Scholarlink

<p>Proper reservoir characterization, including hydrocarbon detection, is the bedrock for efficient subsurface exploration and production. This study presents a robust deep learning framework leveraging artificial neural networks (ANNs) and convolutional neural networks (CNNs) to predict neutron porosity (NEUT) from multivariate well logs in the Kaimiro Field, Taranaki Basin, New Zealand. Our methodology comprises exacting data processing, stratified partitioning, and feature scaling via z-score normalization, applied across feedforward ANN and CNN architectures scaling from two to five hidden layers. Through rigorous RandomizedSearchCV hyperparameter optimization, we empirically demonstrate that a four-hidden-layer architecture optimally calibrates expressive capacity with generalization. The optimized 4-layer ANN achieved superior predictive fidelity on the test set, yielding an R2 of 0.956, Root Mean Square Error (RMSE) of 0.0095, Mean Absolute Error (MAE) of 0.0068, and a Mean Absolute Percentage Error (MAPE) of 2.32%. Correspondingly, the 4-layer CNN achieved a test R2 of 0.783. Furthermore, an error analysis evaluating Percent Bias (PBIAS) yielded a nominal value of − 0.1846%, confirming the absence of systemic overestimation or underestimation in the optimal models and underscoring their exceptional stability. To ensure model transparency, SHapley Additive exPlanations (SHAP) were deployed, elucidating that predictions are physically grounded and predominantly driven by deep resistivity, density, and depth features. This replicable paradigm successfully harmonizes predictive accuracy with explainability, paving the way for trustworthy AI integration in digital petrophysics and actionable subsurface decision-making.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Explainable deep learning analysis of network depth effects on neutron porosity prediction

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月16日

查看全文
  • Minhaz Chowdhury,
  • Md Shofiqul Islam

摘要

Proper reservoir characterization, including hydrocarbon detection, is the bedrock for efficient subsurface exploration and production. This study presents a robust deep learning framework leveraging artificial neural networks (ANNs) and convolutional neural networks (CNNs) to predict neutron porosity (NEUT) from multivariate well logs in the Kaimiro Field, Taranaki Basin, New Zealand. Our methodology comprises exacting data processing, stratified partitioning, and feature scaling via z-score normalization, applied across feedforward ANN and CNN architectures scaling from two to five hidden layers. Through rigorous RandomizedSearchCV hyperparameter optimization, we empirically demonstrate that a four-hidden-layer architecture optimally calibrates expressive capacity with generalization. The optimized 4-layer ANN achieved superior predictive fidelity on the test set, yielding an R2 of 0.956, Root Mean Square Error (RMSE) of 0.0095, Mean Absolute Error (MAE) of 0.0068, and a Mean Absolute Percentage Error (MAPE) of 2.32%. Correspondingly, the 4-layer CNN achieved a test R2 of 0.783. Furthermore, an error analysis evaluating Percent Bias (PBIAS) yielded a nominal value of − 0.1846%, confirming the absence of systemic overestimation or underestimation in the optimal models and underscoring their exceptional stability. To ensure model transparency, SHapley Additive exPlanations (SHAP) were deployed, elucidating that predictions are physically grounded and predominantly driven by deep resistivity, density, and depth features. This replicable paradigm successfully harmonizes predictive accuracy with explainability, paving the way for trustworthy AI integration in digital petrophysics and actionable subsurface decision-making.