Buildings generate huge amounts of data, from the first design models to construction and decades of operation. But collecting data is one thing; turning it into useful information is another. Artificial intelligence could increasingly provide the missing link. That was the central message of Dr. Zhiyu Zheng, researcher at the University of Luxembourg, during the second edition of IngSci’s STEM Spotlight webinar series. Combining a background in mechanical engineering, data science and the built environment, Zheng presented a series of concrete examples showing how AI can support architects and engineers throughout the building life cycle. He was introduced by Anne-Marie Solvi (administrator IngSci) and Anne Marmann (director).
His approach is grounded in practical experience. Before focusing on research, Zheng worked on building simulation and modelling projects in France, including a data centre, where computational fluid dynamics was used to analyse airflow and optimise cooling, and a municipal swimming pool, where modelling helped assess energy consumption and indoor comfort. During his PhD, in collaboration with Schneider Electric, he also developed a data architecture designed to connect the physical building with its virtual representation and manage information across the building life cycle.
Today, at the University of Luxembourg, his research focuses on taking this approach a step further: using AI to extract more value from building data.
From a prompt to a building design
One of the most striking applications concerns the earliest phase of a project: design. Zheng has developed a plugin connecting a Large Language Model (LLM) with Autodesk Revit. Instead of manually creating every element of an initial design, an architect can describe the building in natural language: its purpose, number of floors, architectural style and other requirements. “The more details you give it, the more precise building it can generate,” Zheng explains.
The AI can combine these instructions with information retrieved from previous designs and transmit the result to Revit through its API. In Zheng’s demonstration, different design alternatives could be produced within seconds or minutes. The objective is not necessarily to replace the architect. Rather, AI can automate part of the repetitive work involved in translating an idea into a first digital model, leaving designers more time to explore alternatives and concentrate on creative decisions.
The same principle could be applied beyond architecture. Structural, mechanical or other engineering teams could build databases from their previous projects and allow AI systems to retrieve relevant solutions when developing new designs.
Automating carbon assessment
AI could also remove some of the tedious data-processing work currently required from engineers. Life Cycle Assessment (LCA) is a good example. Assessing the embodied carbon of a building requires information contained in a BIM model to be matched with environmental databases. In practice, this can involve considerable manual work because the two systems may use different names, units or classifications.
Zheng demonstrated another Revit plugin designed to automate much of this process. The software extracts information such as materials and densities from the BIM model and uses AI to align it with the corresponding information in an environmental database. An analysis that might otherwise involve significant manual processing can consequently be accelerated dramatically, with the engineer concentrating on verifying the results.
The principle could equally be applied to cost estimation: BIM components could automatically be matched with entries in price databases, again reducing repetitive manual work.
Checking what has actually been built
Once construction begins, a different problem arises: does the building being constructed actually correspond to the design? Here too, Zheng sees potential for AI. A construction site can be scanned to generate a point cloud and the resulting representation compared with the original BIM model. AI can then help identify discrepancies between the digital design and the physical construction.
The same combination of sensors, data and artificial intelligence could contribute to construction-site safety. Cameras or laser scanners could, for instance, monitor distances between workers and dangerous areas such as unprotected edges. An intelligent system could detect a hazardous situation and warn a worker before an accident occurs. These applications illustrate an important shift: AI is not limited to analysing data after an event. Combined with real-time data, it can potentially help engineers identify problems while there is still time to act.
Making buildings smarter in operation
The operational phase offers perhaps the largest reservoir of data. Modern buildings can continuously generate information on temperature, humidity, CO₂ levels, energy use and equipment operation. AI algorithms can analyse these data streams and optimise control strategies, particularly for heating, ventilation and air-conditioning systems. The aim is straightforward: use the information already generated by a building to operate it more efficiently while maintaining appropriate indoor conditions.
Zheng’s earlier work illustrates how relatively simple design decisions can also emerge from sophisticated modelling. In the case of the municipal swimming pool, excessive solar gains through a glazed façade affected the building’s summer performance. Rather than reducing the glazing, the solution involved extending part of the roof to provide more shading. Digital tools therefore do not necessarily lead to more complex buildings. Sometimes they can reveal relatively simple ways of making them perform better.
AI is only as useful as the data behind it
Yet there is an obstacle to this vision: building data remain highly fragmented. BIM models can be structured differently from one project to another. Environmental databases have their own classifications. Building-management systems come from different manufacturers and can use proprietary formats. Without a coherent way of organising all this information, each AI application risks becoming a bespoke solution. One answer is to develop knowledge graphs and ontologies capable of providing a more standardised representation of information. In other words, before companies can fully exploit AI, they need to think carefully about their data architecture.
This may be one of the webinar’s most important messages. AI does not magically solve poor data management. The ability to create value depends first on making data accessible, structured and understandable.
The right technology for the right problem
Zheng also cautioned against assuming that the current generation of Large Language Models is the answer to every problem. Classical AI and machine-learning techniques remain highly relevant. Predicting future energy consumption from time-series data, for example, may be better handled by specialised machine-learning models than by an LLM. Such approaches can also be less expensive and more efficient. The real challenge is therefore to choose the right technology for the right problem.
At the same time, LLMs offer an important new interface between people and complex digital systems: natural language. An architect no longer necessarily needs to express every instruction through specialised commands. A designer can increasingly describe what is needed in ordinary language and let the software translate that intention into actions.
From artificial intelligence to industrial value
There are still barriers. During the webinar discussion, one participant described the resistance encountered from IT departments when connecting AI-based optimisation tools to HVAC systems in existing buildings. The concern is not necessarily whether the technology works, but what happens to the building data and whether connecting an external AI system creates cybersecurity risks.
For Zheng, such questions show that deploying AI in buildings involves more than algorithms. Data governance, cybersecurity and organisational acceptance will also be essential. This is precisely where cooperation between research and industry can play a role. At the University of Luxembourg, SnT works with industrial partners in fields ranging from the built environment and manufacturing to autonomous systems, mobility and finance, drawing on expertise in AI, data science, software engineering and cybersecurity.
For the construction sector, the opportunities are already becoming tangible. AI can help generate early designs, automate carbon and cost assessments, compare construction with design, identify safety risks and optimise buildings once they are occupied.
The technology is advancing rapidly. But Zheng’s examples suggest that the real transformation may come not simply from having more powerful AI, but from connecting it intelligently to the vast amount of information that buildings already produce. The goal, ultimately, is exactly what the title of the webinar suggests: turning data into value.
- Download Zheng’s présentation: https://ingsci.lu/davinci/wp-content/uploads/2026/09/snt-stem.pdf
- Video of the webinar
