Paper Summary
We present a 60M-parameter well log vision transformer foundation model trained using a masked autoencoding framework (ViT-MAE). This model was pretrained on 1.1 million North American well logs for automated well log imputation and subsequently extended to execute prompt-based geologist-guided formation top interpretation as a downstream task. The formation model was fine-tuned on 271,972 human-interpreted formation tops from 44,062 wells across the Permian Basin, covering 37 formations. Unlike classification-based approaches, this prompt-based method for predicting formation depth facilitates training on a vast collection of existing interpreted wells. The foundation model and geologist-guided interpretation method are crucial for accelerating prospect generation by integrating AI-driven geological analysis into exploration workflows.