ACSA Intersections Research Project

Project Description:

Abstract: The rise of generative artificial intelligence (AI) and diffusion models has redefined architectural representation, augmenting aesthetic speculation and architectural form variability [1], while simultaneously destabilizing established design epistemologies by foregrounding questions of design agency, and the ontological status of the image in performatively driven architecture design production. While those AI models introduce novel aesthetic imaginaries, they remain detached from aspects of building physics as well as climatic and contextual logics, essential to environmental performance evaluation. Their visually compelling outputs lack fidelity, often producing representational illusions, or hallucinations, rather than performative accuracy. By ignoring site-specific parameters, such as orientation, latitude, as well as spatial parameters such as material reflectance, AI-generated designs compromise environmental evaluation accuracy; there is a need for integrating data-driven feedback loops which are fundamental to evaluate design strategies. Thus, AI-generated design must be reframed not as an autonomous outcome, but as a speculative scaffold requiring rigorous pre and post-processing through validated simulation methods to ensure performative consideration. Central to this inquiry is the question: how to develop a method framework for designers to follow, where AI-driven design is informed by environmental logics of daylighting, and how would the framework allow for sustaining successful concept development throughout the generative process? In response, this work is aimed at leveraging the use of diffusion models, with their strategic integration into a hybrid design workflow in which validated daylight simulations are also incorporated at different design phases. In our introduced framework, to consider environmental performance, the design generation is grounded in climate and context-specificity, with geometrically and materially detailed 3D models that serve as the analytical counterpoint to AI-generated designs. Daylight, as a fundamental form-giver in architecture [2], holds immense potential to inform generative design exploration in early phases, and thus was used as an apparatus for space making. Within this framework, “GEN-Daylight”, daylight simulation serves as a design tool along with utilizing generative AI tools of Lookx.ai [3] and Midjourney AI [4]. Both AI platforms support multi-modal inputs including text, sketches, orthographic drawings, and snapshots of massing models. Our method framework enables dynamic interaction between the designers and those model outputs across analytical, exploratory, and development design stages. Particularly, the Lookx.ai tool allows access to pre-trained models alongside custom model training and fine-tuning, further supporting custom and representation objectives. The framework was implemented and critically evaluated within a coordinated design studio over two consecutive years, engaging 115 students in a design-research environment. Following our proposed design workflow, students were able to pursue articulation, interpretation, and refinement of architectural design intentions through strategically curated inputs, along with rationalization and post-processing, as well as engaging in AI-driven and daylight simulation feedback loops. GEN-Daylight sustained conceptual coherence while leveraging the generative affordances of AI, positioning the designer not as a passive user of algorithmic output but as an active choreographer within a hybrid design ecology. By cultivating a productive tension between speculative imagination and analytical rigor, the framework advances a situated design intelligence, one that is creatively agile, and environmentally attuned.