Michigan State University (MSU) researchers have published analysis that concludes that large AI models can produce effective retrofit decisions but are less likely to identify which can deliver the best result most quickly and at the lowest cost. The study is one of the first to look at how large language models (LLMs) perform in determining how to assess efficient and effective building energy retrofits.
Energy efficiency retrofits can span a range of light, medium, or deep retrofit options. Here are some examples of what that looks like:
Light – air sealing small leaks (doors, windows, outlets), weatherstripping and caulking, installing LED lighting, and adding attic insulation.
Medium – comprehensive blower-door guided air sealing, replacing older appliances with ENERGY STAR models, upgrading wall insulation, and replacing an old HVAC system with a high-efficiency furnace or heat pump.
Deep – full exterior insulation or structural insulated panels, triple-pane windows throughout, whole home ventilation with heat recovery, all-electric, high-performance heat pumps for space and water heating, battery storage, or energy management system.
Pinpointing the right retrofit actions to achieve the desired effect can be tough for building owners, especially given the cost. And while building energy retrofits can unlock savings of anywhere from 10-75% (again, scaling from light to deep energy retrofit), the cost and difficulty of the actions are significant considerations.
Study Details
To assess the potential of AI to aid in this work, MSU queried seven LLMs to generate energy retrofit decisions with two parameters: one technically focused on the maximum CO2 reduction, and the other focused on the minimum payback period. The AI models were then rated based on how they aligned with key retrofit measures or if they fell within the top three to five measures.
Using data from ResStock 2024.2 for 49 U.S. states, researchers evaluated the accuracy, consistency, sensitivity, and reasoning of the AI responses based on details issued in the query using building-specific information. The LLMs were then tasked with assessing multiple buildings, comparing costs and efficiency among the options, and identifying which retrofit affected the most significant CO2 reduction and had the lowest payback period.
The analysis determined that while the LLMs did provide effective retrofit options, they had difficulty determining the best one.
Why It Matters
Identifying which retrofit actions to take is essential to achieving the desired building performance and cost savings. While the researchers credit AI tools for generally providing accurate and consistent options, they must be improved before they can be relied upon to make building retrofit decisions.
In the meantime, building energy experts, building scientists, and certified energy auditors will remain a go-to resource for professional guidance on building energy efficiency matters.
Read the complete MSU research study analysis here.