We introduce Symmetric Autoencoder (SymAE), a neural-network architecture designed to automatically extract earthquake information from far-field seismic waves. SymAE represents the measured displacement field using a code that is partitioned into two interpretable components: source and path-scattering information. We achieve this source-path representation using the scale separation principle and stochastic regularization, which traditional autoencoding methods lack. According to the scale separation principle, the variations in far-field band-limited seismic measurements resulting from finite faulting occur across two spatial scales: a slower scale associated with the source processes and a faster scale corresponding to path effects. Once trained, SymAE facilitates the generation of virtual seismograms, engineered to not contain subsurface scattering effects. We present time-reversal imaging of virtual seismograms to accurately infer the kinematic rupture parameters without knowledge of empirical Green’s function. SymAE is an unsupervised learning method that can efficiently scale with large amounts of seismic data and does not require labeled seismograms, making it the first framework that can learn from all available previous earthquakes to accurately characterize a given earthquake. The paper presents the results of an analysis of nearly thirty complex earthquake events, revealing differences between earthquakes in energy rise times, stopping phases, and providing insights into their rupture complexity.