Probabilistic inference of material quantities and embodied carbon in building structures

In an effort to minimise the carbon footprint of building structures, a range of prediction tools and methods have been recently proposed, so to enable design practitioners evaluating how their design choices ultimately affect the carbon embodied in their designs. Such tools are most often targeted for use at the early stage of the design process, that is when exploration of alternative design options is usually undertaken, hence room for potential carbon reductions is greatest and at no extra cost of redesign. The overarching methodology behind existing tools predominantly relies on idealised models to characterise the structural system, usually employing closed-form design equations and/or numerical Finite Element to generate an inventory of material quantity data (that is ultimately required for embodied carbon estimates). Despite the very high level of complexity achieved by some models, the absence of any empirical reference with 'as-built' inventory data of material quantities leaves room for doubt on how accurate such models really are in capturing the complexities and inherent variability of the population of real building structures such models aim to represent. To bypass this limitation, a data-driven probabilistic graphical model is proposed here as alternative to existing approaches. A Bayesian Network was developed and tested as a proof of concept, trained on a dataset of 133 data-points of real building structures, leveraging on six design variables (at most) to fully characterise the entire design space of early design options. Despite the very small set of 'explanatory' design variables, the model exhibited a 73% accuracy (mean average absolute prediction error of 27%) when predicting the embodied carbon on a test sample of unseen real building structures. The study ultimately demonstrates the viability of adopting a probabilistic (data-driven) approach for such an inference task as an inherently robust alternative to data-blind models currently proposed in literature.