Development of a Modular Software Framework for Supporting Architects During Early Design Phases
Early design phases in architecture deal with the conceptualization of a building. During these phases, a rough floor plan layout is designed by an architect based on a high-level description given by a contractor or customer. Traditionally, these phases involve a lot of monotone and repetitive labor. For example, manual research in architectural libraries or dedicated magazines is carried out in order to retrieve reference concepts that may serve as inspiration or to probe the feasibility of the project. Likewise, some aspects of a floor plan design are highly creative and some parts remain rather predictable. One established working method in architecture for turning a high-level description into a specific floor plan layout is the so-called room schedule, in which a set of individual rooms is the focus of interest. Based on this room schedule working method, a smart framework is developed to assist architects with their work: A sketch editor allows for specifying a floor plan concept with a varying degree of abstraction, thus allowing the user to specify every aspect of a floor plan concept as abstract of specific as desired. The sketch editor follows the room schedule working method and aims to support the architect during the entire early design phase. Simultaneously, this sketch editor serves as an input tool for search queries with which the user can search for similar floor plan concepts. The presented framework allows for integrating an arbitrary number of such retrieval systems that utilize a dedicated floor plan database. Currently, three different retrieval systems are integrated, one of them is presented in greater detail here. These retrieval systems rely on so-called semantic fingerprints, which are graph-based abstractions of floor plan concepts. Hence, subgraph matching is employed to these graph-based abstractions of floor plans. Likewise, these semantic fingerprints allow the user to control the retrieval process. In order to accomplish explainability in the result finding, a mapping function is incorporated in the framework that calculates for all found results, how individual rooms in the search query map to individual rooms in the search results. Finally, the sketch editor is connected to a neural network-based predictor that generates automatic suggestions for solving creative problems. These suggestions may perform entire design steps. This helps the user to avoid repetitive and predictable actions like completing a floor plan layout. Likewise, it offers inspiration and provides templates that can be refined by the user. In such a setting, the user and the artificial neural network are in a loop, both manipulating the sketch. In order to evaluate the proposed solution, individual aspects of the framework are tested by different means to demonstrate their usefulness. The user interface is tested by user study that compared it to the traditional approach. The retrieval system is tested by both a stress test and qualitatively analysis. Finally, the design suggestion is both evaluated by established means of artificial intelligence and examined qualitatively.