AI-assisted seismic interpretation toolkit demonstrated on the top of the Dutch Maassluis Formation to support implicit geological modeling
Synopsis
Within the Horizon Europe GO-Forward and MOOI WarmingUP GOO projects, the study tested and implemented a machine-learning toolkit for interpreting onshore seismic data across roughly 300–3500 m depth: self-supervised and semi-supervised contrastive-learning CNNs first improve the signal through noise reduction and interpolation, then horizons and faults are interpreted with (semi-)self-supervised methods using minimal human-generated training data, demonstrated by interpreting the top of the Dutch Maassluis Formation in the Leeuwarden and Waalwijk 3D seismic cubes.
The workflow (Figure 2) starts with AI data conditioning, which performs automatic conditioning of seismic data on large amounts of 2D and 3D seismic data by denoising, interpolation and broadbanding. This is achieved by using Self-supervised learning algorithms (“Noise2Void”, Birnie et al., 2021) and a Generative Adversarial Network (“MDA-GAN”, Dou et al., 2022, (a)).
arXiv · Page 3Interpretation
The work builds an AI-assisted seismic interpretation toolkit for implicit modeling, spanning signal improvement through horizon and fault interpretation. Relative to conventional workflows that rely on large amounts of human-interpreted training data, the toolkit brings self-supervised and semi-supervised learning into the interpretation step, aiming to interpret with minimal human training data. Evidence comes from the authors' description of tested and implemented algorithms and toolkit, plus one demonstration interpretation of the top of the Dutch Maassluis Formation; the text is a summary-level account without quantitative accuracy metrics.
The signal-improvement step applies self-supervised and semi-supervised contrastive-learning CNNs for noise reduction and interpolation. Using contrastive-learning CNNs for seismic denoising and interpolation is a methodological shift relative to conventional signal-processing pipelines. Stated as a method description; no network architecture, training scale, or before/after quantitative comparison is reported.
Horizon and fault interpretation uses (semi-)self-supervised methods to minimize the use of human-generated training data. Reducing dependence on manual labeling is the key orientation of the method for efficiency and reproducibility. The text states the strategy and its implementation in the toolkit; the demonstration is limited to one formation top in two 3D seismic cubes.
The demonstration interpreted the top of the Dutch Maassluis Formation in the Leeuwarden and Waalwijk 3D seismic cubes. This is a first demonstration of the toolkit on real onshore seismic data, covering a shallow-to-deep range of roughly 300–3500 m. The demonstration is a single-formation, two-cube case; no comparison statistics or error magnitudes against manual interpretation are provided.
Perspective
The work targets horizon and fault interpretation of onshore seismic data in roughly the 300–3500 m depth range, serving the input-data preparation needed by implicit modeling and relative geologic time methods; its intended users are teams doing geological modeling and seismic interpretation, in applied projects that need rapid characterization of the shallow-to-deep depth domain. The demonstration is limited to interpreting the top of the Dutch Maassluis Formation in the Leeuwarden and Waalwijk 3D seismic cubes, so current conclusions apply to that demonstration setting, and extension to other formations, other basins, or offshore data requires separate validation.
Readers would still want to know: the specific configurations and training-data scale of the self-supervised and semi-supervised contrastive-learning CNNs for denoising and interpolation; the accuracy of horizon and fault interpretation under minimal human training data, and comparisons with manual interpretation or other methods; how the toolkit performs on formations and data cubes beyond the Maassluis Formation top; and the practical effect of the workflow on uncertainty in implicit modeling results. Because the current text is a summary-level account without figures or quantitative results, these questions cannot be judged from the available material and remain open questions for further reading of the original paper.