可解释性
人工智能
人工神经网络
特征选择
支持向量机
计算机科学
机器学习
分类器(UML)
模式识别(心理学)
灵敏度(控制系统)
混乱
决策树
混淆矩阵
数据挖掘
选择(遗传算法)
图层(电子)
特征(语言学)
变量(数学)
随机森林
梯度下降
主成分分析
线性判别分析
非线性系统
统计分类
数学
预测建模
偏移量(计算机科学)
作者
Sajedeh Rooholamini,Farzin Hosseinifard,Mohsen Salimi,Amirhossein Amirfakhraei,Mona Zamani Pedram,Majid Amidpour
出处
期刊:
[Elsevier BV]
日期:2026-09-29
卷期号:8: 100055-100055
标识
DOI:10.1016/j.aitf.2026.100055
摘要
This study applies machine learning (ML) techniques to classify metal–organic frameworks (MOFs) for adsorption-based desalination considering multiple adsorption-related properties. The compiled dataset includes eight physicochemical input features, while the target variable consists of 11 distinct MOF classes. Several classification algorithms were implemented and compared, including K-Nearest Neighbors (KNN), Logistic Regression, Naïve Bayes, AdaBoost, Support Vector Classifier (SVC), Random Forest, and an Artificial Neural Network (ANN). The ANN architecture consisted of an input layer with eight neurons, hidden layer (s) for nonlinear pattern extraction, and a single output neuron representing multi-class predictions. The training loss curve demonstrated stable convergence with progressive error reduction across epochs, confirming effective learning without instability. Internal weight and bias visualizations further illustrated meaningful feature interactions and stronger hidden-to-output layer connections, improving interpretability of the learned structure. Model evaluation using confusion matrices and ROC–AUC analysis revealed that KNN achieved the highest overall classification performance, outperforming ANN and other models in terms of discrimination capability. AUC values for most classes were significantly higher than 0.50, confirming strong predictive reliability. SHAP analysis identified Surface Area as the most influential feature, followed by Pore Volume and Hydrophobicity Index, highlighting their dominant role in MOF classification. Sensitivity analysis further demonstrated that variations in these parameters significantly shift class probabilities, defining clear decision boundaries. To ensure statistical robustness, MIL-101 (Cr) was excluded due to inconsistent and noisy data.
科研通智能强力驱动
Strongly Powered by AbleSci AI