
Open access
Author
Date
2022Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
In order to reduce anthropogenic environmental impacts, it is important to properly measure them. Life Cycle Assessment (LCA) is a well-established tool for the quantification of the potential environmental impacts of a product or service throughout its complete life cycle. It can be used to compare various alternatives, assess environmental benefits and negative impacts in a systematic manner, integrate environmental sustainability aspects already at the product design and development stages, and support environmental communication and effective decision-making.
LCA models are based on the analysis of complex supply chains and environmental processes. Naturally, there exists variation in the performance of technical systems, differences in efficiencies or output quantities, and changes in equipment efficacy due to varying internal and external factors. A slightly different combination of parameters compared to their default values in simultaneously many processes can significantly affect LCA results. Depending on the objective of LCA, interpretation of the results might change in the presence of selected uncertainties. This raises the question of how sensitive is LCA to variations in the underlying supply chains, or how robust are the results?
To investigate robustness of LCA, Monte Carlo (MC) simulations are employed to numerically propagate uncertainty from the large number of LCA model inputs to the output, and uncertainty analysis (UA) is used to quantify the uncertainty in LCA results. In case it is sufficiently narrow to support robust decision-making, no additional analysis is required. Otherwise, it is important to conduct global sensitivity analysis (GSA) and understand which inputs are responsible for most of the LCA output uncertainty. Then for the identified influential processes, researchers can develop better modeling or invest into additional data collection to reduce output uncertainty, and evaluate reliability of results to make better informed decisions.
The objective of this thesis is to develop new methodologies for validated and robust GSA that are suitable for high-dimensional LCA models with nonlinear and correlated inputs in order to identify influential parameters, potentially reduce uncertainty, and enable more reliable decision making. Since nowadays it is becoming more feasible to include novel data sources in life cycle inventories, and the LCA models are evolving to contain complex nonlinear and correlated elements, the aim was to develop GSA methodologies applicable to a wide spectrum of LCA studies. For this purpose, GSA of large LCA models was separated into screening of non-influential inputs, and ranking of the potentially influential inputs that are remaining. Sensitivity indices - the quantitative measures of inputs' degree of importance - were estimated numerically for each input based on the MC simulations. First, a standardized screening procedure was developed for generic, high-dimensional models, and then tested on a proposed benchmark function, and a case study of climate change impacts from Swiss household consumption. Notably, it monitors robustness and convergence of sensitivity indices, conducts validation of sensitivity results, and contains steps that optimize computational performance of GSA depending on the degree of model linearity. Then Spearman correlations are used as sensitivity indices in case the model is approximately linear, and feature importance metrics from the gradient boosted tree machine learning method otherwise. Applying GSA methods to larger models requires fast and memory-efficient computer codes. The absence of such a code motivated the development of the Python package, gsa_framework, which enhances performance of existing libraries.
The complete GSA protocol tailored for LCA models with many uncertain inputs of various types builds on the robust procedure for high-dimensional screening as previously discussed to filter out unimportant inputs, and then ranks the remaining inputs with Sobol indices. Its performance has been compared against the state-of-the-art approach based on the contributions to the deterministic total life cycle impact assessment score on the case study of Swiss consumption. While our GSA protocol needs more computational resources than the contribution-based method, it also yielded more complete results, and was able to identify potential errors in the ecoinvent life cycle background database. It showed that as few as 20 inputs out of hundreds of thousands are responsible for most of the uncertainty in the carbon footprint of Swiss household consumption.
Finally, we investigated how the correlated and causal dependencies between model inputs can be included in GSA of LCA. Correlations is a long recognized and often neglected issue when performing UA and GSA. Based on the conducted literature review, it seems that no other works performed large scale GSA for complete LCA models in the presence of correlated uncertainties. This work implemented multiple correlated sampling modules that can be included to MC simulations, and then employed Shapley values in combination with gradient boosting to identify important model inputs. Indeed, UA and GSA outcomes change if one accounts for correlations.
Unlike the existing approaches, the GSA protocols proposed in the scope of this thesis are designed to be generic to tackle various complexity of LCA models and data, which includes correlated inputs, model nonlinearities, and real measurements as input samples. At the same time, they contain steps specific for LCA models to optimize computational performance of GSA. They also include systematic validation of sensitivity results on all steps of GSA, allow correlated and nonlinear effects when conducting GSA of LCA, and assess the performance of GSA on the case study of Swiss household consumption. Special focus was placed onto combining classic GSA methods with novel approaches from the fields of dimensionality reduction and interpretability of machine learning models, as well as making it accessible to other LCA practitioners.
This thesis has demonstrated that complex UA and GSA of LCA is nowadays computationally and methodologically feasible, and how it can be rigorously conducted. Such analyses are needed to support more robust and reliable decision-making, and to systematically improve data and modeling quality in commonly used life cycle inventories. GSA of LCA is in principle a well-defined problem that should become part of any LCA study, because it allows us to gain deeper understanding of LCA models, and hence to better identify potentials for informed decisions and policies. Thereby, there is a clear need for advanced modeling paradigms, novel data sources, and better inventory modeling. With this thesis we want to show the potential in conducting new generation UA, GSA and inventory modeling that can be facilitated with open source software solutions and practices supported by community engagement. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000588835Publication status
publishedExternal links
Search print copy at ETH Library
Contributors
Examiner: Hellweg, Stefanie
Examiner: Mutel, Christopher Lucien

Examiner: Cucurachi, Stefano
Examiner: Suh, Sangwon
Publisher
ETH ZurichSubject
Global sensitivity analysis; Life cycle assessment; Uncertainties; Uncertainty analysis; Monte Carlo simulations; Sobol indices; Gradient boosted decision trees; Shapley value; Delta moment-independent indices; Correlated sampling; Global warming potentials; Robustness; Decision-making; Brightway; Python; Open source softwareOrganisational unit
03732 - Hellweg, Stefanie / Hellweg, Stefanie
Related publications and datasets
Is derived from: https://doi.org/10.3929/ethz-b-000522544
Is derived from: https://doi.org/10.3929/ethz-b-000551516
More
Show all metadata
ETH Bibliography
yes
Altmetrics