Research Article

Magma reservoir failure and the onset of caldera collapse at Kīlauea Volcano in 2018

See allHide authors and affiliations

Science  06 Dec 2019:
Vol. 366, Issue 6470, eaaz1822
DOI: 10.1126/science.aaz1822

Caldera collapse and flank eruption

Real-time monitoring of volcanic eruptions involving caldera-forming events are rare (see the Perspective by Sigmundsson). Anderson et al. used several types of geophysical observations to track the caldera-forming collapse at the top of Kīlauea Volcano, Hawai'i, during the 2018 eruption. Gansecki et al. used near–real-time lava composition analysis to determine when magma shifted from highly viscous, slow-moving lava to low-viscosity, fast-moving lava. Patrick et al. used a range of geophysical tools to connect processes at the summit to lava rates coming out of far-away fissures. Together, the three studies improve caldera-collapse models and may help improve real-time hazard responses.

Science, this issue p. eaaz0147, p. eaay9070; p. eaaz1822; see also p. 1200

Structured Abstract


The 2018 rift zone eruption of Kīlauea Volcano, Hawai‘i, drained large volumes of magma from the volcano’s summit reservoir system, causing high-rate subsidence of the ground surface and withdrawal of an active lava lake. Over the span of 1 week, the surface of the lava lake fell more than 300 m. Continued withdrawal of magma caused the rock above the reservoir to fail, triggering the onset of episodic caldera collapse. Surface collapse began near the evacuated lava lake vent, but as the eruption continued over 3 months, the area of the new caldera expanded to ~5 km2 and its volume grew to 0.8 km3. The precursory activity and subsequent growth of the caldera were recorded in far greater detail than was possible at the handful of other caldera collapses observed in the past century. These comprehensive observations permit new insights into the conditions that lead to magma reservoir host rock failure and caldera collapse.


Volcanic caldera collapses can be highly destructive and create prominent topographic features, but little is known about the architecture of subcaldera magma storage zones or the critical decrease in pressure that triggers collapse. Withdrawal of Kīlauea’s lava lake in 2018 can be used to gauge pressure change in the underlying magma reservoir. We developed a model of time-evolving reservoir depressurization to jointly explain lava lake withdrawal rate and the rate and spatial pattern of ground subsidence obtained from radar satellites and a dense local monitoring network.


We tracked the evolution of the magmatic system from steady elastic decompression to inelastic failure. We were able to estimate the location, geometry, volume, and time-evolving pressure within the reservoir as well as conditions required to trigger failure of the overlying crust. Before the onset of collapse, the ground at Kīlauea’s summit was subsiding at nearly 10 cm/day, and the lava lake surface was retreating at ~50 m/day. We found that these phenomena were caused by drainage of magma at a high rate from a storage reservoir centered ~2 km below the surface, with a volume of several cubic kilometers. Drainage rapidly reduced reservoir pressure, stressing the surrounding crust. Two weeks after the rift zone intrusion and eruption began to drain magma from the summit, withdrawal of <4% of the stored magma had reduced pressure in the reservoir by ~17 MPa, causing the host rock above it to begin to fail episodically. The episodic collapses loaded the magma with the weight of the roof, increasing its pressure. The final collapse caldera was closely centered over the magma reservoir, and their horizontal dimensions were comparable. However, the estimated reservoir volume was substantially greater than the caldera volume, indicating incomplete evacuation at the end of the eruption.


Our results tightly constrain the pressure decrease in the magma reservoir before the onset of collapse. Together with geodetic data, this bounds the magma storage volume and the stress changes needed to cause failure of the host rock above the reservoir. Our results demonstrate that a magma reservoir’s roof may begin to fail after withdrawal of only a small fraction of the stored magma. At Kīlauea, this process was likely influenced by a relatively thin and wide reservoir roof and preexisting crustal weaknesses, including an established caldera ring-fault system and the lava lake vent. Roof collapses maintained magma pressure, sustaining the eruption, but they did not (as is sometimes assumed) completely repressurize the reservoir. This indicates residual frictional strength on the collapse-bounding faults. The eruption was not terminated by complete evacuation of stored magma, contrary to assumptions sometimes made when interpreting data from past caldera collapses, and indicates that a different process was responsible for the cessation of the eruption. Joint monitoring of ground deformation and lava lake elevation at other volcanoes, when possible, may yield rich insights into magmatic processes and conditions.

Caldera collapse at Kīlauea in 2018.

(A) Precollapse lava lake on 6 May 2018. The lake surface had fallen ~200 m since the onset of the eruption. (B) Aerial photograph looking west across Kīlauea’s summit on 12 June, after the onset of caldera collapse. Parts of the crater floor had subsided as much as ~180 m as intact blocks. (C) Estimated magma storage zone that partially collapsed to form the caldera. Shown is the isosurface enclosing the region that contained magma in our simulations, at 95% confidence. View is to the southeast.



Caldera-forming eruptions are among Earth’s most hazardous natural phenomena, yet the architecture of subcaldera magma reservoirs and the conditions that trigger collapse are poorly understood. Observations from the formation of a 0.8–cubic kilometer basaltic caldera at Kīlauea Volcano in 2018 included the draining of an active lava lake, which provided a window into pressure decrease in the reservoir. We show that failure began after <4% of magma was withdrawn from a shallow reservoir beneath the volcano’s summit, reducing its internal pressure by ~17 megapascals. Several cubic kilometers of magma were stored in the reservoir, and only a fraction was withdrawn before the end of the eruption. Thus, caldera formation may begin after withdrawal of only small amounts of magma and may end before source reservoirs are completely evacuated.

