Scholarlink

<p>Predicting the shear strength of fault gouge is essential for assessing fault stability. In geomechanics, datasets are typically small and tables compiled from the literature often contain substantial missing features, which constrains conventional modeling. To address these challenges, this study presents a three stage modeling strategy tailored to small-sample settings with extensive missing data. First, five baseline algorithms are benchmarked on a complete feature subset and the multilayer perceptron is identified as the initial best model. Second, a deep tabular model that handles missing inputs, TabNet, is used to exploit additional fault gouge samples with incomplete features. Third, a pretrained tabular foundation model, TabPFN, is adopted and the data from the first two stages are combined with auxiliary loess and clay samples to strengthen learning under limited data. TabPFN attains <i>R</i><sup>2</sup> = 0.88 on the full dataset and <i>R</i><sup>2</sup> = 0.68 on the test set, representing a 19.3% improvement over the multilayer perceptron baseline; inference requires 0.15&#xa0;s and is 4.8 times faster. Model behavior is examined with SHAP and individual conditional response analyses, which verify that the model implicitly learns physically consistent relations such as the Mohr–Coulomb model. The resulting pipeline combines missing aware modeling, pretrained tabular inference, and physics-based checks, and provides a compact and transferable solution for geotechnical prediction when data are limited.</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

Predicting Fault Gouge Shear Strength Under Small-Sample and Missing Feature Conditions: A Three-Stage Framework with Pretrained Tabular Inference

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Weimin Wang,
  • Hejuan Liu,
  • Xilin Shi,
  • Yunhe Su,
  • Haizeng Pan,
  • Shengnan Ban,
  • Hongwei Wang

摘要

Predicting the shear strength of fault gouge is essential for assessing fault stability. In geomechanics, datasets are typically small and tables compiled from the literature often contain substantial missing features, which constrains conventional modeling. To address these challenges, this study presents a three stage modeling strategy tailored to small-sample settings with extensive missing data. First, five baseline algorithms are benchmarked on a complete feature subset and the multilayer perceptron is identified as the initial best model. Second, a deep tabular model that handles missing inputs, TabNet, is used to exploit additional fault gouge samples with incomplete features. Third, a pretrained tabular foundation model, TabPFN, is adopted and the data from the first two stages are combined with auxiliary loess and clay samples to strengthen learning under limited data. TabPFN attains R2 = 0.88 on the full dataset and R2 = 0.68 on the test set, representing a 19.3% improvement over the multilayer perceptron baseline; inference requires 0.15 s and is 4.8 times faster. Model behavior is examined with SHAP and individual conditional response analyses, which verify that the model implicitly learns physically consistent relations such as the Mohr–Coulomb model. The resulting pipeline combines missing aware modeling, pretrained tabular inference, and physics-based checks, and provides a compact and transferable solution for geotechnical prediction when data are limited.