Gaussian Process Regression as a Replicable, Streamlined Approach to Inventory and Uncertainty Analysis in Life Cycle Assessment.

Abstract

Life cycle assessment plays a critical role in quantifying environmental impacts, but its credibility remains challenged when data and uncertainty analysis are lacking. In this study, we propose a data compilation framework to address these two issues. The framework first quantifies the correlations of production activities among existing data in temporal, geographical, and taxonomic dimensions. The framework then introduces covariance functions to convert these correlations to a similarity matrix, and the Gaussian process regression model is adopted to predict new data based on these covariance functions. The associated uncertainty is automatically characterized using the posterior distribution of predictions. The framework is demonstrated on the nitrogen fertilizer application rate for food production─an activity recognized for its environmental burden─with results capable of reflecting temporal and geographical variations. By introducing the concept of phylogenetic distance as a correlation of taxonomy, the framework provides a quantitative basis for predictions in a proxy data usage scenario. The framework can be used in developing temporally and regionally representative life cycle inventories and databases and can facilitate consistent uncertainty quantification in future life cycle assessment methodologies.