Predictive Life Cycle Assessment for Chemical Processes using Machine Learning

Due to the growing awareness of climate change, the chemical industry is increasingly considering not only economic but also ecological criteria in process development. Thus, environmental assessment methods are required that can be applied in early process development stages to support decision-making. An accepted, ISOnormed environmental assessment method is Life Cycle Assessment (LCA). However, LCA requires detailed information on mass and energy balances, which is usually not available in early process development. Furthermore, when assessing emerging technologies, future changes not only of the technology itself but also in the background system, e.g., the energy supply, have to be considered. In this thesis, the example of sector coupling of the steel and chemical industries is first used to investigate how changes in the background system can be considered in the assessment of emerging technologies and how these changes affect the assessment result. Afterwards, a framework for predictive LCA is presented that allows LCA to be integrated at early stages of process development. In contrast to existing approaches from literature using machine learning for component-specific predictions, this work combines machine learning regression models with methods of automated process design. Thereby, process-specific predictions of environmental impacts are enabled only based on the molecular structure of the desired product and the reaction equation. For this purpose, descriptors are determined purely predictively using quantum mechanics and statistical thermodynamics, as well as process shortcuts. In addition, an encoder-decoder neural network is used to increase the information density in the molecular descriptors. As regression models, an artificial neural network and a Gaussian process regression are trained on a consistent data set. The method is exemplarily integrated into computer-aided molecular and process design and used for the design of ecologically optimal solvents. In addition, the process-specific prediction is discussed for the example of CO2-based methanol. The results show that the integration of process-specific features into the LCA prediction increases the prediction accuracy and enables process-specific predictions. The presented method thus enables the integration of ecological criteria in the early development of chemical processes.