Abstract
<title>Abstract</title> <p>Horizontal wellbore geosteering requires continuous estimation of subsurface position, litho-logic confidence, and steering risk from incomplete, noisy, and heterogeneous measurements. These measurements commonly include measured depth, trajectory coordinates, gamma ray logs, interpreted formation surfaces, and typewell references. This paper presents a learning-driven autonomy pipeline for horizontal wellbore geology prediction using a real horizontal-well dataset and an associated typewell reference. The horizontal-well dataset contains 5,278 measured-depth stations, while the typewell dataset contains 1,296 samples. The prediction-start point is inferred from the first missing TVT input value and occurs at row 1,442, corresponding to MD = 12,909 ft. The proposed pipeline integrates geometry-based true-vertical-thickness (TVT) prediction, tree-ensemble baselines, typewell gamma ray interpolation, GR-score lithologic calibration, mul-timodal Random Fourier Feature kernel mean embeddings (RFF-KME), conditional variational autoencoder (CVAE) forecasting, causal block-sparse attention, local uncertainty-aware fusion, facies-state diagnostics, and steering logic. The empirical results indicate that ridge regression is the strongest pure numerical TVT predictor for the studied well, suggesting that TVT is primarily controlled by trajectory geometry and formation-surface relationships. However, ridge regression alone does not provide lithologic confidence, calibrated uncertainty, facies-shift detection, or steering guidance. Therefore, the recommended architecture uses ridge regression as a deterministic geometric backbone and trains a residual sparse-attention CVAE to learn probabilistic corrections around the ridge baseline. The resulting framework combines geometry-driven TVT prediction, lithologic agreement through GR-score, probabilistic residual forecasting, uncertainty-aware fusion, and conservative steering. The proposed approach provides a practical foundation for autonomous geosteering systems that require not only accurate TVT estimates but also interpretable uncertainty and decision support.</p>