Studying the initiation of volcanic eruptions : Time for a petrological perspective

Volcano monitoring is essential for mitigating the risks associated with volcanic activity. As monitoring becomes more sophisticated and widespread, there is a growing need for understanding the relationship between different monitoring records and magmatic processes occurring at depth. This is particularly relevant to the processes that initiate eruptions after the extended periods of repose and inactivity that characterize many volcanic systems. Petrological studies of erupted materials are critical to this effort as they provide direct insight into the physical and chemical changes that occur in erupted magmas prior to their eruption, and allow investigation of past eruptions. Crucially, petrological approaches can also constrain the timing of processes involved in eruption initiation, and the time that might be expected to elapse between remote detection of increased activity and eventual eruption. Despite this, petrological studies of eruption initiation mechanisms are rarely systematically applied to monitored and other high-risk volcanoes. A literature compilation suggests there are significant differences in the composition, volume, style and timescales between eruptions initiated by different mechanisms. Knowledge of the processes that initiate eruptions at a given volcano may thus have significant predictive power.

initiation" and "eruption trigger" being used to describe a broad range of processes and outcomes involved with volcanic eruptions. These include deeper magmatic processes, as well as those that occur more shallowly within a conduit or edifice. Initiation or triggering are also variably applied to changes in eruption intensity or style (such as effusive to explosive transitions) that occur as part of an ongoing eruption, and also surficial events such as collapse events in an eruption column, volcanic edifice, or dome. For this study we consider a useful definition of eruption initiation to be: "the process or set of processes that result in a previously stable or quasi stable accumulation of magmatic material within the crust to commence moving upward and eventually erupt". By "stable or quasi stable", we mean magma bodies or crystal-rich mushes that have recently not been mobile, have not exceeded the critical overpressure or other parameters required to commence upward movement, nor shown previous indication of eruption, with seismicity and other monitoring signals (if known) at baseline levels.
We suggest that a more specific definition of eruption initiation, such as that proposed above, will help focus research on this critical subject, and that the term "eruption triggering" retain a more generic and contextual meaning. The May 18 1980 eruption of Mount St Helens provides an example of this usage. The trigger for the eruption was a magnitude 5.1 earthquake and landslide. However, eruption initiation occurred at several months earlier when magma started to ascend from a crustal reservoir to form a shallow cryptodome after many decades of quiescence.

Processes that initiate volcanic eruptions
Although systematic studies of eruption initiation are rare, there are many studies in the literature that report results and observations that bear on this important topic. We have compiled studies that primarily use petrological, geochemical and related techniques (in some cases this information was also combined with other data from seismicity, ground deformation, gas release and other sources) to infer the processes involved in the initiation of a range of different volcanic eruptions, and that also estimate the time elapsed between the initiation event(s) and eventual eruption. In some cases, these studies do not explicitly identify the eruption initiation mechanism but it is possible to do so from reported data and observations. Our compilation includes volcanoes from a variety of tectonic settings, although it is dominated (90%) by volcanoes from subduction zones. Overall the compilation contains more than 68 individual eruptions, representing 20 different volcanoes. Importantly, although volcanoes exhibit a wide variety of eruption types and styles, the compilation suggests a relatively limited number of eruption initiation mechanisms can be identified ( Fig. 1; Supplementary data), and the majority (> 90%) indicate that intrusion of new magma into an existing resident magma occurred immediately prior to eruption. Overall, we recognize four different eruption initiation mechanisms, three relating to intrusion of new magma and one to accumulation of volatile phases.
We also note that there are other mechanisms that have been suggested to initiate eruptions.
These include near field phenomena such as roof collapse above large magma chambers 34 and build-up of buoyancy forces 35,36 , as well as far field forcing related to large earthquakes 37,38 .
Although these mechanisms may be important in some settings, we do not consider them in detail here as they are less likely to leave distinctive petrological or geochemical signatures in erupted products, other than an absence of evidence for other initiation mechanisms. Future work may be able to identify methods whereby these mechanisms can be recognized from studies of erupted materials.

