Predicting embodied carbon of building structure types through Machine Learning

Résumé En

The building industry is responsible for 38% of total global energy-related CO2 emissions. In the previous years, policies have focused on reducing operational energy, and embodied emissions have become most significant contributors to the total greenhouse gas (GHG) emissions of buildings. Structural systems almost always constitute the largest source of embodied carbon in buildings and largest emissions occur during the production stage. However, embodied CO2 values are strongly influenced by material provenance, and their magnitude varies greatly between countries [1]. Databases gathering information on GHG emissions are therefore essential to improve our understanding of current emissions (baseline), targeted emissions (budget) and to make recommendations from their analysis for the future (benchmarks). The digitalization of the building industry has facilitated the introduction of carbon footprint assessment tools in the structural engineering practice [2]. Choosing appropriate structures in a given project can greatly reduce material consumption and material waste. However, methods to efficiently assess GHG emissions of such structures are lacking. Moreover, they are rarely computed at early design stages, when changes with highest impact are made, but quantitative volumetric and material information – ‘hard’ features - are still unavailable [3]. This research uses machine learning models and investigates strategies to predict the GHG emissions of buildings with two structure typologies, massive structures and frame structures, using descriptive data available in competition briefs - “soft” features. It proposes a methodology to obtain accurate predictions, despite limited data available for the learning. The methodology is tested on the Embodied Carbon of European Buildings EUCB-D database, describing the embodied GHG emissions of buildings. Results prove the potential of the methodology, and motivate its further developments into a tool for data-driven low carbon structural design. Moreover, results identify the importance of soft-features and suggests ways of assembling efficient databases describing building structures, in order to improve the understanding of their environmental impact.