Data first: Training AI solutions for MEP design


MagiCAD Group


Recent developments in AI have quickly made the technology and its various applications a part of our personal and professional lives. At MagiCAD Group we are constantly looking to advance MEP design with the help of the latest technological innovations.  

The question of data 

The immediate benefit of AI technology is the ability to handle and evaluate data at a volume that is far beyond human capability. This holds the promise of more purposeful human work and automated routine work.  

However, in order to develop this capability, AI solutions first need large samples of training data, and this is not readily available. In specialist fields, such as MEP design, this issue becomes even more emphasized. This is why data openness and data sharing may ultimately be needed to speed up the development. 

Overcoming lack of training data with automatic floorplan creation 

Our recent research into potential applications of AI technology came face to face with the key issue of training data. In our case this meant floorplans, since the development of robust machine learning models for MEP design demands exposure to a wide range of floorplan scenarios that are not readily available.  

The scarcity of training material spurred a project that aimed to automate the creation of floorplans—a traditionally manual and resource-intensive task for architects. Although, this was possible in the case of 2D rooms, it would have been far more difficult to do for 3D environments. 

The benefits of diffusion based generative models  

While Generative Adversarial Networks (GANs) have commonly been the method of choice for similar purposes, our project settled on employing diffusion models. Diffusion models employ an iterative process to refine noisy samples, resulting in a stable training environment and the generation of high-quality floorplan images.  

A diffusion-based model for creating floorplans stands out as a multi-purposed solution, simultaneously capable of generating intricate and detailed floorplans from scratch and filling in missing information in existing floorplans through inpainting and predicting room labels. 

Whereas existing work on floorplans has often prioritized room layouts and neglected fenestration and furniture details as irrelevant for end-users, a diffusion-based model retains these elements producing results that are more consistently realistic. This made the method particularly suitable for our purposes.  

Testing the model  

To verify the validity of our approach, sample floorplans were assessed by two user groups, including individuals with architectural expertise and those without. User response was generally positive, and the model floorplans were considered to be sufficiently realistic.  

Although our current diffusion-based model for creating sample room layouts has its limitations when it comes to adjusting design parameters, the impressive performance with modest training data and default parameters indicates that the approach has considerable potential.   

Our floorplan creation project clearly points to the potential of AI and machine learning to streamline and enhance MEP design processes. By keeping pace with new technology and fostering innovation, we strive to deliver customers with the best solutions that meet the evolving needs of the construction industry. 

Examples of floorplans generated by our model. 

Samples of floorplans produced by fixing one part of a floorplan or the outer walls of a floorplan