Strategies for data-driven low carbon structural design.

Today, every building material in a given context is associated to a carbon equivalent factor CO2e, corresponding to its Global Warming Potential (GWP). The digitalization of the building industry has facilitated the introduction of GWP assessment tools in the structural engineering practice. However, it is rarely computed at early design stages, when changes with highest impact are made, as quantitative volumetric and material information – ‘hard’ features – are still unavailable. This research uses machine learning regression models and investigates alternative strategies to predict the GWP of a building structure, using descriptive data available in competition briefs – “soft” features. To this end, we have compared a linear regression model to 9 other regression methods, with different hypothesis and errors functions. The models are ranked based on their predictive accuracy and residual plots. Despite the limited data available, and preliminary results, an accuracy of 70% was reached and residuals had relatively small standard deviations. This indicates that the models are functioning and proves the potential of soft-feature based prediction of GWP. Moreover, a first sensitivity analysis of soft-feature weights on calibrated models helped identify their relative impact on GWP. This understanding could help guiding design decisions at early stages, and could be implemented into an interactive tool for data-driven low carbon structural design.