Rapid structural analysis of prefabricated thin concrete shells using deep learning

: (Alternative Format Thesis)

Abstract

Portland cement, which is the key constituent of concrete, is responsible for approximately 7% of global CO2 emissions. A key target for concrete carbon reductions are floors, which typically contain most of a building’s material mass. To this end, researchers have proposed the use of concrete thin-shells as a flooring system. Using this strategy, the Automating COncrete constRuctioN research project, with which this research is aligned, achieved embodied CO2 emission reductions of up to 48% compared to a flat reinforced concrete slab (37% when the flat levelling structure, which is necessary for concrete thin-shells, is accounted for). However, the use of such systems in practice remains limited, due to the complexity of their design and production. Concrete thin-shells can be prone to buckling, a failure mechanism that necessitates careful consideration, using computationally expensive simulations such as nonlinear Finite Element analysis. This high computational cost prevents the accurate assessment of the buckling behaviour of concrete thin-shells in several situations. During design, the iterative nature of shape optimisation algorithms prevents the use of computationally expensive tasks. As a result, the buckling behaviour is generally only assessed after a promising shape has been pre-selected, which limits design space exploration. During production, the impact of geometric imperfections on the buckling behaviour of concrete thin-shells can only be assessed once concrete has hardened, which prevents rework and can result in waste. Even though concrete is poured in liquid form and could therefore be reworked, the setting time of concrete does not allow for the structural assessment to take place before it hardens.

This thesis proposes a novel approach, based on deep learning surrogate models, to rapidly and accurately estimate the structural behaviour of concrete thin-shells.

To this end, three computational experiments, which test the proposed solution in different settings, were carried out. In the first experiment, a dataset of 20,000 thin-shells with varying spans, heights, thicknesses, and material properties was generated, and their structural behaviour under design loads assessed using linear Finite Element analysis. This data was then used to train and compare three different types of deep learning model – Multilayer Perceptron, Convolutional Neural Network, and Graph Neural Network – for buckling and stress field prediction. In the second experiment, a novel framework for generating geometrically imperfect thin-shells was proposed and used to generate 20,000 imperfect thin-shells. Their structural behaviour under design loads was then re-assessed with linear Finite Element analysis. Multilayer Perceptrons and Convolutional Neural Networks were then trained and compared for the prediction of the buckling factor and stress fields. In the third experiment, a dataset of 5,000 thin-shells without geometric imperfections was generated, and their buckling factors under design loads assessed using geometrically nonlinear Finite Element analyses. To mitigate the computational cost required to generate datasets, a novel Mixed-Fidelity approach that uses linear and nonlinear buckling simulation data for model training was proposed and compared against a baseline Multilayer Perceptron.

The results obtained in the first and third experiments show that deep learning surrogate models are highly accurate at estimating the structural behaviour of as-designed concrete thin-shells, while being many orders of magnitude quicker than linear and nonlinear Finite Element analysis. The models trained in the second experiment are less accurate, reflecting the increased variability of the dataset. While the proposed dataset generation framework appears promising, further research is required to obtain more accurate models. The thesis demonstrates that deep learning surrogate models are a viable approach to estimate rapidly and accurately the structural behaviour of concrete thin-shells. As such, they can facilitate shape optimisation and enable wider design space exploration.

Date of Award	25 Jun 2025
Original language	English
Awarding Institution	University of Bath
Supervisor	Paul Shepherd (Supervisor), Will Hawkins (Supervisor) & Eduardo Costa (Supervisor)