Modern ventilation systems make sure that the air we breathe indoors is safe—no small task when people worldwide spend an average of 60-70% of their time indoors. Ventilation plays a critical role in public health, but comes with a seemingly unavoidable downside: heavy energy consumption.
Solving this trade-off between health and energy efficiency is the focus of Yuanyuan Shi’s research. She is developing AI-accelerated simulations to help ventilation systems modulate their energy use in real time.
“AI can help us operate major physical and energy infrastructure more efficiently, intelligently and sustainably,” says Shi.
Yuanyuan Shi is a 2024 AI2050 Early Career Fellow who develops AI control methods for sustainable energy systems. Her work addresses Hard Problem #4 by tackling major climate issues in energy efficiency and resilience.
I joined the faculty of UC San Diego (UCSD) in 2021, while the university was overseeing a return-to-campus initiative during the COVID-19 pandemic. To ensure better air quality, building ventilation rates were increased to their maximum during working hours. However, this caused energy usage to more than double. Safety came at the cost of sustainability.
While I’d worked with building systems for thermal management during my Ph.D., that experience spurred me to look into the role of ventilation in energy use. Essentially, we’re trying to balance energy usage with an ideal indoor climate, maintaining safe air quality and a comfortable temperature as efficiently as possible.
Often, systems for real-time climate or ventilation control simplify entire rooms into a single point. However, in our simulations using real classrooms, different points in the room experience vastly different air quality. Usually, the criteria for safe carbon dioxide concentrations is around 800 parts per million. If the air quality was measured at the front of the room where the professor is lecturing alone, the air quality appeared to be clean and safe.
In contrast, if it was measured in the center of the room where many students were clustered, the reading was up to twice as high as the safe threshold. By compressing rooms to a point, we overlook how the geometry and the occupancy of the rooms affect air quality. As a result, some occupants might be exposed to poor air quality and risk exposure to pathogens.
The quantities we want to model—indoor air flow, air pollutants or pathogens, or the temperature—are governed by complex physical processes. More specifically, they are governed by partial differential equations or PDEs. PDEs describe how physical quantities like temperature or carbon dioxide levels are distributed over time and space within the room.
Physics-based models are accurate, but simulating them for a large building is time consuming. In one of our recent papers, one simulation within a 400-person classroom at UCSD took one hour. Simulating an entire building could take hundreds of hours. High-resolution PDE models are typically only used when a building is being designed, not for real-time control.
The goal of my research is to develop AI models that retain the underlying physics but accelerate the simulations to make them fast enough for real-time decision making. We also want to maintain safety and reliability guarantees since they’re going to be implemented in real-world settings.
We began by using real-world data to calibrate our models. We started with a dataset from UC Berkeley, from carbon dioxide sensors installed in conference rooms. The second step was to make a 3-D model of a large lecture hall at UCSD, using the class schedule to simulate the student occupancy as well as the temperature, air quality, and energy usage. We’re using a fast neural operator model, which is an AI-based model that can simulate PDEs much faster than conventional methods.
When we used the AI-based model to control ventilation and temperature of the simulated classroom, we reduced energy consumption by around 50% as compared to running the system at full capacity. We also compared our model to simpler models in use in building control today and found that it saved 10-25% more energy.
Currently, we are building a PDE foundation model to address a significant bottleneck. Previously, we had to simulate each room one by one. Converting a 2-D blueprint into a 3-D simulation, generating simulation data, and training an individual AI model for each room currently requires hours of manual processing. Repeating this workflow for thousands of rooms is prohibitively labor-intensive and computationally expensive. Instead of repeating the process for each new room, the foundation model allows us to generalize across diverse room geometries and adapt to new rooms and buildings, dramatically reducing the need for repeated simulation, data generation, and model training.
Once the foundation model is trained, we would simply feed the model the necessary information about a room —its geometry, the layouts of its vents, the HVAC and occupancy settings—and it will predict how temperature and indoor air quality will change over time.
One of the biggest challenges in developing AI methods for sustainable infrastructure is bridging the gap between research and practice. At UCSD, I was fortunate to have access to room blueprints and historical sensor data from diverse campus buildings, as well as opportunities to validate our algorithms in real-world environments. We also had access to DERConnect, UCSD’s large-scale energy systems testbed, which is a platform to test and validate new modeling and control technologies under realistic operating conditions. Without this combination of real building data and real-world testbeds, establishing the necessary data collection and AI model validation pipeline could take years.
The consequences of this research-to-practice gap are already visible in power-grid research, which is another focus of my research. Many powerful AI algorithms have been developed, yet relatively few have made their way into the operation of modern power grids. Closing this gap requires progress on two fronts. AI researchers need to develop methods that are not only accurate, but also trustworthy, explainable, and robust for the system operators to rely on in safety-critical infrastructure. At the same time, researchers and practitioners need to work closely together so that new methods address real operational needs and can be tested, validated, and ultimately deployed in practice.
The challenge, therefore, is not only that better AI methods are needed, but that we need stronger pathways connecting AI innovations with the people who design and operate our sustainable infrastructure systems. Sustained collaboration between AI researchers and practitioners could be an incredibly powerful accelerator toward the goal of building more sustainable, resilient, and intelligent energy infrastructure.