Direct sampling of measured or model data to improve uncertainty analysis in Life Cycle Assessment

Standard practice in Life Cycle Assessment (LCA) is to assume that each uncertain parameter behaves independently under Monte Carlo sampling. This leads to cases which are clearly incorrect, such as engine CO2 emissions being sampled independent from the fuel efficiency, or different providers into a market (such as an electricity market) increasing or decreasing without any regard to the behaviour of other providers. We have developed an open-source toolkit (https://github.com/PascalLesage/brightway2-presamples) that can solve these and other problems through the direct use of measured or pre-computed data and Monte Carlo samples. We demonstrate how this toolkit can provide a number of novel features for uncertainty and sensitivity assessment in LCA: - Monte Carlo samples can be saved and transferred between computers, allowing for perfect reproducibility. - Pre-generated static or stochastic values can be generated by complex, non-linear models, capturing system dynamics more accurately. - Pre-sampled Monte Carlo values can capture correlations between parameters, such as between characterization factors, or between input and outputs (e.g. fuel use and CO2 emissions). - Direct use of population data avoids losses or introduced inaccuracies from fitting data to distributions. We also introduce and demonstrate the idea of "campaigns", an organizational tool for sets of pre-sampled data that allows for quick system variation, guided data acquisition, and prospective LCA.