Rethinking Fault Interpretation: From Manual Picking to Self-Improving AI Workflows What if fault interpretation wasn’t a linear process—but a converging system that improves wi...
Rethinking Fault Interpretation: From Manual Picking to Self-Improving AI Workflows
What if fault interpretation wasn’t a linear process—but a converging system that improves with every iteration?
As part of the URGENT project, dGB Earth Sciences is developing a new paradigm for seismic fault interpretation in OpendTect—one that tightly integrates machine learning, geometric extraction, and interpreter feedback into a continuous improvement loop.
At its core, this workflow challenges the traditional divide between attribute generation, interpretation, and modeling: 🔹A pre-trained ML model (2D/3D) produces a Fault Likelihood volume, encoding spatial fault probabilities 🔹Interpreters interact directly with this volume, placing fault sticks where structural evidence is strongest 🔹These sparse constraints are expanded into continuous fault planes using a tracking algorithm guided by the likelihood field and geometric continuity 🔹Crucially, interpreters remain in the loop—validating, correcting, and enforcing geological realism
But the real shift happens after interpretation. Instead of being an endpoint, interpreted fault planes become training data: Curated, QC’ed interpretations are fed back into the system for transfer learning The ML model is refined to better capture survey-specific fault expressions Re-application produces sharper, more reliable Fault Likelihood volumes, reducing ambiguity and interpretation effort
This creates a closed-loop system where: 🔹Human expertise constrains the model 🔹The model accelerates and standardizes interpretation 🔹Each iteration improves both
The result is not just faster interpretation—but a system that adapts to geology, data quality, and interpreter intent. In structurally complex, noise-prone urban environments—where geothermal success depends on accurate fault characterization—this shift is more than incremental. It’s foundational.
The images show the workflow in action: from likelihood volumes and fault sticks to tracked fault planes and iterative refinement.
Special thanks to Mohammad Fazelzadeh for generating the visuals
dGB Earth Sciences #OpendTect #geoscience #geophysics #AI #seismic #ML
