DATS: A Depth-Aware Transformer Hybrid Framework for Geological Parameter Prediction

Accurate prediction of reservoir permeability constitutes a central task in hydrocarbon field development, but it remains challenging because core measurements are sparse and heterogeneous reservoirs commonly exhibit nonunique logging responses. Conventional machine learning methods typically treat each core measurement as an independent sample for end-to-end permeability prediction, thereby overlooking the stratigraphic context embedded in depth-ordered core sequences. In this study, we propose a depth-aware Transformer-based hybrid modeling framework, termed DATS, for geological parameter prediction from sparsely sampled core data. DATS constructs sequential samples using sliding windows of core measurements and incorporates depth information and stratigraphic priors through three key components: a continuous Fourier-based depth encoding with learnable frequency parameters, an attention mechanism incorporating depth-distance bias penalties, and a two-stage training strategy that bridges continuous well-logging data and discrete core measurements. Applied as a feature extractor to carbonate reservoirs in the study area, DATS was evaluated by using 2856 core samples. Through comparison among three feature combination strategies and 14 downstream models, the best-performing model, DATS-XGBoost_Combined, achieved an R ² of 0.9404. Ablation experiments indicate that each constituent module of DATS contributes positively to the overall model performance. SHAP values and attention matrix analyses further suggest that the performance improvement is associated with the ability of DATS to capture depth-dependent stratigraphic patterns consistent with petrophysical principles. These results suggest that DATS can serve as a supplementary tool for permeability prediction in sparsely cored reservoirs and assist parameter evaluation in uncored or undersampled wells.