A volcanic caldera is a topographic depression formed by fault-bounded subsidence or collapse of Earth’s surface as magma is withdrawn from a crustal storage reservoir, causing the overlying rock to founder (1). Caldera formation can be triggered by magma withdrawal to feed violent explosive eruptions or by intrusion of magma into surrounding rock, sometimes feeding long-lived effusive lava flows. Calderas can be prominent topographic features measuring tens of kilometers in diameter.

Our understanding of volcanic caldera collapses has been strongly limited by a lack of well-documented caldera-forming eruptions. From 1900 to the beginning of 2018, only seven caldera collapses were clearly documented on Earth (2, 3), mostly with limited geophysical and observational networks. Even the well-recorded 2014–2015 collapse at Bárðarbunga, Iceland, occurred beneath hundreds of meters of ice, preventing direct observation (3).

The 825 million m3 caldera collapse at Kīlauea Volcano in 2018 was the largest at the volcano in more than two centuries and was tracked by a dense multiparametric monitoring network and through direct visual observations. These detailed datasets record the transition from steady elastic subsidence to fault-bounded collapse as the roof of Kīlauea’s summit reservoir failed in response to high-rate magma withdrawal to supply the volcano’s East Rift Zone (ERZ) intrusion and eruption. In this study, we modeled ground deformation and lava lake data to infer properties of the magma system as it evolved toward collapse in May 2018. The data offer direct evidence of pressure change in the magma reservoir and present an opportunity to resolve the volcano’s subcaldera magma storage architecture and its relation to collapse timing, style, and volume.

Kīlauea Volcano and the 2018 eruption

Kīlauea Volcano, on the island of Hawai‘i (Fig. 1), is one of the world’s most active volcanoes and erupted almost continuously from 1983 to 2018. For most of that period, Kīlauea’s mantle-derived magma supply largely passed through its summit reservoir system before migrating subhorizontally down the volcano’s ERZ to erupt as lava flows ~20 km from the summit at or near the Pu‘u ‘Ō‘ō vent.

Fig. 1 Kīlauea Volcano and the 2018 eruption.

Photos show a summit explosion on 9 May 2018, the lava lake as it appeared in April 2018, and the primary 2018 LERZ eruptive vent. (A) Shaded topographic map of the island of Hawai‘i; the box shows the extent of the map in (B). (B) During the 2018 eruption, magma flowed >40 km underground subhorizontally from the summit (left) to the LERZ vents (right). See Fig. 2 for an enlargement of the summit area. (C) Schematic cross section (not to scale) showing flow of magma from the summit to the LERZ.

Photos: U.S. Geological Survey

Beginning in 2008, a lava lake was active at the summit of the volcano within Halema‘uma‘u crater; by April 2018 its surface area had grown to more than 40,000 m2. The lava lake was supplied from a shallow magma storage zone (here termed the Halema‘uma‘u reservoir) hypothesized to exist 1 to 2 km beneath Kīlauea’s existing summit caldera (formed in ~1500 CE). Variations in the surface height of the lava lake were strongly correlated with ground deformation, indicating that both were caused by pressure changes in the underlying magma reservoir. Thus, Kīlauea’s lava lake acted as a magma reservoir pressure gauge (46).

Kīlauea’s 35-year-long eruption ended spectacularly on 30 April 2018 with the intrusion of a dike downrift from Pu‘u ‘Ō‘ō into the volcano’s lower ERZ (LERZ) (7) (Fig. 1B). On 3 May, the intrusion emerged in the Leilani Estates subdivision, more than 40 km from the volcano’s summit, ultimately erupting >1 km3 of lava and destroying hundreds of homes. The intrusion and eruption triggered wholesale draining of Kīlauea’s magma system, from the middle ERZ to the summit. Magma drainage from the summit led to lava lake withdrawal and vent collapse, a series of explosions, and ultimately the formation of a new caldera nested within the larger 1500 CE caldera. Summit collapse and most LERZ lava effusion ended in August 2018 after 3 months.

Magma evacuation and the onset of caldera collapse

We recorded subsidence and later collapse of the ground surface at Kīlauea’s summit by visual observations, continuous Global Navigation Satellite System (GPS) stations, borehole tiltmeters, and interferometric synthetic aperture radar (InSAR) interferograms derived from satellite data (8) (Figs. 2 to 4). Variation in lava lake surface height was recorded by laser rangefinder, thermal camera imagery, and structure-from-motion photogrammetry (Figs. 3 and 4) (9).

Fig. 2 Spatial pattern of subsidence at Kīlauea’s summit in 2018.

(A) Ground tilt overlaid on an ascending-mode COSMO-SkyMed interferogram spanning 6 to 10 May 2018 (table S1). Colored dots show observed tilt, and black arrows show best-fitting tilt velocities used for modeling. Each complete InSAR color fringe represents 1.55 cm of displacement in the look direction of the satellite (T symbol, 26.6° from vertical). Small-scale irregularities in the fringe pattern are evident in the caldera. Background shaded digital elevation model (DEM) shows Kīlauea’s summit in 2009, similar to its appearance in April 2018. (B) Observed GPS displacements (colored dots) and best-fitting velocities (black arrows) overlaid on the unwrapped interferogram from (A). An active lava lake was nested within Halema‘uma‘u crater, itself nested in the larger 1500 CE Kīlauea caldera. LoS, line of sight. (C) West-east profiles of LoS COMSO-SkyMed InSAR velocities approximately through the center of Halema‘uma‘u crater. Profiles differ because of different look angles. (D) View of GPS data in (B), looking north.

Fig. 3 Withdrawal of Kīlauea’s lava lake in early May.

