Using machine learning models to estimate ecotoxicity effect factors for various nanomaterials in life cycle...

Although nanomaterials (NMs) have shown great promise across various fields due to their advantageous characteristics, concerns about their environmental consequences before large-scale production and application are increasing. The life cycle assessment (LCA) is a powerful tool for investigating the potential environmental impacts of emerging NMs. However, LCA practices are facing the challenge of lacking characterization factors (CFs) to estimate the environmental implications of NM release. CFs comprise of information on fate factors, exposure factors, and effect factors (EFs) using the USEtox framework. Therefore, this study aimed to address the insufficient toxicity testing data by developing eight machine learning (ML) regression algorithms to predict freshwater ecotoxicity EFs of NMs from their physicochemical properties. The results showed that the AdaBoost model exhibited the best predictive performance, with an R² of 0.786, root-mean-square error (RMSE) of 0.711, and mean absolute error (MAE) of 0.625, and generalized well to an external validation set, achieving an RMSEₑₓₜ of 0.654 and MAEₑₓₜ of 0.550. The Shapley Additive Explanations (SHAP) analysis identified chemical composition, surface area, species, electronegativity, size, and diameter as the important features. Further, the estimated freshwater ecotoxicity EFs and CFs for graphene, derived from the optimal AdaBoost model and Monte Carlo simulation, were 23.15-200.24 potentially affected fraction (PAF)·m³·kg^-1 and 231.49-2002.40 PAF·day·m³·kg^-1, respectively. This study could not only help identify the potential freshwater ecotoxicity of NMs during the early development stage but also inform appropriate decisions for risk management in the absence of reliable data.