Environmental impact level prediction for inorganic and organic chemicals using machine learning models and interpretability analysis.

The rapid proliferation of chemicals in commerce presents an urgent need for efficient and reliable assessment of their environmental impacts to support sustainable manufacturing. However, prior machine learning approaches linking chemical structures to environmental impacts have been restricted to organic compounds. This study represents an initial attempt to address existing gaps in evaluating the environmental impacts of both organic and inorganic chemicals through the integration of molecular descriptors within a two-level classification framework employing a machine learning approach. Results revealed that models with Mordred descriptors had better performance among the utilized features. Promising accuracies existed in five of the established models, such as the model predicting the global warming potential level (0.867) and the freshwater eutrophication level (0.800). The models also exhibited strong generalizability, as predictions for external organic and inorganic chemicals were largely consistent with the expected classifications. Furthermore, the feature importance identified through SHapley Additive exPlanations analysis offered mechanistic insights into the key factors influencing environmental impacts. Together, this study advances the application of machine learning approaches and interpretability analysis for chemical environmental impact level prediction and provides a practical tool to support environmental decision-making.