This practice-led research explores the problem of multi-objective and multi-scale optimisation of modular stadium design based on the quantitative assessment of its performance. Stadiums are one of the largest and more challenging building genre to construct, with its design process traditionally separated by different disciplines. This can prevent a holistic understanding of an interconnected system of drivers that determine the design and performance of stadiums, which is especially challenging in stadiums designed for more than one function. This research examines the benefits and limitations that machine learning might bring to the design process, focusing on the
spectator experience in relation to sightline quality, internal layout, and circulation. The research asks:

1. How can a combined functional and structural stadium design approach based on a modular and parametric design model improve the design process and optimise the performance of multi-use stadiums?

2. How can multi-use stadium design be analysed as a problem of sightline quality, the optimisation of circulation distances and the provision of legally regulated functional spaces and captured in a parametric design model?

3. How can the proposed design method for multi-use stadiums, requiring multi-objective optimisation and performance assessment, benefit from machine learning-based data analysis?

The first part of the research uses machine learning to analyse the spatial and organisational properties of stadiums to identify their key design drivers. Classified by the UK Building Regulations as places for assembly and recreation, the thesis focuses specifically on modular stadiums. It contextualises these through a study of their historical development and increasing design and performance legislation. A typological study, along with the literature review, are used to determine a set of shared parameters that define modular stadium design. Sightline quality, circulation, and layout are found to be critical to the spectator experience, as they define the slope of the seating stands, crowd flow and escape distances, and the overall building organisation and morphology.

The second part of the research uses these design drivers to develop a parametric model to simulate stadium designs. This model captures the typical modular scales of stadiums: 2D section, 3D bay unit, stands, and whole stadium. The model is used to generate, analyse, and compare existing stadium designs and to establish performance benchmarks for key design metrics. As the parametric model is based on shared typological characteristics, the data produced is not detailed enough for construction but useful for the comparison of design performance criteria.

In the third part of the research, the RIBA Plan of Work stages are used to develop a set of workflows to test the parametric model and relevant performance criteria and design optimisation. To overcome the lack of linear relation between the parametric inputs and the resulting performance, the study explores three distinct workflows. These test a multi-objective evolutionary optimisation, semi-supervised learning, and supervised learning approach.

The aim of combining typological analysis, parametric modelling, and an utilisation of the workflows is to facilitate, optimise, and enrich the design process through the study of both existing and previously unattainable design solutions. The research is relevant to both practitioner and academic audiences by expanding the current use of machine learning in design development and evaluation.