SeisBind: Physics-Aware Tri-Modal Representation Binding for Seismic Data via Contrastive Learning

TL;DR

SeisBind addresses the need for physically grounded seismic representations by learning a tri-modal latent space that binds seismic data, velocity models, and explicit physics descriptors via tri-modal contrastive learning. It employs a CLIP-style architecture with three encoders projecting to a shared 256-dimensional embedding, and a hierarchical loss that aligns seismic with velocity and velocity with physics, stabilizing training. On the OpenFWI dataset, SeisBind achieves strong cross-modal retrieval (R@1 up to 0.578 for seismic→velocity and 0.681 for velocity→physics) and enables physics-parameter inference from embeddings with an average relative error of about $9.12 ext{\%}$, including surface velocity error of $2.47 ext{\%}$, illustrating improved interpretability over pixel-wise inversion. This framework provides a flexible, interpretable alternative for subsurface characterization, facilitating expert-guided exploration and potentially faster inference.

Abstract

This letter proposes a physics-aware multi-modal contrastive learning framework designed to transform complex seismic wavefields into human-readable physical representations. Traditional data-driven inversion methods often focus on pixel-wise mapping, which lacks physical grounding and interpretability. To address this, we introduce a novel framework that jointly aligns seismic shot gathers, subsurface velocity models, and explicit physical descriptors (e.g., mean velocity and gradients) in a shared latent space. By introducing these descriptors as a third modality, our approach encourages the learned embeddings to capture intrinsic geological semantics rather than superficial signal correlations. Experiments on the OpenFWI dataset demonstrate that the proposed method not only achieves robust seismic-to-velocity retrieval but also preserves meaningful physical semantics, enabling cross-modal inference of interpretable attributes. This representation-centric perspective provides a flexible foundation for expert-guided subsurface characterization.

Figures (3)

• Figure 1: model architecture
• Figure 2: Qualitative retrieval results. For each query seismic shot gather (first column), the corresponding ground truth (second column) and the top-3 retrieved velocity models are displayed. Similarity scores are provided above each retrieved candidate, demonstrating the model's precision in structural matching.
• Figure 3: Similarity heatmap: Seismic-to-Velocity alignment.