræfajökul calmed down again. You win some, you lose some.
But how did these other volcanoes affect Grimsvötn? The first point to notice is that in both cases, Grimsvötn fully recovered the lost earthquake moment. This means that no magma was lost to these other volcanoes: the interaction was one of pressure or stress, not a direct connection causing magma leakage. When Bardarbunga was preparing to erupt, it reduced the pressure in Grimsvötn. The rising magma in Bardarbunga acted as a safety valve. But as soon as the eruption started, that safety valve shut close again. Now Bardarbunga is not very far from Grimsvötn, and a deep underground connection is not implausible. But Öræfajökull is a completely different system, and very likely completely unconnected. So how could Öræfajökull possibly affect Grimsvötn?
The most plausible suggestion seems to be that there is indeed no direct link. Instead the volcanoes interact through their effect on the spreading rift that runs through Vatnajokull. As magma accumulates in any volcano under Vatnajokull, it changes the stress on the rift. The rift relaxes, and this reduces the squeeze on the Grimsvötn magma reservoir. As the volcanoes settle down, the rift becomes stressed again and the squeeze on Grimsvötn’s inflating magma reservoir is re-instated.
This model explains why there is long-distance interaction between non-communicating volcanoes, and it explains why Grimsvötn fully recovers from its holidays. The precise mechanism remains a bit vague. And the dip around 2003 does not an easy cause: Bardarbunga was becoming active in this period, but so was Katla. Neither was as dominant as the ones last decade. However, Katla is really too far way, and Bardarbunga may be the only plausible cause.
The fit shown above predicts that the eruption is still a year away. This seems surprising given that the monthly number of earthquakes is already so high. Perhaps Grimsvötn will change its mind and go faster. But for now my prediction is 5 May 2021.
Although I predict it, I don’t believe this date myself! The uncertainties are still much too large. We really can’t say much more than ‘probably sometime in the next 18 months’. But to nail my colours to the mast, and to give Carl a target to shoot at (he is very good at shooting, apparently), I have put the date here anyway. It is even in the title of the post. If it is a fail and I need to deny the existence of this date, the only way for me to ever get around it will be to do an IMO and change the past. If the future does not perform to expectations, the answer always lies in the errors of the past.
Albert, March 2020
C’est la lutte finale
Groupons-nous, et demain
Le cratère de Grimvötn
Sera la fin du genre humain.
(A little over the top, I know, but the 19th century was like that.)