Role of Data in AI-based MEP modelling

AI in the Construction industry

Recent years we have witnessed growth in the utilisation of AI within the construction industry. The emergence of LLMs (Large Language Models, for example ChatGPT) has significantly contributed to this expansion. LLMs have demonstrated exceptional capability in creating and analysing building process-related documents, offering substantial help in project management. In addition to LLMs, more traditional AI tools can be applied throughout the building process from the early design phases to hand over to customer. For instance, during early phases of design, AI can help identify choices that optimise buildings sustainability. During the construction, AI can monitor working environments to ensure they meet safety regulations. After the building is handed over, AI can manage building systems to maintain good indoor climates and energy efficiency.

While it may seem like AI is used everywhere in the building process, there are still areas where its use is limited. One such area is MEP (Mechanical, Electrical and Plumbing) modelling. Although the first studies on automatic routing of pipes date back to the early 1990s, there is still no commercially available tool for large scale automatic routing of pipes and ducts in BIM (Building information modelling). So far, the studies of automated routing have mostly focused on various path planning algorithms with limited exploration of AI’s potential in this domain. Could AI be the key to bringing automatic routing tools into hands of designers?

The challenges of AI in MEP modelling

AI holds promise, but a major challenge hinders its full potential in MEP modelling: Data. If we consider the AI applications in building industry mentioned above, they all share one factor in common: the data used in these applications is typically owned and collected by a single party. For example, building automation systems collect indoor environment metrics and the adjustment of valves which can be used for training AI models. Construction companies can capture images from unsafe site and train a model to classify the site’s safety. With LLMs, pre-trained models can be accessed via APIs to extract information from extensive documents. In all these scenarios, the necessary data, the fuel of AI, is achievable by the company using the tool. Nevertheless, data collection in these fields still requires an effort. 

Data in AI driven MEP design

While data challenges are common in AI, they are unique in context of BIM (Building information modelling). This is because the ownership of designs is fragmented. To create a complete model, you need multiple different designs (Architectural, structural and MEP), usually provided by different companies. Additionally, the building owner owns the building and its plans, so their permission is required to share the designs. As a result, multiple parties need to agree on sharing the designs. In addition, the benefits of AI research and data sharing may not be clear to all stakeholders, as automated modelling would primarily benefit the design offices. 

If the necessary data is unavailable, it is possible to create it. However, manually generating thousands designs to train AI is not practical. Automatic way would make it feasible, but this approach presents its own challenges. First, you need a method to draw networks in 3D environment. Then, you need to have a way to assess the quality of these networks. With thousands of designs, how do you train a machine, to distinguish which design is superior? Computers are good with numbers, so you could score the solutions, but you still need a function to calculate the scores. Currently there is no mathematical function that can determine “best possible” network, considering factors like maintenance, install ability, lifecycle costs etc. For example, in engineering of bridges, the adoption of AI is far ahead of MEP modelling presumably because the problem is mathematically understood better. Moreover, In MEP modelling creating the route alone is not enough as you need the routing environments as well. 

Solution for data issue at MagiCAD Group

At MagiCAD Group we have been researching AI-driven modelling for about three years. Our first AI-driven router is now available in MagiCAD for AutoCAD 2025 software solution which can generate UFH (Underfloor Heating) circuits. We have addressed the data challenge with “synthetic data,” which means we generate data ourselves. We first create example environments and designs and use these in the training process. However, synthetic data limits our focus on relatively small routing tasks with limited variation, as the environments and designs need to be validated. Real data could help us to generalise the routing process more effectively, covering more scenarios. In addition, it diminishes the need of synthetic data and therefore speeding up the development. 

Conclusion

In conclusion, AI has made a significant impact in the construction industry, offering valuable assistance in many tasks. That said, one area that remains underdeveloped in AI adoption is MEP modelling. One reason is the fact that complete building model consist of multiple models, created by different companies making it challenging to obtain real-life examples of building designs. The problem can be overcome by generating synthetic data, but its quality and scope are limited, and the mathematical understanding of the problem is still evolving. In addition, generating data is time-consuming which delays the development significantly. By generating a common dataset through collaboration with multiple industry stakeholders, we could boost the AI research in BIM and speed up the development of AI modelling tools.