Scholarlink

<p>The sustainable development of nuclear energy requires efficient and reusable materials for uranium extraction from seawater, where uranyl species exist at ultra-trace levels within complex coordination environments. To overcome the limitations of empirical, trial-and-error design, this study establishes U-Predict v1.0, a structure-performance database for uranium adsorbents, and introduces an interpretable machine learning framework for adsorption performance prediction and mechanistic analysis. Trained on over 220 samples and 54 structural descriptors, the Light Gradient Boosting Machine (LightGBM) model with engineered features and optimized hyperparameters achieved a training <i>R</i><sup>2</sup> of 0.9887 and a test <i>R</i><sup>2</sup> of 0.7501, indicating robust predictive performance under heterogeneous and literature-derived conditions. SHapley Additive exPlanations (SHAP) interpretation revealed that structural and environmental variables, particularly functional group chemistry, solution pH, and surface area, jointly govern uranium the maximum adsorption capacity (<i>q</i><sub>max</sub>) through nonlinear interactions, emphasizing the dual control of material composition and adsorption environment. Experimental validation using a representative high-performance covalent organic frameworks-based adsorbent confirmed the model’s predictive reliability, with the measured <i>q</i><sub>max</sub> of 342.2 mg g<sup>−1</sup> showing reasonable agreement with the predicted value (449.3 mg g<sup>−1</sup>; 23.8% deviation). This study establishes an explainable AI paradigm that bridges modeling and experimentation, providing a transferable foundation to support the intelligent discovery and sustainable design of next-generation uranium adsorbents for high-efficiency extraction.</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

Prediction of uranium adsorption performance by machine learning for sustainable seawater extraction

  • 期刊论文

  • 发表日期：

    2026年3月23日

查看全文
  • Zhenli Sun,
  • Shibo Sun,
  • Sheng Wang,
  • Hanyang Wang,
  • Yiyan Zhang,
  • Jianwei Huang,
  • Fuyou Fan,
  • Yuan Yang,
  • Xiangke Wang

摘要

The sustainable development of nuclear energy requires efficient and reusable materials for uranium extraction from seawater, where uranyl species exist at ultra-trace levels within complex coordination environments. To overcome the limitations of empirical, trial-and-error design, this study establishes U-Predict v1.0, a structure-performance database for uranium adsorbents, and introduces an interpretable machine learning framework for adsorption performance prediction and mechanistic analysis. Trained on over 220 samples and 54 structural descriptors, the Light Gradient Boosting Machine (LightGBM) model with engineered features and optimized hyperparameters achieved a training R2 of 0.9887 and a test R2 of 0.7501, indicating robust predictive performance under heterogeneous and literature-derived conditions. SHapley Additive exPlanations (SHAP) interpretation revealed that structural and environmental variables, particularly functional group chemistry, solution pH, and surface area, jointly govern uranium the maximum adsorption capacity (qmax) through nonlinear interactions, emphasizing the dual control of material composition and adsorption environment. Experimental validation using a representative high-performance covalent organic frameworks-based adsorbent confirmed the model’s predictive reliability, with the measured qmax of 342.2 mg g−1 showing reasonable agreement with the predicted value (449.3 mg g−1; 23.8% deviation). This study establishes an explainable AI paradigm that bridges modeling and experimentation, providing a transferable foundation to support the intelligent discovery and sustainable design of next-generation uranium adsorbents for high-efficiency extraction.