Partial Domain Adaptation for Building Borehole Lithology Model Under Weaker Geological Prior

Geochemistry Earth Sciences Information Science Geophysics Automation Artificial Intelligence Geology
lithology identification partial domain adaptation sample transferability weighting convolutional neural network channel attention mechanism target sample reconstruction geological exploration

Report on the Paper: “Partial Domain Adaptation for Building Borehole Lithology Model Under Weaker Geological Prior”

Background and Research Problem

Lithology identification plays a critical role in stratigraphic characterization and reservoir exploration. However, existing AI and machine learning-based lithology identification methods face substantial challenges when dealing with cross-well data. Specifically, significant data distribution discrepancies arise between wells due to complex sedimentary environments, inconsistent geophysical logging equipment, and measurement technologies. Additionally, target wells may include entirely new lithology categories, resulting in a mismatch between the label spaces (unshared label space) of source and target wells. This exacerbates the difficulty for models to predict accurately in the target well.

This study introduces an innovative framework for partial domain adaptation (PDA) to address cross-well lithology prediction under complex geological conditions. The core challenges are: 1. Significant data distribution differences, which make models trained on source well data unsuitable for direct application to target wells. 2. The possible presence of new lithology categories in target wells, with little to no geological prior knowledge.

To this end, the study proposes a multi-module approach aimed at addressing practical issues arising from unknown label spaces in target wells.

About the Paper

This paper was authored by Jing Li, Jichen Wang, Zerui Li, Yu Kang, and Wenjun Lv, all affiliated with the Department of Automation and the Institute of Advanced Technology at the University of Science and Technology of China (USTC). The paper was published in the December 2024 issue of the IEEE Transactions on Artificial Intelligence. The research was supported by grants from the National Natural Science Foundation of China, the National Key Research and Development Program, and several other funding institutions.

Methods and Research Workflow

The paper proposes a framework called “Sample Transferability Weighting based Partial Domain Adaptation” (ST-PDA), which incorporates multiple steps to address the complexities of cross-well lithology predictions.

1. Method Design

The study developed three core modules: 1. Sample Transferability Weighting Module: This calculates the probability that source well samples belong to shared classes with target wells. It assigns lower weights to unshared class samples, mitigating negative transfer effects. 2. CNN with Channel Attention Mechanism for Feature Extraction (CG2CA): This deep learning backbone effectively extracts wide-ranging discriminative features and uses attention mechanisms to enhance key features, emphasizing lithology differentiation. 3. Target Sample Reconstruction Module: This leverages deconvolution techniques to reconstruct target well samples, enhancing their feature representation and facilitating knowledge transfer between source and target wells.

2. Research Workflow

The experimental workflow is as follows: - Distribution Alignment and Feature Extraction: Domain adversarial neural networks (DANNs) are used to align data distributions and extract domain-invariant features. - Sample Transferability Analysis: A domain classifier dynamically evaluates the transferability weights of source domain samples. - Lithology Classification and Target Optimization: Using sample weights, classifiers and feature extraction networks are adjusted iteratively. - Target Sample Reconstruction: Deconvolution is applied to features in the target domain to improve representation and classification performance.

Datasets and Experiments

Dataset Description

The study utilizes data from 16 wells drilled in the Jiyang Depression of Bohai Bay Basin, China. The input features are based on a combination of six geophysical logging curves: acoustic, gamma-ray, resistivity, neutron, spontaneous potential, and caliper curves. These are commonly used to differentiate lithology classes, including Mudstone (MS), Sandstone (SS), Oil Shale (OS), and Dolomite (DM).

Three datasets were constructed to simulate real-world scenarios: 1. Dataset I: The target wells lack any OS labels. 2. Dataset II: The target wells include rare OS labels. 3. Dataset III: The target wells include a small number of OS labels.

Evaluation Metrics

Performance was evaluated using three primary metrics: - Accuracy (ACC): Measures the proportion of correctly classified samples. - Macro-Recall (Macro-R): Highlights performance across all lithology classes, emphasizing the minority classes. - False-OS (F-OS): Represents the number of misclassified samples as belonging to the OS class.

Results and Analysis

Dataset I Results

The ST-PDA framework exhibited exceptional results across various scenarios: 1. Achieved the highest Macro-R among all methods (e.g., 85.52% in some cases), significantly outperforming the second-best method (ETN at 80.80%). 2. Completely eliminated misclassification of OS samples in the target wells (F-OS = 0), effectively managing unshared classes. 3. Achieved superior alignment of features across lithology categories, especially for minority classes like DM, as verified through visualizations.

Dataset II and III Results

For scenarios where OS samples were rare (Dataset II) or limited (Dataset III), ST-PDA continued to excel: 1. In Dataset II (pair: WG→WH), ST-PDA achieved an OS recall of 85.71%, which is 7.14% higher than the second-best method (ETN). 2. In Dataset III (pair: WI→WJ), ST-PDA achieved the highest Macro-R (84.67%) and correctly predicted the most OS samples (72 out of total).

Ablation Studies

Further experiments evaluated the contributions of individual modules. Key insights include: - Replacing CG2CA with a standard CNN reduced Macro-R by 3.93%. - Removing the sample transferability module caused a 6.72% drop in Macro-R. - Without the target sample reconstruction module, the F-OS metric increased significantly to 2 misclassified samples.

Visualization of Features

Using t-SNE, the study visualized extracted features to reveal the effectiveness of ST-PDA: - ST-PDA demonstrated clear separation of source and target domain features, especially for unshared classes like OS. - Compared to other methods, ST-PDA maximized inter-class distances and minimized intra-class overlap.

Contributions and Future Work

Through extensive experimentation and analysis, this study highlighted the scientific and practical contributions of ST-PDA: 1. Scientific Contributions: ST-PDA solves a novel problem in cross-well lithology prediction under unknown target label spaces. 2. Practical Applications: The proposed framework enhances accuracy and reduces dependency on geological experts, providing actionable insights for reservoir exploration.

Future research can focus on improving the class-distinguishability of domain-invariant features, especially for borderline cases where lithology categories are highly similar.

Key Highlights

  1. Introduced a novel framework that integrates partial domain adaptation with transferability weighting, addressing lithology prediction challenges in target wells with unknown label spaces.
  2. Achieved superior performance in predicting minority lithologies like OS and DM while maintaining class balance.
  3. Established new benchmarks in cross-well lithology prediction, surpassing state-of-the-art methods in multiple metrics.