Abstract:

The study of pollutant transport and transformation in multi-phase environmental systems is directly relevant to the effectiveness of ecological risk assessment and pollution control technologies. However, traditional approaches are limited by data scarcity and challenges in quantifying uncertainties, making the accurate prediction of complex pollutants difficult. In this study, an adaptive evidential deep learning model was developed to predict the environmental transport behavior of emerging pollutants, specifically per- and polyfluoroalkyl substances (PFAS). The proposed model strategically integrates an evidential deep learning framework with advanced molecular representation learning and adaptive mechanisms, employing an improved adaptive evidential loss function with a regularization coefficient of 0.2 for model training and optimization. Experimental results demonstrate that the model achieves high-precision predictions for five critical environmental partition coefficients, including logK_AW (air-water partition coefficient), logK_OA (octanol-air partition coefficient), logK_OC (organic carbon-water partition coefficient), logK_OW (octanol-water partition coefficient), and logS_W (water solubility), with coefficient of determination (R²) values exceeding 0.95 across all endpoints. Benchmark comparisons demonstrate that the evidence-based deep learning framework consistently outperforms conventional Dropout and Ensemble approaches, exhibiting over a 50% performance advantage under high data sparsity (Ratio > 0.8). The model maintains robust performance on the independent test set and accurately identifies regions of high uncertainty, providing a reliable tool for predicting PFAS environmental migration properties and supporting pollution-control strategies.

Key words: prediction, pollution, diffusion, deep learning, per- and polyfluoroalkyl substances

摘要：

污染物在多介质体系中的迁移与转化研究直接关系到环境评估与污染控制技术的有效性。然而，传统方法受限于数据稀缺和不确定性量化，难以应对复杂污染物的预测挑战。本研究开发了一种自适应证据深度学习模型，用于预测新型污染物全氟和多氟烷基物质（PFAS， per- and polyfluoroalkyl substances）的环境迁移行为关键性质。该模型整合证据深度学习框架、分子表示学习与自适应机制，采用改进的自适应证据损失函数（正则化系数0.2）训练。实验结果表明，该模型能够高精度预测五个关键环境迁移性质（R²>0.95）：空气-水分配系数（log K_AW）、辛醇-空气分配系数（log K_OA）、有机碳-水分配系数（log K_OC）、辛醇-水分配系数（log K_OW）以及水溶解度（log S_W）。基准对比显示，证据深度学习方法全面优于传统Dropout和Ensemble方法，在高数据稀疏度（Ratio>0.8）下性能优势超50%。模型在独立测试集上保持稳定性能，能准确识别高不确定性区域，为PFAS环境迁移性质预测和污染防控提供了可靠工具。

关键词: 预测, 污染, 扩散, 深度学习, 全氟和多氟烷基物质