(A) Thermal images of the lava lake surface taken from the south rim of Halema‘uma‘u crater while the lake was draining. (B) Time series of change in lava lake surface height relative to 26 April, and radially outward low-pass–filtered ground tilt at UWD. Time series after 5 May are shown in Fig. 4. Numbers correspond to acquisition times of images in (A). (C) Photograph showing the lava lake on 6 May and the laser rangefinder used to measure its surface height. (D) Relationship between lava lake surface height and radially outward tilt (with Mw 6.9 earthquake offset approximately removed). At all stations, the ratio decreased by ~40% around the time of the Mw 6.9 earthquake, denoted by the horizontal gray line. Correlation coefficients are denoted by r.

Photo: U.S. Geological Survey
Fig. 4 Temporal evolution of summit deflation.

(A) Radial ground tilt at UWD over the full eruption. Positive tilt is consistent with reservoir inflation (pressurization) and negative tilt with deflation. Collapses appear as small sawteeth from 16 to 26 May (nearly invisible at this scale) and as much larger sawteeth during broad-scale collapse (29 May and after). Time series were corrected for certain tectonic offsets. (B) GPS, tilt, lava lake surface height, and vent area time series indicating summit deflation from late April to early June 2018. Stations UWD (tilt) and UWEV (GPS) are approximately colocated (see Fig. 2 for station locations). Lava lake points with boxes were derived from structure-from-motion photogrammetry and are more uncertain. Vent area was inferred from satellite radar (ascending mode in green and descending mode in black) amplitude images as shown in (C); numbers on the time series correspond to these images. Time spans of modeled InSAR data are shown as horizontal bars and denoted with “-a” for ascending mode and “-d” for descending mode. The gray horizontal bar indicates the time span shown in Fig. 3. CSK, COSMO-SkyMed. (C) CSK radar amplitude images showing enlargement of the summit vent. Brighter pixels indicate higher radar reflectivity, so the vent appears black.

Before the onset of the LERZ intrusion, Kīlauea’s lava lake had been overflowing onto the floor of Halema‘uma‘u crater. Deflation began in earnest on 2 May with subsidence and contraction of the ground surface and withdrawal of the lava lake at a rate that reached ~40 m/day (Fig. 3 and fig. S8). On 4 May, an earthquake with moment magnitude (Mw) of 6.9 (M6.9) on the basal decollement between the volcanic pile and the oceanic crust underlying Kīlauea’s south flank (7, 10) shook the volcano and produced long-wavelength extensional strain across the summit. By the end of the day, lake withdrawal had accelerated to 53 m/day, and the ground tilt rate at summit instruments had more than doubled (8) (Fig. 4). Subsidence continued over the following days in a broad, roughly circular region centered near the east rim of Halema‘uma‘u at rates of up to nearly 10 cm/day (Fig. 2). Ground deformation and lava lake surface height were highly correlated (Fig. 3D). Before the Mw 6.9 earthquake, we observed ~5 m of lava lake withdrawal for every microradian of caldera-directed ground tilt at station UWE [located near the U.S. Geological Survey (USGS) Hawaiian Volcano Observatory (HVO); Fig. 2], in agreement with observations made over many years at Kīlauea (4, 5). After the earthquake, this ratio had decreased by ~40%.

Rapid withdrawal of the lava lake was accompanied by sporadic explosions as unsupported conduit wall rock fell into the vent (Fig. 1), gradually increasing its diameter (Fig. 4). By 10 May, after dropping more than 300 m in just over a week (supplementary movie S1), the lava lake had disappeared from view and the vent was blocked by rubble. Ground subsidence continued, however, indicating ongoing depressurization, and HVO became concerned about failure of the rock above the reservoir. From 9 to 15 May, several M ≥ 3 earthquakes per day shook the summit, and tremor [as indicated by the real-time seismic amplitude measurement (RSAM)] was recorded at very high levels. Ground cracks were observed near Halema‘uma‘u crater on 14 May, and by 16 May the GPS network had recorded total subsidence in that area of ~1 m.

On 16 May at 18:16 Hawaii Standard Time (HST), abrupt inflationary (radially outward) ground deformation and very-long-period (VLP) seismic energy (Mw 4.9) were recorded across the summit, an ashy gas plume rose to 20,000 ft, and summit RSAM dropped precipitously. Ground deformation and VLP observations were similar to those previously caused by rockfalls into the lava lake and ascribed to pressurization of the shallow magma system (11) but were much larger in amplitude. They were also similar to observations recorded during caldera collapses at Miyakejima (Japan) and Piton de la Fournaise (La Réunion) volcanoes (1215). Eleven more of these events, informally termed “collapse/explosions” by HVO, occurred before the end of the month. Satellite observations and failure of instruments on the crater rim indicated that the (now empty) lava lake vent was growing more rapidly (Fig. 4) and beginning to cause failure outside of Halema‘uma‘u, but broader-scale, fault-bounded surface collapse was not yet detected. Summit SO2 emission rates increased by two to three times (7), but erupted tephra volumes were much smaller than collapse volumes. Away from the widening vent, the summit continued to subside between collapses in a roughly circular pattern centered on the caldera.

The onset of broader-scale, clearly fault-bounded collapse outside of Halema‘uma‘u crater began in the early morning of 29 May with an abrupt down-dropping of the caldera floor around Halema‘uma‘u, approximately coincident with the onset of higher eruption rates (~150 m3/s) in the LERZ. We measured 1.5 m of subsidence at a GPS station (NPIT) on the northeast rim of the crater during the seconds-long event, and visual observations from HVO revealed subsidence north-northeast and west of Halema‘uma‘u. Away from the subsiding block(s), however, inflationary radially outward deformation and VLP seismicity were observed that were similar to previous events in May but with much larger amplitudes (Fig. 4).

