For much of my career in artificial intelligence, I have focused on how systems make decisions in complex environments. Early AI systems were designed to recognize patterns and make predictions, and over time the field expanded into solutions that could act, learn, and adapt. In industrial settings, like airflow systems, the real challenge often lies in making decisions under constraints, where multiple variables interact and tradeoffs must be managed continuously.
HVAC, chilled water, and ventilation systems are responsible for maintaining stable environments across facilities while consuming a large share of total energy. They operate under constant pressure to meet demand, maintain quality, and control costs, all while responding to changing conditions such as weather, occupancy, and production variability.
Most of these systems are still managed using traditional control strategies. These approaches are effective at maintaining setpoints, but they are not designed to continuously evaluate tradeoffs across the system. Many operations achieve acceptable performance while leaving meaningful efficiency gains unrealized because decisions are made locally rather than in coordination across the system.
The core issue is not a lack of data or instrumentation. Industrial environments generate enormous volumes of information from sensors, control systems, and historical operations. The challenge is how to use that information to guide better decisions as conditions change.
Airflow systems are inherently multi-variable. A decision to adjust a chiller, change a flow rate, or shift load can impact energy consumption, system stability, and downstream processes. These interactions are nonlinear and often difficult to model explicitly. Human operators manage this complexity by drawing on experience, intuition, and an understanding of how the system behaves under different conditions.
One of the most important developments in AI is the ability to capture and apply that kind of expertise. This is where the concept of machine teaching becomes relevant. Rather than relying only on historical data, machine teaching allows domain experts to define objectives, constraints, and acceptable operating ranges. Engineers can encode how they think about tradeoffs, including how to prioritize outcomes and respond to edge cases. This creates a framework where AI systems learn to operate within the same boundaries and priorities that guide human decision-making.
I worked on a system like this with a large enterprise operating a commercial HVAC chiller environment. The system was responsible for delivering chilled water to meet cooling demand at all times while minimizing energy consumption. It had to respond continuously to fluctuations in demand, changes in external temperature, and shifts in system performance throughout the day.
To explore how the system could be improved, we built a simulation environment using historical sensor data and control actions. This created a space where an AI system could practice managing the chiller across a wide range of operating conditions. It monitored demand, power consumption, water temperatures, and flow rates while learning how different decisions affected overall performance.
Using machine teaching, we worked with engineers to define the constraints that could not be violated and the priorities that should guide decision-making. The system learned how to shift load, anticipate changes, and adapt its behavior while continuing to meet cooling requirements. When evaluated against the existing control strategy, it reduced energy costs by 15 percent while maintaining comparable performance in meeting temperature requirements. The gains came from coordinating decisions across the system rather than optimizing individual components in isolation.
This type of approach reflects a broader shift in how industrial systems are being designed. Control remains essential, but as systems become more interconnected and conditions more dynamic, coordinated decision-making becomes a critical factor in performance. Simulation plays a central role in enabling this shift. Industrial systems do not allow for experimentation in live environments, particularly when they are tied to production or safety. By creating high-fidelity simulations, organizations can evaluate strategies, test responses to variability, and validate improvements before deployment.
Human expertise remains essential throughout this process. Operators and engineers understand how systems behave in ways that are difficult to capture through data alone. Their knowledge provides the structure that allows AI systems to operate within real-world constraints and make decisions that align with operational goals.
Airflow systems offer a clear starting point. They are widely deployed, energy-intensive, and deeply connected to overall facility performance. By combining data, simulation, and expert knowledge, organizations can move toward systems that continuously improve how decisions are made.
