DeconNet: An Automated Framework for Resolving Finite-Source Attributes of Small Earthquakes with Array-Based Neural Network Deconvolution

Abstract Detailed analysis of small earthquake ruptures provides important constraints on fault behavior and earthquake physics. A major challenge is separating source signals from complex path effects using conventional time-domain deconvolution, which is typically unstable, labor-intensive, and unable to exploit interstation coherence. We present DeconNet, a standardized and fully automated inversion framework for finite-source attributes of small earthquakes, built around a U-Net-based neural network core that simultaneously recovers apparent source time functions (ASTFs) across multiple stations without manual intervention. By learning spatial correlations during training, DeconNet achieves robust performance, with synthetic tests showing that 98.8% of predicted ASTF durations deviate by <3%. Application to six small earthquakes (ML 1.88–2.82) in a hydraulic-fracturing experimental field in Alberta, Canada, with a dense downhole seismic array resolves rupture parameters and reveals spatial complementarity with surrounding background seismicity. DeconNet enables automated, efficient analysis of second-order rupture characteristics and opens the door to “generalized earthquake catalogs” enriched with additional rupture kinematic and dynamic parameters, with strong potential for application to dense seismic arrays worldwide.