(479c) From Ex-Post to Ex-Ante LCA: A Machine-Learning Approach to Reverse Decision-Making and Innovate in Sustainable Development – the Case of Biorefineries

Life Cycle Analysis is an established technology that is routinely applied to evaluate the sustainability potential of existing and emerging industrial processes. Considering the environmental dimension out of the three aspects of sustainability, Life Cycle Assessment (LCA) is a suitable methodology for the evaluation of environmental impacts because it highlights the stages with the greatest contribution along a production chain. The LCA approach is structured and depends on process inventories (essentially mass and energy balances) that it eventually translates into burdens (mid-point impact categories), damage (end-point categories) and total score metrics (impact). The ex-post LCA approach typically requires a large amount of information, that is usually extracted from detailed flowsheets and/or already completed pilot plants. Concerning new and unsettled processes, LCA studies can be costly, time consuming and unable to efficienctly handle degrees of freedom. Bio-refineries constitute examples of such new processes with largely undefined portfolios of feedstocks, products and processes. They are promising options of chemicals production, capable to produce a wide range of fuels and chemicals equivalent to the conventional fossil-based products. To establish bio-refineries as mature choices and achieve the commercialization of their technologies, the application of sustainable solutions during the design and development stages are crucial. The innovative character of bio-based production and therefore data availability and access on process modelling details, is a challenging point for decision makers to move towards this direction.

The paper advocates the development of ex-ante LCA as an alternative approach to evaluate sustainability. Rather than replacing ex-post LCA, ex-ante LCA, complementary to ex-post, is expanding established functions with capabilities to predict rather than evaluate (anticipatory functions), project rather report (prospective functions), point to missing technologies (consequential), and expand knowledge from new data (dynamic, temporal). To achieve these functions the paper proposes a data driven approach using statistical models and machine learning technology trained by a large amount of background (reactions, chemicals, feedstocks, technologies) and foreground information (integrated flowsheets product portfolios). Machine learning is applied through two sets of developments: ANNs and decision trees. Models are trained from a wide range of biorefineries datasets including 85 products, 10 platform chemicals (e.g., syngas, sugars and lignin), biofuels (e.g., biodiesel, biogas, and alcohols), and biomass sources (e.g., wood chips, wheat straw, vegetable oil). Overall, the training set accounts for 138 datasets, 23 LCA metrics, and 3 allocation methods. Input parameters include descriptors of the molecular structure and process related data associated with production paths of target chemicals. The models have been tested for prediction quality covering most critical aspects of environmental sustainability such as cumulative energy demand (CED) and global warming potential (GWP). The average classification error for decision- tree models are up to 25% whereas for ANN models the average R2 values (coefficient of determination) range between 0.7 and 0.8. The approach is demonstrated in several cases where LCA assesses options, ahead of any inventory calculations: (a) in the development of the biorefinery value chain; (b) to assess and rank technological options given a targeted product and options for feedstock; and (c) in selecting technologies and feedstocks for a given set of products.