Mafic Recharge
Eruption initiation by addition of mafic magma to a more felsic magma reservoir is identified in 37% of the eruptions in our compilation, and is common in volcanoes in arc settings. Although the term "recharge" has a generic connotation of addition of magma, we recommend that the term mafic recharge refer exclusively to cases where significantly more mafic magma (typically basalt or basaltic andesite in composition) is added to a resident more silicic magma (andesite to rhyolite) 39 . The ramifications of this process are known relatively well from analogue and numerical experiments and field and petrological studies, and the petrologic record of this event includes the presence of reversely zoned crystals and different compositional and textural populations of the same mineral derived from distinct mafic and silicic magmas, disequilibrium mineral assemblages (e.g., quartz and olivine), multiple mafic and silicic liquid components in glasses or melt inclusions, and at the field scale hybridized magmas, enclaves, banded pumice, and similar phenomena 21,32, [39][40][41][42][43][44][45][46][47] . Intrusion of mafic material leads to a range of volatile exchange and saturation phenomena, increases in volume and/or internal pressure, and convective overturn. Phenocrysts in this scenario typically show evidence for large temperature contrasts (typically ≥ ~100°C) associated with rim growth 48,49 , often associated with extensive mineral dissolution or reaction rims 39 . The presence of microlites or microphenocrysts with mafic signatures within less mafic magmas also suggests magma mixing immediately prior to eruption 32, 40,45,50. Timing constraints for eruptions initiated via mafic recharge typically derive from estimating the timing elapsed between growth of reversely zoned crystal rims and eruption. Such "step function" zoning geometries in major and trace element abundances are well suited to diffusion modelling 32,49,[51][52][53] . Although an outstanding question is how much time elapses during dissolution before new rim growth occurs, this is probably not significantly longer than the time taken to grow the rims as mineral dissolution rates are typically faster than growth rates.

Rejuvenation
The majority (55%) of eruptions in our compilation result from intrusion of magma of broadly similar composition to the resident magma. We term this rejuvenation, and given that the compositions of magmas associated with rejuvenation also vary, we further recognize both mafic rejuvenation (31% of the compilation) and silicic rejuvenation (24%). We also note that differences in composition between introduced and resident magmas vary on a continuum between mafic recharge and rejuvenation. A fourth potential mechanism -felsic rechargecould occur when felsic magma intrudes a mafic magma reservoir but appears rare 61 .
In contrast to mafic recharge, eruptions initiated via rejuvenation are more likely to result from increased overpressure related to magma addition, together with increased buoyancy forces, and/or decreasing viscosity through changes in temperature and crystallinity. Mineral zoning and mineral populations associated with both mafic rejuvenation and silicic rejuvenation show more subtle differences than in the case of mafic recharge, and the primary difference between the intruding and resident magma may be degree of crystallinity with only minimal temperature differences, as revealed by eruptions that are cryptically zoned in terms of modal crystal proportions 33, 44,62,63 , or contain glomerocrysts, strained crystals and other evidence for disaggregation of crystal-rich cumulates [64][65][66] . Crystals from these include phenocrysts with subtle reverse zoning in the outermost rims, often in phases with more limited compositional stability fields such as olivine, sanidine and quartz, indicating that although the replenishing magma was somewhat less evolved in terms of incompatible trace elements, it was broadly similar with respect to phase stability 33, 67-69 . Mafic rejuvenation is the dominant mechanism in large shield volcanoes in extensional and arc settings, and felsic rejuvenation appears important for many silicic eruptions, including some of the largest known caldera eruptions 33,70 , as well as in arc settings. Eruption timescales for rejuvenation are typically estimated using diffusion in mineral rims. This includes high Ba and high Ti rims on sanidine and quartz for felsic rejuvenation, and high Mg/Fe rims on olivine and orthopyroxene crystals for mafic rejuvenation 31 . In some cases, growth rates of mineral rims can also be used 33, 70 .

