Domain Generalization in Geological Machine Learning

45 concepts45 papers1 proposal

Machine learning models have achieved remarkable success in localized geoscience applications, from predicting water saturation in specific reservoirs to forecasting foundation settlement in urban construction. However, when these models are deployed outside their initial training sites—to a different basin, a new fault line, or an alternative stratigraphic layer—their predictive performance frequently collapses.

This cross-site generalization gap represents a fundamental challenge in geological modeling. Subsurface environments are shaped by millions of years of distinct tectonic, chemical, and depositional histories, making data from one location highly domain-specific. Achieving true model generalizability requires moving away from localized pattern matching toward frameworks that capture invariant physical properties.

Why Geological Models Fail to Transfer

The failure of machine learning models to generalize across geological domains stems from three primary mechanisms: lithological shifts, the memorization of site-specific noise, and environmental hyper-sensitivity.

Lithological and Stratigraphic Heterogeneity

A model optimized for a specific rock type or sedimentary formation cannot easily adapt to a different mineralogical composition. For example, Van-Thuan Truong demonstrates that predictive models for rock mechanical properties trained on specific lithologies like shale or sandstone lose accuracy when applied to significantly different rock types such as carbonates. This occurs because the underlying physical and chemical relationships change entirely. As Xiao Hu et al. observe, the performance of sequence-based deep learning architectures depends heavily on the specific sedimentary patterns and depositional histories of the training site, limiting their immediate transferability to basins with distinct mineralogical compositions.

Memorizing Local Noise vs. Learning Physical Laws

High-capacity neural networks are notorious for finding shortcuts in data. In geophysics, this manifests as models overfitting to the specific noise profiles, acquisition geometries, and labeling biases of their original survey site rather than capturing invariant physical properties. Haowen Bai et al. identify this exact mechanism as a primary driver of the cross-site generalization gap in seismic data processing. When a system overfits to these localized artifacts, it fails to learn the fundamental physical laws that apply across different tectonic settings. As Yumeng Hu et al. argue, without cross-regional validation, it remains unclear whether deep learning models are identifying generalizable seismic patterns or merely memorizing site-specific noise.

Environmental and Hydrogeological Sensitivity

Geotechnical and hydrological models are often hyper-sensitive to local environmental conditions that vary drastically across geographic regions. Yulong Zhang et al. note that structural failure models are frequently limited by their high sensitivity to local hydrogeological conditions, such as specific soil layers and water-rich pockets. Similarly, Jiahao Deng et al. show that the performance of geotechnical prediction models is tightly bound to specific soil characteristics and local environmental factors, preventing their direct transfer to diverse geological profiles.

The Bottleneck of Constant Recalibration

When domain generalization fails, engineers are forced into a cycle of continuous retraining and localized recalibration. This requirement undermines the core benefit of using machine learning surrogates: near-instantaneous speed and scalability.

For instance, the authors of Velocity continuation with Fourier neural operators for accelerated uncertainty quantification point out that neural surrogate models for seismic imaging are often constrained to the specific acquisition geometry and geological characteristics of the training site, necessitating complete retraining for every new survey area.

This operational bottleneck is widespread across the geosciences. Whether predicting soil water infiltration (Guisheng Fan et al.) or determining liquid accumulation states in gas wells (Qinghui He), variations in physical conditions across geographic blocks mean that models optimized for one field require extensive new training data or transfer learning techniques to maintain accuracy in another.

Pathways to Invariant Generalization

To overcome the cross-site generalization gap, recent research focuses on two main strategies: physics-informed hybridization and cross-regional validation.

Physics-Machine Learning Hybridization

One of the most robust ways to ensure a model generalizes across different geographical conditions or climate scenarios is to anchor its predictions within process-based physical laws. By coupling data-driven algorithms with classical physics equations, models are constrained from making physically impossible predictions in unseen domains.

For example, Jiaona Guo et al. successfully mitigate generalization drops across varying basin conditions by coupling a process-based hydrological model with an XGBoost architecture. In drilling engineering, Mengyuan Xiong et al. combine bidirectional temporal deep learning with classical physics equations, allowing their model to maintain predictive accuracy across highly diverse onshore and deep-water geological formations.

Cross-Regional Validation and Transfer Learning

Instead of training models on isolated, localized datasets, researchers are increasingly utilizing massive, multi-site databases to force networks to learn broader structural properties. Anton Morozov et al. demonstrated this by compiling a database spanning 20 different oilfields to expose their model to wide reservoir heterogeneity and geomechanical variations during training. When direct data sharing is impossible, implementing targeted transfer learning workflows—as suggested by J Yeom for adapting borehole breakout detection across diverse lithologies—provides a pathway to update localized capabilities without starting from scratch.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_geological_generalizability_25130ff2,
  title = {geological generalizability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geological-generalizability/rdr_25130ff26e314b2bbe2ec30438f732b0},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_geological_generalizability_25130ff2_v1,
  title = {geological generalizability},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geological-generalizability/rdr_25130ff26e314b2bbe2ec30438f732b0/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake