Improving Real-Time Earthquake Source Characterization Using Diffusion Model Based Broadband Envelope Synthetics

Reliable real-time earthquake source characterization requires the rapid selection of solutions from competing algorithms while minimizing false alarms. To address this challenge, Jozinović et al. (2024) have proposed a ground-motion-envelope-based goodness-of-fit approach that  ranks candidate source solutions using amplitude ratios and cross-correlation between observed and predicted waveform envelopes. In its current implementation, however, this approach relies on the ground motion envelope prediction model of Cua (2005), which is limited to small-to-moderate sized  earthquakes. 

In this work, we explore the benefits and limitations of replacing this empirical model with envelopes derived from machine-learning-generated broadband (up to 50 Hz) synthetic waveforms (Palgunadi et al., 2025). These synthetics are generated using a conditional denoising diffusion model, conditioned on preliminary source parameters (magnitude, hypocentral distance, depth), and site effects. For large magnitude events, we superpose point-source synthetics to produce realistic finite-fault rupture waveforms using the  SWEET workflow (Colosimo, MSc thesis).

We find that the diffusion-based synthetics extrapolate realistically across a broader magnitude range and reproduce observed envelope characteristics as well as, or even better than, the empirical prediction model. This capability has the potential to enable  earlier and more reliable identification of correct source solutions, reduce magnitude and location bias, and improve robustness for larger events.