Mesophotic coral reef ecosystems are remote from coastal stressors, but are still vulnerable to over-exploitation, and remain mostly unprotected. They may be the key to coral reefs resilience, yet little is known about the pattern of larval subsidies from deeper to shallower coral reef habitats. Here we use a biophysical modeling approach to test the hypothesis that fishes from mesophotic coral reef ecosystems may replenish shallow reef populations. We aim at identifying the spatio-temporal patterns and underlying mechanisms of larval connections between Pulley Ridge, a mesophotic reef in the Gulf of Mexico hosting of a variety of shallow-water tropical fishes, and the Florida Keys reefs. A new three-dimensional (3D) polygon habitat module is developed for the open-source Connectivity Modeling System to simulate larval movement behavior of the bicolor damselfish, Stegastes partitus, in a realistic 3D representation of the coral reef habitat. Biological traits such as spawning periodicity, mortality, and vertical migration are also incorporated in the model. Virtual damselfish larvae are released daily from the Pulley Ridge at 80 m depth over 60 lunar spawning cycles and tracked until settlement within a fine resolution (~900 m) hydrodynamic model of the region. Such probabilistic simulations reveal mesophotic-shallow connections with large, yet sporadic pulses of larvae settling in the Florida Keys. Modal and spectral analyses on the spawning time of successful larvae, and on the position of the Florida Current front with respect to Pulley Ridge, demonstrate that specific physical-biological interactions modulate these “perfect storm” events. Indeed, the co-occurrence of (1) peak spawning with frontal features, and (2) cyclonic eddies with ontogenetic vertical migration, contribute to high settlement in the Florida Keys. This study demonstrates that mesophotic coral reef ecosystems can also serve as refugia for coral reef fish and suggests that they have a critical role in the resilience of shallow reef communities.