基于机器学习的生物炭高效吸附全氟化合物效能预测

Machine Learning-Based Prediction of the Efficient Adsorption of Perfluorinated Compounds by Biochar

摘要: 全氟及多氟烷基化合物(PFASs)因其极强的化学稳定性和生物蓄积性，已成为当今环境治理的重大挑战。吸附法因高效、低能耗等优势被认为是去除PFASs最具前景的技术之一。生物炭作为一种低成本、可再生的吸附材料，在PFASs去除领域展现出广阔的应用潜力。然而，传统生物炭研发模式面临实验成本高、参数关联复杂及非线性机制难量化等瓶颈。机器学习方法凭借其处理高维数据、挖掘复杂非线性关系的能力，能够从已有实验数据中快速建立“材料性质–吸附性能”的映射关系，无需进行大量试错实验即可实现性能预测与机制解析，为突破传统研发模式的局限提供了新途径。本研究利用机器学习技术实现生物炭吸附PFASs效能的精准预测，从数据库中筛选并整合了258组实验数据，选取热解温度、(N + O)/C、H/C及比表面积等10种关键理化特征作为输入变量。研究分别采用LR、KNN、SVR和RF四种算法构建模型，并对其效能进行了评估。特征重要性分析显示，(N + O)/C和H/C是影响最大吸附量的核心驱动因子。据此推测，极性官能团可能通过极性和氢键与PFASs发生特异性作用，而高H/C值则暗示疏水作用对吸附的潜在增强效应。相比之下，比表面积和孔体积的重要性排名靠后，初步表明在该体系中化学亲和力可能优于物理孔隙填充机制。上述推测仍需结合实验验证或密度泛函理论等理论计算进一步确认。

Abstract: Per- and polyfluoroalkyl substances (PFASs) have emerged as a significant challenge in contemporary environmental remediation due to their exceptional chemical stability and pronounced bioaccumulation potential. Adsorption is widely regarded as one of the most promising approaches for PFAS removal, owing to its high efficiency and low energy consumption. Biochar, as a low-cost and renewable adsorbent material, demonstrates broad application potential in the field of PFASs removal. However, traditional biochar research and development models face bottlenecks including high experimental costs, complex parameter correlations, and difficulties in quantifying non-linear mechanisms. Machine learning methods, with their capacity to process high-dimensional data and capture complex non-linear relationships, can rapidly establish a mapping between material properties and adsorption performance from existing experimental data. This enables performance prediction and mechanistic analysis without extensive trial-and-error experiments, offering a new approach to overcome the limitations of traditional research and development models. In this study, machine learning was employed to enable accurate prediction of PFASs adsorption performance on biochar. A dataset comprising 258 experimental groups was screened and integrated from the literature. Ten key physicochemical features, including pyrolysis temperature, (N + O)/C ratio, H/C ratio, and specific surface area, were selected as input variables. Four algorithms, namely Linear Regression (LR), K-Nearest Neighbors (KNN), Support Vector Regression (SVR), and Random Forest (RF), were implemented and evaluated for their predictive performance. Feature importance analysis revealed that the (N + O)/C and H/C ratios were the predominant factors governing the maximum adsorption capacity. This finding suggests that polar functional groups may facilitate specific interactions with PFASs via hydrogen bonding and electrostatic forces, whereas an elevated H/C ratio implies a potential enhancement of adsorption through hydrophobic interactions. In contrast, specific surface area and pore volume exhibited relatively lower importance, preliminarily indicating that chemical affinity may play a more critical role than physical pore filling in this adsorption system. These proposed mechanisms require further validation through experimental approaches or theoretical calculations such as density functional theory (DFT). Overall, this study not only enables efficient prediction of adsorption performance but also provides a scientific foundation for the rational design of biochar tailored for PFAS remediation.