Modeling Fine-Scale Acoustic Variability in Stratified Flows Using Diffusion Models

This work tackles the challenge of modeling fine-scale acoustic variability driven by gravity-wave (GW) fields in stratified atmospheric flows. While Fourier Neural Operators (FNOs) offer an efficient framework for surrogate modeling of PDE-governed systems, they exhibit a spectral bias that hampers their ability to resolve high-frequency features associated with subgrid-scale atmospheric variability. To address this limitation, we introduce a hybrid modeling framework that combines an FNO with a conditional diffusion model. Acting as a generative corrector, the diffusion model refines FNO predictions by injecting high-resolution structures consistent with the learned statistical distribution of GW-induced perturbations. This two-stage architecture is applied to a decade of infrasound recordings collected from stations located hundreds of kilometers from explosive sources. The hybrid model successfully reconstructs high-frequency wavetrains and significantly improves the match between predicted and observed power spectral densities, outperforming the FNO baseline. Additionally, modal decomposition reveals that the hybrid approach restores higher-order modes associated with fine-scale variability that remain unresolved in FNO-only outputs. Our findings establish a new paradigm for integrating generative models with neural operators, paving the way for data-driven surrogate models of multi-scale infrasound propagation in complex, stochastic atmospheric environments.