Machine learning-driven optimization of arsenic phytoextraction using amendments

Exogenous amendments are crucial for enhancing the remediation efficiency of arsenic-contaminated soils by Pteris vittata. However, their effectiveness is unstable due to various factors, and neglecting their economic costs hinder broader application. In this study, we analyzed 2299 data points from 121 published datasets and used machine learning to predict and optimize the performance of amendments to enhance the phytoextraction efficiency. Using a random forest model, we predicted changes in As accumulation in P. vittata in response to specific amendments, considering 18 parameters across four categories: changes in P. vittata, amendments, soil properties, and cultivation conditions. The model achieved an R² value of 0.846. Using %IncMSE to quantify parameter contribution, we found that the biomass of P. vittata had a greater influence than the As concentration. Additionally, amendment type, application time, cultivation duration, and soil-available As were key factors in enhancing As accumulation in P. vittata. Regarding economic cost, different amendments required an investment ranging from 0.57 to 3903.86 CNY to enhance 1 g of As accumulation in P. vittata. Among these, phosphate fertilizers had the lowest cost, whereas calcium acetate, ethylenediamine-N,N'-disuccinic acid, and glutathione did not have economic advantages as amendments. This study offers guidance on the development of amendments, providing an important reference for the practical application of phytoextraction in As-contaminated soils.