Vapor saturation and exsolution
This mechanism has long been considered important [71][72][73][74][75] but is identified in only ~7% of the eruptions in our compilation, all of which occur in arc settings. Petrologic modelling suggests that volatile accumulation during progressive igneous evolution may be an important eruption initiation mechanism for large silicic magma bodies 72,73,75 . Vapor saturation and increased overpressure can occur related to decompression ("first boiling"), or more commonly when the magma attains vapor saturation during crystallization ("second boiling") 74 . In addition, upward movement of vapor exsolved deeper in a magmatic system or assimilation of hydrothermallyaltered wallrocks, could also produce increased vapor pressure and vapor saturation 72,73,75 . Vapor accumulation may be more challenging to definitively identify using petrological means, as most major and accessory phases record normal zoning and other changes corresponding to subtly decreasing temperatures and increased crystallinity. However, minerals that incorporate volatile species, such as amphibole, biotite, or apatite; fluid or melt inclusions hosted in a variety of phases; and/or mineral zoning in trace elements that preferentially partition into an exsolved vapor, can record progressive changes in vapor saturation and vapor composition during progressive crystallization [76][77][78][79][80] . Eruption initiation timescales for volatile accumulation in arc magmas have been estimated using diffusion of volatiles or elements with an affinity for the vapor phase or from mineral rims associated with vapor accumulation 79,80 , and from reequilibration (or lack thereof) of melt inclusions 79,80 .

Comparison between different eruption mechanisms
Our compilation allows us to compare some key eruption characteristics -the eruption style, erupted volume, erupted composition, and eruption initiation timescale -between eruptions initiated by the different mechanisms we identify above. We classify these characteristics using a categorical variables approach, either because they are already categorical (erupted rock type, eruption style), or because the large uncertainties warrant such an approach (erupted volume, eruption initiation timescale) (supplementary methods). Results, summarized in Figures 1-3, suggest there are systematic differences between eruptions initiated by the different eruption initiation mechanisms we identify. In some cases, these differences are obvious (for example, mafic rejuvenation and felsic rejuvenation are constrained to produce eruptions of mafic and felsic magmas), but for other parameters these systematic variations suggest there are consistent differences related to the processes of eruption initiation. From these data we hypothesize that there are general trends in increasing eruption volumes, longer initiation timescales, more silicic compositions, and more explosive eruption style going from mafic rejuvenation to mafic recharge, silicic rejuvenation and volatile accumulation, as summarized in Figure 3.
To make more rigorous comparisons we apply Fisher's exact test, a statistical significance test for categorical data 81,82 , an approach that minimizes the effect of large uncertainties in timescales and erupted volumes (supplementary methods). Using a simplified "dichotomized" categorical variable approach, we test both 4 x 2 contingency tables that compare the distribution of a given characteristic across all of the four different eruption initiation mechanisms, and 2 x 2 contingency tables to directly compare a single characteristic between pairs of eruption initiation mechanisms. In all cases we explore the null hypothesis that no differences exist in the distribution of eruption characteristics between different initiation mechanisms. Results are expressed in terms of P values (Supplemental Data Tables 4-8), which represents the probability of the null hypothesis being true given the observed distribution. We use a P value of ≤0.05 as a guide that the null hypothesis can be rejected, but caution against applying this rigidly 83 , and simply consider low P values as good candidates for exploration with more comprehensive data sets. Comparison of each characteristic across the four different initiation mechanisms using a 4 x 2 contingency table shows low P values for erupted composition, erupted volume, eruption style and initiation timescale, and thus the probability of the null hypothesis explaining the observed distribution for each of these parameters is considered low. The 2 x 2 contingency tables show P values < 0.05 for comparisons of mafic rejuvenation vs. felsic rejuvenation, mafic rejuvenation vs. mafic recharge and mafic rejuvenation vs. volatile accumulation for erupted volume, erupted composition, and eruption style. Mafic rejuvenation vs. felsic rejuvenation and mafic rejuvenation vs. mafic recharge also have P < 0.05 for the eruption initiation timescale.
Other comparisons also have relatively low P values (< 0.4): mafic rejuvenation vs. volatile accumulation for eruption initiation timescale; mafic recharge vs. felsic rejuvenation, and vs.
volatile accumulation for eruption style; and volatile accumulation vs. mafic recharge and vs.
felsic rejuvenation for erupted composition. This may suggest these comparisons are also worth further exploration -particularly those associated with volatile accumulation as the number of studies in the compilation is low (n = 5).

