AUSTIN, Texas – The large language models that underlie generative AI offer an opportunity to help everyone involved in buildings — architects, engineers, construction professionals and facility managers — access building systems that converse with one another on optimizing development, maintenance and compliance issues.
“It’s time for us to think about not looking at predictive buildings, but more towards conversational buildings,” Nan Ma, assistant professor of architectural engineering at Worcester Polytechnical Institute, said at the ASHRAE annual conference this week.
Ma, founding director of the Laboratory for Healthy, Environmental, and Resilient Buildings, or HERB-Lab, at WPI, led a group of global researchers on a Building and Environment study earlier this year that looks at how LLMs can help make sustainable, intelligent and human-centric buildings a reality.
“Things like energy use, comfort, occupancy, maintenance logs, control behavior [and] control logics” are the kinds of factors that LLMs could process to enable building systems to better help managers maintain operations better, she said.
For example, a building manager, maintenance team or facility manager could ask an LLM why the eastern zone of a building is so hot in the afternoon, and receive an answer based on information that has been logged over years. The LLMs will process previously siloed data on temperature, energy consumption and occupancy patterns, among other things, and then share that information across building systems to produce an answer along with its context, Ma said.
The idea of using LLMs to enable building systems to converse with one another solves one of the biggest problems faced by building operators who want to use AI to help them manage their facilities better, what’s known as a “great handoff problem.” The term refers to the separation among sources of data created in a building’s lifecycle and the designs, principles, ideas and equations produced by architects and engineers. The data isn’t easily communicable across disciplines, impeding the ability of stakeholders to use the information in a meaningful way, Ma said.
But LLMs sharing the information with one another could allow building systems to more easily communicate their information, helping managers to maintain buildings better, she said.
The process could also enable LLMs to streamline regulatory compliance and building code interpretation for design and construction, according to the research Ma presented. The LLMs could accomplish this compliance help by translating natural language building regulations into a machine-interpretable and computable representation, then extracting information from building information models, or BIMs, and enriching and matching concepts across codes and BIMs. The LLMs then automate rule execution, among other steps, to generate reports that detail non-compliance and suggest potential fixes, according to the research.
“It doesn’t have to be perfect, but we don’t have to start from scratch,” Ma said.
There remain areas where more effort is needed before LLMs could pull the most value from the data, said Ma, pointing to HVAC control and intelligent building control as two examples.
“Buildings have become a knowledge system,” Ma said. “There are just so many data points. They’re in very different formats, they’re in very different structures, and it’s very hard to come up with a pipeline or schema to connect all of them.”
To reach these conclusions, the researchers examined a number of questions concerning the role of LLMs in shaping sustainable, intelligent and human-centric buildings, Ma said. Among them: How can LLMs streamline regulatory compliance and enable intelligent digital twins across the building lifecycle? The researchers also looked at the ethical responsibilities posed by LLMs in the built environment.
