Environmental impact assessment of biomass process chains at early design stages using decision trees
P. Karka, S. Papadokonstantakis, A. C. Kokossis
The International Journal of Life Cycle Assessment, Volume 24, pp.1675–1700, 2019
Bio-based products; Early-stage assessment; Life cycle assessment (LCA); ReCiPe method
Purpose Life cycle assessment (LCA) is generally considered as a suitable methodology for the evaluation of environmental impacts of processes. However, it requires large amount and often inaccessible process data at early design stages. The present study provides an approach to streamline LCA for a broad set of biomass process chains. The proposed method breaks away from conventional LCA work in that the purpose is to support decision at early stages assuming minimal use of data available and points to most dominant LCA impacts, providing useful feedback to process design. Methods The prediction mechanism employs decision trees, which form “if-then rules” using a set of critical parameters of the process chain with respect to various environmental impacts. The models classify products into three classes, namely having low, medium, and high environmental impact. Data for model development were obtained from early design stages and include descriptors of the molecular structure of the product and process chain-related variables corresponding to chemistry, complexity, and generic process conditions. Twenty-three LCA metrics were selected as target attributes, according to the ReCiPe and the cumulative energy demand (CED) methods. A broad set of process chains is derived from the work of Karka et al. (Int J Life Cycle Assess 22(9):1418–1440, 2017). Results and discussion Results demonstrate that the average classification error for the decision trees ranges between 13.4 and 43.8% for the various LCA metrics and multifunctionality approaches. Allocation approaches present a better classification performance (up to 25% error) compared with the substitution approach for LCA metrics, such as climate change, CED, and human health. For the majority of models, low- and high-output classes are characterized by better predictive performance compared with the medium class. The interpretability of selected decision trees is analyzed in terms of pruning levels and “irrational” branches. The results of the application of the decision tress for recently published case studies show for instance that 8 out of 13 cases were correctly classified for CED. Conclusions The proposed approach provides a first generation of models in the form of computationally inexpensive and easily interpretable decision trees that can be used as pre-screening tools for the environmental assessment of bio-based production ahead of detailed design and conventional LCA approaches. The transparent structure of the decision trees facilitates the identification of critical decision variables providing insights for improvement in terms of process parameters, biomass feedstock, or even targeted product.