Discussion -The utility of petrological studies in understanding eruption initiation
Our results emphasize two important points. Firstly, although further refinements are certainly possible, petrological approaches are one of the best means we have to characterize a given volcanic system in terms of the process or processes that have initiated eruptions, but are also underutilized. Observations from volcano monitoring also provide important insight, but petrological studies of erupted materials allow for the identification of physical and chemical changes associated with eruption, and the associated timescales, and can be applied throughout the available eruptive record. Secondly, having some knowledge of the initiation processes that are likely to occur in a given system offers significant potential for insight into forecasting the initiation timescale and some other characteristics of an eventual eruption. Thus, systematic studies of eruption initiation mechanisms by petrological and other techniques, would seem to offer tremendous promise in some cases for assessing the nature of future eruptions and their associated hazards, and would add value to extant or planned monitoring programs.
To improve and expand on this we make four recommendations for priority areas of work over the next decade.
Increased emphasis on understanding eruption initiation. It is time for a renewed emphasis on the critical subject of eruption initiation mechanisms, including understanding the processes that take volcanic systems from quiescence to eruption and how these processes are recorded (or not) in erupted magmatic products. This research will be multidisciplinary, but there is much to gain from further systematic development and use of petrological and related approaches to identify and constrain initiation mechanisms. Our review reveals that many studies of volcanic systems report sufficient petrological, textural and other information to infer the initiation mechanism, but do not explicitly do so. Introducing consistent nomenclature and classification of eruption initiation mechanisms, as we recommend above, will also help. It may also be possible to further constrain the types of eruption initiation mechanisms beyond the simple categorization we present above.
More data is better than better. We need constraints on the initiation processes and associated timescales for more eruptions. Although we should also aim to improve the accuracy of timescale estimates based on diffusion chronometry and other methods, we argue that progress Greater integration with monitoring. Monitoring active and potentially active volcanoes saves lives and property, particularly when integrated with effective hazard mitigation planning.
However, there is a need for greater understanding of the relationship between the signals gained from volcanic monitoring methods, and the signals of underlying magmatic processes recorded in the rocks themselves. Petrologic data from modern eruptions is key here, as they allow direct comparison between monitoring signals and the physical and chemical changes recorded by magma in a magma reservoir undergoing initiation. These relationships are actively being explored, aided by increasingly common application of diffusion chronometry 84-86 but much progress remains to be made. To illustrate the importance of this approach, we show a compilation of the timescales of unrest based on various monitoring signals for eruptions with corresponding petrologic initiation timescales from diffusion chronometry or similar approaches in Figure 4. Although we are limited by available data, the results suggest there is not a uniformly simple 1:1 relationship. Mafic rejuvenation is the most frequently identified eruption initiation mechanism in studies where initiation timescales and unrest timescales are both documented, and also appears the most likely to show agreement in the general magnitude of these timescales. However, the limited data suggests the same may not be true for other eruption initiation mechanisms. Improved understanding of the relationship between these two signals will also improve the ability to assess and forecast volcanic hazards.
Increased petrological monitoring. Petrologic studies offer a cost-effective way to understand more about volcanic hazards in understudied volcanoes and to leverage existing monitoring resources. It is relatively common to map the compositions, type, and extent of eruptions through time at a given volcano to gauge hazards, but it is less common to combine this with systematic studies of eruption initiation mechanisms through petrologic studies. In relatively poorly monitored volcanoes this might be one way to understand the likely nature of future unrest episodes. Such observations could be used to augment the monitoring record, and provide a deep time historical context for recognizing the likely mechanism for future eruption initiation and onset of potential eruption episodes, as well as offering the specific potential for understanding the timing of new eruptions. In systems with more comprehensive monitoring programs, such data could also help refine traditionally problematic aspects of monitoring such as recognizing the causes of "failed eruptions" -episodes of instrumental and other unrest that do not result in eruptions. Petrologic research of this type is also a perfect forum for constructive cooperation between academic and government agencies. Importantly, the time to conduct such studies is in the early stages of a monitoring program and prior to the onset of a new eruptive episode.
In conclusion we reiterate that petrologic studies of eruption initiation mechanisms show great promise for improving our understanding of the magmatic processes associated with initiation of volcanic eruptions. A systematic effort at integrating petrologic insights into eruption initiation with monitoring signals offers an excellent opportunity to capitalize on this promise and to make rapid progress into the understanding of volcanic behavior.        As an initiation mechanism, mafic rejuvenation has the most data available for this comparison, as well the best agreement between the two timescales, suggesting the monitoring signals are more likely recording the same event(s) as the petrologic signals. The limited data for mafic recharge suggests the monitoring signals records events prior to the petrologic signals.
There is insufficient data on eruptions initiated by either felsic rejuvenation or volatile accumulation to make a similar assessment.  As an initiation mechanism, mafic rejuvenation has the most data available for this comparison, as well the best agreement between the two timescales, suggesting the monitoring signals are more likely recording the same event(s) as the petrologic signals. The limited data for mafic recharge suggests the monitoring signals records events prior to the petrologic signals.

