PaperlayerView exact revision on arXiv

Source-linked AI summary

An AI-assisted conditioning and geological interpretation workflow for usage in implicit geological modeling

Stefan Carpentier, Jan Diederik van Wees, Eva de Boever, Jan Niederau, Camille Chapeland, Suzanne Atkins, Boris Boullenger, Jens Wollenweber

arXiv:2610.09871v1physics.geo-phcs.AIcs.LG

TL;DR

Geothermal and ATES projects face shared seismic-data challenges, and Dutch ATES remains at pilot scale. This study integrates AI-based seismic conditioning with semi-supervised horizon interpretation and fault detection, reporting more efficient, spatially consistent, higher-resolution interpretations and supporting their use in applied geological modeling.

  • Problem

    Geothermal and ATES projects face shared seismic-data challenges; Dutch ATES remains at pilot scale and requires further experience and knowledge sharing.

    Source: Introduction · p. 2

  • Method

    The workflow integrates AI-based seismic conditioning with semi-supervised horizon interpretation and fault detection.

    Source: Introduction · p. 3, Discussion · p. 13

  • Results

    The integrated workflow substantially improves interpretation efficiency, spatial consistency, and resolution; semi-supervised predictions use only a small fraction of manually interpreted inlines.

    Source: Discussion · p. 13, Conclusions · p. 17

  • Takeaways & Limitations

    The workflow supports re-use of existing seismic data and integration of consistent 3D interpretations into implicit geological modeling and decision-making.

    Source: Conclusions · p. 17

    Field-data training and fine-tuning are more variable than synthetic-data training, and predicted units and faults still require validation by experienced interpreters.

    Source: Discussion · p. 15

Abstract

from arXiv · show

Implicit modeling and Relative Geologic Time are geological modeling techniques that enable more efficient, faster, less biased and more reproducible modeling results. For optimal operation, these techniques require many well-constrained input data. In the framework of the Horizon Europe GO-Forward and MOOI WarmingUP GOO projects and to accelerate Implicit modeling, Machine Learning (ML) methods have been tested and implemented in a toolkit for the interpretation of (onshore) seismic data from the shallow to deep range (+- 300 - 3500 m). The goal is to rapidly characterise this depth domain by efficient interpretation of horizons and faults in seismic data. The first step is to improve the signal by applying AI techniques like self-supervised and semi-supervised contrastive learning CNN's for noise reduction and interpolation. Next, horizons and faults are interpreted with minimal use of human-generated training data by using (semi-) self-supervised methods. The resulting developed toolkit supports the application of the implemented algorithms in an efficient workflow. As a first demonstration, the top of the Dutch Maassluis Formation has been interpreted in the Leeuwarden and Waalwijk 3D seismic cubes. Overall, this study demonstrates that AI-assisted interpretation workflows have reached a level of maturity that allows their integration into applied geological modeling and decision-making.

Introduction

The introduction situates the workflow in the move toward implicit modeling and describes data and interpretation challenges for geothermal and ATES applications. It presents an AI-assisted tool for conditioning seismic data and interpreting horizons and faults.

Introduction

  • Implicit modeling represents relative geological time as a 3D scalar potential field, with interpreted horizons as equipotential surfaces and discontinuities permitted at faults and unconformities.The RGT field preserves 3D genetic context and supports sequence- and chronostratigraphic modeling.Source: Introduction · p. 2
  • Geothermal and ATES projects face seismic-data challenges; shallow-target data are often scarce, and ATES in the Netherlands remains at pilot scale.The introduction also notes the need to gain and share experience and knowledge as ATES scales up.Source: Introduction · p. 2
  • Manual picking and tracking interpret only a limited portion of 3D surveys, motivating ML-based horizon and fault interpretation across the full survey.Existing ML predictions can be faster but may remain inconsistent or simplified relative to refined human interpretation.Source: Introduction · p. 2
  • The workflow uses CNNs, specifically U-nets, and neural networks learn mappings from inputs to targets by minimizing output error.The stated applications include classifying lithology, facies, horizons, and faults.Source: Introduction · p. 3, Figure 1 · p. 3
  • The ready-to-use tool conditions seismic data through denoising, interpolation, and broadbanding, then uses semi-supervised methods to interpret horizons and detect faults.The workflow names Noise2Void and MDA-GAN for conditioning, CONSS for horizon interpretation, and FaultNet and FaultSSL for fault detection.Source: Introduction · p. 3, Figure 2 · p. 4
  • The paper applies mostly AI-based automated workflows to existing 2D and 3D seismic data to improve subsurface characterization for ATES and geothermal energy.The stated intended improvement includes characterization of layers and fault density.Source: Introduction · p. 4