On 1 June, enabled by a marked reduction of Kīlauea’s summit plume, an unoccupied aerial vehicle took the first clear photos of Halema‘uma‘u since mid-May. The photos showed major collapse and widening of the vent, ~30 m of subsidence of the western floor of Halema‘uma‘u, and faulting and subsidence of the 1500 CE caldera floor more than 1 km northwest of the former lava lake. As more collapses occurred through June, the surface expression and area of slumping expanded greatly. Collapse events were roughly periodic in time (Fig. 4A), preceded by marked increases in earthquake rate (7), and sometimes followed by surges in effusion rate at the LERZ vent ~40 km distant (16). The final collapse geometry was not fully established until mid- to late June, with clockwise propagation of a fault scarp through the center of the older 1500 CE caldera. By the time the new caldera stopped growing in early August, 62 collapses had occurred, producing as much as ~500 m of subsidence and a total collapse area of ~5 km2.

Modeling lava lake and ground deformation data

Our goals were to estimate the subcaldera magma reservoir geometry; infer the conditions under which the reservoir’s host rock began to fail; and evaluate how these parameters related to the style, location, and volume of subsequent caldera collapse. We used data from the period of near-constant high-rate subsidence after the Mw 6.9 earthquake and preceding the first collapse event on 16 May (Fig. 4), which we treated as the effective onset of caldera collapse. Observations suggest that during this time, rock at the summit responded elastically to changing stresses and slip on buried ring faults was minimal (8). We hypothesized that ground deformation and changes in lava lake surface height were generated by pressure change at constant rate p˙ in a magma reservoir beneath Kīlauea’s summit (4, 6). We constructed a model that relates p˙ to the rate of lava lake surface height change, assuming a magmastatic relationship, and to observed ground deformation velocities by using a continuum-mechanical model of a spheroidal magma reservoir embedded in an elastic half-space (Fig. 5) (8). The deformation model was implemented using the finite element method and then employed to construct a fast numerical surrogate suitable for Markov chain Monte Carlo (MCMC) estimation (8, 17). Primary model parameters are shown in Fig. 5.

Fig. 5 Model geometry and estimated parameters.

(A) Conceptual model geometry including instruments that recorded observations used in this study. The reservoir centroid is shown for simplicity directly beneath the lava lake, but this is not required in our model. (B) Marginal posterior PDFs of primary estimated model parameters (8), excluding “nuisance” parameters associated with data uncertainties (fig. S17). East and north positions are relative to 19.4073°N, 155.2784°W (the east rim of precollapse Halema‘uma‘u crater), and depth is approximately relative to the volcano’s summit.

We performed a joint Bayesian parameter estimation using the lava lake withdrawal rate together with GPS, ground tilt, and InSAR velocities (8). We also used independent information from previous studies to constrain lava lake density and host rock rigidity, and we placed limits on the proximity of the top of the magma reservoir to the surface. We directly estimated reservoir location, geometry, and pressure change rate, and allowed “nuisance” parameters (including host rock shear modulus and magma density) to vary to account for their uncertainties. From the MCMC results and additional independent information we computed other parameters of interest, such as the rate of magma outflow from the reservoir. Parameter estimates take the form of probability density functions (PDFs), which account for uncertainties in data and prior information. We found that model output is consistent with the withdrawal rate of Kīlauea’s lava lake and the first-order temporal and spatial pattern of ground deformation preceding caldera collapse (Fig. 6). We discuss our modeling results and implications throughout the following sections.

Fig. 6 Fit of model to observations.

Shown are predictions from the mean of the posterior distribution. We do not show lava lake data, which the model is able to fit “exactly” (to within an arbitrary precision). (A) Sentinel-1 ascending- and descending-mode interferograms (see fig. S15 for COSMO-SkyMed). The variance of the InSAR data is reduced by more than 95% after subtracting model predictions. Residuals in and south of the caldera do remain (the images in the rightmost column have a different color scale to highlight these effects). (B) Vertical GPS velocities. (C) Horizontal GPS velocities. Formal 95% data uncertainty ellipses are shown but are too small to be easily visible; in the estimation, these uncertainties are scaled using data-weighting hyperparameters (8). (D) Ground tilt rates.

Location and geometry of subcaldera magma storage

Magma reservoir depth, volume, and geometry play a direct role in the onset, style, and duration of caldera collapse (15, 1821), but magma storage beneath most calderas is poorly understood and subject to controversy (22, 23). Investigations at volcanoes that have hosted historic caldera-forming eruptions suggest that storage zones may be complex and occur over a range of depths (3, 2428). We found that geodetic data preceding caldera collapse at Kīlauea in 2018 are consistent with evacuation of magma from a storage reservoir centered at ~2 km depth just east of Halema‘uma‘u crater (Fig. 5 and table S2). The estimated magma reservoir is somewhat vertically elongated, as required to explain the observed ratio of vertical to horizontal displacements. The reservoir’s depth implies an initial (pre-eruptive) magma pressure of ~45 MPa on the basis of the magmastatic lava lake relationship together with prior constraint on magma density (8). To the extent that magma density and lithostatic density were similar, the open lava lake vent precludes large magmatic overpressures before the onset of the eruption (8).