Figures
There is insufficient data on eruptions initiated by either felsic rejuvenation or volatile accumulation to make a similar assessment.

Statistical Comparison of Eruption Initiation Mechanisms
The details of our compilation and sources are shown in Supplemental Data and initiation timescale as categorical, even where they are nominally continuous, as this minimizes the effects of the large uncertainties that are often apparent in these quantities. For the timescale, there was still some overlap between some categories, so each study was assigned the six points, and these were assigned to relevant categories (i.e., a timing estimate that ranged from weeks to years was given two points in each of the 'years', 'months' and 'weeks' categories).
Points were then summed for each category and expressed in percent of total.
To investigate whether there are significant differences in timescales, eruption type, and erupted volume between different initiation mechanisms we have investigated our categorical data using the two tailed Fisher Exact Test, a statistical significance test used for the analysis of categorical data in contingency tables (Fisher, 1922, Hall andRichardson, 2016). This test is valid over a range of sample sizes and is preferred over the 2 test where, as in our case, individual categories may be small (n < 5-10). To do this we reassigned our data for all parameters into two "dichotomized" categories, selected to minimize overlap between categories for individual studies (Supplemental Data and an alternate hypothesis (H1) states: Where π represents the proportion of eruptions initiated by mafic recharge and felsic rejuvenation that have volume < 1 km 3 . We then determine from our observed data if we have enough evidence to reject the null hypothesis at a reasonable level of significance.
We have implemented this approach using both 2 x 4 contingency tables to compare all identified eruption initiation mechanisms, and have also conducted further focused hypothesis testing between pairs of eruption mechanisms for a specific characteristic using 2 x 2 contingency tables (Supplemental Data Tables 4-7). Calculations for 2 x 2 and 2 x 4 contingency tables were done using the 'fishertest' routine and 'MyFisher24' function (Cardillo, 2020) respectively in MATLAB TM . Results are reported in terms of P values in Supplemental Data Tables 4-7, where P represents the probability of getting the observed distribution, assuming that the null hypothesis is correct. In accordance with recommended usage (Wasserstein and Lazar, 2016) we do not use P < 0.05 as a rigid criterion to reject the null hypothesis, but as a guide to suggest where important relationships may exist. Where comparisons show P values that are relatively low, but not less than 0.05, these may also be further tested with more data.

Parameters Categories
Volume < 1 km 3 1 -10 km 3 10 -100  Table 7. Summary of P values determined for 2 x 2 contingency tables for erupted composition (mafic vs. silicic). P represents the probability that the observed data supports accepting the null hypothesis that no difference in erupted composition exists between the pairs of eruption initiation mechanisms shown. Grey highlights comparisons where P < 0.05