<p>Earthquake science and seismology rely on phase association: grouping seismic arrivals recorded across multiple stations into the earthquakes that generated them. Modern deep-learning detectors now routinely identify many more small events than before, providing information about fault dynamics on increasingly fine spatiotemporal scales. But the resulting high-rate arrival sequences are increasingly difficult to associate, especially when local wave speeds are heterogeneous or poorly known. Here we introduce HARPA, a phase-association framework that represents observed and predicted arrival sequences as probability distributions and compares them using an optimal-transport metric. HARPA jointly estimates earthquake locations, origin times and a low-dimensional representation of the wave-speed field using travel-time neural fields, neural networks that map coordinates to travel times. On standard low-rate datasets with simple wave speed, HARPA performs comparably to state-of-the-art associators. At high rates and with laterally heterogeneous or unknown wave speed, HARPA outperforms existing methods. Our results show that adaptive travel-time modeling becomes important when seismicity is dense and suggest a route toward joint association and passive-source tomography from unassociated arrivals.</p>