In the past two millennia, two long-lived, deep calderas have existed at the summit of Kīlauea: one from ~200 BCE to ~1000 CE, and the modern caldera, which formed in ~1500 CE and began refilling in ~1800 CE (29). Magma storage beneath Kīlauea’s 1500 CE caldera was inferred in the first written records of the volcano nearly two centuries ago (30) and explains subsidence associated with rift zone intrusions and eruptions. At least two persistent magma reservoirs—the Halema‘uma‘u reservoir just east of Halem‘uma‘u crater and another at greater depth beneath the south part of the 1500 CE caldera—have been hypothesized on the basis of geodetic and other observations (6, 3138). Several transient storage zones may also have existed (36), and VLP seismic energy frequently emitted from a source ~1 km beneath the northeast rim of Halema‘uma‘u (39) has been interpreted as the intersection of north- and east-trending dikes (11, 40). The geometries and relationships between these various magma storage regions have been difficult to interpret, and in some cases appear to change over time. The reservoir location and geometry we estimate here lead us to conclude that magma withdrawal from the Halema‘uma‘u reservoir was responsible for observed ground subsidence in 2018.

Misfits between model predictions and geodetic data provide additional insight into magma storage (Fig. 6). Our model closely fits lava lake withdrawal rate data but cannot account for small-scale features observed in the InSAR data (fig. S7), nor can it explain the very-high-quality GPS data to within formal uncertainties. Material heterogeneity such as preexisting faults and altered rocks, localized shallow magma storage, or irregularities in the top of the reservoir itself may be responsible for these features [we scale data uncertainties to account for these limitations (8)]. The model also inadequately accounts for subsidence observed south of the caldera. This likely reflects the early stages of magma drainage from Kīlauea’s deeper and more enigmatic south caldera reservoir. Ground deformation believed to be due to magma evacuation from this reservoir increased in cumulative magnitude and spatial extent through June and July and continued after the cessation of the eruption (presumably as magma drained to refill the ERZ). However, most of the deformation during our modeled time period can be attributed to the Halema‘uma‘u reservoir (predicted deformation from the model reduces variance in modeled InSAR scenes by 93 to 96%).

Volume of magma storage

The volume of magma stored beneath a volcano exerts a primary control on nearly all aspects of volcanic activity, including limiting the size of an eruption and any possible caldera collapse. Yet, magma storage volumes are very poorly known at almost all of Earth’s volcanoes. Intensive study at Kīlauea over previous years has yielded estimates for the Halema‘uma‘u reservoir varying over two orders of magnitude [from 0.2 to >20 km3 (6, 34, 4144)].

In general, geodetic data can be used to resolve the quantity Vp˙/μ for a magma reservoir, where V is reservoir volume, p˙ is pressure change rate, and μ is host rock shear modulus, but not these terms independently (45). Our parameter estimation resolved V by using constraints on p˙ from the lava lake data (below) and on μ from previous studies (6, 41). Because p˙ is much more tightly constrained than μ, we were able to resolve the ratio V/μ1.3±0.15 m3/Pa (8) (fig. S16). This implies that reservoir volume should be of the same order as the rigidity of the host rock. The combination of spatially dense geodetic data with the finite-source model used in our study provided additional constraint on reservoir volume (45), and the maximum size of the reservoir was geometrically limited by its depth and shape (both resolved geodetically).

We found that 2.5 to 7.2 km3 of magma (at 68% credible bounds) was stored beneath the summit of the volcano in the Halema‘uma‘u reservoir at the beginning of May 2018. The upper bound should be considered only approximate; volumes of 10 km3 or even larger cannot strictly be ruled out by the data, particularly if we relax a priori limits on the presence of magma storage at very shallow depths (<750 m) (8). On the other hand, volumes of <1 km3 are improbable, because smaller reservoirs cannot explain the high rate of observed ground deformation without requiring an unreasonably weak host rock (pressure change rate is tightly constrained by the lava lake data). Precollapse storage volumes for other basaltic calderas are not well known, but our calculated volume is far smaller than that of reservoirs inferred to have supplied large silicic caldera-forming eruptions.

Rate of magma depressurization and drainage

Reservoir pressure change rate p˙ is constrained in our parameter estimation by the observed rate of lava lake withdrawal, the prior distribution on lava lake density, and the magmastatic assumption (8). Thus, p˙ is insensitive to geodetic data and modeling. We estimated that pressure in the reservoir decreased at 1.25 ± 0.09 MPa/day (Fig. 5B) after the Mw 6.9 earthquake. At this rate, pressure at the reservoir’s centroid would have decreased to atmospheric (an impossibility) by early June. Continuation of the eruption at a high rate for 3 months therefore required an increase of reservoir pressure through collapse of the overlying rock. This mechanism is also consistent with surges in effusion rate after collapses later in the eruption (16).

The volumetric rate of contraction V˙ of the magma reservoir and the volumetric rate q at which magma exited the reservoir are important to the timing of caldera collapse and the dynamics of summit draining and its relation with processes in the ERZ (19, 21). We computed V˙=1.3×106±0.1 m3/day (15 m3/s) using estimated model parameters together with a numerical model for the elastic compressibility of the magma reservoir (8). This estimate is tightly constrained by the geodetic data. Combined with our posterior distribution for p˙, we found that each pascal of pressure reduction in the reservoir reduced its volume by ~1 m3 (dV/dp = 1.0 ± 0.1 m3/Pa). Because of the rigidity of the host rock, the reservoir itself was contracting at only ~0.03% per day while its internal centroid pressure was decreasing at ~3% per day. At shallower depths in the reservoir, the relative pressure change rate would have been even greater.

Because magma is compressible, the rate at which the reservoir contracted was likely not equal to the rate of magma withdrawal. Using our distribution for V˙ and independent constraint on compressibility (8), we estimated a net magma outflow rate q from the Halema‘uma‘u reservoir of 2.3 million to 5.4 million m3/day (27 to 62 m3/s) at 68% credible bounds. This rate exceeds the average supply to Kīlauea from the mantle by an order of magnitude (37, 46, 47) and thus should approximate the total rate of flow to the ERZ from the contracting reservoir. Adding another ~5 to 10 m3/s from the draining lava lake and its feeder conduit (8) yields a combined outflow rate of ~35 to 70 m3/s from the lava lake and Halema‘uma‘u reservoir. This is much higher than the time-averaged eruption rate from 3 to 18 May (7 m3/s) (48), indicating that summit magma was entering the rift without erupting in order to feed deflation of the middle ERZ and growth of the LERZ intrusion. By June, after the onset of collapse events, LERZ eruption rates had increased by at least an order of magnitude (7), and the time-averaged rate of caldera collapse was ~two to five times larger than our estimated magma outflow rate. These observations strongly suggest a large increase in magma withdrawal rate from the summit in association with caldera collapse.

