| The biological high toxicity and environmental accumulation of persistent organic pollutants（POPs）in the atmospheric environment are a common concern of the international community.It is generally believed that passive sampling technology can effectively monitor the distribution and migration of pollutants in the atmosphere.Polyurethane foam（PUF）has been widely used in passive sampling as a mainstream adsorption material in recent years.The accurate identification of PUF-air partition coefficient（KPA）is a key indicator to measure the successful application of samples.A large number of studies show that KPA is often affected by actual environmental factors（such as wind speed,temperature and humidity）in the monitoring process.In addition,considering the wide variety of organic pollutants and the endless emergence of new pollutants,it is difficult to meet the needs of environmental risk assessment and air pollution control only through experiments.Therefore,it is particularly important to develop a simple and rapid theoretical prediction method to estimate the equilibrium partition coefficient of organic matter.In order to master the partition behavior of different organic pollutants onto PUF,it is a good choice to use the multi-parameter model to predict KPA.Based on multi-parameter linear free energy relations（pp-LFERs）and quantitative structure property relations（QSPR）,12 different linear and nonlinear prediction models are established by three machine learning algorithms.These models were comprehensively verified,evaluated and characterized,and the potential mechanism of the partition behavior of organic pollutants onto PUF were discussed from the molecular point of view.In this study,by consulting a large number of literatures,the KPA experimental values of 362 organic pollutants were collected,including 14 categories of organic compounds（such as benzene,polycyclic aromatic hydrocarbons,polychlorinated biphenyls and pesticides）,covering the distribution characteristics of four different ambient temperatures（-12℃,15℃,25℃ and 35℃）.The specific research contents are as follows:（1）pp-LFERs based on Abraham descriptors.In this work,the Abraham descriptors（E,S,A,B,L）were screened and processed by stepwise regression at four different ambient temperatures（-12℃,15℃,25℃ and 35℃）,and four pp-LFERs models were developed combined with multiple linear regression（MLR）.After a series of mathematical statistical evaluation and cross validation,the four optimal prediction models have good goodness-of-fit（Radj2=0.713～0.996）,robustness（QLOO2:0.680～0.993;QBOOT2:0.738～0.791）and generalization ability（Rext2=0.708～0.998;Qext2:0.704～0.997）.A series of statistical parameters（CCC,SE,MAE and RMSE,etc.）show that the models have good prediction performance,The application domain characterization shows that the modeled experimental data are well representative.In addition,the mechanism explanation reveals that the main factors affecting the partition coefficient of organic matter between PUF membrane and gas phase are solute solvent interaction,molecular polarizability,molecular bond,polarization dipole interaction and hydrogen bond interaction.（2）QSPR based on AlvaDesc descriptors.Based on the study of pp-LFERs,in order to further explore the potential adsorption mechanism of organic pollutants on PUF,two QSPR prediction models were established according to the KPA experimental values of 362 organic pollutants.2-5 optimal variables were selected from 5290 AlvaDesc descriptors by stepwise regression method.Eight optimal KPA prediction models were constructed by using classical linear algorithm（MLR）and popular nonlinear algorithm（artificial neural network,ANN）.The results of internal and external validation show that the eight prediction models have good goodness-of-fit,robustness and prediction performance,meet the threshold standards of R2>0.7 and Q2>0.6,and can be used to predict the KPA values of other organic pollutants in the application domain.Compared with pp-LFERs model,the results show that the performance of QSPR-MLR and QSPR-ANN models based on AlvaDesc descriptors are better than ppLFERs（such as pp-LFERs Radj2:0.713 |
