Bulk rock and matrix glass compositions of all the samples follow a classical arc calc-alkaline series ranging from a basaltic-andesite with 53–57 wt% SiO₂ and 3.5–5.0 wt% alkalis for the CCI to a trachyte with 64–66 wt% SiO₂ and 7.0–8.0 wt% alkalis for the GKI (see Figure 2A). In between, the Marapi PDC has a mafic andesitic bulk rock composition of 57 wt% SiO₂ and 5.0 wt% Na₂O+ K₂O, which is quite different from the trachy-andesitic to trachytic interstitial melt composition that spans from 59 to 65 wt% SiO₂ and 6.5–8.5 wt% Na₂O+ K₂O. The CCI composition falls in the same range as recent Llaima deposits that were produced by strombolian and effusive eruptions (Bouvet de Maisonneuve et al., 2012, 2013; Ruth et al., 2016; Schonwalder-Angel et al., 2018). This is not the case for the Marapi PDC that is relatively more mafic in composition than the two most recent eruptions of 2014–2017 for which the glass composition was rhyolitic (Nurfiani et al., 2021). The trachytic composition of the black PDC unit of the GKI is similar to but slightly more mafic than the lower (earlier) units of the same ignimbrite (see Figure 2A) and very different from the current activity of Batur that is mostly basaltic to basaltic-andesitic (Marinelli and Tazieff, 1968; Sutawidjaja, 1990, 2009; Reubi and Nicholls, 2004a, 2004b, 2005).

The crystal-poor CCI contains 0.4–4 vol% plagioclase, 0–1 vol% pyroxene, 0–1 vol% olivine, <0.5 vol% Fe-Ti oxide as phenocrysts with a total phenocryst content of 1–4 vol% on a vesicle-free basis (Table 1; Figure 3A). Plagioclase crystals are euhedral presenting some oscillatory zoning and span a wide range of compositions from An40—An85 (An–Anorthite content = Ca/(Ca+Na+K)*100 in mol%; Figure 2B). Some lower An-content plagioclases were analysed but they were probably from enclaves from the bedrock, as there is a high quantity of granitic lithics inside the deposit. Pyroxene crystals are euhedral and vary in composition with a majority of clinopyroxene (Augite-Diopside) and fewer orthopyroxene (Enstatite; Figure 2C). Fe-Ti oxides also span a wide range of compositions and follow two distinct trends in a TiO₂ vs. FeO_T diagram (FeO_T = Total iron as FeO; Figure 2D): One Fe and Ti-rich group that consists of magnetite to titanomagnetite phenocrysts, and one Cr-rich group that consists of a few chromites included in olivine. Olivine crystals are euhedral with some fast-growing textures at the rim (Figure 3A). They are normally zoned from core to rim and yield a Mg# of 75 to 50 (Mg# = Mg/(Mg+Fe)*100 in mol%; Figure 2E).

The crystal-rich Marapi PDC contains 15–23 vol% plagioclase, 10–20 vol% clino- and orthopyroxene in relatively similar amounts, and 1–3 vol% Fe-Ti oxide as phenocrysts for a total phenocryst content ranging from 32 to 40 vol% on a vesicle-free basis (—Figure 3). Plagioclase phenocrysts display a wide range of textures with the largest phenocrysts (0.5–1.5 mm) being euhedral to blocky with sieved high-An (An70–85) cores and lower An rims (An50–70) and smaller euhedral phenocrysts (0.25–0.5 mm) that are either An-rich (An > 70) or An poor with a sieved core (An <70). Cpx (Clinopyroxene) and Opx (Orthopyroxene) phenocrysts occur either side by side in clusters or as single phenocrysts in the matrix. Both Cpx and Opx are euhedral and have a narrow compositional range (Figure 2C and Figure 3). Fe-Ti oxides are also euhedral and are all low-Ti magnetite with a very narrow compositional range.

