Predicting perceptions of workplace design: a multi-attribute machine learning approach

The connection between workplace environments and worker experiences is a rapidly growing field, yet comprehensive evidence-based studies examining multiple spatial attributes remain limited. Unlike prior research focusing on isolated design elements, this study introduces a user-focused method that quantifies the relative and combined influence of 10 spatial attributes to perceptions of productivity, comfort, and social connection. Using computer-generated 3D office images created through parametric modeling, we developed a database of 1024 labeled scenes. Our digital equivalent of a controlled experiment included carefully designed “control” and “test” conditions, indicating the presence or absence of each attribute, such as window size and operability, visible sunlight and view features, types of furniture, indoor plants, and natural materials. Using a multi-image ranking survey, we crowdsourced 24,888 rankings from 2100 participants and derived a novel perception score for each image (a QM-score). We trained machine learning models (SVMs) to predict scores based on single and combined attributes. Larger windows and visible sunlight predicted higher scores across all perceptions. Window view depth was a stronger predictor of high scores than natural views. Wood floors emerged as a strong predictor of increased productivity and comfort. Multi-attribute models with combined terms outperformed single-attribute models, and the connection model achieved the highest prediction accuracy (68.8 %). This study advances understanding of how workplace design affects worker performance and well-being and introduces a scalable, AI-driven methodology for architects, facility managers, and workplace strategists designing effective office environments. It also lays the foundation for further investigations in evidence-based architectural design generation.