A Neural Network-Based Framework To Predict Process-Specific Environmental Impacts

Growing environmental concern and strict regulations led to an increasing effort of the chemical industry to develop greener production pathways. To ensure that this development indeed improves environmental aspects requires an early-stage estimation of the environmental impact in early process design. An accepted method to evaluate the environmental impact is Life Cycle Assessment (LCA). However, LCA requires detailed data on mass and energy balances, which is usually limited in early process design. Therefore, predictive LCA approaches are required. Current predictive LCA approaches estimate the environmental impacts of chemicals only based on molecular descriptors. Thus, the predicted impacts are independent from the chosen production process. A potentially greener process cannot be distinguished from the conventional route. In this work, we propose a fully predictive, neural network-based framework, which utilizes both molecular and process descriptors to distinguish between production pathways. The framework is fully automatized and includes feature selection, setup of the network architecture, and predicts 17 environmental impact categories. The pathway-specific prediction is illustrated for two examples, comparing the CO2-based production of methanol and formic acid to their respective fossil production pathway. The presented framework is competitive to LCA predictions from literature but can now also distinguish between process alternatives. Thus, our framework can serve as initial screening tool to identify environmentally beneficial process alternatives.