WindowGraphNet: Graph Neural Networks for Daylight Factor Prediction

Abstract

Assessing daylight in architecture increasingly depends on Radiance-based simulations that, although highly accurate, are too computationally demanding for rapid iteration during early-stage design. As layouts shift, openings are repositioned, or geometries become more complex, the turnaround time of physically based simulation limits the designer’s ability to explore alternatives. Surrogate models offer a promising solution, yet most existing approaches rely on artificial neural networks (ANNs) trained on tightly controlled parametric datasets and evaluated only within their training distribution. As a result, their reliability for real, highly variable design geometries remains uncertain.

This thesis investigates graph neural networks (GNNs) as a new surrogate modelling paradigm for daylight prediction. Unlike vector-based ANNs, GNNs represent spatial relationships explicitly, allowing the model to “reason” about how light propagates through space rather than inferring patterns solely from coordinate inputs. To the best of the author’s knowledge, this is the first formulation of daylight factor (DF) prediction as a graph-learning task, representing rooms as heterogeneous graphs in which window and sensor nodes exchange geometric and photometric information such as distance, direction, and relative orientation.

A deliberately restricted training setup, limited to simple square rooms, is used to test whether GNNs can generalise beyond the conditions they are trained on. The models are then evaluated on a wide range of unseen geometries, including elongated, asymmetric, rotated, and self-occluding layouts, alongside ANN baselines from the daylight literature. The proposed model, WindowGraphNet, consistently preserves the spatial structure of DF distributions across these unseen cases, whereas coordinate-anchored ANNs often fail when room form or orientation deviates from their training domain. The GNN’s main limitation occurs in deeply recessed or fully occluded areas, where long light paths fall beyond
the model’s message-passing range, leading to systematic overestimation of low-illumination regions.

These findings demonstrate that generalisation in surrogate daylight modelling is governed primarily by the representation: embedding relational and geometric structure enables robustness that conventional ANNs cannot achieve through data alone. WindowGraphNet therefore establishes a new methodological foundation for surrogate daylight tools, with an extensible graph structure that can be expanded to incorporate surface nodes, external obstructions, or climate-based inputs. With further training on
diverse datasets, such graph-based surrogates have the potential to deliver fast, geometry-aware daylight feedback directly within architectural design environments, supporting more informed and iterative decision-making in early-stage design