A Novel Diffusion Model for Pairwise Geoscience Data Generation with Unbalanced Training Dataset

TL;DR

UB-Diff tackles the challenge of generating paired multi-modal geoscience data under data imbalance by introducing a co-latent diffusion framework built on a 1-in-2-out encoder-decoder. The method leverages a matched two-step training scheme to fully utilize abundant data from the majority modality while integrating scarce minority data, enabling reliable simultaneous generation of velocity maps and seismic waveforms. Experiments on OpenFWI show substantial improvements in both macro and pairwise data quality over state-of-the-art diffusion baselines, with clear gains in downstream FWI tasks. The approach promises broad applicability to data-scarce scientific domains and other multi-modal synthesis tasks where modality imbalance is prevalent.

Abstract

Recently, the advent of generative AI technologies has made transformational impacts on our daily lives, yet its application in scientific applications remains in its early stages. Data scarcity is a major, well-known barrier in data-driven scientific computing, so physics-guided generative AI holds significant promise. In scientific computing, most tasks study the conversion of multiple data modalities to describe physical phenomena, for example, spatial and waveform in seismic imaging, time and frequency in signal processing, and temporal and spectral in climate modeling; as such, multi-modal pairwise data generation is highly required instead of single-modal data generation, which is usually used in natural images (e.g., faces, scenery). Moreover, in real-world applications, the unbalance of available data in terms of modalities commonly exists; for example, the spatial data (i.e., velocity maps) in seismic imaging can be easily simulated, but real-world seismic waveform is largely lacking. While the most recent efforts enable the powerful diffusion model to generate multi-modal data, how to leverage the unbalanced available data is still unclear. In this work, we use seismic imaging in subsurface geophysics as a vehicle to present ``UB-Diff'', a novel diffusion model for multi-modal paired scientific data generation. One major innovation is a one-in-two-out encoder-decoder network structure, which can ensure pairwise data is obtained from a co-latent representation. Then, the co-latent representation will be used by the diffusion process for pairwise data generation. Experimental results on the OpenFWI dataset show that UB-Diff significantly outperforms existing techniques in terms of Fréchet Inception Distance (FID) score and pairwise evaluation, indicating the generation of reliable and useful multi-modal pairwise data.

Figures (11)

• Figure 1: Illustration of FWI: (a) photo of a flat rock layer; (b) velocity map used to show the subsurface structure; (c) seismic waveform obtained from the receivers placed on the surface; and (d) wave propagation in the velocity map.
• Figure 2: Overview of UB-Diff, which utilizes all available data, benefiting the whole process, especially the diffusion process.
• Figure 3: 1-in-2-out network for seismic waveform and velocity map. Using the example with the velocity map as the majority data and the seismic waveform as the minority.
• Figure 4: Diffusion process of UB-Diff, generating latent of $\mathbf{ma}$ and $\mathbf{mi}$ simultaneously.
• Figure 5: Generated velocity map by baselines and UB-Diff.