Optimisation framework for sustainable design of concrete buildings

The building industry is identified as the largest single contributor to climate change, due to extensive consumption of natural resources and the discharge of high volumes of carbon emissions. It is consequently imperative for the whole sector to work towards sustainable design and construction. Although structural engineers have the greatest potential to enhance buildings' sustainability by means of structural optimisation and/or material efficiency, they often play a restricted role in the sustainable design of a project. While many studies separately investigate the eco-friendly potentials of horizontal frames and vertical systems, most of them have not thoroughly considered all major components together for the whole structure, making it challenging for structural designers to incorporate and apply their findings into design projects. In addition, many design factors that are decided in early design stages have tremendous impacts on a building's life cycle carbon footprint. Therefore, a comprehensive op imisation methodology that allows for a thorough environmental impact assessment and a quick investigation of the design solutions space is extremely essential. To facilitate sustainable designs of buildings in conceptual and preliminary designs, this research attempts to develop an innovative optimisation framework, combining an advanced deterministic optimisation algorithm and a data-driven surrogate model. Overall, the framework is comprised of two main phases: the design optimisation phase and the surrogate Artificial Neural Networks (ANN) modelling phase. In the design optimisation phase, the carbon-minimised design problems are formulated in accordance with relevant Australian design standards and solved with deterministic Branch-and-Reduce algorithm. Three types of concrete buildings are investigated, namely flat plate, flat slab with drop panels, and beam-slab systems. To verify the effectiveness and reliability of the formulated problems and adopted algorithm, sample building problems are solved and compared with their conventionally designed counterparts. Accordingly, the optimised buildings have shown to be environmentally superior to the conventional designs, with a reduction in EC of 0.8-22.6%, 1.1-32.3%, and 1.8-26.6%, respectively for flat plate, flat slab, and beam-slab buildings. Regardless of the type of buildings, most of the optimised designs were solved within two days, demonstrating significant time efficiency. In the surrogate ANN modelling phase, hundreds of building optimisation problems with different structural heights, spans, and column grids are randomly generated and solved for minimum CO2 emissions. These numerical applications are subsequently used to develop ANNs for the predictions of optimal design solutions. The input variables are the basic information of a building, including the building height, numbers of spans, column spacings, and concrete strengths for slabs and columns. The outputs are the essential design solutions, namely the slab thickness, drop panel depth, beam dimensions, column size, amounts of reinforcement for slabs, beams and columns, and the resultant carbon footprint. Thousands of ANNs with different hyperparameters and configurations are investigated to determine the best performing models. The networks are evaluated based on three statistical metrics: the Root Mean Squared Error (RMSE), Mean Absolute Percentage Error (MAPE), and adjusted Coefficient of Determination (?̅2). Compared to statistical multiple linear regression models, ANNs have shown to possess outstanding prediction capability. Most of the best predicting ANNs produce highly accurate results with small RMSEs, MAPEs of less than 10%, and high goodness of fit (?̅2>0.9). While flat plate and flat slab buildings require only 1-2 days to tune the network, the tuning times of surrogate ANNs for beam slab buildings were 3-4 days, which is still short in comparison with the time frame available at initial design phases. Once the models are properly trained, they can predict the design solut ons in seconds. Given the reliability of the dataset generated from the optimisation phase as well as the high efficiency and accuracy of the developed ANNs, this innovative framework can assist structural engineers to deliver the most sustainable designs for entire buildings, especially in the short time frames of early design stages.