| id |
acadia23_v2_520 |
| authors |
Ampanavos, Spyridon; Bernal, Marcelo; Okhoya, Victor |
| year |
2023 |
| title |
Daylight ML: A General-Purpose Deep-Learning Surrogate Model for Annual Daylight Distribution |
| source |
ACADIA 2023: Habits of the Anthropocene: Scarcity and Abundance in a Post-Material Economy [Volume 2: Proceedings of the 43rd Annual Conference for the Association for Computer Aided Design in Architecture (ACADIA) ISBN 979-8-9891764-0-3]. Denver. 26-28 October 2023. edited by A. Crawford, N. Diniz, R. Beckett, J. Vanucchi, M. Swackhamer 520-531. |
| summary |
Building performance simulation, such as daylight evaluation can lead to better quality designs. However, time constraints are currently limiting its use for design exploration. Surrogate modeling can offer drastic speed improvements to simulation processes, but existing models are either project specific or offer limited flexibility to design inputs, while requiring a significant initial investment for their training. This research introduces a method for predicting spatial distribution of annual daylight metrics using a raytrac- ing-based encoding of the inputs, and a deep-learning surrogate model. The method can operate on spaces of any shape. Using synthetic data, surrogate models for Atlanta, Georgia, and Boston, Massachusetts, were trained, and achieved low average errors on the test set for all daylight metrics considered. Furthermore, models trained on simple datasets of rectangular spaces were able to predict accurate results for L-shaped, circular, and courtyard-shaped spaces, and for sensors that had twice the density of the ones in the training set. Overall, the results suggest that trained models can be used to evaluate the daylight quality of any project or design within their respective locations. |
| series |
ACADIA |
| type |
paper |
| email |
|
| full text |
 file.pdf (2,967,667 bytes) |
| references |
Content-type: text/plain
|
| last changed |
2024/12/20 09:13 |
|
