Early-stage environmental impact forecasting of chemicals and processes with machine learning and data analytics tools

Early-stage environmental evaluation of chemicals and process technologies is often constrained by the lack of detailed life cycle inventory (LCI) data and limited access to commercial life cycle assessment (LCA) tools. This work presents a predictive framework for estimating the environmental impacts of chemicals and processes during the early design stages when such information is unavailable. The framework combines Machine Learning (ML) models, Artificial Neural Networks (ANN), and eXtreme Gradient Boosting (XGBoost), trained on molecular descriptors and thermodynamic properties to predict four cradle-to-gate (production phase) endpoint metrics: Human Health Impact (HHI), Ecosystem Quality Impact (EQI), Global Warming Potential (GWP), and Resource Utilization Impact (RUI). To extend the analysis to the gate-to-gate (use phase) and gate-to-grave (end-of-life) phases, a power law regression model was developed to estimate the GWP scaling exponents for technologies as a function of the capacity/processing rate and energy consumption. Feature importance analysis using SHapley Additive exPlanations (SHAP) indicates that both molecular and thermodynamic descriptors contribute significantly to the prediction of environmental impacts. Two case studies were selected to demonstrate and validate the applicability of this framework. Overall, the results demonstrate the combined approach provides a practical tool for early-stage environmental screening of chemicals and processes. Supplementary Information: The online version contains supplementary material available at 10.1007/s10098-026-03479-8.