Prediction of Lithology Type Using Artificial Neural Networks

Accurate estimation of subsurface lithology is critical for the implementation of shallow geothermal energy systems. In Cyprus, an eastern Mediterranean island, where geological formations vary significantly across short distances, predicting lithological profiles from limited borehole data presents a major challenge. The study presented here investigates the potential of Artificial Neural Networks (ANNs) to predict lithology categories using spatial and geological attributes such as depth, fault distance, and formation type. Three ANN architectures were tested: a standard feed-forward neural network, an optimized fully connected deep learning model, and a feature-engineered model incorporating regional division and new geological parameters. The obtained results indicate that the feed-forward ANN achieved a correlation coefficient (R) of 64%, while the optimized deep learning model reached 65%. After feature engineering and data filtering, prediction accuracy improved to 74%, demonstrating the strong potential of ANN-based models for lithology classification and subsurface profiling in geothermal energy applications. For amelioration, future work can then focus on Convolutional Neural Networks (CNNs) and hybrid models.