Artificial neural networks for parametric daylight design

In parametric design environments, the use of Artificial Neural Networks (ANNs) promises greater feasibility than simulations in exploring the performance of solution spaces due to a reduction in overall computation time. This is because ANNs, once trained on selected input and output patterns, enable instantaneous predictions of expected outputs for new unseen input in the recall mode. In this study, ANNs were trained on simulation data to learn the relationship between design parameter and the resulting daylight performance. The ANNs were trained with selected input-output patterns generated from a reduced set of simulations in order to predict daylight performance for a hypercube of design solutions. This work demonstrates the integration of ANNs in a case study exploring designs for the central atrium of a school building. The study discusses the obtained design results and highlights the efficacy of the proposed method. Conclusions are drawn on the advantages of brute-force based daylight design explorations and the potential of an ANN-integrated design approach