Optimizing the benefit–risk trade-off in nano-agrochemicals through explainable machine learning: beyond concentration†

Balancing the benefits and undesirable environmental impacts is essential for ensuring successful applications of emerging nano-agrochemicals. However, there is a lack of transparent and explainable trade-off methodologies in this safety-sensitive field. Here, an explainable machine learning-driven multi-objective optimization approach is proposed to maximize the performance and minimize undesirable implications of seed nanopriming. The root dry weight under salinity stress and the relative concentration of the constituent elements of the used nanoparticles in shoots are considered potential indicators of the benefit and risk, respectively. An ensemble strategy of model explanation, based on self-explainable models, is employed to obtain more reliable, unbiased, and trustworthy results with small datasets. Multi-objective optimization is employed to select potential treatments among numerous generated candidates based on the predictions of explainable machine learning models. Furthermore, model explanations are combined with prior knowledge to explain this selection process and elucidate the factors' effects on the benefit and risk. The explanation results highlight the importance of considering the well-known concentration-dependent effect of nanoparticles in conjunction with other factors such as zeta potential and surface area, which is further verified by statistical analysis. Together, this study provides a promising approach for accelerating the discovery, assessment, and regulation of nanomaterials and may facilitate their sustainable applications in agriculture and the environment.