Under what circumstances is the final size of a laboratory earthquake predictable at the onset of the P-wave?

• We generate laboratory earthquakes with seismograms and dynamic rupture measurements. • Seismograms combined with imaging validate basic far field seismic assumptions. • On average, large events initially accelerate faster than small ones. • Magnitude predictability is strong except for full ruptures that follow a partial rupture. How do earthquakes begin and what information about this process is contained in a far field seismogram? We present a quantitative analysis of laboratory earthquakes incorporating both laboratory-scale seismic measurements coupled with high-speed imaging of the controlled dynamic ruptures that generated them. We generated variations in the rupture properties by imposing sequences of controlled artificial barriers along the laboratory fault. We first demonstrate that direct measurements of imaged slip events correspond to established seismic analysis of acoustic signals; the seismograms correctly record the rupture moments and maximum moment rates. We then investigate the ruptures’ early growth by comparing their measured seismogram velocities to their final size. Due to higher initial elastic energies imposed prior to nucleation, larger events accelerate more rapidly at the rupture onset. We find that the corresponding seismogram velocities are therefore predictive of the final rupture size. This observation holds in the presence of barriers with one notable exception. Rupture events that overtake a previously arrested rupture are less magnitude predictable, likely because of the stress heterogeneity (and resulting stored elastic energy) induced by the earlier event. For all other events, the higher elastic energy at nucleation results in faster and larger ruptures, and hence the initial seismogram velocity and ultimate size correlate well. This degree of magnitude predictability is consistent with some, but not all recent natural observations. For early warning purposes, we suggest that confining the observational database to the conditions most conducive to magnitude predictability may provide stronger correlations.