CAREER: Active Designer-Computer Collaborative Conceptual Design Using Data-driven Models

This Faculty Early Career Development Program (CAREER) grant supports research that will create, evaluate, and disseminate new methods for incorporating decision support through artificial intelligence during early building design. The project addresses challenges in structuring and managing the data used to train predictive models that augment design, reveal relationships between key early design decisions and the predicted performance of the building, and communicate suggestions for improving the design in a way that is intelligible and actionable for professionals in architecture and engineering. Parametric modeling is increasingly used in architectural engineering to consider the performance benefits of design options through early simulation. Benefits of exploring different building geometries, configurations, and systems include reduced energy use and embodied carbon as well as increased daylight availability. However, common methods for systematically searching through design options have historically required specialized modeling knowledge, long simulation times, and automated optimization algorithms that are difficult to control. Such workflows can be tedious to formulate and often fail to deliver clear, actionable directions during a natural design process. In response, this research will create and interrogate a system that learns the anticipated behavior of building forms, predicts these values for a new form, and provides suggestions for significant drivers and how to improve an initial design. The research will advance interactive parametric design while reducing repeated simulation, a strategy which could promote national welfare by generating high-performance, sustainable future buildings. Guided by an industry advisory board, the integrated research and education component of the project involves educating pipeline and undergraduate design students as well as mid-career architectural engineers in the use of computational tools in design space exploration. The research plan involves curating building data, using it to train flexible machine learning models that are widely applicable in design, and then exploring and testing new methods for interacting with such models during the design process. The first step is to create and consolidate digital libraries of building forms and corresponding performance models at early design resolution and then segment them into typology-specific “design spaces." The libraries will come from existing datasets, models of fully designed buildings, and self-generated simulation data. Next, the design models will be reduced to representative features that are universal for each design space segment. The data will then be used to identify performance patterns and train surrogate models that predict key performance metrics for new forms based on the underlying data structures, which can be updated as more data is generated. Finally, several methods will be synthesized into a new flexible design procedure to read in potential geometries, visualize expected performance, and provide improvement suggestions while combining sensitivity analysis, dimensionality reduction, and interactive optimization. Design suggestions will be based on simplified design latent spaces and user preferences, enabling true, interactive human-computer collaboration. The methods will be validated by comparing prediction results to full simulations and databases of existing buildings, as well as an initial design study of the interface assessing designer behavior while engaging with design suggestion. Interactive computational tools will be integrated into coursework to develop student intuition about building behavior and help them think creatively and synthetically during technical design tasks. Students will participate in programs to educate practicing engineers in computational thinking, parametric design, and optimization, which are critical to their engagement with an industry that increasingly manages design information and makes decisions using code.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.