DEEP STRATIGRAPHIC INFERENCE: A TWO-STAGE TRAINING CURRICULUM AND HEURISTIC GATE FOR HIGH-PRECISION CHANGE-POINT DETECTION

Shwetha Salimath, Francesca Bugiotti, Sylvain Wlodarczyk, Sohaib Ouzineb

Keywords: Time Series Transformers, Change point detection, Rotary Positional Embedding, Hybrid Models, Carbon Sequestration, Curriculum Learning

TL;DR: End-to-end training for time series localization is unstable. We solve this with a two-stage curriculum and a simple heuristic gate, achieving SOTA precision in geological mapping from well-log data.

Abstract:

Accurate geological characterization of subsurface reservoirs using well log data is essential for high-impact applications such as carbon sequestration and environmental monitoring. This task, which we term Deep Stratigraphic Inference, requires the high-precision localization of change-points within noisy time series. While transformers are powerful, a naive end-to-end regression approach fails due to training instabilities. To address this, we propose CURT-Point (Curriculum-trained Regression Transformer for Point Localization), a comprehensive framework for robust time series localization. CURT-Point's core is a Two-Stage Training Curriculum that first pre-trains the transformer as an expert classifier, then fine-tunes a specialized regression head. To maximize robustness, the framework is completed by a post-processing Hybrid System incorporating a Heuristic Gate, which achieves the best overall performance by intelligently ensembling an attention-based regression with a robust peak-finding heuristic, both derived from the same unified Transformer backbone. The effectiveness of this framework hinges significantly on two additional advancements: we show that a fusion of specific data preprocessing with an innovative constrained data augmentation tactic is crucial for dealing with real-world signal flaws, and we establish that Rotary Positional Embeddings (RoPE) play a crucial role in attaining high performance.
Our final Hybrid System, validated on three real-world well-log datasets of increasing complexity, achieves state-of-the-art recall and median errors, providing a generalizable workflow for high-precision time series localization.

Primary Area: learning on time series and dynamical systems

Submission Number: 6360