LithoFormer AI Improves Geological Characterization from Well Log Data.

Accurate geological characterization of subsurface reservoirs is crucial for projects like carbon capture, geothermal energy, and natural resource extraction. Current automated methods often use a limited "sliding-window" approach, which fails to capture the broader geological context, leading to misaligned formation layers and inconsistencies. Researchers have developed LithoFormer, a novel framework that utilizes a Seq2Seq transformer model to overcome these limitations. LithoFormer processes entire multivariate well logs in a single pass, allowing it to understand long-range geological dependencies. It incorporates a channel-independent PatchTST backbone with rotary positional embeddings (RoPE) and a decoupled multi-task head to jointly predict geological zonation and precise boundary probabilities. A key innovation is a geology-informed loss function that enforces physical constraints, such as the Law of Superposition, preventing illogical stratigraphic order violations. Validated on real-world datasets, LithoFormer achieved a 90% reduction in median boundary error and eliminated all stratigraphic order violations. It also reduced manual expert labor by 80%, offering a scalable and reliable solution for large-scale subsurface modeling.