Bayesian network modelling for predicting the environmental hazard of silver nanomaterials in soils.

Abstract

In alignment with the European Union's Green Deal, which directs safe and sustainable practices for all chemicals, including nanomaterials (NMs) and advanced materials (AdMa), this study addresses the environmental hazard of silver NMs to terrestrial ecosystems. In the context of safe and sustainable by design (SSbD) framework, there is a need for methodologies that integrate pHysicochemical characteristics and experimental conditions to reliably predict their hazards to exposed species. Bayesian Networks (BN) represent a pivotal machine-learning (ML) tool with the potential to accelerate the SSbD process by leveraging predictive capabilities. In this study, we employed BN models trained on a literature-derived dataset capturing the ecotoxicity of silver (Ag) NMs in soils, focusing on predicting chronic no-observed effect concentrations (chronic NOECs). The model incorporates physicochemical characteristics such as surface coating, nominal particle diameter and particle shape as provided by manufacturers, species information such as life stage and taxonomic class, and exposure medium characteristics. The BN, refined through expert insights, achieved an average predictive accuracy of approximately 82 % across the output labels. The study also extracted interpretable rules from the BN, outlining environmental safety criteria and identified key factors influencing NM hazard for terrestrial organisms. The critical need for experimental datasets that provide fuller details of physiochemical characteristics and experimental conditions, as well as current limitations, are highlighted. This modelling approach facilitates the rapid screening of the potential hazards of AgNMs to terrestrial ecosystems, with the potential to accelerate safety evaluations and rationalise experimental demands.