Extraction of Stable Empirical Green's Functions From Short‐Duration Ambient Noise Using a Physics‐Constrained Self‐Supervised Network

Abstract

Ambient noise interferometry is a key seismological method that extracts stable empirical Green's functions (EGFs) from seismic noise for subsurface imaging without requiring earthquakes. However, its performance is often challenged by non‐diffuse and persistent directional noise sources, which hinder the retrieval of EGFs and introduce spurious arrivals. The conventional method with year‐long stacking of cross‐correlation functions (CCFs) is impractical for short‐duration deployments and often fails to suppress the spurious arrivals arising from the uneven distribution of noise energy. To overcome these limitations, we develop a self‐supervised neural network weighting stacking (NNWS) method that learns optimal stacking weights for CCFs by jointly optimizing signal‐to‐noise ratio (SNR) and waveform symmetry, both key proxies of a diffuse wavefield. This physics‐constrained, label‐free approach effectively suppresses acausal signals from strong directional noise. Using synthetic tests, we demonstrate that NNWS successfully mitigates contamination even when interfering noise arrives near the true surface wave. Applied to USArray data, our method recovers stable, high‐quality EGFs using only 10 days of continuous data. The resulting SNRs surpass those achieved by conventional 1‐year linear stacking. The resulting EGFs also exhibit enhanced symmetry and significantly suppressed acausal noise, enabling reliable dispersion measurements from short‐duration data. NNWS thus provides an efficient and robust strategy for rapid retrieval of EGFs under realistic, non‐uniform noise conditions, considerably expanding the application scope of ambient noise interferometry.