ANN-Based approach for selection of optimum floor system of multi-storied buildings in the preliminary design stage

Abstract

Eco-friendly and climate-resilient building concepts are globally emerging with increasing Greenhouse Gas (GHG) emissions from the building sector. Addressing the embodied carbon of floor systems becomes essential as these contribute more than 50% of the embodied energy of the building footprint. Whereas, in the Preliminary Design Stage (PDS), striving for the best alternative satisfying performance, economic and environmental criteria is challenging within a limited time while decision-making is generally from the expertise’s design sense and intuition. Engineers in design firms have been encountering these recursive design processes over the years; however, the rate of reutilization of the previous design information and experience is very rare. Machine Learning (ML) techniques can offer smart decision-making tools in the PDS. Nevertheless, only a few addressed the optimization in the PDS while none of the studies comprehensively addressed the optimum floor system considering all three criteria simultaneously. This study utilized a human brain mimicking algorithm, Artificial Neural Network (ANN) in the optimization of reinforced-concrete floor systems of a multi-storied building in the PDS encapsulating the performance, economic and environmental considerations with the aid of data collection from the constructed buildings in Sri Lanka. The eight distinct models are proposed to predict each quantity in both Beam-Slab System (BSS) and Flat-Slab System (FSS). The proposed models in the BSS database are recorded predictions with R2 values of above 0.8 demonstrating reasonable performance in the heterogeneous real-world data while the models in the FSS database show superior performance which yields R2 values greater than 0.95 with synthetic data. Also, the predictions from all the models fall below a 30% error margin and the models in the FSS database demonstrated the higher performance with significantly lower error margins of 10 to 20%. Further, a conducted comparative analysis with the range of alternative ML models demonstrates the performance and capabilities of human-brain-inspired ANN under heterogeneous and high-dimensional data with inherently qualitative in nature. SHAP explanations demonstrated the relative contribution of each feature in the quantity predictions which in turn streamlined the confidence in the end-users. The trained ANN model is finally transformed into a design tool to facilitate the engineers to predict quantities that offer a platform to select the optimized floor system in the PDS