The crystal-poor GKI contains 4–8 vol% plagioclase, 0.5–1.5 vol% Cpx, 0.5–1 vol% olivine and <0.5 vol% titanomagnetite as phenocrysts for a total phenocryst content ranging from 6 to 10 vol% on a vesicle-free basis (Table 1; Figure 3). Plagioclase, Cpx and titanomagnetite are all euhedral and span a narrow compositional range with plagioclase crystals ranging from An30—An50, Cpx being mostly Ca-rich augite, and titanomagnetite containing 17–22 wt% TiO₂ and 69–72 wt% FeO_T. Olivine crystals are euhedral or present in clusters with Cpx and plagioclase, rarely zoned, and mostly Fe-rich with Mg# ranging from 27 to 62 (Figure 2E).

There are important textural differences between the groundmasses of the three deposits. The groundmass of the CCI is almost entirely crystallized with 80–90 vol% microlites on a vesicle-free basis, that are plagioclase (55–65 vol%), Cpx and olivine (20–30 vol%), and Fe-Ti oxides (0–2 vol%; Table 1 and Figures 3B,C). All microlites display fast-growth textures with clear plagioclase-pyroxene-magnetite intergrowth (Figure 3C) and the presence of occasional dendrites on the plagioclase tips. Crystal sizes vary from <1 µm for magnetite and 1–30 µm for pyroxenes to 10–200 µm for plagioclase. The most elongated plagioclases are often broken perpendicularly to the long axis with vesicles growing in the cracks, showing a “boudinage” structure (see Figure 4). Variations in crystal sizes are visible between samples indicating varying dynamics of crystal nucleation and growth, however, the total vesicle-free microlite content always stays in the 80–90 vol% range.

The groundmass of Marapi’s PDC is also almost completely crystallized with 50–80 vol% microlites on a vesicle-free basis that are vastly dominated by plagioclase (49–70 vol%) with the occurrence of some pyroxene (<<1 vol%—8 vol%; Table 1 and Figures 3E,F). There are also nanolites whose nature was not characterized in the remaining matrix (Figure 3F). All microlites present fast-growth textures with the presence of dendrites on most plagioclase crystals, plagioclase-pyroxene intergrowths, and radial-crystallization of plagioclase from a central nucleus (Figure 3F). Microlite sizes are relatively smaller than in the CCI with Fe-Ti oxides and pyroxene that are <<1–10 µm and plagioclase that is 1–20 µm in length.

The groundmass of the GKI is microlite-poor with 2–8 vol% crystals on a vesicle-free basis that are plagioclase (2–7 vol%), pyroxene (0.1–1 vol%) and Fe-Ti oxides (<0.5 vol%; Table 1). Plagioclase microlites are usually euhedral/blocky and very elongated (Figure 3H) with skeletal extremities (Figure 3I), while pyroxene and oxide microlites are mostly blocky. Plagioclase sizes range from 10 to 200 μm, while pyroxene and Fe-Ti oxide sizes range from 1 to 10 µm. Some of the most elongated plagioclase crystals are bent (Figure 4).

Pre-eruptive magma temperatures were assessed using the variety of analyses performed on plagioclase, pyroxene, and olivine coupled with average groundmass composition (EDS data), average interstitial melt composition or average melt inclusion compositions for each sample (Figure 9A). The variation between samples for the usage of either groundmass/interstitial melt or melt inclusion is function of the equilibrium between the pairs phenocrysts/liquid. We focused on using a maximum number of pairs at the equilibrium. For the basaltic-andesitic CCI, the plagioclase-liquid, the Cpx-liquid and the olivine-liquid geothermometers (Putirka et al., 2007; Putirka, 2008) yield very similar temperatures (within model uncertainties) of 1,068–1,096 ± 36°C, 1,045–1,100 ± 42°C, and 1,047–1,110 ± 29°C, respectively (Figure 9). These values largely overlap with most of the temperatures retrieved using Sc-Y partitioning between olivine and melt, with a comparable average temperature of 1,114 ± 15°C despite higher temperatures (up to 1,240 °C calculated for olivines with a very high Mg-core) having been recorded too. The few pairs of pyroxenes phenocrysts found to be in equilibrium give cooler temperatures of 946–1,013 ± 38°C. All these pre-eruptive magma temperature estimates are approximately 25–50°C cooler than the temperatures calculated for the most recent eruptions of Llaima (Figure 9A; Bouvet de Maisonneuve et al., 2012).

