A new data-efficient, deep learning-based methodology for geological subsurface reconstructions

Deep learning approaches for geological subsurface reconstruction typically require extensive training datasets, limiting their practical application in geosciences where data acquisition is costly and sparse. We present a methodology using sparse convolutional autoencoders that effectively learns from synthetically generated training data while maintaining strong generalization to real-world scenarios. Our model is trained exclusively on synthetic basin boundary configurations and corresponding forward-modeled Vertical Electrical Sounding (VES) responses, thereby eliminating reliance on extensive real-world training datasets. Through transfer learning, the model achieves high reconstruction accuracy with as few as 1000 synthetic training examples. Systematic tests reveal the model preserves strong performance beyond its training distribution, suggesting it learns robust heuristic approximations and remains effective beyond the training range of 3–50 input points.

The trained model was applied to the Huancayo tectonic basin in the Peruvian Andes. There, the 300 to 350-m deep subsurface geometry of the tectonic basin was sucessfully modeled on basis of data input from 41, newly acquired VES logs along two cross sections of 12- and 14-km long.  Surprisingly, the reconstruction also revealed previously unidentified fold and thrust systems, for which the model was not explicitely trained, while also maintaining physical consistency with field measurements.

Our results demonstrate that sparse convolutional autoencoders, when trained on synthetic datasets, can effectively bridge the gap between data-hungry deep learning methods and data-sparse geological applications.