On 31 August a new eruption began from the same fissure and is still ongoing at the time of writing. After 4 September the movement associated with the dyke was minor, suggesting an approximate equilibrium between inflow of magma into the dyke and magma flowing out of it feeding the eruption. Minor eruptions may have occurred under Vatnajškull; shallow ice depressions marked by circular crevasses (ice cauldrons) were discovered in the period 27/08-07/09, indicating leakage of magma or magmatic heat to the glacier causing basal melting ( Fig. 1 and 2b). On 5 September, aircraft radar profiling showed that the ice surface in the centre of the B ‡r!arbunga caldera had subsided 16 m relative to the surroundings, resulting in a 0.32±0.08 km 3 subsidence bowl ( can be compared to a 1 day interferogram over the ice surface spanning 27 -28 August (Fig. 1), that has maximum line-of-sight (LOS) increase of 57 cm, indicating 55-70 cm of subsidence, during 24 hours. From 24 August to 6 September 16 M≥5 earthquakes occurred on the caldera boundary.Over 22000 earthquakes were automatically detected 16/08-06/09 2014, 5000 of which have been manually checked. Four thousand of these have been relatively relocated, defining the dyke segments. Ground deformation in areas outside the Vatnajškull ice cap, and on nunataks within the ice cap, is well mapped by a combination of InSAR, continuously recording GPS sites, and campaign GPS measurements. The GPS observations and analysis give the temporal evolution of the three-dimensional displacements used in the modelling (Fig. 1). Interferometric analysis of synthetic aperture radar images from the COSMO-SkyMed, RADARSAT-2 and TerraSAR-X satellites was used to form 11 interferograms showing LOS change spanning different time intervals (Supplementary Fig. 2). The analysis of seismic and geodetic data is described in Methods.Initial modelling of the dyke, with no a priori constraints on position, strike or dip, show the deformation data require the dyke to be approximately vertical and line up with the seismicity (Extended Data item 4). We therefore fixed the dip to be vertical and the lateral position of the dyke to coincide with the earthquake locations.We modelled the dyke as a series of rectangular patches and estimated the opening and slip on each patch ( Fig. 3a; see Supplementary Figures 3-4 for slip and standard deviations of opening). We used a Markov-chain Monte Carlo approach to estimate 7 the multivariate probability distribution for all model parameters (Methods) on each day 16/08-06/09 2014 (Fig. 2d). The results suggest that most of the magma injected into the dyke is shallower than the seismicity, which mostly spans the depth range from 5 to 8 km below sea level (see Fig. 2c and Methods). While magma may extend to depths greater than 9 km near the centre of the ice cap, towards the edge of the ice cap where constraints from InSAR and GPS are much better, significant opening is all shallower than 5 km (Fig. 3a). The total volume intruded into the dyke by 28 August was 0.48-0...