Deep Learning-Based Reconstruction and Representation Learning of Open-Hole Well Logs Using Machine Learning

Well-log measurements provide essential information for lithological identification, petrophysical characterization, and reservoir interpretation but incomplete log intervals may occur because of technical, operational, or economic constraints. This study investigates machine-learning-based reconstruction of missing open-hole well-log curves using a depth-indexed dataset from a single well in the Bărbuncești oil and gas field, Romania. Seven predictors (DEPTH, AC, GR, RFOC, RILD, RILM, and SP) were used to estimate ten target properties (CALI, CGR, NPHI, NPOR, PEF, POTA, RHOB, SGR, THOR, and URAN) using Random Forest, Extra Trees, and Artificial Neural Networks. Approximately 2350 valid observations were used for model development and evaluation. Robustness was first assessed using 20 repeated 70/15/15 training/validation/testing partitions and an additional 80/10/10 sensitivity analysis. Under repeated 70/15/15 random partitioning, mean R2 values ranged from 0.809 for URAN to 0.935 for RHOB, with relatively small standard deviations, while the 80/10/10 configuration produced only limited changes in performance. However, a nested 10-fold depth-blocked cross-validation procedure, in which model-family selection was performed exclusively on training/validation data and each outer-test block remained completely untouched during model development, yielded negative mean outer-test R2 values for all ten target properties when DEPTH was included as a predictor. This marked discrepancy demonstrates that random hold-out validation can substantially overestimate predictive performance in densely sampled, depth-indexed, single-well data and that strong random-split performance should not be interpreted as evidence of generalization to unseen contiguous depth intervals. The reconstructed missing intervals are therefore presented as model-based estimates showing qualitative geological and petrophysical consistency rather than as quantitatively validated reconstructions, because measured target values are unavailable within those intervals. These findings highlight the importance of depth-structured validation for assessing machine-learning-based well-log reconstruction and the need for independent validation using additional wells and geological settings.