Scholarlink

<p>Metal–organic frameworks (MOFs) are tunable porous materials that have been widely studied for gas adsorption, separation, and catalysis. Predicting and classifying the pore-limiting diameter (PLD) are valuable for rational MOF screening and design. Here, we develop an ensemble machine-learning pipeline: Random Forest (RF), XGBoost (XGB), and Light Gradient Boosting Machine (LGBM) to classify pore classes and regress PLD using a curated dataset derived from the Cambridge Structural Database. To address class imbalance across pore classes, we evaluate TomekLinks, Synthetic Minority Oversampling Technique (SMOTE), and SMOTETomek. Among the resampling strategies, RF provided the strongest overall performance, and the RF+TomekLinks configuration delivered the best balance between classification and regression. On the held-out test set, the best model achieved <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> <EquationSource Format="MATHML"><math> <msup> <mi>R</mi> <mn>2</mn> </msup> </math></EquationSource> </InlineEquation> up to 0.999 with RMSE as low as 0.124 Å&#xa0; for PLD regression. Feature-importance analysis indicated that both metal-center elemental properties and linker-related topological descriptors contribute substantially to PLD prediction. Applicability-domain assessment using leverage/Williams-plot analysis shows that most test structures lie within the modeled domain, whereas out-of-domain cases are explicitly flagged. Overall, the proposed workflow enables accurate, domain-aware PLD estimation to support high-throughput MOF screening; nevertheless, external validation of independent MOF collections and uncertainty quantification remain important for broader deployment.</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

Robust ensemble learning for classification and accurate prediction of pore-limiting diameter in metal–organic frameworks

  • 期刊论文

  • 发表日期：

    2026年3月10日

查看全文
  • Wahyu Aji Eko Prabowo,
  • Muhamad Akrom,
  • Supriadi Rustad,
  • Wise Herowati,
  • Harun Al Azies,
  • Hermawan Kresno Dipojono,
  • Subagjo,
  • Febdian Rusydi,
  • Hideaki Kasai

摘要

Metal–organic frameworks (MOFs) are tunable porous materials that have been widely studied for gas adsorption, separation, and catalysis. Predicting and classifying the pore-limiting diameter (PLD) are valuable for rational MOF screening and design. Here, we develop an ensemble machine-learning pipeline: Random Forest (RF), XGBoost (XGB), and Light Gradient Boosting Machine (LGBM) to classify pore classes and regress PLD using a curated dataset derived from the Cambridge Structural Database. To address class imbalance across pore classes, we evaluate TomekLinks, Synthetic Minority Oversampling Technique (SMOTE), and SMOTETomek. Among the resampling strategies, RF provided the strongest overall performance, and the RF+TomekLinks configuration delivered the best balance between classification and regression. On the held-out test set, the best model achieved \(R^2\) R 2 up to 0.999 with RMSE as low as 0.124 Å  for PLD regression. Feature-importance analysis indicated that both metal-center elemental properties and linker-related topological descriptors contribute substantially to PLD prediction. Applicability-domain assessment using leverage/Williams-plot analysis shows that most test structures lie within the modeled domain, whereas out-of-domain cases are explicitly flagged. Overall, the proposed workflow enables accurate, domain-aware PLD estimation to support high-throughput MOF screening; nevertheless, external validation of independent MOF collections and uncertainty quantification remain important for broader deployment.