@article{fdi:010094157,
  title = {{AI}-based geological subsurface reconstruction using sparse convolutional autoencoders},
  author = {{U}ribe-{V}entura, {R}. and {B}arriga-{B}errios, {Y}. and {B}arriga-{G}amarra, {J}. and {B}aby, {P}atrice and {V}iveen, {W}.},
  editor = {},
  language = {{ENG}},
  abstract = {{S}ubsurface reconstruction is critical for geological modeling and resource exploration. {C}onventional spatial interpolation methods are limited by stationarity and spatial isotropy assumptions, while advanced geostatistical techniques require specialized datasets. {D}eep learning approaches often need large datasets, which is impractical for geoscientific applications. {T}his study presents an {AI}-based methodology using a sparse convolutional autoencoder for robust subsurface modeling under data constraints and integrating secondary data sources such as {V}ertical {E}lectrical {S}ounding ({VES}) data. {A} four-stage testing framework was implemented: (1) emulating conventional interpolation for baseline performance; (2) reconstructing subsurface geometries from synthetic data; (3) incorporating geophysical constraints through {VES} forward modeling; and (4) validating the methodology using a real-world case study from the {H}uancayo tectonic basin in the {P}eruvian {A}ndes, using 41 {VES} measurements across two cross-sections (12 and 14 km long). {R}esults demonstrate that the proposed model effectively emulates kriging interpolation (mean squared error: 1.5 x 10-3 to 1.2 x 10-3 with 100-800 training examples) through transfer learning from an inverse-distance, pre-trained model. {I}n subsurface reconstruction, the model outperforms kriging (37.4-61.7 % improvement across 1-15 % sampling densities) through its ability to adapt to non-stationary conditions. {W}hen incorporating synthetic {VES} data, the model effectively reconstructed subsurface geometries with error reduction from 4.1 x 10-1 to 9.1 x 10-3 as stations increased from 1 to 40, demonstrating diminishing returns beyond this point. {A}pplication to the {H}uancayo basin case study validated the model's practical applicability by successfully identifying previously unmapped features including the contact between basement and sedimentary infill, folds and faults. {T}he methodology demonstrates the {AI}'s capability to enhance geological understanding in complex tectonic settings, revealing subtle features and refining existing assumptions about subsurface architecture.},
  keywords = {{D}eep learning ; {E}lectrical resistivity ; {G}eophysics ; {A}ndes ; {H}uancayo basin ; {B}asin structure ; {PEROU} ; {ANDES}},
  booktitle = {},
  journal = {{C}omputers and {G}eosciences},
  volume = {204},
  numero = {},
  pages = {105981 [17 p.]},
  ISSN = {0098-3004},
  year = {2025},
  DOI = {10.1016/j.cageo.2025.105981},
  URL = {https://www.documentation.ird.fr/hor/fdi:010094157},
}