Starling murmurations under predation

European starlings perform a great diversity of patterns of collective behaviour when hunted by aerial predators; their large flocks continuously change shape, size, and internal structure. How these complex patterns emerge by self-organization through rules at the individual level is still unknown. Here, we combine video footage of large flocks of starlings that are pursued by a robotic predator, the Robot-Falcon, with simulations in a new data-driven 3-dimentional agent-based model (StarEscape) in order to unravel the emergence of these aerial displays. In vivo, we identified that flock members often differed in their evasive manoeuvres and that different patterns of collective escape arose simultaneously in different parts of the flock. In silica, we investigated how rules of coordination and escape at the individual level generate the group density, internal dynamics, and patterns of collective escape we observed in the field. We show that key factors generating these complex patterns are the type of escape reaction of a few early responders, the speed with which the escape information propagates through the flock, the positions of the evading flock members relative to the predator, and the previous state of the flock (hysteresis). The processes we reveal are key to understanding starling murmurations, while StarEscape is a biologically-relevant tool for the study of both mechanistic and functional aspects of spatiotemporal patterns across many species. Our study highlights the importance of investigating fine-scale dynamics and feedback between the micro and macro levels, such as individual reaction frequency and group geometry, when studying self-organized adaptive systems.