Lotte Kat – Visual comfort l(AI)outs

Winnaar Dutch Daylight Student Award 2024
Project type: Research

Visual comfort l(AI)outs

Lotte explored the use of machine learning to ensure daylight compliance in residential apartments. By training algorithms on a dataset of Swiss apartments’ daylight performance, she developed a proof of concept demonstrating how machine learning can filter out non-compliant designs. This allows designers to focus on creative aspects rather than the tedious task of checking regulations.

• Topic: machine learning for daylight access compliance of residential apartment.
• Method: training of machine learning algorithms on existing dataset of daylight performance in Swiss apartments
• Impact: proof of concept of how ML can help designers with the ‘boring stuff’ by filtering out only compliant designs and let the designer focus on design

Optimizing the layout of residential buildings based on daylight performance and view quality is crucial for the visual comfort and well-being of building occupants. Natural daylight not only enhances visual comfort but also reduces the need for artificial lighting. Additionally, combining good natural (day)lighting with views of greenery can significantly improve the health and well-being of building occupants. The interior layout and orientation of rooms play a crucial role in maximizing exposure to daylight and greenery views, enhancing overall quality of life, and reducing energy consumption. The EN17037 guideline outlines essential standards for daylight and view quality in residential spaces, ensuring that specific room requirements are met.

Despite these factors, there is currently a lack of Machine Learning (ML) methods to support designers in making informed decisions regarding early interior design decisions that affect daylight and view quality. ML methods offer valuable support for performance-based decision-making process at the early-stage building design. A novel workflow has been introduced to integrate ML models into the architectural design process. With the designer’s input floor layout designs, the ML models predict daylight provision and view quality, which are then translated into practical visual representations by a post-processing step. This approach allows input designs to be evaluated by the ML model, leading to improved design decisions while maintaining the designer’s independence.

A multimodal ML model, which combines a ResNet50 and a fully connected network, has been trained using one image and five numerical features to predict daylight illuminance and view quality. The best-performing model achieved a Mean Square Error (MSE) of 0.0440, and an R2 score of 0.7411 and 0.7815 for the daylight and view metrics, respectively. The results of the predictive models for daylight and view are further analysed for different apartment categories and at various resolutions. A ranking method is used to identify the best apartment layout based on specific requirements. This framework represents a significant advancement in integrating ML models into architectural workflows. It systematically evaluates daylight, view quality, and room orientation, providing quick visual feedback and offering performance-based ranking that aligns with contemporary design standards and requirements.