Method

The workflow conditions seismic data with self-supervised denoising and GAN interpolation, then applies semi-supervised methods to interpret horizons and faults with limited labeled data. Pilot results support its use for scalable interpretation, while field labels and processing choices shape its application.

Method

  • In the pilot 3D cube, conditioning increased reflector coherence and clarified Maassluis and Oosterhout reflections for later interpretation.The workflow combines self-supervised blind-spot denoising with generative-adversarial interpolation to enhance continuity while preserving reflector character.Source: Method · pp. 4–5
  • Self-supervised denoising and GAN interpolation suppress random noise, reconstruct acquisition gaps, and preserve reflector character.The blind-spot network estimates a central sample from surrounding seismic noise; the GAN learns to fill gaps with patterns matching surrounding traces.Source: Method · pp. 4–5
  • CONSS classifies horizon overburden and underburden using representative conditioned lines, requiring only 1% of cube inlines or crosslines.In a pilot cube, Maassluis and Oosterhout horizons were predicted across all inlines and crosslines.Source: Method · pp. 5–6

2.3 AI seismic fault detection

  • FaultNet is described as a U-net autoencoder, and fault-detection training includes special weighted labels for faults parallel to inlines and crosslines.FaultSSL and FaultNet are applied through semi-supervised segmentation for fault detection in seismic data.Source: Figure 6 · p. 6, Method · p. 7, §2.3 · p. 8
  • FaultSSL is a semi-supervised fault-detection method using a teacher-student structure, with supervised training on synthetic data and 2D labels and unsupervised proxy tasks to incorporate unlabeled field seismic data.TNO-GDN resampled volumes to 12.5 m × 12.5 m CDP spacing and 4 ms sampling, used 256 × 256 × 256-sample tiles with half-tile diagonal blending, applied time-direction mean and inline/crossline median filters, and thresholded probabilities at 0.5.Source: §2.3 · pp. 7–8

3. Results

The workflow produced plausible 3D interpretations of horizons and faults, including in the shallow Maassluis and Oosterhout formations. Conditioning improved seismic continuity, while human correction addressed mismatches between inherited horizon labels and local reflections.

3. Results

  • In Leeuwarden and Waalwijk, AI denoising suppressed noise and interpolation reconstructed gaps while preserving signal character and improving reflector continuity.The Leeuwarden denoised example reports minimal signal leakage and maximum amplitude and phase preservation; the Waalwijk interpolation restores shallow gaps using surrounding 3D context.Source: §3 · p. 8, Figure 9 · p. 9, Figure 10 · p. 9
  • Leeuwarden horizon predictions used interpretations from 11 inlines, but borehole-based DGM Shallow horizons sometimes missed local seismic reflections, prompting a human-in-the-loop correction tool.The workflow's facies figure reports that only 1% labeled facies were needed for training.Source: §3 · p. 10, Figure 15 · p. 13
  • After snap-to-peak corrections were retrained, Top Maassluis followed its reflector, an Elsterian Glacial Valley was identified, and Waalwijk predictions captured clinoforms and an unseen fault.The authors describe the resulting interpretation as capable of delineating major stratigraphic units.Source: §3 · p. 12
  • In Waalwijk, the pretrained fault workflow produced a consistent, plausible 3D interpretation, detecting faults in shallow Maassluis and Oosterhout strata to about 200 ms.The detected faults matched seismic reflections.Source: §3 · p. 12
  • Across several datasets, the workflow is described as mature, with plausible and usable 3D interpretations even in shallow Maassluis and Oosterhout formations.The authors report variable field performance because input labels may be inconsistent, incorrect, or scale-variant.Source: §3 · p. 13

