This project proposes a structural prediction tool tailored for the early stages of architectural design. Its goal is to provide architects with a structural assistant that offers meaningful insights, enabling them to make more informed decisions and, in turn, support their creative process from the outset.
DEMO
As part of the course, a website was developed that allows users to upload their 3D model for analysis. Once the model is uploaded, the system generates a suggested structural layout. The architect can review the proposal, delete any unwanted elements, and re-run the analysis to receive an updated configuration. Additionally, a chatbot is available to provide insights, answer questions, and assist with decision-making throughout the process.
PROBLEM VALIDATION
As is widely recognized, architects must collaborate with a range of specialists throughout the design process. Among the most critical—and often most challenging—is their collaboration with structural engineers. This relationship has long been marked by recurring issues within the AEC industry. To better understand and address this gap, we engaged in conversations with both architects and structural engineers to identify where they believe the core problems lie.
Interestingly, we found a shared sense of frustration between the two disciplines. Structural engineers often feel that architects propose designs that are structurally unfeasible, while architects feel forced to compromise their vision later in the process. Structural engineers emphasized that being involved earlier in the design phase would significantly improve outcomes and reduce such conflicts
Another noteworthy statistic, reported in the Construction Research Congress 2024 publication, is that 80% of projects experience significant delays due to a lack of early collaboration between architects and engineers.
Most structural engineering tools are geared toward the later stages of design and are primarily tailored to the needs of structural engineers, rather than supporting the architect’s perspective in the early design phase.
“A lot of research has been done in this domain. An interesting paper that helped us ground our logic in a functional approach is Automated Column and Beam Placement on Single-storeyed Convex Orthogonal Floor Plans (Conference Paper, November 2018)
SOLUTION
Our approach is twofold: we aim to provide a tool that respects the architect’s desire to experiment with form while also addressing the structural engineer’s need for a calculated and reliable methodology.
On the technical side, the process begins with the architect importing a 3D model, with the constraint that it must consist of orthogonal volumes. The pipeline starts with mesh segmentation, where walls and floors are automatically identified based on their geometry—specifically their size and surface normals. Following this, a structural grid is generated, guided by the maximum allowable spacing between columns, which is determined by the chosen construction material. These maximum distances are automatically retrieved from a curated expert knowledge database. Columns located beneath cantilevers are removed, as are those that end up being unnecessarily close to each other, to avoid overdesign. Finally, the system intelligently adds additional columns only where further structural support is needed.
As for the analysis component, a graph-based representation (RAG) is created. This structure retrieves information such as section types based on beam spans and other relevant material properties needed to evaluate the structural reliability of each member. The result is a finite element analysis (FEA) output in CSV format, which can be used to interface with structural engineering software or shared directly with a structural engineer for further development.
We conducted over 50 experiments using different appropriate models to validate the logic.
FUTURE SCOPE
The future scope of the project includes expanding to support more complex shapes and geometries, integrating a wider range of structural typologies, and exploring the use of Graph Neural Networks (GNNs), a promising and emerging trend in the field.
