Enhancing visual comfort prediction in space layouts using deep learning models

In designing performance-based facilities, such as educational buildings, assessing visual comfort is crucial. The computational cost and resource-intensive modeling necessitate an efficient analytical tool for diverse layout assessment. In the current study, predictive models were developed using deep learning and machine learning to predict visual comfort in typical elementary schools in Tehran. Firstly, layouts were modeled, considering influential parameters. Subsequently, simulation-based datasets were analyzed and labeled. The VGG16 and VGG19 Architectures from convolutional neural networks (CNNs), along with the pix2pix model from generative adversarial networks for forecasting respectively, numerical and pictorial indices regarding each visual comfort metric. The pix2pix model performs approximately SSIM of 0.9 for sDA, ASE, and UDIexceed3000 and 0.86 for DAp. The extracted features by CNNs were harnessed in training models. Eventually, models with the Bayesian Ridge algorithm had a promising performance which exhibited acceptable R-2 values for metrics in the range of 0.90-0.96. Toward a depth analysis, parameter sensitivity using the Shapley additive explanations method was evaluated with XGBoost models. Additionally, the Spearman Correlation underscores the substantial impact of factors like WWR and aspect ratio on metrics.