Hassan Raouf, Kazemi Mohammad Reza
Civil Engineering Department, College of Engineering, Imam Mohammad Ibn Saud Islamic University (IMSIU), Riyadh, 11432, Saudi Arabia.
Process Engineering Department, Bandar Imam Petrochemical Company (BIPC), Mahshahr, Iran.
Sci Rep. 2025 Apr 30;15(1):15157. doi: 10.1038/s41598-025-99759-2.
Adsorption prediction of organic components on biochar and resins is essential for advancing industrial and energy technologies. This study utilized a dataset of 1750 adsorption isotherms comprising adsorption data for 73 organic materials on 50 biochar samples and 30 polymer resins. Machine learning models were developed using eight input parameters, including five Abraham solvation descriptors, total pore volume (Vt), specific surface area (BET), and equilibrium concentration (logCe), with the output parameter being adsorption degree (logKd). The dataset was split into training (1225 data points), testing (262), and validation (263). Various machine learning methods were evaluated, including Linear Regression, Ridge Regression, Lasso Regression, Elastic Net, Support Vector Regression (SVR), k-Nearest Neighbors (KNN), Decision Trees, Random Forests, Gradient Boosting Machines, Artificial Neural Networks (ANNs), Convolutional Neural Networks (CNNs), Gaussian Processes, as well as ensemble algorithms such as XGBoost, LightGBM, and CatBoost. Among these, XGBoost achieved superior accuracy with an R² of 0.974 and a mean squared error (MSE) of 0.0343, followed by LightGBM (R²=0.964, MSE = 0.0484) and CatBoost (R²=0.984, MSE = 0.0212). Simpler models such as Linear Regression and Elastic Net showed lower performance, with R² values ranging from 0.678 to 0.875 and higher MSE values. Sensitivity and SHAP analyses identified equilibrium concentration and specific surface area as the most critical factors influencing adsorption. The findings underscore the effectiveness of machine learning methods, particularly XGBoost, LightGBM, and CatBoost, in forecasting adsorption levels with high precision while offering actionable insights into key variables driving adsorption mechanisms.
预测有机成分在生物炭和树脂上的吸附情况对于推进工业和能源技术至关重要。本研究使用了一个包含1750个吸附等温线的数据集，该数据集涵盖了73种有机材料在50个生物炭样品和30种聚合物树脂上的吸附数据。利用包括五个亚伯拉罕溶剂化描述符、总孔体积（Vt）、比表面积（BET）和平衡浓度（logCe）在内的八个输入参数开发了机器学习模型，输出参数为吸附度（logKd）。该数据集被分为训练集（1225个数据点）、测试集（262个）和验证集（263个）。评估了各种机器学习方法，包括线性回归、岭回归、套索回归、弹性网络、支持向量回归（SVR）、k近邻（KNN）、决策树、随机森林、梯度提升机、人工神经网络（ANN）、卷积神经网络（CNN）、高斯过程，以及诸如XGBoost、LightGBM和CatBoost等集成算法。其中，XGBoost表现出卓越的准确性，R²为0.974，均方误差（MSE）为0.0343，其次是LightGBM（R² = 0.964，MSE = 0.0484）和CatBoost（R² = 0.984，MSE = 0.0212）。线性回归和弹性网络等较简单的模型表现较差，R²值在0.678至0.875之间，且MSE值较高。敏感性和SHAP分析确定平衡浓度和比表面积是影响吸附的最关键因素。研究结果强调了机器学习方法，特别是XGBoost、LightGBM和CatBoost，在高精度预测吸附水平方面的有效性，同时为驱动吸附机制的关键变量提供了可操作的见解。
