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

Junhuan Yang, Yuzhou Zhang, Yi Sheng, Youzuo Lin, Lei Yang

TL;DR

This paper introduces UB-Diff, a novel diffusion model designed to generate paired geoscience data (like seismic waveforms and velocity maps) even when one type of data is much scarcer than the other. By cleverly leveraging the abundant data to train a shared latent representation, UB-Diff significantly improves the quality of generated paired data compared to existing methods, addressing a critical bottleneck in scientific data generation.

What Problem Does This Paper Address?

Generative AI has revolutionized many fields, but its application in scientific domains like geophysics faces unique challenges. A major hurdle is data scarcity, particularly the lack of paired multi-modal data. In geophysics, for instance, velocity maps (representing rock properties) can often be simulated or are more readily available, while actual seismic waveforms (the data used to infer these properties) are much harder and more expensive to obtain. This data imbalance means that most existing data-driven methods, which rely on large, perfectly paired datasets, struggle to perform effectively. The paper aims to bridge this gap by developing a generative model that can produce high-quality, paired geoscience data despite significant imbalances in the availability of different data modalities.

Methods

The core of UB-Diff is a novel diffusion model adapted for multi-modal generation. The key innovation lies in its encoder-decoder architecture and a two-step training strategy.

  • One-in-Two-Out Encoder-Decoder: The model uses a single encoder that takes data from the majority modality (e.g., velocity maps) and maps it into a co-latent space. This co-latent representation is designed to capture shared information between the modalities. Two separate decoders then reconstruct both the majority modality and the minority modality (e.g., seismic waveforms) from this co-latent space. This ensures that the generated data is inherently paired.
  • Two-Step Training: To handle data imbalance, the training is split into two phases:
  • Self-supervised learning on the majority data: The encoder and decoders are first trained using only the abundant majority modality data. This step focuses on learning a robust representation from the plentiful data.
  • Fine-tuning with minority data: The model is then fine-tuned using the scarce minority modality data, leveraging the previously learned co-latent representation. This allows the model to learn how to generate the less available data while still benefiting from the information learned from the majority data.
  • Diffusion Process: The diffusion process operates in the learned co-latent space, enabling the generation of paired data by denoising a latent representation.

Datasets

The experiments were conducted using five geoscience datasets from the OpenFWI benchmark: FlatVel-A, CurveVel-A, FlatFault-A, CurveFault-A, and Style-A. The paper notes that for training, a large proportion of velocity maps were used (e.g., 24,000 to 60,000 per dataset), while only a much smaller number of paired seismic waveforms were available (1,000 or 5,000). No new benchmark datasets were introduced; instead, existing ones were utilized to demonstrate the model's effectiveness on imbalanced real-world data.

Key Insights and Ideas

  • The power of a shared co-latent space: The paper's central idea is that by learning a common latent representation from the abundant data, the model can effectively guide the generation of the scarcer data, creating well-aligned pairs.
  • Two-step training is crucial for imbalance: The proposed two-step training strategy is not just an optimization; it's essential for UB-Diff to overcome the limitations imposed by data scarcity and imbalance, leading to significantly better results than a single-stage training.
  • Seismic waveform generation benefits from velocity maps: The results suggest that even when generating only seismic waveforms (a single-task scenario), incorporating information from velocity maps (via the UB-Diff framework) leads to superior performance compared to methods that only use seismic data. This highlights the synergistic benefit of multi-modal learning.
  • Diffusion models can be adapted for scientific data: While diffusion models have excelled in natural image generation, this work demonstrates their potential and adaptability for complex scientific data with specific physical constraints and data availability issues.

Key Results

UB-Diff demonstrated superior performance across various metrics compared to existing methods like MT-Diffusion, DDPM, and SD.

  • FID Scores: UB-Diff achieved significantly lower FID scores (indicating better generation quality) for both velocity maps and seismic waveforms, especially in scenarios with limited seismic data (1k Seis). For example, on FlatVel-A with 1k Seis, UB-Diff achieved FID scores of 16.45 for velocity and 97.03 for seismic data, drastically outperforming MT-Diff.
  • Pairwise Evaluation: When trained on generated data and tested for its ability to aid Full-Waveform Inversion (FWI) tasks using metrics like SSIM, MAE, and MSE, UB-Diff consistently produced better results than MT-Diff, indicating more reliable and accurate paired data generation.
  • Qualitative Results: Visualizations showed that UB-Diff could generate more accurate reflected seismic waves compared to MT-Diffusion, which struggled with these subtle features.

Limitations and Future Work

  • Computational Cost: While diffusion models are powerful, they can be computationally intensive. The paper doesn't explicitly detail the training or inference times but implies a trade-off for improved quality.
  • Generalizability to other scientific domains: The current work focuses on seismic data. While the framework is designed to be generalizable, its performance on other scientific modalities (e.g., medical imaging, climate data) would require further validation.
  • Future Work: The authors suggest exploring broader applications in scientific data generation and further investigating the model's adaptability to different data scenarios and modalities. The potential for UB-Diff to aid in other physics-based inverse problems is also an avenue for future research.

Summary updated

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.
BibTeX
@inproceedings{yang2025novel,
  author = {Junhuan Yang and Yuzhou Zhang and Yi Sheng and Youzuo Lin and Lei Yang},
  title = {{A Novel Diffusion Model for Pairwise Geoscience Data Generation with Unbalanced Training Dataset}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i20.35504},
}
Figure