LithoFormer: Transformer-Based Stratigraphic Inference from Well Logs
A new framework called LithoFormer uses a Seq2Seq transformer model to perform stratigraphic inference from multivariate well log data in a single pass. Unlike existing sliding-window classification methods, it captures long-range geological dependencies by employing a channel-independent PatchTST backbone enhanced with rotary positional embeddings (RoPE). A decoupled multi-task head jointly predicts geological zonation and boundary probabilities. The approach aims to improve accuracy for subsurface characterization in carbon capture and storage (CCS), geothermal development, and resource extraction. The paper is available on arXiv under ID 2607.22804.
Key facts
- LithoFormer uses a Seq2Seq transformer model for stratigraphic inference.
- It processes entire multivariate well logs in a single pass.
- The framework employs a channel-independent PatchTST backbone with RoPE.
- A decoupled multi-task head predicts zonation and boundary probabilities.
- It addresses limitations of sliding-window classification methods.
- Applications include CCS, geothermal development, and resource extraction.
- The paper is published on arXiv with ID 2607.22804.
- The method captures long-range geological dependencies.
Entities
Institutions
- arXiv