Generative Diffusion Models for Geophysical Classification Problems

Summary

Reliable seismic source classification requires distinguishing physically distinct yet observationally overlapping signals whose diagnostic information is distributed across multiple temporal and spectral scales. Here we present DiffClassNet, a diffusion-based framework for geophysical classification that leverages the multi-scale feature-learning properties of denoising diffusion probabilistic models. By extracting and aggregating features across multiple noise levels during the denoising process, the model learns hierarchical representations in which different diffusion timesteps emphasize complementary signal characteristics, from high-frequency onset patterns at low noise levels to broader energy distributions at higher noise levels. We evaluate the framework on two representative seismic classification problems in Texas: discrimination between natural earthquakes and quarry blasts, and causal-factor identification of induced seismicity in the Delaware Basin involving hydraulic fracturing, shallow saltwater disposal, and deep saltwater disposal. Across the benchmark settings considered here, DiffClassNet achieves competitive performance relative to established baselines. Analyses of feature evolution, latent-space organization across diffusion levels, and architectural ablations further suggest that the diffusion mechanism helps preserve complementary class-relevant structure and supports a more interpretable feature space. These results indicate that diffusion-based learning provides a useful framework for seismic source classification, with potential value for both predictive performance and interpretability.