From Marapi’s mafic andesitic PDC, calculated pre-eruptive magma temperatures are about 100 °C cooler, ranging from 972 to 994 ± 36°C, 984–1,016 ± 42°C and 901–990 ± 38°C using the plagioclase-liquid, Cpx-liquid, and two-pyroxene geothermometers respectively. The two-pyroxene thermometer gives very similar crystallization temperatures to the ones found for the slightly more evolved 2014 and 2017 eruptions (Nurfiani et al., 2021), with differences being within uncertainties of the model.

For the trachytic GKI, the plagioclase-liquid, Cpx-liquid and olivine-liquid geothermometers give very similar pre-eruptive magma temperatures of 930–960 ± 36°C, 897–931 ± 42°C and 955–956 ± 29°C (N=4) respectively. Temperatures calculated from Sc-Y partitioning between olivine and melt show a bimodal distribution with a lower range of 905–930 ± 15°C, in agreement with the three other methods and given by the most Fe-rich olivine that are in equilibrium with the interstitial melt (Mg#27–40; Figure 2E). The higher temperature range of 1,023–1,079 ± 15°C is recorded by more magnesian olivines (Mg# > 40) that are likely xenocrysts or antecrysts given that they are not in equilibrium with the interstitial melt. Average pre-eruptive temperatures of 900–950°C are higher by 50°C than the temperatures calculated by Reubi and Nicholls (2004a) for the lower part of the sequence of the GKI. They are also 50 and 150°C lower than for the mafic andesitic Marapi PDC and the basaltic-andesitic CCI respectively.

Due to the inherently challenging nature of determining pre-eruptive water contents from phenocryst-poor, melt inclusion-free, and microlite-rich samples, the diversity of methods applied often yields large ranges and uncertainties, and poorly overlaps. The estimated pre-eruptive water contents are summarized in Figure 9B. For the CCI, we estimated 2.54–2.92 wt% H₂O in the melt with the plagioclase-liquid equilibria (Waters and Lange, 2015), 0.6–1.23 ± 0.4 wt% H₂O with F-Cl-OH partitioning between apatite and melt (Li and Costa, 2020), 1.0–4.1 wt% H₂O with water partitioning between Cpx and melt (Edmonds et al., 2016) and 1.4–7.3 wt% H₂O with water partitioning between Olivine and melt (Aubaud et al., 2004; Kohn and Grant, 2006). Results given by apatite need to be treated carefully as only three crystals could be measured due to their scarcity, and the apatite measured were small (<20 um on the long axis) and inside the matrix so probably underwent degassing. Water estimates from NAMs come with large error bars due to the big uncertainties on the partition coefficients of H₂O between melt and NAMs, especially for olivine where it varies threefold between studies (Aubaud et al., 2004; Kohn and Grant, 2006). Therefore, we will place more trust in the water contents estimated with the plagioclase-liquid hygrometer. These water contents are ∼0.5 wt% greater than for the most recent eruptions of Llaima (Bouvet de Maisonneuve et al., 2012).

For Marapi’s PDC deposit, we estimated 1.74–1.93 wt% H₂O with the plagioclase-liquid hygrometer, 0.73–2.03 wt% H₂O with apatite, 1.53–5.80 wt% H₂O with water partitioning between Opx and melt, and 1.03–2.83 wt% H₂O with water partitioning between Cpx and melt. For the same reasons as for the CCI samples, apatite in Marapi’s PDC are likely degassed and might not reflect accurately the magma’s pre-eruptive water content. Nevertheless, the other methods overlap quite well and suggest a pre-eruptive dissolved water content of ∼1.5–2.5 wt%, slightly lower than for the CCI.

