Paper submitted to IMAGE 2026, by Sujitha Martin, Manoj Alwani, Yuan Tian, Jared Kramer, Arun Ramanathan (Amazon Web Services), Ben Lasscock, Alejandro Valenciano (TGS).
Abstract
Raw seismic data consists of complex signals captured by specialized instruments that geophysicists must process through carefully sequenced filters to extract meaningful geological information. When designing these processing workflows, experts frequently encounter technical challenges ranging from incorrect configurations to problematic information flow patterns. Traditionally, these errors remain undetected until workflow execution, resulting in significant resource waste and project delays. Our research explored how large language models (LLMs) can process graph-based workflow representations to predict execution failures before implementation. We developed two complementary approaches: an embedding-based supervised contrastive learning method for classifying workflows, and a fine-tuned LLM approach for generating error explanations. Our findings indicate that the embedding-based approach achieves 59% precision across a diverse set of error types, while the fine-tuned LLM, though comparable to pretrained models in failure identification (52.87% vs. 49.32% precision), significantly outperforms in explanation capabilities—achieving 32.61% explanation accuracy compared to 0% for the base model. These results demonstrate the potential for AI-assisted workflow validation to reduce computational costs and accelerate geophysical data processing in production environments.
Read the full article here.