A computational framework for evaluating sensory-sensitive buildings using graph-based artificial intelligence and evolutionary algorithms

Al-Harasis, Dania ORCID: https://orcid.org/0000-0001-9102-8882 2026. A computational framework for evaluating sensory-sensitive buildings using graph-based artificial intelligence and evolutionary algorithms. PhD Thesis, Cardiff University. Item availability restricted.

	PDF - Accepted Post-Print Version Restricted to Repository staff only until 7 September 2027 due to copyright restrictions. Download (73MB)
	PDF (Cardiff University Electronic Publication Form) - Supplemental Material Restricted to Repository staff only Download (1MB)

Abstract

Designing spatial environments for individuals with autism requires control over how spaces connect, separate, and sequence. These requirements are documented in guidelines and peer-reviewed research but have never been computationally encoded. No reproducible framework translates autism-friendly spatial requirements into machine-readable constraints for graph-based layout evaluation and repair. This thesis addresses that gap by developing a five-phase computational framework that formalises autism-friendly design knowledge into quantitative spatial constraints. These constraints are then integrated within a generative and predictive tool for evaluating and repairing graph-encoded architectural layouts. Phase One extracted 80 design indicators and classified 37 as Spatial Design Qualities across 11 thematic clusters and three spatial levels of impact. Phase Two translated these into four computable specifications: a weighted adjacency matrix, occupancy-dependent geometric parameters, external wall requirements, and a Malleability-Rigidity prioritisation scale. Phase Three employed a hybrid Genetic Algorithm and Simulated Annealing framework to generate a labelled, graph-encoded synthetic dataset of 17,839 spatial layout graphs across 263 scenarios. Phase Four trained and evaluated 3 Graph Neural Network architectures against 3 non-graph baselines under a scenario-aware splitting protocol, with GraphConv achieving R² = 0.9735 on the held-out test set and R² = 0.9553 on out-of-distribution validation. Phase Five integrated all preceding outputs into a prototypical tool processing IFC-based floor plans through semantic normalisation, benchmark-guided alignment, and constraint-guided graph repair. Evaluated across ten projects under 200 experimental runs spanning four stress-test clusters, the tool produced benchmark-consistent graphs with space-by-space area compliance under concurrent semantic failure, absent connectivity, and geometric distortion. The framework demonstrates that autism-friendly design knowledge can be formalised and operationalised within a tool that evaluates and repairs spatial layouts at the graph level. The framework is scoped to autism centres and operates on graph-based representations rather than geometric drawings. Constraint-compliance scores reflect adherence to formalised spatial constraints derived from the literature.

Item Type:	Thesis (PhD)
Date Type:	Completion
Status:	Unpublished
Schools:	Schools > Architecture
Date of First Compliant Deposit:	8 September 2026
Last Modified:	08 Sep 2026 11:21
URI:	https://orca.cardiff.ac.uk/id/eprint/189440

Actions (repository staff only)

Edit Item