For the GKI, we estimated 2.91–3.24 wt% H₂O with the plagioclase-liquid equilibria, 2.2–3.9 wt% H₂O with apatite measured at the SIMS, 1.6–6.3 wt% H₂O with apatite measured at the EPMA, 0.5–4.62 wt% H₂O with water partitioning between Cpx and melt, and 1.0–7.3 wt% H₂O with water partitioning between olivine and melt. Apatite measured by SIMS and EPMA were large and included in pyroxene or plagioclase phenocrysts and thus record the reservoir’s volatile content before eruption. The average water contents are 3.07 and 3.46 wt% for apatite measured at the SIMS and the EPMA respectively, which is consistent. These averages are also in good agreement with the plagioclase-liquid hygrometer (3.1 wt% in average) and the water contents derived from olivine (3.0 wt% in average), suggesting that the GKI ignimbrite was the most hydrous of the three units studied but only by slightly.

Melt CO₂ was also determined for the GKI ignimbrite, where direct CO₂ measurements could be done on apatite and melt-CO₂ content could be retrieved using the H₂O-CO₂ partition coefficient between apatite and melt (Riker et al., 2018). We find melt-CO₂ contents of 47–62 ppm.

Pre-eruptive storage pressures were estimated using the H₂O-CO₂ saturation model of Papale et al. (2006) for the CCI and GKI and two-pyroxene equilibria for the CCI and Marapi’s PDC (Figure 9C). For the CCI we obtain pressures of 0.5–0.8 kbar s, which is in the same range as pressures calculated for recent Llaima products using the same H₂O-CO₂ saturation model (Bouvet de Maisonneuve et al., 2012). However, as we have no independent measurement of CO₂ for the CCI, it may be because CO₂ values are the same between both studies. The two-pyroxene geobarometer gives pressures of 1.6–3.8 kbar, which is significantly deeper and potentially more likely given the greater magma volume involved in the ignimbrite than in recent strombolian eruptions.

For Marapi’s PDC we obtain the deepest storage pressure range of 1.3–5.5 kbars based on two-pyroxene equilibria. These magma storage pressures are overlapping but slightly lower and narrower in range than the ones obtained by (Nurfiani et al., 2021) for the 2014 and 2017 eruption products. For the GKI, H₂O-CO₂ saturation pressures range from 0.5 to 0.9 kbar s, similarly to the H₂O-CO₂ saturation pressures for the CCI.

Despite very similar appearances in the field and quite similar mineral assemblages, the three units studied have substantially different pre-eruptive magma properties and/or storage conditions. The phenocryst-poor CCI and GKI were fed by rather shallow magma reservoirs typical of caldera-forming systems (Cashman and Giordano, 2014; Edmonds et al., 2019—Figure 10). The phenocryst-rich Marapi PDC was fed by a somewhat deeper reservoir, although maybe not as deep as the one feeding current magmatic activity (Nurfiani et al., 2021—Figure 10). As expected, pre-eruptive storage temperatures decrease with increasing magma/groundmass silica content. However, substantial degrees of undercooling are recorded for the CCI magma, while lower degrees of undercooling are recorded for the Marapi PDC and for the GKI magmas (Figure 9A). Pre-eruptive water contents do not vary linearly with melt silica content nor temperature. Instead, the CCI and GKI magmas are quite hydrous (>2.5 wt% H2O), while the Marapi PDC magma is dryer (<2 wt% H2O) despite being stored at greater depths.

The CCI magma viscosity is low, as expected for a phenocryst-poor basaltic-andesite (Figure 11). In contrast, the mafic andesitic Marapi PDC magma viscosity is higher by 2–3 orders of magnitude given its higher phenocryst-content, more silicic groundmass, and lower water content. It is of the same order of magnitude as the phenocryst-poor, alkali- and water-rich GKI magma viscosity (Figure 11). These noticeably different initial conditions set in the reservoir will influence the timing and feedbacks between magma ascent rate, volatile exsolution, and microlite nucleation, ultimately determining whether and when fragmentation will occur.

Vesicle number densities (N_v) are similar for all three units and fall within the range of what has been observed for Plinian eruptions of each magma composition (Figure 8A). We find slightly higher numbers of N_v for the CCI than the one found by Valdivia et al. (2021) - (0.1–3.10¹³ m⁻³ - Figure 7A) that were computed solely from X-ray microCT using isolated vesicles. The CCI falls in the upper range of N_v for mafic Plinian eruptions (10¹²–10¹⁴ m⁻³) with the examples of Etna, Tarawera, Fontana Lapilli, and Izu-Oshima that produced Plinian fall deposits but no PDCs (Coltelli et al., 1998; Houghton et al., 2004; Toramaru, 2006; Sable et al., 2006,2009; Costantini et al., 2010). Similarly, for the Marapi and GKI products, we find typical VNDs (10¹³–10¹⁶ m⁻³) for more evolved Plinian eruptions (both fall and PDC deposits) such as Askja in 1875, Quilotoa or Novarupta (Alfano et al., 2011 and references therein).

