BIMScript: Material-Aware Structured Scene Programs for BIM Ingestion
BIMScript introduces a material-aware structured language model that reconstructs scenes into editable, semantically explicit programs with precise geometry and material attributes, achieving high-speed generation and direct, end-to-end ingestion into BIM tools like Revit for sustainable asset management.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
The built environment around us—the offices we work in, the homes we live in, and the schools we attend—is responsible for a massive share of global material use and carbon emissions. Most of the buildings that will exist in the year 2050 are already standing today, which means the future of our planet depends heavily on how we manage, renovate, or eventually take apart these existing structures. To do this effectively, architects and engineers need a precise digital map of what is actually there. This map is called a Building Information Model, or BIM. Unlike a simple photograph or a basic 3D drawing, a BIM is a smart, detailed description of every wall, door, and window, including what they are made of and how worn they are. Creating these models for existing buildings is currently a slow, tedious job that requires surveyors to scan a room and then modelers to manually trace every single object by hand, a process that can take days or weeks for a single building.
Researchers at Aalborg University in Denmark have developed a new approach to speed this up, turning the slow manual work into an automatic process that reads a building like a computer program. They call their system BIMScript. Instead of trying to build a mesh of triangles to represent a room, their system looks at a 3D scan of a space and writes a short list of instructions, much like a recipe, that tells a computer exactly how to rebuild the room. These instructions say things like "make a wall here," "put a door there," and "finish this wall with wood paneling." The breakthrough is that the system doesn't just guess the shape of the room; it also figures out the materials and the condition of every object, such as whether a wall is painted plaster in good shape or if a door is made of composite material that is new. This level of detail is crucial because knowing what a building is made of determines how much energy it takes to heat, how much carbon was used to build it, and how easily it can be recycled.
The team faced three major hurdles to make this work. First, they needed the system to know what things were made of, not just where they were. Previous tools could draw a wall but couldn't tell if it was brick or drywall. To solve this, the researchers created a massive library of over 1.9 million examples of building parts, where every piece was labeled with its material and condition. They trained their system using a powerful artificial intelligence that could look at photos of the scanned rooms and guess the materials, effectively teaching the computer to recognize the difference between wood paneling and wallpaper just by looking at the texture and color in the scan.
Second, the process needed to be fast enough to be useful. The old way of generating these instructions took several seconds per room, which is too slow for large projects. The researchers discovered that the delay wasn't because the computer was doing too much math, but because it was wasting time starting and stopping tiny tasks over and over again. By reorganizing how the computer handled these tasks, they cut the time needed to generate the instructions by more than three times, bringing it down to roughly one second per room. This change allows the system to process entire buildings in a timeframe that makes interactive design possible.
Third, the system needed to be precise enough to be trusted by professional builders. The computer initially placed walls and doors on a grid where each square was 5 centimeters wide. While this is good for a rough sketch, professional construction requires millimeter-level accuracy. The researchers added a final step where the system makes tiny adjustments to the position of every element, fine-tuning the location so it fits perfectly within the scan. They tested this by feeding the generated instructions directly into professional building software used by architects. The software successfully created real, usable 3D models with the correct materials and dimensions from synthetic data, proving that the computer-generated instructions are accurate enough to replace manual tracing in controlled environments.
The result is a system that can take a raw 3D scan of a building and turn it into a smart, detailed digital twin in about a second. This digital twin is not just a picture; it is a structured list of facts that a computer can read and understand. Because the building is described in a language that both humans and machines can read, it opens the door for artificial intelligence to help with complex decisions. For example, a city planner could ask the system to find all the wooden doors in a district that are in poor condition, or calculate the total carbon footprint of a building's materials without ever visiting the site. The researchers validated that this method works on synthetic data and that the output can be exported to standard industry formats, though generalization to real-world scans remains an open challenge. By automating the creation of these detailed models, the system offers a practical way to understand, manage, and sustainably improve the vast stock of buildings that already exist, turning a slow, manual chore into a fast, intelligent process.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.