Multi-Stage Lithological-Petrophysical Interpretation Using Machine Learning

Abstract We present an integrated, multi-stage ML workflow for automated lithological-petrophysical interpretation of well-log data. Rather than isolated models, we couple lithology classification with petrophysical regression via physically grounded intermediate targets. This design enforces consistency across electrical, acoustic, and radioactive logs, yielding cleaner formation boundaries, depth-consistent classes, and improved robustness to noise. The approach explicitly encodes geophysical constraints and leverages feature engineering and cross-well validation to enhance transferability. Automatically depth-matched logs feed a three-stage cascade: (1) preliminary lithology; (2) petrophysical regression (porosity, permeability, hydrocarbon saturation) conditioned on (1); (3) refined lithology that reconciles the intermediate physical parameters. Residual 1D CNNs with dilations learn local textures; a Transformer with positional encoding captures long-range dependencies over 500-point windows. We train a chain of classifiers/regressors with multi-task heads and auxiliary losses using normalized features, presence masks, and upweighted key channels; optionally, an additional network generates synthetic fragments to span gaps. Post-processing merges overly thin intervals and aligns boundaries, regularization and validation monitoring control overfitting. Across multiple wells and fields (within the basin), the cascade produced cleaner, geologically readable profiles: sharper, more stable formation boundaries; improved depthwise consistency across the hierarchy (lithotype – reservoir quality type – fluid saturation type); and reduced sensitivity to missing or noisy channels. Conditioning petrophysical regression on preliminary lithology, then refining lithology using those intermediate physical parameters, limited error propagation and increased robustness to gaps via presence masks and channel weighting, optional synthetic fragments further stabilized predictions. The hierarchical 1D CNN+Transformer consistently outperformed single-stage multiclass baselines in boundary coherence and class stability. Quality was assessed against reference petrophysical formulas/classification schemes and cross-checked with core analyses and expert interpretations, regularization and metric monitoring mitigated overfitting. The approach depends on sufficient data volume and quality at each stage. Novelty lies in a physically guided cascade implemented with a hierarchical CNN+Transformer and embeddings of prior decisions. An optional pre-cascade step uses an additional network to synthesize log fragments from existing measurements to mitigate gaps. Coupled with depth matching, presence masks, channel weighting, and geology-informed post-processing, this design advances deployable subsurface ML beyond single-stage pipelines.