Applying heterogeneous vesicle nucleation as the main process (Eq. 14 in Shea, 2017) and applying Toramaru (2006)’s model to our VNDs, we find pre-fragmentation decompression rates in the range of 0.4–1.6 MPa/s (Figure 8B). These estimated decompression rates fall in the middle of the spectrum for large explosive eruptions (Figure 8) and are consistent with decompression rates (0.1–10 MPa/s) calculated using numerical conduit modelling for other Plinian eruptions (Papale and Dobran, 1993; Kaminski and Jaupart, 1997; Papale et al., 1998; Neri et al., 2022). Such large decompression rates would prevent volatiles from escaping by bubbles decoupling from the melt or through the formation of a bubble network, which then further enhances magma ascent rate and strain rate, leading eventually to fragmentation.

Brittle fragmentation occurs when the melt relaxation time is longer than the time necessary for vesicles to expand (Dingwell 1996; Papale et al., 1999; Gonnermann and Manga 2007). Traces of these high strain rates have been observed in the GKI samples, with the presence of bent plagioclase microlites (Figure 4). Similarly, in the CCI samples, varying degrees of “boudinage” are visible on the most elongated and largest plagioclase microlites/microphenocrysts as well as varying dynamics of crystal and vesicle nucleation and growth due to intense shearing (Figure 4). Interestingly in the Marapi PDC samples, no obvious traces of high-strain rates are visible despite the very similar calculated melt viscosities (5.6 × 10⁴–1.1 × 10⁶ for Marapi vs. 2.1.10⁵–7.5.10⁵ for the GKI deposit; Figure 9). Vesicles kept their dominantly spherical shape without any obvious deformation due to shearing in the melt. It is also possible for the Marapi samples analysed in this study to have been issued from the centre of the conduit and as such to not show evidences of intense strain rate as has been suggested in other systems (Polacci et al., 2003; Rosi et al., 2004; Bouvet de Maisonneuve et al., 2009). However, in these other studies, there was a clear macroscopical difference between the types of clasts representing distinct regions of the conduit, which were not seen during sampling for Marapi.

The high pre-fragmentation decompression rates found for the magmas responsible for the three studied deposits are a consequence of very high vesicle number densities. High VNDs reflect the magma’s ability to keep volatiles until fragmentation, thus resulting in strain-rates due to bubble expansion during decompression that are too important for the melt structure to relax (Dingwell and Webb, 1989; Gardner et al., 2000). In the case of the Marapi and GKI magmas, the relatively high magma viscosity has the dual role of resisting fluid flow due to the high relaxation time of its melt structure but also retaining volatiles more efficiently by resisting bubble expansion and coalescence (Dingwell and Webb, 1989).