Reservoir failure thresholds

Placing bounds on the thresholds at which magma reservoirs begin to fail is important for determining the collapse hazard of an ongoing eruption (49), interpreting the geological record, and understanding the mechanical processes that lead to caldera collapse. Reservoir failure is triggered by stresses imparted to the host rock by changes in internal pressure. Kīlauea’s lava lake provided a window into changing magma system pressure but disappeared from view ~1 week before the first collapse event. However, by assuming that pressure continued to decrease at rate p˙ between the end of the modeled time period (14 May) and the first collapse (16 May), as suggested to first order by geodetic data, we estimated a pressure change at failure Δpf = −17.2 ± 1.1 MPa (8).

We also used tilt data as a direct empirical proxy for pressure change, using the scaling relationship established while the lava lake was active (at UWD, 0.078 ± 0.006 MPa per microradian of radial tilt). This approach does not rely on any model except for the magmastatic relationship used to establish the scaling ratio, nor does it require an assumption of constant rates, but it can be affected by ground deformation caused by processes other than reservoir pressure change. We used this approach to estimate pressure changes after 16 May under the assumption that ground tilt during collapse events was caused entirely by changes in reservoir pressure [this likely overestimates pressure changes somewhat owing to faulting processes (50)]. With this approach, we obtained pressure changes of ~17.8 and ~25.0 MPa immediately before the first collapse event on 16 May (similar to the model-based results) and the first broad-scale collapse on 29 May, respectively (Fig. 7 and fig. S10) (8). These estimates imply a relative pressure reduction exceeding 30% at the reservoir’s centroid by 16 May. They can also be related to shear stresses in the host rock, although the conditions required to trigger failure are complex and poorly understood. Using simple geometrical arguments, we computed the shear stress that the deflating reservoir imparted to an overlying cylindrical ring fault and estimated a stress change of between ~8 and 13 MPa (8, 18).

Fig. 7 Pressure change in the magma reservoir.

(A) Time series of reservoir pressure change derived from scaled tilt at UWD. The time span is similar to that in Fig. 4. Uncertainties are due to lava lake density and the lake-tilt ratio (Fig. 3). Certain offsets not apparently related to magmatic processes were removed from UWD tilt data. (B) Marginal distributions for pressure change immediately preceding the first collapse (16 May) and the first large collapse (29 May). We combined marginal distributions for tiltmeters UWD, UWE, SDH, and SMC to produce the distribution in (C).

Although it is pressure changes that trigger collapse, due to the lack of observations at natural systems failure criteria are more typically formulated in terms of volume changes. Reservoir volume change may be tracked nearly in real-time using geodetic data, and erupted volume may be tracked directly or with geophysical observations. We defined critical fractions Vcrit=ΔVf/V and fcrit=Δqf/V (19), where ΔVf and Δqf are the reservoir volume change and total magma extraction volume at the time of failure, respectively. To estimate ΔVf, we scaled the model-based estimate of Δpf at the first collapse by the ratio dV/dp obtained from the Bayesian estimation results. Because dV/dp ≈ 1, the magnitudes of pressure and volume changes were comparable. Scaling by reservoir volume yielded Vcrit = 0.27 to 0.66%, and further scaling by system compressibility yielded fcrit = 0.68 to 2.2%, both at 68% credible bounds (table S2). At 95% confidence, we concluded that <3.5% of magma was evacuated before the onset of collapse at Kīlauea.

Geometry of the roof block

The aspect ratio of the roof block above a magma reservoir (Fig. 8, C and D) influences not only the timing of collapse onset but also its subsequent structural development and style (20, 51, 52). In general, low-aspect roof blocks [Ra < 1, where Ra is the thickness T of the crust above the magma reservoir divided by the reservoir diameter D (52)] tend to favor a central coherent collapse “piston” bounded by reverse faults, whereas high-aspect (Ra > 1) blocks favor incoherent subsidence through migration of fractures upward from the reservoir. However, observational constraints on Ra from real-world caldera collapses are limited, owing to poor knowledge of the geometry of subcaldera magma reservoirs. Caldera diameter must generally be used as a proxy for reservoir diameter and roof thickness inferred roughly from geological or geophysical data (18, 19, 53, 54).

Fig. 8 Probabilistic magma storage in the Halema‘uma‘u reservoir beneath Kīlauea’s summit.

Contours and shading indicate estimated probability of magma storage based on the range of model geometries inferred in the parameter estimation (8). (A and B) Results for a horizontal slice near the reservoir centroid at 2 km depth. (C) Probability along an east-west slice at the reservoir centroid. Model depths are converted to vertical elevations using the approximate mean geodetic observation elevation [1100 m above sea level (asl)]. Colors indicate relative probability (red, more likely; blue and white, less likely). Red circles show geometry predicted by the median of the posterior distribution. Shaded DEMs in (A) and (B) show the summit as it appeared before and after the 2018 caldera collapse, respectively. The dashed rectangle above the storage zone in (C) shows the rough geometry of the roof block. The bulk of magma was stored below sea level and the subaerial ERZ vents (Fig. 1). (D) Posterior PDFs of roof aspect ratio and the probability of complete reservoir evacuation given the observed caldera collapse volume, along with complementary cumulative distribution.