Discussion

The workflow combines AI-based seismic conditioning with semi-supervised horizon and fault interpretation, improving interpretation efficiency and spatial consistency in shallow to mid-depth onshore settings. Its results also show that geological relevance of training labels and expert validation remain important.

Discussion

  • The integrated workflow substantially improves interpretation efficiency, spatial consistency, and resolution in shallow to mid-depth onshore settings relevant to geothermal energy and ATES.It combines AI-based conditioning, semi-supervised horizon interpretation, and semi-supervised fault detection.Source: Discussion · p. 13
  • Leeuwarden facies classification used 1% labeled facies and recognized the Elsterian Valley infill, reinterpreted as Top Maassluis.The figure identifies the meandering Maassluis infill within the Oosterhout Formation.Source: Figure 16 · p. 14
  • Self-supervised denoising and GAN-based interpolation improve reflector continuity and coherence in vintage land seismic, clarifying Maassluis and Oosterhout reflections, especially in noisy shallow data.Input-specific training-window and post-processing parameters are still required.Source: Discussion · p. 15
  • Semi-supervised horizon interpretation can reduce human-labeled training needs, but predictions may diverge from local seismic signatures when labels do not honor reflectivity; human correction remains essential.Regional models based primarily on borehole data are one stated example of this mismatch.Source: Discussion · p. 15
  • FaultSSL shows convincing 3D fault continuity and realistic geometries with limited labeled data, detecting faults into the shallow Maassluis and Oosterhout formations.Its lower apparent sensitivity to noise and processing artefacts is conditional on sufficient conditioning and representative training data.Source: Discussion · p. 15
  • Field-data training and fine-tuning show greater variability than synthetic-data training, and predicted units and faults still require validation by experienced interpreters.The passage attributes field-data challenges to label consistency, scale dependency, and geological ambiguity.Source: Discussion · p. 15
  • For geothermal and ATES applications, the workflow offers scalable, reproducible characterization in data-limited settings; continued work should improve robustness, geological constraints, and implicit-model integration.The application discussion links shallow stratigraphic and fault-architecture uncertainty to project feasibility and public acceptance.Source: Discussion · p. 15

Conclusions

The study combines AI-based seismic conditioning with semi-supervised horizon interpretation and fault detection to interpret vintage 2D and 3D seismic data. It reports improved signal interpretability and consistent 3D predictions, while retaining expert correction and oversight.

Conclusions

  • Self-supervised denoising and GAN-based interpolation improve land-seismic quality, with enhanced reflector continuity and coherence supporting more reliable subsequent horizon and fault interpretation.The improved visibility of Maassluis and Oosterhout reflections supports consistent three-dimensional mapping.Source: Conclusions · pp. 16–17
  • Semi-supervised horizon and fault interpretation reduces dependence on extensive human-labeled datasets, achieving consistent 3D predictions from a small fraction of manually interpreted inlines.Label geological relevance and consistency remain important, particularly when regional models are not fully constrained by seismic reflectivity.Source: Conclusions · p. 17
  • Human-in-the-loop correction remains essential for geological plausibility and adapting predictions to local seismic and stratigraphic conditions; AI accelerates interpretation rather than replacing experts.Source: Conclusions · p. 17
  • The workflow is presented as mature for applied geothermal and ATES projects, enabling reuse of existing seismic data and supporting integration with implicit geological modeling and decision-making.The authors position it for subsurface characterization in data-limited settings.Source: Conclusions · p. 17
Loading 2610.09871v1…