The CCI magma’s low melt viscosity should not have efficiently kept volatiles coupled with it and should have led to intense lava fountaining at high decompression rates due to the formation of a two-phase (gas + phenocryst-poor melt) flow (Gonnermann and Manga, 2012; Houghton et al., 2016) resulting in inertial-fragmentation (Namiki and Manga, 2008; Namiki et al., 2021). However, the very low initial phenocryst content (<4 vol%), the relatively high-water content for a mafic magma (2.5–3 wt%) and the low silica content of the magma before eruption helped to significantly delay microlite crystallization and growth, despite the melt’s sub-liquidus, supersaturated state (Arzilli et al., 2019; Figure 10). In addition, the high ascent rate promoted intense volatile exsolution, which drove the whole system’s undercooling degree even higher. The large amount of diffusion-controlled crystal textures such as dendrites, spherulites and plagioclase/pyroxene/oxides intergrowth show that very high degrees of undercooling were at play during the ascent of the CCI magma (Lofgren, 1974; Brugger and Hammer, 2010b; Shea and Hammer, 2013; Marshall et al., 2022). In addition, the square-like shapes of vesicles suggest that this crystallization occurred during ascent simultaneously with bubble growth. Therefore, despite part of the volatile budget probably escaping during ascent, the large undercooling of the magma triggered its intense and very rapid crystallization after reaching a critical point of volatile exsolution during magma decompression (Figure 10). The intense strain rate derived from the important decompression rate could have been a deciding factor in the crystallization onset (Tripoli et al., 2019). This mechanism resulted in a major viscosity jump of several orders of magnitude, as is the case for a magma going from 0 to 10 vol% to almost 80–90 vol% rigid particles within minutes (Costa et al., 2009; Faroughi and Huber, 2015; Moitra et al., 2018; Arzilli et al., 2019), resulting in a sudden resistance to fluid flow and causing the magma to be disrupted into particles (Figure 11). At magma ascent rates on the order of what would be expected for hawaiian eruptions, the crystallization of 80 vol% of the melt within minutes would result in magma fragmenting in the same brittle way as more evolved and more viscous magmas that generate Plinian eruptions (Figure 10).

In the Marapi PDC, the observed extensive microlite crystallization, with microlite shapes being essentially dominated by fast-growth textures (Figure 3), are hard to reconcile with the highly-viscous/silicic melts as diffusivities of species are impeded by viscosity (Mungall, 2002). In addition, the absence of any deformed vesicles or microlites does not support a simultaneous growth of bubbles with microlite crystallization. The low undercooling in the reservoir compared to the CCI is also not sufficient to explain such extensive and fast groundmass crystallization. Therefore, it is likely that the high-crystallization rate of the Marapi PDC matrix occurred post-emplacement, due to high-cooling rates. As a result, fragmentation mechanisms for this eruption would solely be the result of bubble overpressure fostered by high melt viscosity (Figure 10).

Overall, conduit processes played a major role with the intense microlite crystallization in the matrix for the CCI magma (80 vol% microlites on a vesicle-free basis) that changed the rheology of the melt, eventually leading to strain-induced fragmentation. On the other hand, the Marapi PDC and the GKI show classical signs of brittle fragmentation driven mostly by bubble overpressure and/or high-strain rates promoted by the melt viscosity that was already high in the reservoir.

In the case of large caldera-forming eruptions, PDCs form usually by pyroclastic-fountaining, which implies that the mixture of gas and pyroclasts formed after fragmentation is not buoyant enough to rise and form an umbrella plume but instead falls to the ground and flows radially around the vent (Druitt, 1998; Branney and Kokelaar, 2002; Sulpizio et al., 2014; Trolese et al., 2019). Both the GKI and the CCI deposits, as they are large volume ignimbrites, fall in this category (Figure 10). Field and textural observations of these deposits indicate the presence of a large quantity of lithics and enclaves that were entrained during the rise of the magma in the conduit and the emplacement of PDCs. As both these deposits are part of the ultimate sequence of their respective ignimbrites, the emplacement processes could be due to the widening of the conduit because of the caldera collapse, causing the eruptive column to become less-buoyant. Texturally, the fact that these deposits both fall on the lower end of observed vesicularities (45–65 vol%) for Plinian deposits could be an indication of their lower buoyancy and another reason for why they were emplaced as PDCs (Colombier et al., 2017).

The Marapi PDC is of clearly lower volume compared to the CCI and the GKI. However, the generally round aspect of the clasts inside the matrix and the general texture of the pyroclasts (high vesicle number density, highly spherical vesicles) suggest a process that is associated with column collapse rather than a dome or flank collapse (Branney and Kokelaar, 2002). It is worth stressing that the clast chemistry, texture, crystal content and volatile content within this PDC are extremely similar to the very recently emplaced deposits of the 2014 and 2017 phreatomagmatic eruptions (Nurfiani et al., 2021), except for the much more primitive matrix glass in the PDC deposit. Therefore, Marapi is likely capable of producing moderate volume PDCs such as the one investigated if/when a larger volume eruption occurs.