The set of finite-source geodetic models derived from our MCMC analysis allowed us to estimate Ra. Taking T to be the distance between the surface and the top depth of each magma reservoir in the posterior probability distribution, we found that the roof block at Kīlauea was thin and wide, with Ra ≈ 0.4 (Fig. 8). Ra would be smaller if we were to relax our minimum reservoir top depth (8) but would be larger if we measured height from a point other than its very top. Small reservoirs from our probability distribution yield aspect ratios closer to 1, but in general Ra > 1 appears unlikely.

Reservoir evacuation and the end of the eruption

It is often assumed that caldera-forming eruptions are terminated by the near-complete evacuation of their source reservoirs (3, 49, 54, 55), as suggested by some models (56) and perhaps indicated by long repose periods after some collapses (55). This hypothesis has implications for hazards during ongoing eruptions. It also allows for interpreting data from past events because it implies that erupted volume is approximately equal to reservoir volume. Although there is evidence that this assumption may not be valid (20, 56), it has been difficult to evaluate because of limited knowledge of subcaldera magma reservoir volumes.

Taking the total 2018 summit collapse volume (7) as a proxy for the total volume change of the shallow reservoir during the eruption, we used our posterior PDF for reservoir volume to estimate that only 11 to 33% of Kīlauea’s shallow magma reservoir was evacuated by the end of the eruption. The probability of complete drainage is very small; we estimated <5% probability that even half of the reservoir was drained (Fig. 8). This inference is consistent with the relative constancy of collapse-related geophysical signals from June to August (7), which might have changed in character if the reservoir had neared complete evacuation, and also with the post-eruptive return of episodic days-long ground deformation cycles at the summit, which are believed to be caused by pressure perturbations in the shallow magma reservoir (6). Our results suggest caution in assuming that magma reservoirs (at least basaltic ones) fully evacuate during caldera-forming eruptions.


Caldera collapse at Kīlauea in 2018 was caused by high-rate magma evacuation from a roughly equant storage zone of several cubic kilometers at shallow depth (~2 km), centered just east of the former Halema‘uma‘u crater. Many previous studies have inferred magma storage in this area, but 2018 data provide new insights. Our simple geodetic model cannot account for magma withdrawal from other reservoirs or the fine-scale topology of magma storage [for instance, we likely cannot rule out magma stored in a broad plexus of interconnected magma-filled cracks (57) with similar magma volume], but it well explains the observed overall spatial pattern of ground deformation. Likewise, the rate of magma system depressurization estimated by our model can explain the observed rate of lava lake withdrawal and ground deformation.

When did caldera collapse begin? Seismicity after the Mw 6.9 earthquake might have indicated the early stages of caldera-fault propagation at depth (58), but there appeared to be relatively little effect on surface deformation during the first half of May (8), and there was no geophysical evidence for collapse of rock into the deeper magmatic system. Quasi-periodic VLP seismic and geodetic signals recorded from 16 to 26 May were associated with vent widening, volume loss, and ejection of ash, but not surface faulting over a broad area. Yet, InSAR data from this time showed a more complex deformation pattern in the caldera than that present earlier in the month, suggestive of the early-stage surface expression of slip on buried caldera faults. Furthermore, geophysical signals were similar to those recorded during caldera collapses at other volcanoes and at Kīlauea after 29 May, when broadscale collapse was visually observed. Thus, the events of 16 to 26 May were evidently related to collapse of rock into the magmatic system, although the extent to which these collapses occurred into the lava lake feeder conduit and/or shallow dike-like storage bodies, as opposed to the Halema‘uma‘u reservoir, remains an open question. Also unclear is the extent to which any propagation of buried caldera faults during this time related to geophysical observations. Nonetheless, we conclude that caldera collapse effectively began on 16 May, accelerated and enlarged on 29 May (when we were able to closely tie visual observations of broader-scale collapse to geophysical signals), and did not reach its full surface expression until late June.

The critical thresholds required for caldera collapse are thought to be controlled by many factors, including the shape (aspect ratio) of the roof rock above the reservoir (18, 19); exsolved magmatic volatiles, which buffer pressure drop due to magma extraction (56, 59, 60); and preexisting faults and weaknesses (49). At Kīlauea, the 2018 collapse occurred within an older, larger caldera and, in some areas, appeared to proceed along preexisting faults. We speculate that both the empty lava lake vent and the relatively thin and wide roof block might have promoted failure (18, 19). It is also possible that, at shallow depths, the retreating magma surface could have encountered a flared conduit geometry, leading to instability. An open question is how critical failure thresholds might differ between small nested-caldera basaltic systems, such as Kīlauea, and large silicic systems.

Caldera collapse began at Kīlauea after the elastic reservoir had contracted only very slightly (Vcrit < 1.1%), caused by withdrawal of only a very small fraction of its stored magma (fcrit < 4%). Geological observations and models have suggested that fcrit may range from <10% to >90% (18, 19), but direct evidence has been lacking (note that many studies do not distinguish between Vcrit and fcrit, which are equal only if magma is incompressible). Geophysical observations from basaltic collapses at Piton de la Fournaise, Fernandina (Galápagos), Miyakejima, and Bárðarbunga volcanoes yielded fcrit of ~8 to 20%, in some cases much lower than values suggested by analog models (3, 49) but still much higher than we found for Kīlauea. Although it is possible that collapse began unusually quickly at Kīlauea, these previous estimates had to rely on assumptions that the volumes of initial collapse events were comparable to precollapse magma withdrawal volumes and that eruptions completely drained their magma reservoirs (3, 49, 54). As we have shown here, these assumptions are not always valid and could lead to a substantial overestimation of fcrit. These discrepancies indicate that calderas may fail more quickly than previously understood.

