Frameworks for the application of machine learning in life cycle assessment for process modeling

In the face of escalating emission reduction demands and heightened public awareness, the imperative for transparent assessments is fundamental to responsible and sustainable development. The use of life cycle assessment (LCA) is instrumental in identifying environmental hotspots in intricate systems and guiding the design and selection of environmentally conscious production methods. However, LCA's comprehensive approach demands substantial data, fundamentally material and energy flows for individual materials or processes, consolidated within a life cycle inventory (LCI). The amount of time, resources, and expertise required to compile an accurate LCI dataset are among the greatest concerns for LCA practitioners. During the early design phase for industrial-scale production, process simulation is a useful tool for estimating LCI data; however, first-principle models can sometimes be unfeasible. This has prompted researchers and engineers to advocate for simplified or surrogate versions of these intricate models in some particular cases. In contrast to first-principle models, machine learning (ML) models efficiently manage extensive datasets and complex systems without rigorous model equations. This work assesses the current state of ML-LCA integration through literature and bibliometric analysis, categorizing works into three clusters and identifying publication trends. Furthermore, this analysis yielded three frameworks aimed at facilitating the integration of ML techniques into LCA workflows, enhancing precision and efficiency in environmental impact assessment. The first framework revealed the interest in abstracting a complete process into a surrogate ML model for fast LCI predictions. Conversely, the second one focused on substituting a complex part of the process for an ML surrogate model based on data from experiments or literature. Finally, in the third framework, LCA performance was directly correlated with a system characteristic, enabling direct and fast predictions of LCIs or LCA performance indicators, and optimization in not yet designed systems.