Although it is changes in magma pressure that drive host rock failure and caldera collapse, robust estimates of precollapse pressure changes have previously been unavailable. Magma extraction volumes are far more readily measured in nature but are only relevant to collapse to the extent that they influence reservoir pressure (an effect modulated by the compressibility of magma in the reservoir). Data from Kīlauea allowed us to move beyond reliance on fcrit and directly estimate precollapse pressure change. Knowledge of the pressure change makes it possible to compute stress changes on the roof block and thus tie the observations to the failure process.

Once failure began, episodic roof block collapse transferred the load of the overlying rock to the magma, increasing its pressure. This process may explain similar episodic geophysical observations at other basaltic caldera collapses (14, 15, 61). By using ground tilt as a proxy for reservoir pressure change, we estimated that inflationary deformation during the first collapse event on 16 May was caused by a pressure increase of ~1.3 MPa in the reservoir, only a fraction of the preceding deflation. Because reservoir pressure was likely near lithostatic at the onset of the eruption, this result indicates incomplete repressurization of the reservoir after the onset of collapse and implies residual frictional strength on the walls of the collapsing block(s) such that the weight of the roof was not entirely supported by the magma. This finding stands in contrast to assumptions that roof collapses reestablish lithostatic pressure in the reservoir (56, 59) but supports the results of some numerical models (62).

The surface expression of caldera collapse was complex, asymmetric, and evolving, consisting of funnel-like gravitational failure into the evacuated lava lake vent and piston-like slumping of coherent blocks as large as ~150 ha, in some cases clearly bounded by preexisting faults. Taken as a whole, these events were consistent with collapse of roof rock into a shallow reservoir, governed not only by the aspect ratio of the roof but also by preexisting caldera faults and structural weaknesses, and possibly shallow unmodeled magma storage [e.g., (11, 63)]. These observations are consistent with geological investigations and numerical experiments that demonstrate the complex diversity of collapse styles that can occur during caldera formation (51, 64).

The location and lateral extent of magma storage inferred from our model are similar to the final geometry of the 2018 caldera collapse (Fig. 8). To first order, the relationship between the range of plausible reservoir geometries and observed caldera dimensions favors primary collapse faults ranging from near-vertical to inward dipping. Results indicate that the shallow subcaldera magma storage system spanned only a portion of the caldera in existence from 1500 CE to the present. The larger magma storage body required to explain the 1500 CE collapse may have been partially destroyed then or in a subsequent event (such as a large collapse that occurred at the volcano in 1868) or may have involved failure of deeper parts of the summit magma system.

Globally, lava lakes are rare. Where they do exist, close observation during magma draining events may bear rich dividends, particularly if relayed in real time. Some of the data used in this study were evaluated in rapid-response mode internally by the USGS during the eruption with a preliminary form of our model. Resulting parameter estimates were used to better understand the possible course of the eruption and guided our thinking about hazards as the eruption progressed, highlighting the importance of near–real-time data and modeling capabilities at the world’s volcano observatories.


Despite insights into volcanic calderas afforded over the past two decades by well-documented collapses at Miyakejima, Piton de la Fournaise, and Bárðarbunga volcanoes, the conditions that trigger the onset of collapse remain only poorly understood. Draining of Kīlauea’s summit lava lake in 2018 yielded a window into changing pressure in the volcano’s shallow magma reservoir. We tracked the evolution of the magmatic system as it underwent steady high-rate elastic decompression due to magma withdrawal, followed by episodic fault-bounded caldera collapse. We were able to quantify the changing pressure in the reservoir, which, together with geodetic data, made it possible to estimate the volume of magma storage and the critical thresholds that preceded the onset of collapse. Caldera collapse began due to a relatively large decrease in the magma reservoir’s internal pressure caused by withdrawal of only a small fraction of stored magma. Episodic fault-bounded subsidence of the roof block above the reservoir increased magma pressure, sustaining the flow of magma and thus representing a critical turning point in the evolution of the eruption.

Supplementary Materials

Materials and Methods

Figs. S1 to S17

Tables S1 and S2

References (71117)

Movie S1

References and Notes

  1. Materials and methods are available as supplementary materials.
Acknowledgments: E. Rumpf analyzed vent collapse from HVO webcam photos. P. Cervelli assisted with implementation of the analytical deformation model. M. McLay and Y. Zheng assisted with interferogram processing. This work benefited from numerous discussions with scientists at the Hawaiian Volcano Observatory and throughout the USGS. Any use of trade, firm, or product names is for descriptive purposes only and does not imply endorsement by the U.S. government. Funding: This work was funded by the USGS Volcano Hazards Program. Author contributions: K.R.A. conceptualized the project, analyzed data, developed the model, performed inversions, and coordinated manuscript writing. I.A.J. operated geodetic instruments, analyzed geodetic data, and contributed to modeling. M.R.P. installed and operated lava lake instrumentation and analyzed lava lake data. M.G. implemented the emulator, analyzed data uncertainties, and contributed to the Bayesian inversion. P.S. contributed to conceptualization, modeling, and validation of results. M.P.P. processed and analyzed InSAR data. E.K.M.-B. interpreted results and contributed to modeling. A.M. and all other authors contributed to data interpretation and manuscript production, and all USGS authors contributed to the eruption response and data collection. Competing interests: The authors declare no competing interests. Data and materials availability: Sentinel SAR data are available from (65); COSMO-SkyMed SAR data from (66); DEM data from (67, 68), tilt data from (69), GPS data from (70), and lava lake data from (9).

Stay Connected to Science

Navigate This Article